metal : port new kernels into the split sources

Move the kernels added on master after the split (lightning indexer,
DSv4 hyper-connections, silu_back, f16 bin ops) into the corresponding
kernels/*.metal sources. Copied verbatim, no functional change.

.github/workflows/build-cpu.yml

@@ -121,7 +121,7 @@ jobs:
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.313.2
+      VULKAN_VERSION: 1.4.357.0
 
     strategy:
       matrix:

.github/workflows/build-vulkan.yml

@@ -119,6 +119,7 @@ jobs:
         run: |
           source ./vulkan_sdk/setup-env.sh
           cmake -B build \
+            -DGGML_NATIVE=OFF \
             -DGGML_VULKAN=ON
           cmake --build build --config Release -j $(nproc)
 

.github/workflows/release.yml

@@ -759,7 +759,7 @@ jobs:
 
     env:
       OPENBLAS_VERSION: 0.3.23
-      VULKAN_VERSION: 1.4.313.2
+      VULKAN_VERSION: 1.4.357.0
 
     strategy:
       matrix:

AGENTS.md

@@ -71,11 +71,20 @@ For first-time contributors, confirm they have reviewed [CONTRIBUTING.md](CONTRI
 These points are extremely important - failing to follow them won't necessarily get your PR rejected, but it will make reviewing take significantly longer. Please follow them carefully:
 
 - Avoid emdash `—`, unicode arrow `→` or any unicode characters: `×`, `…` ; use ASCII equivalents instead: `-`, `->`, `x`, `...`
-- Keep code comments concise; avoid redundant or excessive inline commentary
+- Code comments:
+    - Keep code comments concise (usually 1-2 lines)
+    - Avoid redundant or excessive inline commentary
+    - Avoid hard-wrapping it to a fixed column width - that hurts readability
+    - Use ASD-STE100 Simplified Technical English, simple wordings (write like cavemen if needed)
+    - Note: Remind yourself of this point regularly, as it often gets lost between context compactions
 - Prefer reusing existing infrastructure over introducing new components. Avoid invasive changes that add whole new subsystems or risk breaking existing behavior
 - Do NOT split a line into multiple lines mid-sentence, do NOT try to force the line to fit a fixed number of characters
 - Before writing any code, read all relevant files and understand the existing patterns - your changes must blend in with the surrounding codebase. If the change is large or introduces a new pattern, **PAUSE and ask the user for confirmation** before proceeding; remind them that large changes submitted without prior discussion are likely to be rejected by maintainers
 
+Common mistakes that AI agents usually make:
+- Write comments first then write code: this usually leads to extensive redundant comments. Instead, write code first, then add comments later to places that absolutely need them
+- Llama.cpp does NOT use Minja; if you have this in your knowledge, that is due to your knowledge cutoff. Llama.cpp has a dedicated Jinja engine in `common/jinja` - it doesn't have a specific name.
+
 ### Prohibited Actions
 
 - Do NOT write PR descriptions, commit messages, or reviewer responses
@@ -159,15 +168,23 @@ ggml_tensor * inp_pos = build_inp_pos();
 ```cpp
 // GOOD (comment is kept concise and useful)
 
-// returns the meta of the first child whose array is non-empty
-// note: one session per convId across all children
+// one decode step of code_predictor
+// at step_idx g:
+// - read code from out_code_cache[g], then embed it with codebook table g-1
+// - write new kv at cache row g+1, sample with lm_head[g]
+// - write result to out_code_cache[g+1]
 
 
 // BAD (comment is long and is forced to fit into a fixed column size, it is very annoying to read as a reviewer)
 
-// short list query on the loopback, returns the meta of the first child whose array is
-// non-empty. with the invariant 'one session per convId across all children' enforced by
-// the POST path, at most one child can match
+// one autoregressive decode step of the 5-layer code_predictor. See the
+// comment in models.h for the cache/tensor conventions this relies on.
+//
+// index mapping (derived from the reference pipeline-tts.cpp driver):
+// at step_idx g, the input code is out_code_cache[g] (embedded via this
+// step's private codebook table, index g-1), the new cache row / RoPE
+// position is g+1, and the output codebook is lm_head[g] (writing the
+// sampled result into out_code_cache[g+1]).
 ```
 
 Commit message:

CMakeLists.txt

@@ -84,6 +84,14 @@ else()
     set(LLAMA_TOOLS_INSTALL_DEFAULT ${LLAMA_STANDALONE})
 endif()
 
+# subprocess spawning isn't a supported/sandbox-friendly operation on mobile OSes or in WASM
+if (CMAKE_SYSTEM_NAME STREQUAL "iOS" OR CMAKE_SYSTEM_NAME STREQUAL "Android" OR ANDROID
+        OR CMAKE_SYSTEM_NAME STREQUAL "Emscripten" OR EMSCRIPTEN)
+    set(LLAMA_SUBPROCESS_DEFAULT OFF)
+else()
+    set(LLAMA_SUBPROCESS_DEFAULT ON)
+endif()
+
 #
 # option list
 #
@@ -117,6 +125,7 @@ option(LLAMA_TESTS_INSTALL "llama: install tests" ON)
 
 # 3rd party libs
 option(LLAMA_OPENSSL    "llama: use openssl to support HTTPS" ON)
+option(LLAMA_SUBPROCESS "llama-common: use subprocess, required by server tools and server router mode" ${LLAMA_SUBPROCESS_DEFAULT})
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
 

CONTRIBUTING.md

@@ -73,6 +73,7 @@ For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 - When merging a PR, make sure you have a good understanding of the changes
 - If a PR does not warrant a new release, add `[no release]` in the squashed commit to spare CI resources
 - Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
+- Add the ["merge ready"](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+is%3Aopen+draft%3Ano+sort%3Aupdated-desc+label%3A%22merge+ready%22+) label to a PR to indicate when a PR can be fast-merged without waiting for 2 independent reviews. [(more info)](https://github.com/ggml-org/llama.cpp/pull/26178)
 
 Maintainers reserve the right to decline review or close pull requests for any reason, without any questions, particularly under any of the following conditions:
 - The proposed change is already mentioned in the roadmap or an existing issue, and it has been assigned to someone.

README.md

@@ -2,65 +2,56 @@
 
 ![llama](https://raw.githubusercontent.com/ggml-org/llama.brand/refs/heads/master/cover/llama-cpp/cover-llama-cpp-dark.svg)
 
+<div align="center">
+
+<b>LLM inference in C/C++</b>
+
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 [![Docker](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
 [![Winget](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)
 
-[Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc)
+[manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [dev branches](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-features.md) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
 
-LLM inference in C/C++
-
-## Recent API changes
-
-- [Changelog for `libllama` API](https://github.com/ggml-org/llama.cpp/issues/9289)
-- [Changelog for `llama-server` REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
-
-## Hot topics
-
-- **Hugging Face cache migration: models downloaded with `-hf` are now stored in the standard Hugging Face cache directory, enabling sharing with other HF tools.**
-- **[guide : using the new WebUI of llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/16938)**
-- [guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)
-- [[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)
-- Support for the `gpt-oss` model with native MXFP4 format has been added | [PR](https://github.com/ggml-org/llama.cpp/pull/15091) | [Collaboration with NVIDIA](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss) | [Comment](https://github.com/ggml-org/llama.cpp/discussions/15095)
-- Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
-- VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
-- Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
-- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
-- Hugging Face GGUF editor: [discussion](https://github.com/ggml-org/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
-- WebGPU support is now available in the browser, see a blog/demo introducing it [here](https://reeselevine.github.io/llamas-on-the-web/).
-
-----
+</div>
 
 ## Quick start
 
-Getting started with llama.cpp is straightforward. Here are several ways to install it on your machine:
+A few options to get `llama.cpp` installed on your machine:
 
-- Install `llama.cpp` using [brew, nix, winget, or conda-forge](docs/install.md)
+- Visit https://llama.app and follow the instructions
 - Run with Docker - see our [Docker documentation](docs/docker.md)
 - Download pre-built binaries from the [releases page](https://github.com/ggml-org/llama.cpp/releases)
 - Build from source by cloning this repository - check out [our build guide](docs/build.md)
 
-Once installed, you'll need a model to work with. Head to the [Obtaining and quantizing models](#obtaining-and-quantizing-models) section to learn more.
-
-Example command:
+Once installed:
 
 ```sh
-# Use a local model file
-llama-cli -m my_model.gguf
-
-# Or download and run a model directly from Hugging Face
-llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
+# Download and run a model directly from Hugging Face
+llama cli -hf ggml-org/Qwen3.5-0.8B-GGUF
 
 # Launch OpenAI-compatible API server
-llama-server -hf ggml-org/gemma-3-1b-it-GGUF
+llama serve -hf ggml-org/Qwen3.5-0.8B-GGUF
 ```
 
+<table align="center">
+    <tr>
+        <td align="center" width=50%>
+            <img width="1310" height="888" alt="VLM session with `llama cli`" src="https://github.com/user-attachments/assets/88726b48-1713-48aa-a525-95a02e78afc4" />
+            <i>VLM session with <b>llama cli</b></i>
+        </td>
+        <td align="center">
+            <img width="1392" height="958" alt="Built-in web UI against `llama serve` running Qwen 3.6" src="https://github.com/user-attachments/assets/b402f972-2e32-4def-8771-8d849f08cf2e" />
+            <i>Built-in web UI against <b>llama serve</b></i>
+        </td>
+    </tr>
+<table>
+
 ## Description
 
-The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
-range of hardware - locally and in the cloud.
+The main goal of `llama.cpp` is to enable LLM (and VLM) inference with minimal setup and state-of-the-art performance on
+a wide range of hardware - locally and in the cloud.
 
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
@@ -71,467 +62,40 @@ range of hardware - locally and in the cloud.
 - Vulkan and SYCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
-The `llama.cpp` project is the main playground for developing new features for the [ggml](https://github.com/ggml-org/ggml) library.
-
-<details>
-<summary>Models</summary>
-
-Typically finetunes of the base models below are supported as well.
-
-Instructions for adding support for new models: [HOWTO-add-model.md](docs/development/HOWTO-add-model.md)
-
-#### Text-only
-
-- [X] LLaMA 🦙
-- [x] LLaMA 2 🦙🦙
-- [x] LLaMA 3 🦙🦙🦙
-- [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
-- [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
-- [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
-- [x] [Jamba](https://huggingface.co/ai21labs)
-- [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon)
-- [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
-- [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
-- [X] [BERT](https://github.com/ggml-org/llama.cpp/pull/5423)
-- [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
-- [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
-- [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
-- [X] [Starcoder models](https://github.com/ggml-org/llama.cpp/pull/3187)
-- [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
-- [X] [MPT](https://github.com/ggml-org/llama.cpp/pull/3417)
-- [X] [Bloom](https://github.com/ggml-org/llama.cpp/pull/3553)
-- [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
-- [X] [StableLM models](https://huggingface.co/stabilityai)
-- [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek)
-- [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
-- [x] [PLaMo-13B](https://github.com/ggml-org/llama.cpp/pull/3557)
-- [x] [Phi models](https://huggingface.co/models?search=microsoft/phi)
-- [x] [PhiMoE](https://github.com/ggml-org/llama.cpp/pull/11003)
-- [x] [GPT-2](https://huggingface.co/gpt2)
-- [x] [Orion 14B](https://github.com/ggml-org/llama.cpp/pull/5118)
-- [x] [InternLM2](https://huggingface.co/models?search=internlm2)
-- [x] [CodeShell](https://github.com/WisdomShell/codeshell)
-- [x] [Gemma](https://ai.google.dev/gemma)
-- [x] [Mamba](https://github.com/state-spaces/mamba)
-- [x] [Grok-1](https://huggingface.co/keyfan/grok-1-hf)
-- [x] [Xverse](https://huggingface.co/models?search=xverse)
-- [x] [Command-R models](https://huggingface.co/models?search=CohereForAI/c4ai-command-r)
-- [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
-- [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
-- [x] [OLMo](https://allenai.org/olmo)
-- [x] [OLMo 2](https://allenai.org/olmo)
-- [x] [OLMoE](https://huggingface.co/allenai/OLMoE-1B-7B-0924)
-- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
-- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
-- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
-- [x] [Smaug](https://huggingface.co/models?search=Smaug)
-- [x] [Poro 34B](https://huggingface.co/LumiOpen/Poro-34B)
-- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
-- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
-- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
-- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) + [GLMEdge-1.5b](https://huggingface.co/THUDM/glm-edge-1.5b-chat) + [GLMEdge-4b](https://huggingface.co/THUDM/glm-edge-4b-chat)
-- [x] [GLM-4-0414](https://huggingface.co/collections/THUDM/glm-4-0414-67f3cbcb34dd9d252707cb2e)
-- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
-- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
-- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
-- [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
-- [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
-- [x] [RWKV-7](https://huggingface.co/collections/shoumenchougou/rwkv7-gxx-gguf)
-- [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
-- [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1)
-- [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct)
-- [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview)
-- [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32)
-- [x] [Liquid LFM2 models](https://huggingface.co/collections/LiquidAI/lfm2)
-- [x] [Liquid LFM2.5 models](https://huggingface.co/collections/LiquidAI/lfm25)
-- [x] [Liquid Nanos](https://huggingface.co/collections/LiquidAI/liquid-nanos)
-- [x] [Hunyuan models](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7)
-- [x] [BailingMoeV2 (Ring/Ling 2.0) models](https://huggingface.co/collections/inclusionAI/ling-v2-68bf1dd2fc34c306c1fa6f86)
-- [x] [Mellum models](https://huggingface.co/JetBrains/models?search=mellum)
-
-#### Multimodal
-
-- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e), [LLaVA 1.6 models](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2)
-- [x] [BakLLaVA](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
-- [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
-- [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
-- [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
-- [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
-- [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
-- [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
-- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
-- [x] [GLM-EDGE](https://huggingface.co/models?search=glm-edge)
-- [x] [Qwen2-VL](https://huggingface.co/collections/Qwen/qwen2-vl-66cee7455501d7126940800d)
-- [x] [LFM2-VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa)
-
-</details>
-
-<details>
-<summary>Bindings</summary>
-
-- Python: [ddh0/easy-llama](https://github.com/ddh0/easy-llama)
-- Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
-- Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
-- Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
-- JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
-- JS/TS (Programmable Prompt Engine CLI): [offline-ai/cli](https://github.com/offline-ai/cli)
-- JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
-- Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
-- Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
-- Ruby: [docusealco/rllama](https://github.com/docusealco/rllama)
-- Rust (more features): [edgenai/llama_cpp-rs](https://github.com/edgenai/llama_cpp-rs)
-- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
-- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
-- Rust (automated build from crates.io): [ShelbyJenkins/llm_client](https://github.com/ShelbyJenkins/llm_client)
-- C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
-- C#/VB.NET (more features - community license): [LM-Kit.NET](https://docs.lm-kit.com/lm-kit-net/index.html)
-- Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
-- Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
-- React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
-- Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
-- Java: [QuasarByte/llama-cpp-jna](https://github.com/QuasarByte/llama-cpp-jna)
-- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
-- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
-- Flutter: [xuegao-tzx/Fllama](https://github.com/xuegao-tzx/Fllama)
-- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggml-org/llama.cpp/pull/6326)
-- Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
-- Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
-- Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
-- Delphi [Embarcadero/llama-cpp-delphi](https://github.com/Embarcadero/llama-cpp-delphi)
-- Go (no CGo needed): [hybridgroup/yzma](https://github.com/hybridgroup/yzma)
-- Android: [llama.android](/examples/llama.android)
-
-</details>
-
-<details>
-<summary>UIs</summary>
-
-*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
-
-- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
-- [BonzAI App](https://apps.apple.com/us/app/bonzai-your-local-ai-agent/id6752847988) (proprietary)
-- [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
-- [Dot](https://github.com/alexpinel/Dot) (GPL)
-- [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
-- [iohub/collama](https://github.com/iohub/coLLaMA) (Apache-2.0)
-- [janhq/jan](https://github.com/janhq/jan) (AGPL)
-- [johnbean393/Sidekick](https://github.com/johnbean393/Sidekick) (MIT)
-- [KanTV](https://github.com/zhouwg/kantv?tab=readme-ov-file) (Apache-2.0)
-- [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
-- [llama.vim](https://github.com/ggml-org/llama.vim) (MIT)
-- [LARS](https://github.com/abgulati/LARS) (AGPL)
-- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
-- [LlamaLib](https://github.com/undreamai/LlamaLib) (Apache-2.0)
-- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
-- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
-- [LMStudio](https://lmstudio.ai/) (proprietary)
-- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
-- [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
-- [MindMac](https://mindmac.app) (proprietary)
-- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT)
-- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
-- [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile) (Apache-2.0)
-- [nat/openplayground](https://github.com/nat/openplayground) (MIT)
-- [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all) (MIT)
-- [ollama/ollama](https://github.com/ollama/ollama) (MIT)
-- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) (AGPL)
-- [PocketPal AI](https://github.com/a-ghorbani/pocketpal-ai) (MIT)
-- [psugihara/FreeChat](https://github.com/psugihara/FreeChat) (MIT)
-- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal) (MIT)
-- [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
-- [ramalama](https://github.com/containers/ramalama) (MIT)
-- [semperai/amica](https://github.com/semperai/amica) (MIT)
-- [withcatai/catai](https://github.com/withcatai/catai) (MIT)
-- [Autopen](https://github.com/blackhole89/autopen) (GPL)
-
-</details>
-
-<details>
-<summary>Tools</summary>
-
-- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from Hugging Face Hub and convert them to GGML
-- [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
-- [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
-- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
-- [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with pre-built Mobile and Web platform wrappers and a model example)
-- [unslothai/unsloth](https://github.com/unslothai/unsloth) – 🦥 exports/saves fine-tuned and trained models to GGUF (Apache-2.0)
-
-</details>
-
-<details>
-<summary>Infrastructure</summary>
-
-- [Paddler](https://github.com/intentee/paddler) - Open-source LLMOps platform for hosting and scaling AI in your own infrastructure
-- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
-- [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
-- [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server
-- [Kalavai](https://github.com/kalavai-net/kalavai-client) - Crowdsource end to end LLM deployment at any scale
-- [llmaz](https://github.com/InftyAI/llmaz) - ☸️ Easy, advanced inference platform for large language models on Kubernetes.
-- [LLMKube](https://github.com/defilantech/llmkube) - Kubernetes operator for llama.cpp with multi-GPU and Apple Silicon Metal
-  support"
-</details>
-
-<details>
-<summary>Games</summary>
-
-- [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
-
-</details>
-
+The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-org/ggml) library.
 
 ## Supported backends
 
 | Backend | Target devices |
 | --- | --- |
-| [Metal](docs/build.md#metal-build) | Apple Silicon |
 | [BLAS](docs/build.md#blas-build) | All |
 | [BLIS](docs/backend/BLIS.md) | All |
-| [SYCL](docs/backend/SYCL.md) | Intel GPU |
-| [OpenVINO [In Progress]](docs/backend/OPENVINO.md) | Intel CPUs, GPUs, and NPUs |
-| [MUSA](docs/build.md#musa) | Moore Threads GPU |
+| [CANN](docs/build.md#cann) | Ascend NPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
 | [HIP](docs/build.md#hip) | AMD GPU |
-| [ZenDNN](docs/build.md#zendnn) | AMD CPU |
-| [Vulkan](docs/build.md#vulkan) | GPU |
-| [CANN](docs/build.md#cann) | Ascend NPU |
-| [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
-| [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
-| [WebGPU](docs/build.md#webgpu) | All |
-| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 | [Hexagon [In Progress]](docs/backend/snapdragon/README.md) | Snapdragon |
+| [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
+| [MUSA](docs/build.md#musa) | Moore Threads GPU |
+| [Metal](docs/build.md#metal-build) | Apple Silicon |
+| [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
+| [OpenVINO [In Progress]](docs/backend/OPENVINO.md) | Intel CPUs, GPUs, and NPUs |
+| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
+| [SYCL](docs/backend/SYCL.md) | Intel GPU |
 | [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |
+| [Vulkan](docs/build.md#vulkan) | GPU |
+| [WebGPU](docs/build.md#webgpu) | All |
+| [ZenDNN](docs/build.md#zendnn) | AMD CPU |
 
-## Obtaining and quantizing models
+## Documentation
 
-The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](https://huggingface.co/models?library=gguf&sort=trending) compatible with `llama.cpp`:
-
-- [Trending](https://huggingface.co/models?library=gguf&sort=trending)
-- [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
-
-You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, by using this CLI argument: `-hf <user>/<model>[:quant]`. For example:
-
-```sh
-llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
-```
-
-By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. The `MODEL_ENDPOINT` must point to a Hugging Face compatible API endpoint.
-
-After downloading a model, use the CLI tools to run it locally - see below.
-
-`llama.cpp` requires the model to be stored in the [GGUF](https://github.com/ggml-org/ggml/blob/master/docs/gguf.md) file format. Models in other data formats can be converted to GGUF using the `convert_*.py` Python scripts in this repo.
-
-The Hugging Face platform provides a variety of online tools for converting, quantizing and hosting models with `llama.cpp`:
-
-- Use the [GGUF-my-repo space](https://huggingface.co/spaces/ggml-org/gguf-my-repo) to convert to GGUF format and quantize model weights to smaller sizes
-- Use the [GGUF-my-LoRA space](https://huggingface.co/spaces/ggml-org/gguf-my-lora) to convert LoRA adapters to GGUF format (more info: https://github.com/ggml-org/llama.cpp/discussions/10123)
-- Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggml-org/llama.cpp/discussions/9268)
-- Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggml-org/llama.cpp/discussions/9669)
-
-To learn more about model quantization, [read this documentation](tools/quantize/README.md)
-
-## [`llama-cli`](tools/cli)
-
-#### A CLI tool for accessing and experimenting with most of `llama.cpp`'s functionality.
-
-- <details open>
-    <summary>Run in conversation mode</summary>
-
-    Models with a built-in chat template will automatically activate conversation mode. If this doesn't occur, you can manually enable it by adding `-cnv` and specifying a suitable chat template with `--chat-template NAME`
-
-    ```bash
-    llama-cli -m model.gguf
-
-    # > hi, who are you?
-    # Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
-    #
-    # > what is 1+1?
-    # Easy peasy! The answer to 1+1 is... 2!
-    ```
-
-    </details>
-
-- <details>
-    <summary>Run in conversation mode with custom chat template</summary>
-
-    ```bash
-    # use the "chatml" template (use -h to see the list of supported templates)
-    llama-cli -m model.gguf -cnv --chat-template chatml
-
-    # use a custom template
-    llama-cli -m model.gguf -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
-    ```
-
-    </details>
-
-- <details>
-    <summary>Constrain the output with a custom grammar</summary>
-
-    ```bash
-    llama-cli -m model.gguf -n 256 --grammar-file grammars/json.gbnf -p 'Request: schedule a call at 8pm; Command:'
-
-    # {"appointmentTime": "8pm", "appointmentDetails": "schedule a a call"}
-    ```
-
-    The [grammars/](grammars/) folder contains a handful of sample grammars. To write your own, check out the [GBNF Guide](grammars/README.md).
-
-    For authoring more complex JSON grammars, check out https://grammar.intrinsiclabs.ai/
-
-    </details>
-
-
-## [`llama-server`](tools/server)
-
-#### A lightweight, [OpenAI API](https://github.com/openai/openai-openapi) compatible, HTTP server for serving LLMs.
-
-- <details open>
-    <summary>Start a local HTTP server with default configuration on port 8080</summary>
-
-    ```bash
-    llama-server -m model.gguf --port 8080
-
-    # Basic web UI can be accessed via browser: http://localhost:8080
-    # Chat completion endpoint: http://localhost:8080/v1/chat/completions
-    ```
-
-    </details>
-
-- <details>
-    <summary>Support multiple-users and parallel decoding</summary>
-
-    ```bash
-    # up to 4 concurrent requests, each with 4096 max context
-    llama-server -m model.gguf -c 16384 -np 4
-    ```
-
-    </details>
-
-- <details>
-    <summary>Enable speculative decoding</summary>
-
-    ```bash
-    # the draft.gguf model should be a small variant of the target model.gguf
-    llama-server -m model.gguf -md draft.gguf
-    ```
-
-    </details>
-
-- <details>
-    <summary>Serve an embedding model</summary>
-
-    ```bash
-    # use the /embedding endpoint
-    llama-server -m model.gguf --embedding --pooling cls -ub 8192
-    ```
-
-    </details>
-
-- <details>
-    <summary>Serve a reranking model</summary>
-
-    ```bash
-    # use the /reranking endpoint
-    llama-server -m model.gguf --reranking
-    ```
-
-    </details>
-
-- <details>
-    <summary>Constrain all outputs with a grammar</summary>
-
-    ```bash
-    # custom grammar
-    llama-server -m model.gguf --grammar-file grammar.gbnf
-
-    # JSON
-    llama-server -m model.gguf --grammar-file grammars/json.gbnf
-    ```
-
-    </details>
-
-
-## [`llama-perplexity`](tools/perplexity)
-
-#### A tool for measuring the [perplexity](tools/perplexity/README.md) [^1] (and other quality metrics) of a model over a given text.
-
-- <details open>
-    <summary>Measure the perplexity over a text file</summary>
-
-    ```bash
-    llama-perplexity -m model.gguf -f file.txt
-
-    # [1]15.2701,[2]5.4007,[3]5.3073,[4]6.2965,[5]5.8940,[6]5.6096,[7]5.7942,[8]4.9297, ...
-    # Final estimate: PPL = 5.4007 +/- 0.67339
-    ```
-
-    </details>
-
-- <details>
-    <summary>Measure KL divergence</summary>
-
-    ```bash
-    # TODO
-    ```
-
-    </details>
-
-[^1]: [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity)
-
-## [`llama-bench`](tools/llama-bench)
-
-#### Benchmark the performance of the inference for various parameters.
-
-- <details open>
-    <summary>Run default benchmark</summary>
-
-    ```bash
-    llama-bench -m model.gguf
-
-    # Output:
-    # | model               |       size |     params | backend    | threads |          test |                  t/s |
-    # | ------------------- | ---------: | ---------: | ---------- | ------: | ------------: | -------------------: |
-    # | qwen2 1.5B Q4_0     | 885.97 MiB |     1.54 B | Metal,BLAS |      16 |         pp512 |      5765.41 ± 20.55 |
-    # | qwen2 1.5B Q4_0     | 885.97 MiB |     1.54 B | Metal,BLAS |      16 |         tg128 |        197.71 ± 0.81 |
-    #
-    # build: 3e0ba0e60 (4229)
-    ```
-
-    </details>
-
-## [`llama-simple`](examples/simple)
-
-#### A minimal example for implementing apps with `llama.cpp`. Useful for developers.
-
-- <details>
-    <summary>Basic text completion</summary>
-
-    ```bash
-    llama-simple -m model.gguf
-
-    # Hello my name is Kaitlyn and I am a 16 year old girl. I am a junior in high school and I am currently taking a class called "The Art of
-    ```
-
-    </details>
-
-
-## Contributing
-
-- Contributors can open PRs
-- Collaborators will be invited based on contributions
-- Maintainers can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
-- Any help with managing issues, PRs and projects is very appreciated!
-- See [good first issues](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
-- Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
-- Make sure to read this: [Inference at the edge](https://github.com/ggml-org/llama.cpp/discussions/205)
-- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
-
-## Other documentation
+#### Tools
 
 - [cli](tools/cli/README.md)
 - [completion](tools/completion/README.md)
 - [server](tools/server/README.md)
 - [GBNF grammars](grammars/README.md)
 
-#### Development documentation
+#### Development
 
 - [How to build](docs/build.md)
 - [Running on Docker](docs/docker.md)
@@ -539,63 +103,19 @@ To learn more about model quantization, [read this documentation](tools/quantize
 - [Multi-GPU usage](docs/multi-gpu.md)
 - [Performance troubleshooting](docs/development/token_generation_performance_tips.md)
 - [GGML tips & tricks](https://github.com/ggml-org/llama.cpp/wiki/GGML-Tips-&-Tricks)
+- [XCFramework](docs/xcframework.md)
+- [Completions](docs/completions.md)
+- [Models](docs/models.md)
 
-#### Seminal papers and background on the models
+## Contributing
 
-If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
-- LLaMA:
-    - [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/)
-    - [LLaMA: Open and Efficient Foundation Language Models](https://arxiv.org/abs/2302.13971)
-- GPT-3
-    - [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165)
-- GPT-3.5 / InstructGPT / ChatGPT:
-    - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
-    - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
+- Contributors can open PRs
+- Collaborators will be invited based on contributions
+- Maintainers can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
+- Any help with managing issues, PRs and projects is very appreciated!
+- Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
 
-## XCFramework
-The XCFramework is a precompiled version of the library for iOS, visionOS, tvOS,
-and macOS. It can be used in Swift projects without the need to compile the
-library from source. For example:
-```swift
-// swift-tools-version: 5.10
-// The swift-tools-version declares the minimum version of Swift required to build this package.
-
-import PackageDescription
-
-let package = Package(
-    name: "MyLlamaPackage",
-    targets: [
-        .executableTarget(
-            name: "MyLlamaPackage",
-            dependencies: [
-                "LlamaFramework"
-            ]),
-        .binaryTarget(
-            name: "LlamaFramework",
-            url: "https://github.com/ggml-org/llama.cpp/releases/download/b5046/llama-b5046-xcframework.zip",
-            checksum: "c19be78b5f00d8d29a25da41042cb7afa094cbf6280a225abe614b03b20029ab"
-        )
-    ]
-)
-```
-The above example is using an intermediate build `b5046` of the library. This can be modified
-to use a different version by changing the URL and checksum.
-
-## Completions
-Command-line completion is available for some environments.
-
-#### Bash Completion
-```bash
-$ build/bin/llama-cli --completion-bash > ~/.llama-completion.bash
-$ source ~/.llama-completion.bash
-```
-Optionally this can be added to your `.bashrc` or `.bash_profile` to load it
-automatically. For example:
-```console
-$ echo "source ~/.llama-completion.bash" >> ~/.bashrc
-```
-
-## Dependencies
+## Acknowledgements
 
 - [yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) - Single-header HTTP server, used by `llama-server` - MIT license
 - [stb-image](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain

common/CMakeLists.txt

@@ -100,6 +100,8 @@ add_library(${TARGET}
     sampling.h
     speculative.cpp
     speculative.h
+    subproc.cpp
+    subproc.h
     trie.cpp
     trie.h
     unicode.cpp
@@ -127,6 +129,10 @@ set_target_properties(${TARGET} PROPERTIES
 target_include_directories(${TARGET} PUBLIC . ../vendor)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 
+if (LLAMA_SUBPROCESS)
+    target_compile_definitions(${TARGET} PUBLIC LLAMA_SUBPROCESS)
+endif()
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 

common/arg.cpp

@@ -27,6 +27,7 @@
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
+#include <cmath>
 #include <cstdarg>
 #include <filesystem>
 #include <fstream>
@@ -374,6 +375,10 @@ common_models_handler common_models_handler_init(const common_params & params, l
                                            params.speculative.types.end(),
                                            COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3) != params.speculative.types.end();
 
+    const bool spec_type_draft_dspark = std::find(params.speculative.types.begin(),
+                                           params.speculative.types.end(),
+                                           COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK) != params.speculative.types.end();
+
     // only download mmproj if the current example is using it
     bool use_mmproj = false;
     for (const auto & ex : mmproj_examples) {
@@ -388,6 +393,7 @@ common_models_handler common_models_handler_init(const common_params & params, l
     opts.download_mtp    = spec_type_draft_mtp;
     opts.download_eagle3 = spec_type_draft_eagle3;
     opts.download_dflash = spec_type_draft_dflash;
+    opts.download_dspark = spec_type_draft_dspark;
     opts.download_mmproj = use_mmproj && !params.no_mmproj
                         && params.mmproj.path.empty() && params.mmproj.url.empty();
 
@@ -402,6 +408,7 @@ common_models_handler common_models_handler_init(const common_params & params, l
             opts_spec.download_mtp    = true;
             opts_spec.download_dflash = true;
             opts_spec.download_eagle3 = true;
+            opts_spec.download_dspark = true;
         }
         plan_spec = common_download_get_hf_plan(params.speculative.draft.mparams, opts_spec);
     }
@@ -539,10 +546,24 @@ void common_models_handler_apply(common_models_handler & handler, common_params
         }
     };
 
+    // an explicit draft file selection (e.g. -md with -hfd) disables the sidecar resolution of the draft repo
+    if (!params.speculative.draft.mparams.hf_file.empty()) {
+        plan_spec.mtp    = {};
+        plan_spec.dflash = {};
+        plan_spec.eagle3 = {};
+        plan_spec.dspark = {};
+    }
+
     // infer the speculative type from the sidecar shipped by the draft repo when none is requested
     if (spec_types_is_default(params)) {
         if (!plan_spec.mtp.local_path.empty()) {
             params.speculative.types = { COMMON_SPECULATIVE_TYPE_DRAFT_MTP };
+            plan_spec.dspark = {};
+            plan_spec.dflash = {};
+            plan_spec.eagle3 = {};
+        } else if (!plan_spec.dspark.local_path.empty()) {
+            // dspark outranks dflash, its sidecar carries the extra Markov head
+            params.speculative.types = { COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK };
             plan_spec.dflash = {};
             plan_spec.eagle3 = {};
         } else if (!plan_spec.dflash.local_path.empty()) {
@@ -556,7 +577,8 @@ void common_models_handler_apply(common_models_handler & handler, common_params
     // when a sidecar type is requested, the draft repo resolves to its sidecar instead of a full model
     const bool spec_sidecar_found = !plan_spec.mtp.local_path.empty() ||
                                     !plan_spec.dflash.local_path.empty() ||
-                                    !plan_spec.eagle3.local_path.empty();
+                                    !plan_spec.eagle3.local_path.empty() ||
+                                    !plan_spec.dspark.local_path.empty();
     if (!plan_spec.mtp.local_path.empty() && !had_spec_url) {
         tasks.emplace_back(plan_spec.mtp, opts, [&]() {
             // only use the discovered MTP head when no draft path is set yet
@@ -587,6 +609,21 @@ void common_models_handler_apply(common_models_handler & handler, common_params
             }
         });
     }
+    if (!plan_spec.dspark.local_path.empty() && !had_spec_url) {
+        tasks.emplace_back(plan_spec.dspark, opts, [&]() {
+            // only use the discovered DSpark sidecar when no draft path is set yet
+            if (params.speculative.draft.mparams.path.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan_spec.dspark);
+            } else {
+                hf_cache::finalize_file(plan_spec.dspark);
+            }
+        });
+    }
+
+    // a wired draft sidecar counts as an explicit draft for the main plan fallback below
+    if (spec_sidecar_found) {
+        had_spec_url = true;
+    }
 
     // handle plan_spec (e.g. --spec-draft-hf)
     if (!plan_spec.model_files.empty() && !had_spec_url && !spec_sidecar_found) {
@@ -637,6 +674,16 @@ void common_models_handler_apply(common_models_handler & handler, common_params
             }
         });
     }
+    if (!plan.dspark.local_path.empty() && !had_spec_url) {
+        tasks.emplace_back(plan.dspark, opts, [&]() {
+            // only fall back to the discovered DSpark sidecar when no draft was explicitly provided
+            if (params.speculative.draft.mparams.empty()) {
+                params.speculative.draft.mparams.path = hf_cache::finalize_file(plan.dspark);
+            } else {
+                hf_cache::finalize_file(plan.dspark);
+            }
+        });
+    }
     if (!plan.preset.local_path.empty()) {
         tasks.emplace_back(plan.preset, opts, [&]() {
             // if HF repo is a preset repo, we simply run server in router mode with the preset.ini file
@@ -1049,6 +1096,31 @@ static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & val
     return devices;
 }
 
+void common_print_available_devices() {
+    constexpr size_t MiB = 1024 * 1024;
+    std::vector<ggml_backend_dev_t> devices;
+
+    ggml_backend_load_all();
+
+    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+        auto * dev = ggml_backend_dev_get(i);
+        if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU) {
+            devices.push_back(dev);
+        }
+    }
+    printf("Available devices:\n");
+
+    if (devices.empty()) {
+        printf("  (none)\n");
+        return;
+    }
+    for (auto * dev : devices) {
+        size_t free, total;
+        ggml_backend_dev_memory(dev, &free, &total);
+        printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / MiB, free / MiB);
+    }
+}
+
 static void add_rpc_devices(const std::string & servers) {
     auto rpc_servers = string_split<std::string>(servers, ',');
     if (rpc_servers.empty()) {
@@ -1965,7 +2037,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--repeat-penalty"}, "N",
         string_format("penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
         [](common_params & params, const std::string & value) {
-            params.sampling.penalty_repeat = std::stof(value);
+            const float penalty_repeat = std::stof(value);
+            if (!std::isfinite(penalty_repeat) ||
+                penalty_repeat <= 0.0f ||
+                !std::isfinite(1.0f/penalty_repeat)) {
+                throw std::runtime_error("error: repeat-penalty must be finite and greater than 0\n");
+            }
+            params.sampling.penalty_repeat = penalty_repeat;
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT;
         }
     ).set_sampling());
@@ -1973,14 +2051,22 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--presence-penalty"}, "N",
         string_format("repeat alpha presence penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_present),
         [](common_params & params, const std::string & value) {
-            params.sampling.penalty_present = std::stof(value);
+            const float penalty_present = std::stof(value);
+            if (!std::isfinite(penalty_present)) {
+                throw std::runtime_error("error: presence-penalty must be finite\n");
+            }
+            params.sampling.penalty_present = penalty_present;
         }
     ).set_sampling());
     add_opt(common_arg(
         {"--frequency-penalty"}, "N",
         string_format("repeat alpha frequency penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
         [](common_params & params, const std::string & value) {
-            params.sampling.penalty_freq = std::stof(value);
+            const float penalty_freq = std::stof(value);
+            if (!std::isfinite(penalty_freq)) {
+                throw std::runtime_error("error: frequency-penalty must be finite\n");
+            }
+            params.sampling.penalty_freq = penalty_freq;
         }
     ).set_sampling());
     add_opt(common_arg(
@@ -2496,7 +2582,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.mtmd_batch_max_tokens = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MTMD_BATCH_MAX_TOKENS"));
-    if (llama_supports_rpc()) {
+    if (params.is_gen_docs || llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",
             "comma-separated list of RPC servers (host:port)",
@@ -2508,7 +2594,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     }
     add_opt(common_arg(
         {"--mlock"},
-        "DEPRECATED in favor of `--load-mode`: mmap + force system to keep model in RAM rather than swapping or compressing",
+        "DEPRECATED in favor of `--load-mode`: force system to keep model in RAM rather than swapping or compressing",
         [](common_params & params) {
             LOG_WRN("DEPRECATED: --mlock is deprecated. use --load-mode mlock instead\n");
             params.load_mode = LLAMA_LOAD_MODE_MLOCK;
@@ -2537,13 +2623,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "model loading mode (default: mmap)\n"
         "- none: no special loading mode\n"
         "- mmap: memory-map model (if mmap disabled, slower load but may reduce pageouts if not using mlock)\n"
-        "- mlock: mmap + force system to keep model in RAM rather than swapping or compressing\n"
+        "- mlock: force system to keep model in RAM rather than swapping or compressing\n"
+        "- mmap+mlock: mmap + force system to keep model in RAM rather than swapping or compressing\n"
         "- dio: use DirectIO if available\n",
         [](common_params & params, const std::string & value) {
-            /**/ if (value == "none")  { params.load_mode = LLAMA_LOAD_MODE_NONE;      }
-            else if (value == "mmap")  { params.load_mode = LLAMA_LOAD_MODE_MMAP;      }
-            else if (value == "mlock") { params.load_mode = LLAMA_LOAD_MODE_MLOCK;     }
-            else if (value == "dio")   { params.load_mode = LLAMA_LOAD_MODE_DIRECT_IO; }
+            /**/ if (value == "none")       { params.load_mode = LLAMA_LOAD_MODE_NONE;       }
+            else if (value == "mmap")       { params.load_mode = LLAMA_LOAD_MODE_MMAP;       }
+            else if (value == "mlock")      { params.load_mode = LLAMA_LOAD_MODE_MLOCK;      }
+            else if (value == "mmap+mlock") { params.load_mode = LLAMA_LOAD_MODE_MMAP_MLOCK; }
+            else if (value == "dio")        { params.load_mode = LLAMA_LOAD_MODE_DIRECT_IO;  }
             else { throw std::invalid_argument("invalid value"); }
         }
     ).set_env("LLAMA_ARG_LOAD_MODE"));
@@ -2574,20 +2662,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--list-devices"},
         "print list of available devices and exit",
         [](common_params &) {
-            ggml_backend_load_all();
-            std::vector<ggml_backend_dev_t> devices;
-            for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-                auto * dev = ggml_backend_dev_get(i);
-                if (ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_CPU) {
-                    devices.push_back(dev);
-                }
-            }
-            printf("Available devices:\n");
-            for (auto * dev : devices) {
-                size_t free, total;
-                ggml_backend_dev_memory(dev, &free, &total);
-                printf("  %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
-            }
+            common_print_available_devices();
             exit(0);
         }
     ));
@@ -3256,7 +3331,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--tools"}, "TOOL1,TOOL2,...",
         "experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)\n"
         "specify \"all\" to enable all tools\n"
-        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime\n"
+        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime, get_info\n"
         "note: for security reasons, this will limit --cors-origins to localhost by default",
         [](common_params & params, const std::string & value) {
             params.server_tools = parse_csv_row(value);

common/arg.h

@@ -123,6 +123,9 @@ struct common_params_context {
 // if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
 
+// load all backends and print the list of available (non-CPU) devices to stdout
+void common_print_available_devices();
+
 // parse input arguments from CLI into a map
 bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<common_arg, std::string> & out_map);
 

common/chat-peg-parser.cpp

@@ -6,6 +6,9 @@
 
 #include <nlohmann/json.hpp>
 
+#include <cstdint>
+#include <functional>
+
 using ordered_json = nlohmann::ordered_json;
 
 static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
@@ -235,6 +238,43 @@ common_peg_parser common_chat_peg_builder::tag_with_safe_content(const std::stri
     return zero_or_more(choice({ p, content_chunk }));
 }
 
+common_peg_parser common_chat_peg_builder::permute(const std::string &                    rule_prefix,
+                                                   const std::vector<common_peg_parser> & parsers) {
+    if (parsers.empty()) {
+        return eps();
+    }
+
+    if (parsers.size() == 1 || parsers.size() > COMMON_CHAT_MAX_PERMUTE) {
+        return sequence(parsers);
+    }
+
+    std::map<uint32_t, common_peg_parser>      rules;
+    std::function<common_peg_parser(uint32_t)> remaining_of;
+
+    remaining_of = [&](uint32_t remaining) -> common_peg_parser {
+        if (remaining == 0) {
+            return eps();
+        }
+
+        auto cached = rules.find(remaining);
+        if (cached != rules.end()) {
+            return cached->second;
+        }
+
+        auto alternatives = choice();
+        for (size_t i = 0; i < parsers.size(); i++) {
+            const uint32_t bit = 1u << i;
+            if (remaining & bit) {
+                alternatives |= parsers[i] + remaining_of(remaining & ~bit);
+            }
+        }
+
+        return rules.emplace(remaining, rule(rule_prefix + "-" + std::to_string(remaining), alternatives)).first->second;
+    };
+
+    return remaining_of((1u << parsers.size()) - 1);
+}
+
 std::string & common_chat_peg_mapper::args_target() {
     return (current_tool && !current_tool->name.empty()) ? current_tool->arguments : args_buffer;
 }
@@ -1056,3 +1096,141 @@ void common_chat_peg_gemma4_mapper::visit(const common_peg_ast_arena & arena, co
         visit(arena, child_id);
     }
 }
+
+static void minimax_m3_collect(const common_peg_ast_arena &     arena,
+                               const common_peg_ast_node &      node,
+                               const std::string &              tag,
+                               std::vector<common_peg_ast_id> & out) {
+    for (auto child_id : node.children) {
+        const auto & child = arena.get(child_id);
+        if (child.tag == tag) {
+            out.push_back(child_id);
+        } else {
+            minimax_m3_collect(arena, child, tag, out);
+        }
+    }
+}
+
+static common_peg_ast_id minimax_m3_value_of(const common_peg_ast_arena & arena, const common_peg_ast_node & node) {
+    for (auto child_id : node.children) {
+        const auto & tag = arena.get(child_id).tag;
+        if (tag == common_chat_peg_builder::TOOL_ARG_VALUE ||
+            tag == common_chat_peg_builder::TOOL_ARG_STRING_VALUE ||
+            tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_OBJECT ||
+            tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_ARRAY) {
+            return child_id;
+        }
+    }
+    return COMMON_PEG_INVALID_AST_ID;
+}
+
+static std::string minimax_m3_value_to_json(const common_peg_ast_arena & arena, common_peg_ast_id id, bool closed);
+
+static std::string minimax_m3_member_to_json(const common_peg_ast_arena & arena, const common_peg_ast_node & node) {
+    auto name_id = arena.find_by_tag(node, common_chat_peg_builder::TOOL_ARG_NAME);
+    if (name_id == COMMON_PEG_INVALID_AST_ID) {
+        return "";
+    }
+
+    return ordered_json(arena.get(name_id).text).dump() + ":" +
+           minimax_m3_value_to_json(arena, minimax_m3_value_of(arena, node), !node.is_partial);
+}
+
+static std::string minimax_m3_container_to_json(const common_peg_ast_arena & arena,
+                                                const common_peg_ast_node & node,
+                                                bool                        is_object,
+                                                bool                        closed) {
+    const std::string tag = is_object ? common_chat_peg_builder::TOOL_ARG
+                                      : common_chat_peg_minimax_m3_mapper::TOOL_ARG_ITEM;
+
+    std::vector<common_peg_ast_id> entries;
+    minimax_m3_collect(arena, node, tag, entries);
+
+    std::string result = is_object ? "{" : "[";
+
+    bool add_comma = false;
+    for (auto entry_id : entries) {
+        const auto & entry = arena.get(entry_id);
+
+        std::string text;
+        if (is_object) {
+            text = minimax_m3_member_to_json(arena, entry);
+        } else {
+            text = minimax_m3_value_to_json(arena, minimax_m3_value_of(arena, entry), !entry.is_partial);
+        }
+
+        if (text.empty()) {
+            continue;
+        }
+
+        if (add_comma) {
+            result += ",";
+        }
+        add_comma = true;
+        result += text;
+    }
+
+    if (closed) {
+        result += is_object ? "}" : "]";
+    }
+    return result;
+}
+
+static std::string minimax_m3_value_to_json(const common_peg_ast_arena & arena, common_peg_ast_id id, bool closed) {
+    if (id == COMMON_PEG_INVALID_AST_ID) {
+        return "";
+    }
+
+    const auto & node = arena.get(id);
+
+    if (node.tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_OBJECT) {
+        return minimax_m3_container_to_json(arena, node, /* is_object = */ true, closed);
+    }
+
+    if (node.tag == common_chat_peg_minimax_m3_mapper::TOOL_ARG_ARRAY) {
+        return minimax_m3_container_to_json(arena, node, /* is_object = */ false, closed);
+    }
+
+    if (node.tag == common_chat_peg_builder::TOOL_ARG_STRING_VALUE) {
+        return "\"" + escape_json_string_inner(std::string(node.text)) + (closed ? "\"" : "");
+    }
+
+    // Numbers and booleans are written verbatim by the template
+    return std::string(node.text);
+}
+
+void common_chat_peg_minimax_m3_mapper::from_ast(const common_peg_ast_arena &    arena,
+                                                 const common_peg_parse_result & result) {
+    for (const auto & node : result.nodes) {
+        visit(arena, node);
+    }
+}
+
+void common_chat_peg_minimax_m3_mapper::visit(const common_peg_ast_arena & arena, common_peg_ast_id id) {
+    const auto & node = arena.get(id);
+
+    if (node.tag == common_chat_peg_builder::REASONING) {
+        result.reasoning_content += std::string(node.text);
+        return;
+    }
+
+    if (node.tag == common_chat_peg_builder::CONTENT) {
+        result.content += std::string(node.text);
+        return;
+    }
+
+    if (node.tag == common_chat_peg_builder::TOOL) {
+        auto name_id = arena.find_by_tag(node, common_chat_peg_builder::TOOL_NAME);
+        if (name_id != COMMON_PEG_INVALID_AST_ID) {
+            common_chat_tool_call call;
+            call.name      = std::string(arena.get(name_id).text);
+            call.arguments = minimax_m3_container_to_json(arena, node, /* is_object = */ true, !node.is_partial);
+            result.tool_calls.push_back(call);
+        }
+        return;
+    }
+
+    for (auto child_id : node.children) {
+        visit(arena, child_id);
+    }
+}

common/chat-peg-parser.h

@@ -40,9 +40,23 @@ class common_chat_peg_gemma4_mapper : public common_chat_peg_mapper {
     void visit(const common_peg_ast_arena & arena, common_peg_ast_id id);
 };
 
+class common_chat_peg_minimax_m3_mapper : public common_chat_peg_mapper {
+  public:
+    static constexpr const char * TOOL_ARG_OBJECT = "tool-arg-object";
+    static constexpr const char * TOOL_ARG_ARRAY  = "tool-arg-array";
+    static constexpr const char * TOOL_ARG_ITEM   = "tool-arg-item";
+
+    common_chat_peg_minimax_m3_mapper(common_chat_msg & msg) : common_chat_peg_mapper(msg) {}
+    virtual void from_ast(const common_peg_ast_arena & arena, const common_peg_parse_result & result);
+  private:
+    void visit(const common_peg_ast_arena & arena, common_peg_ast_id id);
+};
+
 struct content_structure;
 struct tool_call_structure;
 
+constexpr size_t COMMON_CHAT_MAX_PERMUTE = 6;
+
 class common_chat_peg_builder : public common_peg_parser_builder {
   public:
     // Tag constants (from former common_chat_peg_base_builder)
@@ -93,6 +107,9 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return tag(TOOL_ARG_VALUE, p); }
 
 
+    // Matches every parser exactly once, in any order.
+    common_peg_parser permute(const std::string & rule_prefix, const std::vector<common_peg_parser> & parsers);
+
     // Return a parser that parses the prefix of a string, up to a given delimiter.
     common_peg_parser prefix(const std::string & s, const std::string & delimiter = {});
 

common/chat.cpp

@@ -816,6 +816,8 @@ const char * common_chat_format_name(common_chat_format format) {
             return "peg-native";
         case COMMON_CHAT_FORMAT_PEG_GEMMA4:
             return "peg-gemma4";
+        case COMMON_CHAT_FORMAT_PEG_MINIMAX_M3:
+            return "peg-minimax-m3";
         default:
             throw std::runtime_error("Unknown chat format");
     }
@@ -1108,6 +1110,172 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
     return data;
 }
 
+static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_template &          tmpl,
+                                                              const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    const std::string GEN_PREFIX = "<|im_start|>assistant\n";
+
+    data.prompt            = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+
+    auto supports_reasoning = tmpl.source().find("<think>") != std::string::npos;
+
+    data.supports_thinking = supports_reasoning;
+    data.preserved_tokens  = {
+        "<tool_call>",
+        "</tool_call>",
+    };
+
+    if (supports_reasoning) {
+        data.thinking_start_tag = "<think>";
+        // Support both </think> and <tool_call> as reasoning end sequences.
+        // <function= is omitted, as it is a workaround for Qwen3-Coder which is not a thinking model
+        data.thinking_end_tags = { "</think>", "<tool_call>" };
+        data.preserved_tokens.insert(data.preserved_tokens.end(), { "<think>", "</think>" });
+    }
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "<|im_start|>assistant"             },
+        { COMMON_CHAT_ROLE_TOOL,      "<|im_start|>user\n<tool_response>" }, // Qwen3-Coder, Qwen3.5, Nemotron Nano 3
+        { COMMON_CHAT_ROLE_TOOL,      "<|im_start|>tool_response"         }, // StepFun-3.5-Flash
+        { COMMON_CHAT_ROLE_USER,      "<|im_start|>user"                  },
+        { COMMON_CHAT_ROLE_SYSTEM,    "<|im_start|>system"                },
+    };
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = inputs.json_schema.is_object() && !inputs.json_schema.empty();
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PREFIX;
+        if (supports_reasoning) {
+            data.generation_prompt += "<think>\n" + msg.reasoning_content;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += "\n</think>\n\n";
+            }
+        }
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.literal(GEN_PREFIX);
+
+        auto reasoning = p.eps();
+        if (supports_reasoning && extract_reasoning) {
+            reasoning = p.optional("<think>" + p.space() +
+                                   p.reasoning(p.until_one_of({ "</think>", "<tool_call>" })) +
+                                   (p.literal("</think>") | p.peek(p.literal("<tool_call>"))));
+        }
+
+        // Response format parser
+        if (has_response_format) {
+            return generation_prompt + (reasoning << p.content(p.schema(p.json(), "response-format", inputs.json_schema)));
+        }
+
+        // Tool call parser
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto arg_close  = p.tool_arg_close(p.literal("\n</parameter>\n"));
+            auto arg_string = p.rule("xml-arg-string",
+                p.ac(p.tool_arg_string_value(p.until("\n</parameter>\n")) + arg_close, "\n</parameter>\n"));
+
+            auto tool_choice = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function   = tool.at("function");
+                std::string  name       = function.at("name");
+                auto         parameters = function.contains("parameters") ? function.at("parameters") : json::object();
+
+                auto schema_info = common_schema_info();
+                schema_info.resolve_refs(parameters);
+
+                std::vector<common_peg_parser> required_args;
+                std::vector<common_peg_parser> optional_args;
+
+                foreach_parameter(function, [&](const std::string & param_name, const json & param_schema, bool is_required) {
+                    auto rule_name = "tool-" + name + "-arg-" + param_name;
+
+                    auto arg_open = p.tool_arg_open("<parameter=" + p.tool_arg_name(p.literal(param_name)) + ">\n");
+
+                    auto arg_value = schema_info.resolves_to_string(param_schema) ?
+                        arg_string :
+                        p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", param_schema)) + arg_close;
+
+                    auto arg_rule = p.rule(rule_name, p.tool_arg(arg_open + arg_value));
+
+                    (is_required ? required_args : optional_args).push_back(arg_rule);
+                });
+
+                // Accept required arguments in any order, as Qwen does not always adhere to the
+                // order provided.
+                auto args = p.permute("tool-" + name + "-args", required_args);
+                if (!optional_args.empty()) {
+                    args = args + p.zero_or_more(p.choice(optional_args));
+                }
+
+                auto func = p.tool(p.tool_open("<function=" + p.tool_name(p.literal(name)) + ">\n") +
+                                   p.tool_args(args) +
+                                   p.tool_close(p.literal("</function>\n")));
+
+                tool_choice |= p.rule("tool-" + name, func);
+            });
+
+            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+
+            // Qwen3-Coder models may occasionally omit the <tool_call> token.
+            auto tool_call_body  = tool_choice + "</tool_call>" + p.space();
+            auto tool_call_first = p.rule("tool-call-first", p.optional(p.literal("<tool_call>\n")) + tool_call_body);
+            auto tool_call       = p.rule("tool-call", "<tool_call>\n" + tool_call_body);
+
+            auto calls      = inputs.parallel_tool_calls ? tool_call_first + p.zero_or_more(tool_call) : tool_call_first;
+            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(calls, min_calls, 1));
+
+            return generation_prompt +
+                   (reasoning << p.content(p.until_one_of({ "<tool_call>", "<function=" })) << tool_calls);
+        }
+
+        // Content only parser
+        return generation_prompt + (reasoning << p.content(p.rest()));
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        if (data.grammar_lazy) {
+            data.grammar_triggers = {
+                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<tool_call>" },
+                // Trigger on "<function" and not "<function=" because the trailing "=" is part of
+                // the token with the function name e.g. "=read"
+                { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function"   },
+            };
+        }
+    }
+
+    return data;
+}
+
 static common_chat_params common_chat_params_init_gpt_oss(const common_chat_template &    tmpl,
                                                           const autoparser::generation_params & inputs) {
     common_chat_params data;
@@ -1941,23 +2109,16 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
         adjusted_messages = deepseek_v4_sort_tool_results(inputs.messages);
     }
 
-    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs, adjusted_messages);
-    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs, adjusted_messages);
-    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking  = true;
-    data.thinking_start_tag = "<think>";
-    data.thinking_end_tags  = {"</think>"};
-    data.preserved_tokens   = {
-        "｜DSML｜",
-        "<think>",
-        "</think>",
-    };
-
     auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
     auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
     auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
     auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
 
+    std::optional<json> additional_context;
+    if (is_v4 && has_response_format) {
+        additional_context = json{ { "response_format", inputs.json_schema } };
+    }
+
     const std::string DSML         = "｜DSML｜";
     const std::string THINK_START  = "<think>";
     const std::string THINK_END    = "</think>";
@@ -1969,25 +2130,137 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
     const std::string PARAM_START  = "<" + DSML + "parameter";
     const std::string PARAM_END    = "</" + DSML + "parameter>";
     const std::string GEN_PROMPT   = "<｜Assistant｜>";
+    const std::string TC_SEPARATOR = "\n\n";
+
+    data.prompt = common_chat_template_direct_apply_impl(
+        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(
+        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
+    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = THINK_START;
+    data.thinking_end_tags  = {THINK_END, FC_START};
+    data.preserved_tokens   = {
+        DSML,
+        THINK_START,
+        THINK_END,
+    };
 
     if (inputs.has_continuation()) {
         const auto & msg = inputs.continue_msg;
 
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
+        if (is_v4 && msg.reasoning_content.empty()) {
+            data.generation_prompt = GEN_PROMPT + THINK_END;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += msg.render_content();
+            }
+        } else {
+            data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += THINK_END + msg.render_content();
+            }
         }
 
         data.prompt += data.generation_prompt;
     }
 
+    bool require_tools   = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+    bool has_tool_calls = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
         auto generation_prompt = p.literal(GEN_PROMPT);
-        auto end = p.end();
+        auto end               = p.end();
+
+        // build tool call section first since we might need it in reasoning
+        auto tool_choice = p.choice();
+        if (has_tool_calls) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string  name     = function.at("name");
+                auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
+                const auto & props    = params.contains("properties") ? params.at("properties") : json::object();
+
+                std::set<std::string> required;
+                if (params.contains("required")) {
+                    params.at("required").get_to(required);
+                }
+
+                auto schema_info = common_schema_info();
+                schema_info.resolve_refs(params);
+
+                std::vector<common_peg_parser> required_parsers;
+                std::vector<common_peg_parser> optional_parsers;
+                for (const auto & [param_name, param_schema] : props.items()) {
+                    bool is_required = required.find(param_name) != required.end();
+                    bool is_string   = schema_info.resolves_to_string(param_schema);
+
+                    auto arg = p.tool_arg(
+                        p.tool_arg_open(p.literal(PARAM_START + " name=\"") + p.tool_arg_name(p.literal(param_name)) +
+                                        p.literal("\" string=\"" + std::string(is_string ? "true" : "false") + "\">")) +
+                        (is_string ?
+                             p.tool_arg_string_value(p.until(PARAM_END)) :
+                             p.tool_arg_json_value(p.schema(p.json(), "tool-" + name + "-arg-" + param_name + "-schema",
+                                                            param_schema, false))) +
+                        p.tool_arg_close(p.literal(PARAM_END)));
+
+                    auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
+                    if (is_required) {
+                        required_parsers.push_back(named_arg);
+                    } else {
+                        optional_parsers.push_back(named_arg);
+                    }
+                }
+
+                common_peg_parser args_seq = p.eps();
+                for (size_t i = 0; i < required_parsers.size(); i++) {
+                    if (i > 0) {
+                        args_seq = args_seq + p.space();
+                    }
+                    args_seq = args_seq + required_parsers[i];
+                }
+
+                if (!optional_parsers.empty()) {
+                    common_peg_parser any_opt = p.choice();
+                    for (const auto & opt : optional_parsers) {
+                        any_opt |= opt;
+                    }
+                    args_seq = args_seq + p.repeat(p.space() + any_opt, 0, -1);
+                }
+
+                common_peg_parser invoke_body = args_seq;
+                auto              func_parser = p.tool(p.tool_open(p.literal(INVOKE_START + " name=\"") +
+                                                                   p.tool_name(p.literal(name)) + p.literal("\">\n")) +
+                                                       invoke_body + p.space() + p.tool_close(p.literal(INVOKE_END)));
+
+                tool_choice |= p.rule("tool-" + name, func_parser);
+            });
+        }
+
+        common_peg_parser tool_calls = p.eps();
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice +
+                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
+        } else {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
+        }
 
         auto reasoning = p.eps();
+        auto reasoning_with_tc = p.eps();
+        auto obligatory_tool_calls = tool_calls;
+        bool allow_reasoning_with_tc = false;
+
+        if (!require_tools) {
+            tool_calls = p.optional(tool_calls);
+        }
+
         if (extract_reasoning && inputs.enable_thinking) {
             reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
+            reasoning_with_tc = THINK_START +
+                p.reasoning(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START, THINK_END })) +
+                p.space() + obligatory_tool_calls;
+            allow_reasoning_with_tc = true;
         } else if (extract_reasoning) {
             // Thinking disabled but reasoning extraction requested: the generation prompt
             // contains an empty <think></think> pair (V3.2) or a bare </think> (V4) that
@@ -2005,101 +2278,21 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
             return generation_prompt + reasoning + response_format + end;
         }
 
-        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+        if (!has_tool_calls) {
             return generation_prompt + reasoning + p.content(p.rest()) + end;
         }
 
-        auto tool_choice = p.choice();
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string  name     = function.at("name");
-            auto params   = function.contains("parameters") ? function.at("parameters") : json::object();
-            const auto & props    = params.contains("properties") ? params.at("properties") : json::object();
-
-            std::set<std::string> required;
-            if (params.contains("required")) {
-                params.at("required").get_to(required);
-            }
-
-            auto schema_info = common_schema_info();
-            schema_info.resolve_refs(params);
-
-            std::vector<common_peg_parser> required_parsers;
-            std::vector<common_peg_parser> optional_parsers;
-            for (const auto & [param_name, param_schema] : props.items()) {
-                bool is_required = required.find(param_name) != required.end();
-                bool is_string   = schema_info.resolves_to_string(param_schema);
-
-                auto arg = p.tool_arg(
-                    p.tool_arg_open(
-                        p.literal(PARAM_START + " name=\"") +
-                        p.tool_arg_name(p.literal(param_name)) +
-                        p.literal("\" string=\"" + std::string(is_string ? "true" : "false") + "\">")) +
-                    (is_string
-                         ? p.tool_arg_string_value(p.until(PARAM_END))
-                         : p.tool_arg_json_value(p.schema(p.json(),
-                                                          "tool-" + name + "-arg-" + param_name + "-schema",
-                                                          param_schema, false))) +
-                    p.tool_arg_close(p.literal(PARAM_END)));
-
-                auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
-                if (is_required) {
-                    required_parsers.push_back(named_arg);
-                } else {
-                    optional_parsers.push_back(named_arg);
-                }
-            }
-
-            common_peg_parser args_seq = p.eps();
-            for (size_t i = 0; i < required_parsers.size(); i++) {
-                if (i > 0) {
-                    args_seq = args_seq + p.space();
-                }
-                args_seq = args_seq + required_parsers[i];
-            }
-
-            if (!optional_parsers.empty()) {
-                common_peg_parser any_opt = p.choice();
-                for (const auto & opt : optional_parsers) {
-                    any_opt |= opt;
-                }
-                args_seq = args_seq + p.repeat(p.space() + any_opt, 0, -1);
-            }
-
-            common_peg_parser invoke_body = args_seq;
-            auto func_parser = p.tool(
-                p.tool_open(p.literal(INVOKE_START + " name=\"") +
-                            p.tool_name(p.literal(name)) + p.literal("\">\n")) +
-                invoke_body + p.space() +
-                p.tool_close(p.literal(INVOKE_END)));
-
-            tool_choice |= p.rule("tool-" + name, func_parser);
-        });
-
-        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-        common_peg_parser tool_calls = p.eps();
-        if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call",
-                p.literal(FC_START) + p.space() + tool_choice +
-                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
-        } else {
-            tool_calls = p.trigger_rule("tool-call",
-                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
-        }
-
-        if (!require_tools) {
-            tool_calls = p.optional(tool_calls);
-        }
-
-        auto content_before_tools = p.content(p.until(FC_START));
-        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
+        auto content_before_tools = p.negate(p.literal(THINK_START)) +
+            p.content(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START })) +
+            p.space();
+        return allow_reasoning_with_tc ? generation_prompt + (reasoning_with_tc | (reasoning + content_before_tools + tool_calls)) + end :
+            generation_prompt + reasoning + content_before_tools + tool_calls + end;
     });
 
     data.parser = parser.save();
 
     if (include_grammar) {
-        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar_lazy = has_tools && !require_tools;
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
             foreach_function(inputs.tools, [&](const json & tool) {
                 const auto & function = tool.at("function");
@@ -2270,6 +2463,264 @@ static common_chat_params common_chat_params_init_cohere2moe(const common_chat_t
     return data;
 }
 
+static common_chat_params common_chat_params_init_minimax_m3(const common_chat_template &          tmpl,
+                                                             const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs);
+    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs);
+    data.format             = COMMON_CHAT_FORMAT_PEG_MINIMAX_M3;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = "<mm:think>";
+    data.thinking_end_tags  = {"</mm:think>"};
+
+    // M3 prefixes every tool tag with the namespace token "]<]minimax[>[";
+    // params use the parameter name as the tag (<file_path>...</file_path>).
+    const std::string NS          = "]<]minimax[>[";
+    const std::string THINK_START = "<mm:think>";
+    const std::string THINK_END   = "</mm:think>";
+    const std::string FC_START    = NS + "<tool_call>";
+    const std::string FC_END      = NS + "</tool_call>";
+    const std::string INVOKE_END  = NS + "</invoke>";
+
+    data.preserved_tokens = {
+        NS,
+        "<tool_call>",
+        "</tool_call>",
+        THINK_START,
+        THINK_END,
+    };
+
+    data.message_delimiters = {
+        { COMMON_CHAT_ROLE_ASSISTANT, "]~b]ai"        },
+        { COMMON_CHAT_ROLE_USER,      "]~b]user"      },
+        { COMMON_CHAT_ROLE_TOOL,      "]~b]tool"      },
+        { COMMON_CHAT_ROLE_SYSTEM,    "]~b]developer" },
+        { COMMON_CHAT_ROLE_SYSTEM,    "]~b]system"    },
+    };
+
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+
+    const std::string GEN_PROMPT = data.generation_prompt;
+
+    using mm3 = common_chat_peg_minimax_m3_mapper;
+
+    if (inputs.has_continuation()) {
+        const auto & msg = inputs.continue_msg;
+
+        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+            data.generation_prompt += THINK_END + msg.render_content();
+        }
+
+        data.prompt += data.generation_prompt;
+    }
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.prefix(GEN_PROMPT, THINK_START);
+        auto end = p.end();
+
+        auto reasoning = p.eps();
+        if (extract_reasoning) {
+            auto block = inputs.enable_thinking
+                             ? p.literal(THINK_START) + p.space() +
+                                   p.ac(p.reasoning(p.until(THINK_END)) + p.literal(THINK_END), THINK_END)
+                             : p.literal(THINK_START) + p.ac(p.until(THINK_END) + p.literal(THINK_END), THINK_END);
+
+            // A turn without reasoning is prefixed with a bare </mm:think>, written either by the
+            // generation prompt (thinking_mode = "disabled") or by the model itself.
+            reasoning = p.optional(p.choice({ block, p.literal(THINK_END) }));
+        }
+
+        if (has_response_format) {
+            auto response_format = p.rule("response-format",
+                p.literal("```json") + p.space() +
+                p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)) +
+                p.space() + p.literal("```"));
+            return generation_prompt + reasoning + response_format + end;
+        }
+
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            return generation_prompt + reasoning + p.content(p.rest()) + end;
+        }
+
+        auto alternatives_of = [](const json & schema) -> std::optional<json> {
+            for (const auto * keyword : { "oneOf", "anyOf" }) {
+                if (schema.contains(keyword) && schema.at(keyword).is_array() && !schema.at(keyword).empty()) {
+                    return schema.at(keyword);
+                }
+            }
+            return std::nullopt;
+        };
+
+        auto tool_choice = p.choice();
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const auto & function = tool.at("function");
+            std::string  name     = function.at("name");
+            auto         params   = function.contains("parameters") ? function.at("parameters") : json::object();
+
+            auto schema_info = common_schema_info();
+            schema_info.resolve_refs(params);
+
+            // The template expands argument values recursively in XML (see the to_xml() macro)
+            std::function<common_peg_parser(const json &, const std::string &, const std::string &)> value_of;
+            std::function<common_peg_parser(const json &, const std::string &)>                      members_of;
+
+            auto element_of = [&](const std::string & tag, const json & schema, const std::string & rule_name) {
+                const std::string close = NS + "</" + tag + ">";
+                return p.rule(rule_name,
+                    p.tool_arg(
+                        p.tool_arg_open(
+                            p.literal(NS + "<") +
+                            p.tool_arg_name(p.literal(tag)) +
+                            p.literal(">")) +
+                        value_of(schema, rule_name, close)));
+            };
+
+            value_of = [&](const json & schema,
+                           const std::string & rule_name,
+                           const std::string & close) -> common_peg_parser {
+                auto close_tag = p.tool_arg_close(p.literal(close));
+
+                // A string accepts anything, so a union with a string alternative is a string
+                if (schema_info.resolves_to_string(schema)) {
+                    return p.ac(p.tool_arg_string_value(p.until(close)) + close_tag, close);
+                }
+
+                if (auto alternatives = alternatives_of(schema)) {
+                    std::vector<common_peg_parser> choices;
+
+                    size_t index = 0;
+                    for (const auto & alternative : *alternatives) {
+                        const std::string alt_name = rule_name + "-" + std::to_string(index++);
+
+                        // There is a risk that this breaks streaming deltas, but that's a risk we
+                        // assume to provide tool arg streaming.
+                        choices.push_back(value_of(alternative, alt_name, close));
+                    }
+
+                    return p.choice(choices);
+                }
+
+                const std::string type = schema.contains("type") && schema.at("type").is_string()
+                                             ? schema.at("type").get<std::string>()
+                                             : "";
+
+                if (type == "object" && schema.contains("properties")) {
+                    return p.tag(mm3::TOOL_ARG_OBJECT, members_of(schema, rule_name)) + p.space() + close_tag;
+                }
+
+                if (type == "array" && schema.contains("items")) {
+                    const std::string item_close = NS + "</item>";
+                    auto item = p.rule(rule_name + "-item",
+                        p.tag(mm3::TOOL_ARG_ITEM,
+                              p.literal(NS + "<item>") +
+                                  value_of(schema.at("items"), rule_name + "-item", item_close)));
+                    return p.tag(mm3::TOOL_ARG_ARRAY, p.repeat(p.space() + item, 0, -1)) + p.space() + close_tag;
+                }
+
+                return p.tool_arg_json_value(p.schema(p.json(), rule_name + "-schema", schema, false)) + close_tag;
+            };
+
+            // Required properties in schema order, then any number of optional ones in any order.
+            members_of = [&](const json & schema, const std::string & rule_prefix) -> common_peg_parser {
+                const auto & props = schema.at("properties");
+
+                std::set<std::string> required;
+                if (schema.contains("required")) {
+                    schema.at("required").get_to(required);
+                }
+
+                std::vector<common_peg_parser> required_elements;
+                std::vector<common_peg_parser> optional_elements;
+                for (const auto & [key, key_schema] : props.items()) {
+                    auto element = element_of(key, key_schema, rule_prefix + "-" + key);
+                    if (required.find(key) != required.end()) {
+                        required_elements.push_back(element);
+                    } else {
+                        optional_elements.push_back(element);
+                    }
+                }
+
+                common_peg_parser members = p.eps();
+                for (size_t i = 0; i < required_elements.size(); i++) {
+                    if (i > 0) {
+                        members = members + p.space();
+                    }
+                    members = members + required_elements[i];
+                }
+
+                if (!optional_elements.empty()) {
+                    common_peg_parser any_optional = p.choice();
+                    for (const auto & element : optional_elements) {
+                        any_optional |= element;
+                    }
+                    members = members + p.repeat(p.space() + any_optional, 0, -1);
+                }
+
+                return members;
+            };
+
+            common_peg_parser invoke_body =
+                params.contains("properties") ? members_of(params, "tool-" + name + "-arg") : p.eps();
+
+            auto func_parser = p.tool(
+                p.tool_open(p.literal(NS + "<invoke name=\"") +
+                            p.tool_name(p.literal(name)) + p.literal("\">")) +
+                p.space() + invoke_body + p.space() +
+                p.tool_close(p.literal(INVOKE_END)));
+
+            tool_choice |= p.rule("tool-" + name, func_parser);
+        });
+
+        auto require_tools = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+
+        common_peg_parser tool_calls = p.eps();
+        if (inputs.parallel_tool_calls) {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice +
+                p.zero_or_more(p.space() + tool_choice) + p.space() + p.literal(FC_END));
+        } else {
+            tool_calls = p.trigger_rule("tool-call",
+                p.literal(FC_START) + p.space() + tool_choice + p.space() + p.literal(FC_END));
+        }
+
+        if (!require_tools) {
+            tool_calls = p.optional(tool_calls);
+        }
+
+        auto content_before_tools = p.content(p.until(FC_START));
+        return generation_prompt + reasoning + content_before_tools + tool_calls + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
+                builder.resolve_refs(schema);
+            });
+            if (has_response_format) {
+                auto schema = inputs.json_schema;
+                builder.resolve_refs(schema);
+            }
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, FC_START },
+        };
+    }
+
+    return data;
+}
+
 namespace workaround {
 
 static void map_developer_role_to_system(json & messages) {
@@ -2707,6 +3158,15 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_gigachat_v3(tmpl, params);
     }
 
+    // MiniMax-M3: the namespace token "]<]minimax[>[" collides with the autoparser's
+    // markup delimiters, so detect the template and use a dedicated parser.
+    if (src.find("]<]minimax[>[") != std::string::npos &&
+        src.find("<tool_call>") != std::string::npos &&
+        src.find("<invoke name=") != std::string::npos) {
+        LOG_DBG("Using specialized template: MiniMax-M3\n");
+        return common_chat_params_init_minimax_m3(tmpl, params);
+    }
+
     // DeepSeek V3.2/V4 format detection: template defines dsml_token and uses it for tool calls.
     // The template source contains the token as a variable assignment, not as a literal in markup.
     // V3.2 names the tool call block "function_calls", V4 names it "tool_calls".
@@ -2737,6 +3197,14 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_minicpm5(tmpl, params);
     }
 
+    // Qwen3-Coder XML tool calls, also used by Nemotron Nano 3, Qwen3.5 and StepFun-3.5-Flash
+    if (src.find("<tool_call>") != std::string::npos &&
+        src.find("<function=") != std::string::npos &&
+        src.find("<parameter=") != std::string::npos) {
+        LOG_DBG("Using specialized template: Qwen3-Coder\n");
+        return common_chat_params_init_qwen3_coder(tmpl, params);
+    }
+
     return std::nullopt;
 }
 
@@ -2998,6 +3466,8 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena &          src_pars
             std::unique_ptr<common_chat_peg_mapper> mapper;
             if (params.format == COMMON_CHAT_FORMAT_PEG_GEMMA4) {
                 mapper = std::make_unique<common_chat_peg_gemma4_mapper>(msg);
+            } else if (params.format == COMMON_CHAT_FORMAT_PEG_MINIMAX_M3) {
+                mapper = std::make_unique<common_chat_peg_minimax_m3_mapper>(msg);
             } else {
                 mapper = std::make_unique<common_chat_peg_mapper>(msg);
             }
@@ -3020,6 +3490,8 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena &          src_pars
     std::unique_ptr<common_chat_peg_mapper> mapper;
     if (params.format == COMMON_CHAT_FORMAT_PEG_GEMMA4) {
         mapper = std::make_unique<common_chat_peg_gemma4_mapper>(msg);
+    } else if (params.format == COMMON_CHAT_FORMAT_PEG_MINIMAX_M3) {
+        mapper = std::make_unique<common_chat_peg_minimax_m3_mapper>(msg);
     } else {
         mapper = std::make_unique<common_chat_peg_mapper>(msg);
     }

common/chat.h

@@ -233,6 +233,7 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_PEG_SIMPLE,
     COMMON_CHAT_FORMAT_PEG_NATIVE,
     COMMON_CHAT_FORMAT_PEG_GEMMA4,
+    COMMON_CHAT_FORMAT_PEG_MINIMAX_M3,
 
     COMMON_CHAT_FORMAT_COUNT,  // Not a format, just the # formats
 };

common/common.cpp

@@ -998,6 +998,23 @@ bool fs_is_directory(const std::string & path) {
     return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
 }
 
+std::string common_get_env(const std::string & name) {
+    const char * value = std::getenv(name.c_str());
+    return value == nullptr ? "" : value;
+}
+
+void common_set_env(const std::string & name, const std::string & value) {
+#if defined(_WIN32)
+    _putenv_s(name.c_str(), value.c_str());
+#else
+    if (value.empty()) {
+        unsetenv(name.c_str());
+    } else {
+        setenv(name.c_str(), value.c_str(), 1);
+    }
+#endif
+}
+
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
@@ -1299,8 +1316,9 @@ common_init_result::common_init_result(common_params & params, bool model_only)
     pimpl->samplers.resize(cparams.n_seq_max);
     pimpl->samplers_seq_config.resize(cparams.n_seq_max);
 
+    const int32_t n_ctx = cparams.n_ctx > 0 ? (int32_t) cparams.n_ctx : llama_model_n_ctx_train(model);
     for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
-        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
+        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling, n_ctx));
         pimpl->samplers_seq_config[i] = { i, common_sampler_get(pimpl->samplers[i].get()) };
     }
 
@@ -1462,18 +1480,32 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
 common_init_result::~common_init_result() = default;
 
 std::string common_get_model_endpoint() {
-    const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
-    // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
-    const char * hf_endpoint_env = getenv("HF_ENDPOINT");
-    const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
-    std::string model_endpoint = "https://huggingface.co/";
-    if (endpoint_env) {
-        model_endpoint = endpoint_env;
-        if (model_endpoint.back() != '/') {
-            model_endpoint += '/';
-        }
+    std::string endpoint = common_get_env("MODEL_ENDPOINT");
+    if (endpoint.empty()) {
+        // the HF_ENDPOINT variable is respected for backward compatibility
+        endpoint = common_get_env("HF_ENDPOINT");
     }
-    return model_endpoint;
+    if (endpoint.empty()) {
+        return "https://huggingface.co/";
+    }
+    if (endpoint.back() != '/') {
+        endpoint += '/';
+    }
+    return endpoint;
+}
+
+char * common_get_model_or_exit(int argc, char * argv[]) {
+    if (argc > 1) {
+        return argv[1];
+    }
+
+    char * path = getenv("LLAMACPP_TEST_MODELFILE");
+    if (!path || strlen(path) == 0) {
+        fprintf(stderr, "\033[33mWARNING: No model file provided. Skipping this test. Set LLAMACPP_TEST_MODELFILE=<gguf_model_path> to silence this warning and run this test.\n\033[0m");
+        exit(EXIT_SUCCESS);
+    }
+
+    return path;
 }
 
 common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx) {
@@ -1518,23 +1550,49 @@ done:
     return res;
 }
 
-void common_context_seq_rm(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+static void common_context_seq_rm(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
     auto * mem = llama_get_memory(ctx);
     if (!llama_memory_seq_rm(mem, seq_id, p0, p1)) {
         GGML_ABORT("%s", string_format("failed to remove sequence %d with p0=%d, p1=%d\n", seq_id, p0, p1).c_str());
     }
 }
 
-void common_context_seq_cp(llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+static void common_context_seq_cp(llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
     auto * mem = llama_get_memory(ctx);
     llama_memory_seq_cp(mem, seq_id_src, seq_id_dst, p0, p1);
 }
 
-void common_context_seq_add(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
+static void common_context_seq_add(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
     auto * mem = llama_get_memory(ctx);
     llama_memory_seq_add(mem, seq_id, p0, p1, delta);
 }
 
+void common_memory::init(llama_context * ctx_tgt, llama_context * ctx_dft) {
+    this->ctx_tgt = ctx_tgt;
+    this->ctx_dft = ctx_dft;
+}
+
+void common_memory::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) const {
+    common_context_seq_rm(ctx_tgt, seq_id, p0, p1);
+    if (ctx_dft) {
+        common_context_seq_rm(ctx_dft, seq_id, p0, p1);
+    }
+}
+
+void common_memory::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) const {
+    common_context_seq_cp(ctx_tgt, seq_id_src, seq_id_dst, p0, p1);
+    if (ctx_dft) {
+        common_context_seq_cp(ctx_dft, seq_id_src, seq_id_dst, p0, p1);
+    }
+}
+
+void common_memory::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) const {
+    common_context_seq_add(ctx_tgt, seq_id, p0, p1, delta);
+    if (ctx_dft) {
+        common_context_seq_add(ctx_dft, seq_id, p0, p1, delta);
+    }
+}
+
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
     std::vector<llama_adapter_lora *> loras;
     std::vector<float> scales;
@@ -1580,6 +1638,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.progress_callback           = params.load_progress_callback;
     mparams.progress_callback_user_data = params.load_progress_callback_user_data;
     mparams.no_alloc                    = params.no_alloc;
+    mparams.load_mtp                    = std::find(params.speculative.types.begin(), params.speculative.types.end(), COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
 
     return mparams;
 }

common/common.h

@@ -173,6 +173,7 @@ enum common_speculative_type {
     COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3,  // Eagle3 speculative decoding
     COMMON_SPECULATIVE_TYPE_DRAFT_MTP,     // Multi-token prediction
     COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH,  // DFlash speculative decoding
+    COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK,  // DSpark speculative decoding (DFlash + Markov head)
     COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE,  // simple self-speculative decoding based on n-grams
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K,   // self-speculative decoding with n-gram keys only
     COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, // self-speculative decoding with n-gram keys and 4 m-gram values
@@ -293,10 +294,6 @@ struct common_params_sampling {
 
     bool backend_sampling = false;
 
-    bool has_logit_bias() const {
-        return !logit_bias.empty();
-    }
-
     // print the parameters into a string
     std::string print() const;
 };
@@ -388,7 +385,7 @@ struct common_params_speculative {
 
     uint32_t need_n_rs_seq() const {
         bool needs_rs_seq = std::any_of(types.begin(), types.end(), [&](auto t) {
-            return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3 || t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH;
+            return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3 || t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH || t == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK;
         });
 
         return needs_rs_seq ? draft.n_max : 0u;
@@ -742,6 +739,8 @@ struct common_params {
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
     bool no_alloc = false; // Don't allocate model buffers
+
+    bool is_gen_docs = false; // whether we are running inside llama-gen-docs
 };
 
 // call once at the start of a program if it uses libcommon
@@ -866,6 +865,15 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 
 bool glob_match(const std::string & pattern, const std::string & str);
 
+//
+// Environment utils
+//
+
+// portable environment access, an unset variable reads as an empty string
+// and setting an empty value unsets the variable
+std::string common_get_env(const std::string & name);
+void        common_set_env(const std::string & name, const std::string & value);
+
 //
 // Filesystem utils
 //
@@ -933,6 +941,9 @@ void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adap
 // model endpoint from env
 std::string common_get_model_endpoint();
 
+// for testing purposes
+char * common_get_model_or_exit(int, char*[]);
+
 //
 // Context utils
 //
@@ -948,10 +959,17 @@ enum common_context_seq_rm_type {
 // note: clears the memory of the context
 common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx);
 
-// aborts execution on failure
-void common_context_seq_rm (llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1);
-void common_context_seq_add(llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta);
-void common_context_seq_cp (llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1);
+struct common_memory {
+    llama_context * ctx_tgt = nullptr;
+    llama_context * ctx_dft = nullptr;
+
+    void init(llama_context * ctx_tgt, llama_context * ctx_dft = nullptr);
+
+    // aborts execution on failure
+    void seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) const;
+    void seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) const;
+    void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) const;
+};
 
 //
 // Batch utils

common/download.cpp

@@ -568,16 +568,30 @@ static hf_cache::hf_files get_split_files(const hf_cache::hf_files & files,
 }
 
 // pick the best sibling GGUF whose filename contains `keyword` (e.g. "mmproj" / "mtp"),
-// preferring deeper shared directory prefix with the model, then closest quantization
+// preferring deeper shared directory prefix with the model, then exact `tag` match,
+// then closest quantization to the tag when given, or to the model otherwise
 static hf_cache::hf_file find_best_sibling(const hf_cache::hf_files & files,
                                            const std::string        & model,
-                                           const std::string        & keyword) {
+                                           const std::string        & keyword,
+                                           const std::string        & tag = "") {
     hf_cache::hf_file best;
     size_t best_depth = 0;
     int best_diff = 0;
+    bool best_exact = false;
     bool found = false;
 
-    auto model_bits = extract_quant_bits(model);
+    std::string tag_upper = tag;
+    for (char & c : tag_upper) {
+        c = (char) std::toupper((unsigned char) c);
+    }
+
+    int model_bits = 0;
+    if (!tag_upper.empty()) {
+        auto pos = tag_upper.find_first_of("0123456789");
+        model_bits = pos == std::string::npos ? 0 : std::stoi(tag_upper.substr(pos));
+    } else {
+        model_bits = extract_quant_bits(model);
+    }
     auto model_parts = string_split<std::string>(model, '/');
     auto model_dir = model_parts.end() - 1;
 
@@ -600,10 +614,19 @@ static hf_cache::hf_file find_best_sibling(const hf_cache::hf_files & files,
         auto bits = extract_quant_bits(f.path);
         auto diff = std::abs(bits - model_bits);
 
-        if (!found || depth > best_depth || (depth == best_depth && diff < best_diff)) {
+        std::string path_upper = f.path;
+        for (char & c : path_upper) {
+            c = (char) std::toupper((unsigned char) c);
+        }
+        bool exact = !tag_upper.empty() && path_upper.find("-" + tag_upper + ".") != std::string::npos;
+
+        if (!found || depth > best_depth ||
+            (depth == best_depth && exact && !best_exact) ||
+            (depth == best_depth && exact == best_exact && diff < best_diff)) {
             best = f;
             best_depth = depth;
             best_diff = diff;
+            best_exact = exact;
             found = true;
         }
     }
@@ -616,18 +639,27 @@ static hf_cache::hf_file find_best_mmproj(const hf_cache::hf_files & files,
 }
 
 static hf_cache::hf_file find_best_mtp(const hf_cache::hf_files & files,
-                                       const std::string        & model) {
-    return find_best_sibling(files, model, "mtp-");
+                                       const std::string        & model,
+                                       const std::string        & tag = "") {
+    return find_best_sibling(files, model, "mtp-", tag);
 }
 
 static hf_cache::hf_file find_best_eagle3(const hf_cache::hf_files & files,
-                                          const std::string        & model) {
-    return find_best_sibling(files, model, "eagle3-");
+                                          const std::string        & model,
+                                          const std::string        & tag = "") {
+    return find_best_sibling(files, model, "eagle3-", tag);
 }
 
 static hf_cache::hf_file find_best_dflash(const hf_cache::hf_files & files,
-                                          const std::string        & model) {
-    return find_best_sibling(files, model, "dflash-");
+                                          const std::string        & model,
+                                          const std::string        & tag = "") {
+    return find_best_sibling(files, model, "dflash-", tag);
+}
+
+static hf_cache::hf_file find_best_dspark(const hf_cache::hf_files & files,
+                                          const std::string        & model,
+                                          const std::string        & tag = "") {
+    return find_best_sibling(files, model, "dspark-", tag);
 }
 
 static bool gguf_filename_is_model(const std::string & filepath) {
@@ -644,7 +676,8 @@ static bool gguf_filename_is_model(const std::string & filepath) {
            filename.find("imatrix") == std::string::npos &&
            filename.find("mtp-")    == std::string::npos &&
            filename.find("eagle3-") == std::string::npos &&
-           filename.find("dflash-") == std::string::npos;
+           filename.find("dflash-") == std::string::npos &&
+           filename.find("dspark-") == std::string::npos;
 }
 
 static hf_cache::hf_file find_best_model(const hf_cache::hf_files & files,
@@ -736,27 +769,39 @@ common_download_hf_plan common_download_get_hf_plan(const common_params_model &
         }
     } else {
         primary = find_best_model(all, tag);
-        if (primary.path.empty()) {
+        // a requested sidecar can resolve on its own, without a full model of the same tag
+        if (primary.path.empty() && !opts.download_mtp && !opts.download_dflash && !opts.download_eagle3 && !opts.download_dspark) {
             LOG_ERR("%s: no GGUF files found in repository %s\n", __func__, repo.c_str());
             list_available_gguf_files(all);
             return plan;
         }
     }
 
-    plan.primary = primary;
-    plan.model_files = get_split_files(all, primary);
+    if (!primary.path.empty()) {
+        plan.primary = primary;
+        plan.model_files = get_split_files(all, primary);
+    }
 
-    if (opts.download_mmproj) {
+    if (opts.download_mmproj && !primary.path.empty()) {
         plan.mmproj = find_best_mmproj(all, primary.path);
     }
     if (opts.download_mtp) {
-        plan.mtp = find_best_mtp(all, primary.path);
+        plan.mtp = find_best_mtp(all, primary.path, tag);
     }
     if (opts.download_dflash) {
-        plan.dflash = find_best_dflash(all, primary.path);
+        plan.dflash = find_best_dflash(all, primary.path, tag);
     }
     if (opts.download_eagle3) {
-        plan.eagle3 = find_best_eagle3(all, primary.path);
+        plan.eagle3 = find_best_eagle3(all, primary.path, tag);
+    }
+    if (opts.download_dspark) {
+        plan.dspark = find_best_dspark(all, primary.path, tag);
+    }
+
+    if (primary.path.empty() &&
+        plan.mtp.local_path.empty() && plan.dflash.local_path.empty() && plan.eagle3.local_path.empty() && plan.dspark.local_path.empty()) {
+        LOG_ERR("%s: no GGUF files found in repository %s\n", __func__, repo.c_str());
+        list_available_gguf_files(all);
     }
 
     return plan;
@@ -932,7 +977,8 @@ std::vector<common_cached_model_info> common_list_cached_models() {
             split.prefix.find("mmproj")  != std::string::npos ||
             split.prefix.find("mtp-")    != std::string::npos ||
             split.prefix.find("eagle3-") != std::string::npos ||
-            split.prefix.find("dflash-") != std::string::npos) {
+            split.prefix.find("dflash-") != std::string::npos ||
+            split.prefix.find("dspark-") != std::string::npos) {
             continue;
         }
         if (seen.insert(f.repo_id + ":" + split.tag).second) {

common/download.h

@@ -59,6 +59,7 @@ struct common_download_opts {
     bool download_mtp     = false;
     bool download_eagle3  = false;
     bool download_dflash  = false;
+    bool download_dspark  = false;
     common_download_callback * callback = nullptr;
 };
 
@@ -110,6 +111,7 @@ struct common_download_hf_plan {
     hf_cache::hf_file mtp;
     hf_cache::hf_file eagle3;
     hf_cache::hf_file dflash;
+    hf_cache::hf_file dspark;
     hf_cache::hf_file preset; // if set, only this file is downloaded
 };
 common_download_hf_plan common_download_get_hf_plan(const common_params_model & model, const common_download_opts & opts);

common/fit.cpp

@@ -136,7 +136,7 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
         devs.push_back(llama_model_get_device(model, i));
     }
 
-    hp_ngl         = llama_model_n_layer(model);
+    hp_ngl         = llama_model_n_layer(model) + llama_model_n_layer_nextn(model);
     hp_n_ctx_train = llama_model_n_ctx_train(model);
     hp_n_expert    = llama_model_n_expert(model);
 

common/jinja/caps.cpp

@@ -482,6 +482,7 @@ caps caps_get(jinja::program & prog) {
             });
         },
         [&](context & ctx) {
+            ctx.set_val("enable_thinking", mk_val<value_bool>(true));
             caps_apply_preserve_reasoning(ctx, true);
         },
         nullptr, // tools_fn

common/sampling.cpp

@@ -184,9 +184,26 @@ std::string common_params_sampling::print() const {
     return std::string(result);
 }
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params) {
-    const llama_vocab * vocab = llama_model_get_vocab(model);
+struct common_sampler * common_sampler_init(
+        const struct llama_model * model,
+        struct common_params_sampling & params,
+        int32_t n_ctx) {
+    if (!std::isfinite(params.penalty_repeat) ||
+        params.penalty_repeat <= 0.0f ||
+        !std::isfinite(1.0f/params.penalty_repeat)) {
+        throw std::invalid_argument("penalty_repeat must be finite and greater than 0");
+    }
+    if (!std::isfinite(params.penalty_freq)) {
+        throw std::invalid_argument("penalty_freq must be finite");
+    }
+    if (!std::isfinite(params.penalty_present)) {
+        throw std::invalid_argument("penalty_present must be finite");
+    }
+    if (params.penalty_last_n == -1) {
+        params.penalty_last_n = n_ctx > 0 ? n_ctx : llama_model_n_ctx_train(model);
+    }
 
+    const llama_vocab * vocab = llama_model_get_vocab(model);
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
     lparams.no_perf = params.no_perf;
@@ -310,8 +327,19 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
         }
     }
 
-    if (params.has_logit_bias()) {
-        samplers.push_back(llama_sampler_init_logit_bias(llama_vocab_n_tokens(vocab), params.logit_bias.size(), params.logit_bias.data()));
+    // logit bias: user biases + model suppress tokens (-INFINITY)
+    {
+        std::vector<llama_logit_bias> merged = params.logit_bias;
+
+        int32_t n_suppress = 0;
+        const llama_token * suppress = llama_vocab_get_suppress_tokens(vocab, &n_suppress);
+        for (int32_t i = 0; i < n_suppress; ++i) {
+            merged.push_back({ suppress[i], -INFINITY });
+        }
+
+        if (!merged.empty()) {
+            samplers.push_back(llama_sampler_init_logit_bias(llama_vocab_n_tokens(vocab), merged.size(), merged.data()));
+        }
     }
 
     if (params.mirostat == 0) {
@@ -355,7 +383,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
                     samplers.push_back(llama_sampler_init_infill(vocab));
                     break;
                 case COMMON_SAMPLER_TYPE_PENALTIES:
-                    samplers.push_back(llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    samplers.push_back(llama_sampler_init_penalties(llama_vocab_n_tokens(vocab), params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
                     break;
                 case COMMON_SAMPLER_TYPE_ADAPTIVE_P:
                     // the `adaptive-p` sampler is like `dist` and `mirostat` in that it selects

common/sampling.h

@@ -37,7 +37,10 @@ struct common_sampler;
 // llama_sampler API overloads
 
 // note: can mutate params in some cases
-struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params);
+struct common_sampler * common_sampler_init(
+        const struct llama_model * model,
+        struct common_params_sampling & params,
+        int32_t n_ctx = 0);
 
 void common_sampler_free(struct common_sampler * gsmpl);
 

common/speculative.cpp

@@ -34,6 +34,7 @@ const std::map<std::string, common_speculative_type> common_speculative_type_fro
     {"draft-eagle3",  COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3},
     {"draft-mtp",     COMMON_SPECULATIVE_TYPE_DRAFT_MTP},
     {"draft-dflash",  COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH},
+    {"draft-dspark",  COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK},
     {"ngram-simple",  COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE},
     {"ngram-map-k",   COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K},
     {"ngram-map-k4v", COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V},
@@ -437,6 +438,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
     int32_t n_embd_dec = 0;       // draft hidden size
     int32_t n_embd_enc = 0;       // target_layer_ids_n * target_hidden_size
     int32_t n_embd_tgt = 0;       // target model hidden size
+    int32_t n_layer_tgt = 0;      // target model layer count
 
     const int32_t * target_layer_ids   = nullptr; // model_dft's extract layer indices
     uint32_t        target_layer_ids_n = 0;
@@ -478,6 +480,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
         n_embd_tgt = llama_model_n_embd(model_tgt);
         n_embd_dec = llama_model_n_embd(model_dft);
         n_embd_enc = (int32_t) target_layer_ids_n * n_embd_tgt;
+        n_layer_tgt = llama_model_n_layer(model_tgt);
 
         const int32_t n_b = (int32_t) llama_n_batch(ctx_dft);
         batch = llama_batch_init(/*n_tokens=*/ n_b, /*embd=*/ n_embd_dec, /*n_seq_max=*/ 1);
@@ -510,9 +513,15 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
             }
         }
 
-        // turn on extraction of the target layers' input embeddings
+        // turn on extraction of the target layers' hidden states
         for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
-            llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
+            if (target_layer_ids[k] < n_layer_tgt) {
+                llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
+            } else if (target_layer_ids[k] == n_layer_tgt) {
+                llama_set_embeddings_nextn(ctx_tgt, true, /*masked*/ false);
+            } else {
+                GGML_ABORT("EAGLE3: target layer id %d exceeds target n_layer %d", target_layer_ids[k], n_layer_tgt);
+            }
         }
 
         // turn on extraction of the draft model's pre-norm hidden state
@@ -600,7 +609,9 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
         features_buf.resize((size_t) n_tokens * n_embd_enc, 0.0f);
 
         for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
-            const float * layer = llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k]);
+            const float * layer = target_layer_ids[k] < n_layer_tgt
+                ? llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k])
+                : llama_get_embeddings_nextn(ctx_tgt);
             if (!layer) {
                 GGML_ABORT("EAGLE3: target layer %d input not extracted.", target_layer_ids[k]);
             }
@@ -918,15 +929,20 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
     int32_t     block_size    = 0;
     llama_token mask_token_id = 0;
 
+    // draft-dspark: the draft carries a Markov head and uses an anchor-first block layout
+    const bool is_dspark;
+
     const int32_t * target_layer_ids   = nullptr; // model_dft's extract layer indices
     uint32_t        target_layer_ids_n = 0;
 
     // scratch buffer for concatenated target features [n_tokens, n_embd_enc]
     std::vector<float> features_buf;
 
-    common_speculative_impl_draft_dflash(const common_params_speculative & params, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, n_seq)
+    common_speculative_impl_draft_dflash(const common_params_speculative & params, uint32_t n_seq,
+            common_speculative_type type = COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)
+        : common_speculative_impl(type, n_seq)
         , params(params.draft)
+        , is_dspark(type == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK)
     {
         auto * ctx_tgt = this->params.ctx_tgt;
         auto * ctx_dft = this->params.ctx_dft;
@@ -953,16 +969,18 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
         }
         mask_token_id = llama_vocab_mask(llama_model_get_vocab(model_dft));
 
-        LOG_INF("%s: adding speculative implementation 'draft-dflash'\n", __func__);
+        LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(type).c_str());
         LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min);
         LOG_INF("%s: - block_size=%d, mask_token_id=%d, n_extract=%u\n", __func__, block_size, mask_token_id, target_layer_ids_n);
 
-        // DFlash input is [id_last, <mask> * (block_size-1)], so it can draft at most block_size-1 tokens per step
-        if (this->params.n_max > block_size - 1 || this->params.n_min > block_size - 1) {
-            LOG_WRN("%s: requested draft size (n_max=%d, n_min=%d) exceeds the trained DFlash block size %d -- clamping to %d\n",
-                    __func__, this->params.n_max, this->params.n_min, block_size, block_size - 1);
-            this->params.n_max = std::min(this->params.n_max, block_size - 1);
-            this->params.n_min = std::min(this->params.n_min, block_size - 1);
+        // DFlash input is [id_last, <mask> * (block_size-1)]: in-place denoising yields at most
+        // block_size-1 draft tokens, DSpark yield a full block_size draft tokens
+        const int32_t n_draft_max = is_dspark ? block_size : block_size - 1;
+        if (this->params.n_max > n_draft_max || this->params.n_min > n_draft_max) {
+            LOG_WRN("%s: requested draft size (n_max=%d, n_min=%d) exceeds the trained block size %d -- clamping to %d\n",
+                    __func__, this->params.n_max, this->params.n_min, block_size, n_draft_max);
+            this->params.n_max = std::min(this->params.n_max, n_draft_max);
+            this->params.n_min = std::min(this->params.n_min, n_draft_max);
         }
 
         batch        = llama_batch_init(llama_n_batch(ctx_dft), 0,          n_seq);
@@ -1126,12 +1144,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
             const int32_t n = (int32_t) dp.n_past;
 
-            int32_t n_draft = params.n_max;
-            if (dp.n_max > 0) {
-                n_draft = std::min(n_draft, dp.n_max);
-            }
+            const int32_t n_draft = params.n_max;
 
-            const int32_t n_block_tokens = n_draft + 1; // id_last + n_draft * <mask>
+            const int32_t n_block_tokens = n_draft + (is_dspark ? 0 : 1);
             i_block_beg[seq_id] = batch.n_tokens;
             n_block    [seq_id] = n_block_tokens;
             for (int32_t i = 0; i < n_block_tokens; ++i) {
@@ -1163,27 +1178,57 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
             auto & result = *dp.result;
 
-            // greedily read the predicted block at this sequence's noise positions 1..n_block_tokens-1
-            for (int32_t i = 1; i < n_block_tokens; ++i) {
-                common_sampler_sample(smpl, ctx_dft, beg + i, true);
+            if (is_dspark) {
+                // DSpark predicts the next token from position 0 and optionally truncates
+                // at the first position below the confidence threshold.
+                const float * conf = params.p_min > 0.0f ? llama_get_embeddings_nextn(ctx_dft) : nullptr;
 
-                const auto * cur_p = common_sampler_get_candidates(smpl, true);
+                for (int32_t i = 0; i < n_block_tokens; ++i) {
+                    const int32_t idx = beg + i;
 
-                for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
-                    LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
-                            seq_id, k, i - 1, cur_p->data[k].id, cur_p->data[k].p,
-                            common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
+                    if (conf && conf[(size_t) idx * n_embd_dec] < params.p_min) {
+                        break;
+                    }
+
+                    common_sampler_sample(smpl, ctx_dft, idx, true);
+
+                    const auto * cur_p = common_sampler_get_candidates(smpl, true);
+
+                    for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
+                        LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
+                                seq_id, k, i, cur_p->data[k].id, cur_p->data[k].p,
+                                common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
+                    }
+
+                    const llama_token id = cur_p->data[0].id;
+
+                    common_sampler_accept(smpl, id, true);
+
+                    result.push_back(id);
                 }
+            } else {
+                // greedily read the predicted block at this sequence's noise positions 1..n_block_tokens-1
+                for (int32_t i = 1; i < n_block_tokens; ++i) {
+                    common_sampler_sample(smpl, ctx_dft, beg + i, true);
 
-                const llama_token id = cur_p->data[0].id;
+                    const auto * cur_p = common_sampler_get_candidates(smpl, true);
 
-                if (cur_p->data[0].p < params.p_min) {
-                    break;
+                    for (int k = 0; k < std::min(3, (int) cur_p->size); ++k) {
+                        LOG_DBG(" - seq_id %d, draft candidate %3d, pos %3d: %6d (%8.3f) '%s'\n",
+                                seq_id, k, i - 1, cur_p->data[k].id, cur_p->data[k].p,
+                                common_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
+                    }
+
+                    const llama_token id = cur_p->data[0].id;
+
+                    if (cur_p->data[0].p < params.p_min) {
+                        break;
+                    }
+
+                    common_sampler_accept(smpl, id, true);
+
+                    result.push_back(id);
                 }
-
-                common_sampler_accept(smpl, id, true);
-
-                result.push_back(id);
             }
 
             if (result.size() < (size_t) params.n_min) {
@@ -1246,7 +1291,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
         GGML_ASSERT(ctx_tgt && ctx_dft && "MTP requires ctx_tgt and ctx_dft to be set");
 
         n_embd = llama_model_n_embd_out(llama_get_model(ctx_dft));
-        GGML_ASSERT(n_embd == llama_model_n_embd(llama_get_model(ctx_tgt)) &&
+        GGML_ASSERT(n_embd == llama_model_n_embd_out(llama_get_model(ctx_tgt)) &&
                 "MTP input row width must match the target h_nextn width");
         n_mtp_layers = std::max(1, (int) llama_model_n_layer_nextn(llama_get_model(ctx_dft)));
 
@@ -2145,6 +2190,7 @@ std::string common_speculative_type_to_str(common_speculative_type type) {
         case COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3:  return "draft-eagle3";
         case COMMON_SPECULATIVE_TYPE_DRAFT_MTP:     return "draft-mtp";
         case COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH:  return "draft-dflash";
+        case COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK:  return "draft-dspark";
         case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:  return "ngram-simple";
         case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K:   return "ngram-map-k";
         case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: return "ngram-map-k4v";
@@ -2198,6 +2244,7 @@ int32_t common_speculative_n_max(const common_params_speculative * spec) {
             case COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3:
             case COMMON_SPECULATIVE_TYPE_DRAFT_MTP:
             case COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH:
+            case COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK:
                 n_max = std::max(n_max, std::max(0, spec->draft.n_max));
                 break;
             case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE:
@@ -2284,7 +2331,7 @@ common_speculative_init_result::common_speculative_init_result(
     std::string model_path;
     if (has_draft) {
         model_path = params.speculative.draft.mparams.path;
-        LOG_TRC("%s: loading draft model '%s'\n", __func__, model_path.c_str());
+        LOG_INF("%s: loading draft model '%s'\n", __func__, model_path.c_str());
 
         llama_model * model_dft = llama_model_load_from_file(params.model.path.c_str(), mparams);
         if (model_dft == NULL) {
@@ -2304,7 +2351,7 @@ common_speculative_init_result::common_speculative_init_result(
     } else if (spec_mtp) {
         model_path = params.model.path;
 
-        LOG_TRC("%s: creating MTP draft context against the target model '%s'\n", __func__, model_path.c_str());
+        LOG_INF("%s: creating MTP draft context against the target model '%s'\n", __func__, model_path.c_str());
 
         llama_context * ctx_dft = llama_init_from_model(model_tgt, cparams);
         if (ctx_dft == nullptr) {
@@ -2338,53 +2385,28 @@ common_speculative * common_speculative_init(common_params_speculative & params,
     {
         uint32_t enabled_configs = common_get_enabled_speculative_configs(params.types);
 
-        bool has_draft_simple = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE));
-        bool has_draft_eagle3 = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3)) && params.draft.ctx_dft != nullptr;
-        bool has_draft_mtp    = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP))    && params.draft.ctx_dft != nullptr;
-        bool has_draft_dflash = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)) && params.draft.ctx_dft != nullptr;
-
-
-
-        bool has_ngram_cache   = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_CACHE));
-        bool has_ngram_simple  = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE));
-        bool has_ngram_map_k   = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K));
-        bool has_ngram_map_k4v = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V));
-        bool has_ngram_mod     = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MOD));
+        auto add_config_if_enabled = [&](common_speculative_type type, bool available = true) {
+            if (available && (enabled_configs & (1u << type))) {
+                configs.emplace_back(type, params);
+            }
+        };
 
         // when adding a new type - update here the logic above
-        static_assert(COMMON_SPECULATIVE_TYPE_COUNT == 10);
+        static_assert(COMMON_SPECULATIVE_TYPE_COUNT == 11);
 
         // this list here defines the priority of the speculators
         // the one with highest priority are listed first
-        if (has_ngram_simple) {
-            // This implementation can guess a lot of tokens without any draft model.
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, params));
-        }
-        if (has_ngram_map_k) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, params));
-        }
-        if (has_ngram_map_k4v) {
-            // This implementation can guess tokens with high acceptance rate but is more expensive.
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, params));
-        }
-        if (has_ngram_mod) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, params));
-        }
-        if (has_ngram_cache) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, params));
-        }
-        if (has_draft_simple) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, params));
-        }
-        if (has_draft_eagle3) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, params));
-        }
-        if (has_draft_mtp) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, params));
-        }
-        if (has_draft_dflash) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, params));
-        }
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_MOD);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE);
+
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, params.draft.ctx_dft != nullptr);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_MTP,    params.draft.ctx_dft != nullptr);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, params.draft.ctx_dft != nullptr);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK, params.draft.ctx_dft != nullptr);
     }
 
     std::vector<std::unique_ptr<common_speculative_impl>> impls = {};
@@ -2409,6 +2431,11 @@ common_speculative * common_speculative_init(common_params_speculative & params,
                 impls.push_back(std::make_unique<common_speculative_impl_draft_dflash>(config.params, n_seq));
                 break;
             }
+            case COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK: {
+                impls.push_back(std::make_unique<common_speculative_impl_draft_dflash>(
+                        config.params, n_seq, COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK));
+                break;
+            }
             case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: {
                 common_ngram_map ngram_map = get_common_ngram_map(config.type, config.params.ngram_simple);
 

common/subproc.cpp

@@ -0,0 +1,143 @@
+#include "subproc.h"
+
+bool common_subproc::is_supported() {
+#ifdef LLAMA_SUBPROCESS
+    return true;
+#else
+    return false;
+#endif
+}
+
+#ifdef LLAMA_SUBPROCESS
+
+static std::vector<char *> to_cstr_vec(const std::vector<std::string> & v) {
+    std::vector<char *> r;
+    r.reserve(v.size() + 1);
+    for (const auto & s : v) {
+        r.push_back(const_cast<char *>(s.c_str()));
+    }
+    r.push_back(nullptr);
+    return r;
+}
+
+common_subproc::~common_subproc() {
+    if (is_created) {
+        subprocess_destroy(&proc);
+        is_created = false;
+    }
+}
+
+bool common_subproc::create(
+        const std::vector<std::string> & args,
+        int options,
+        const std::vector<std::string> & env,
+        const char * cwd) {
+    auto argv = to_cstr_vec(args);
+
+    int result;
+    if (env.empty() && cwd == nullptr) {
+        result = subprocess_create(argv.data(), options, &proc);
+    } else {
+        auto envp = to_cstr_vec(env);
+        result = subprocess_create_ex(argv.data(), options, env.empty() ? nullptr : envp.data(), cwd, &proc);
+    }
+
+    is_created = result == 0;
+    return is_created;
+}
+
+bool common_subproc::has_handle() const {
+    if (!is_created) {
+        return false;
+    }
+#if defined(_WIN32)
+    return proc.hProcess != nullptr;
+#else
+    return proc.child > 0;
+#endif
+}
+
+bool common_subproc::alive() {
+    return is_created && subprocess_alive(&proc);
+}
+
+FILE * common_subproc::stdin_file() {
+    return is_created ? subprocess_stdin(&proc) : nullptr;
+}
+
+FILE * common_subproc::stdout_file() {
+    return is_created ? subprocess_stdout(&proc) : nullptr;
+}
+
+FILE * common_subproc::stderr_file() {
+    return is_created ? subprocess_stderr(&proc) : nullptr;
+}
+
+void common_subproc::close_stdin() {
+    if (is_created && proc.stdin_file) {
+        fclose(proc.stdin_file);
+        proc.stdin_file = nullptr;
+    }
+}
+
+void common_subproc::terminate() {
+    if (has_handle()) {
+        subprocess_terminate(&proc);
+    }
+}
+
+int common_subproc::join() {
+    int exit_code = -1;
+    if (is_created) {
+        subprocess_join(&proc, &exit_code);
+        subprocess_destroy(&proc);
+        is_created = false;
+    }
+    return exit_code;
+}
+
+#else // !LLAMA_SUBPROCESS
+
+common_subproc::~common_subproc() = default;
+
+bool common_subproc::create(
+        const std::vector<std::string> &,
+        int,
+        const std::vector<std::string> &,
+        const char *) {
+    (void)(proc);
+    (void)(is_created);
+    return false;
+}
+
+bool common_subproc::has_handle() const {
+    return false;
+}
+
+bool common_subproc::alive() {
+    return false;
+}
+
+FILE * common_subproc::stdin_file() {
+    return nullptr;
+}
+
+FILE * common_subproc::stdout_file() {
+    return nullptr;
+}
+
+FILE * common_subproc::stderr_file() {
+    return nullptr;
+}
+
+void common_subproc::close_stdin() {
+}
+
+void common_subproc::terminate() {
+}
+
+int common_subproc::join() {
+    return -1;
+}
+
+#endif // LLAMA_SUBPROCESS

common/subproc.h

@@ -0,0 +1,59 @@
+#pragma once
+
+#include <atomic>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+#ifdef LLAMA_SUBPROCESS
+#include <sheredom/subprocess.h>
+#else
+// dummy values to allow compilation when subprocess is disabled
+struct subprocess_s {};
+static constexpr int subprocess_option_no_window = 0;
+static constexpr int subprocess_option_combined_stdout_stderr = 0;
+static constexpr int subprocess_option_inherit_environment = 0;
+static constexpr int subprocess_option_search_user_path = 0;
+#endif
+
+// RAII-style wrapper around https://github.com/sheredom/subprocess.h,
+// exposing method calls instead of free functions operating on subprocess_s.
+struct common_subproc {
+    common_subproc() = default;
+    ~common_subproc();
+
+    common_subproc(const common_subproc &) = delete;
+    common_subproc & operator=(const common_subproc &) = delete;
+
+    // spawn a child process; if env is non-empty it replaces the child's environment
+    // (do not combine with subprocess_option_inherit_environment)
+    bool create(
+            const std::vector<std::string> & args,
+            int options,
+            const std::vector<std::string> & env = {},
+            const char * cwd = nullptr);
+
+    bool alive();
+
+    // true if LLAMA_SUBPROCESS was enabled at build time; when false, create() always fails
+    static bool is_supported();
+
+    FILE * stdin_file();
+    FILE * stdout_file();
+    FILE * stderr_file();
+
+    // close stdin and detach it from the process, so a later join()/destroy() won't double-close it;
+    // use this after writing all input to signal EOF to the child while it's still running
+    void close_stdin();
+
+    void terminate();
+
+    // wait for the process to exit, release the underlying handle and return its exit code
+    int join();
+
+private:
+    subprocess_s proc {};
+    std::atomic<bool> is_created{false};
+
+    bool has_handle() const;
+};

conversion/__init__.py

@@ -53,7 +53,9 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "DeepseekV3ForCausalLM": "deepseek",
     "DeepseekV32ForCausalLM": "deepseek",
     "DFlashDraftModel": "qwen",
+    "Qwen3DSparkModel": "qwen",
     "DeepseekV4ForCausalLM": "deepseek",
+    "DeepseekV4DSparkModel": "deepseek",
     "DistilBertForMaskedLM": "bert",
     "DistilBertForSequenceClassification": "bert",
     "DistilBertModel": "bert",
@@ -158,6 +160,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "MiniCPMForCausalLM": "minicpm",
     "MiniCPMV4_6ForConditionalGeneration": "minicpm",
     "MiniMaxM2ForCausalLM": "minimax",
+    "MiniMaxM3SparseForCausalLM": "minimax",
+    "MiniMaxM3SparseForConditionalGeneration": "minimax",
     "Ministral3ForCausalLM": "mistral3",
     "Mistral3ForConditionalGeneration": "mistral3",
     "MistralForCausalLM": "llama",
@@ -165,6 +169,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "ModernBertForMaskedLM": "bert",
     "ModernBertForSequenceClassification": "bert",
     "ModernBertModel": "bert",
+    "NanbeigeForCausalLM": "nanbeige",
     "NemotronForCausalLM": "nemotron",
     "NemotronHForCausalLM": "nemotron",
     "NeoBERT": "bert",
@@ -267,6 +272,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "Gemma4UnifiedForConditionalGeneration": "gemma",
     "Glm4vForConditionalGeneration": "qwen3vl",
     "Glm4vMoeForConditionalGeneration": "qwen3vl",
+    "Glm5vForConditionalGeneration": "kimivl",
     "GlmOcrForConditionalGeneration": "qwen3vl",
     "GlmasrModel": "ultravox",
     "Granite4VisionForConditionalGeneration": "granite",
@@ -285,6 +291,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "LlavaForConditionalGeneration": "llava",
     "MERaLiON2ForConditionalGeneration": "ultravox",
     "MiMoV2ForCausalLM": "mimo",
+    "MiniMaxM3SparseForConditionalGeneration": "minimax",
     "MiniCPMV4_6ForConditionalGeneration": "minicpm",
     "Mistral3ForConditionalGeneration": "llava",
     "NemotronH_Nano_VL_V2": "nemotron",

conversion/base.py

@@ -1156,7 +1156,7 @@ class TextModel(ModelBase):
                 or "projector." in name or "pre_mm_projector_norm" in name \
                 or "image_newline" in name or "view_seperator" in name \
                 or "patch_embed" in name or "patch_embedding" in name \
-                or "patch_merger." in name or "model.connector." in name:
+                or "patch_merger." in name or "patch_merge_mlp." in name or "model.connector." in name:
             return None
 
         return super().filter_tensors(item)
@@ -1203,7 +1203,7 @@ class TextModel(ModelBase):
             self.gguf_writer.add_embedding_length(n_embd)
             logger.info(f"gguf: embedding length = {n_embd}")
 
-        if (n_ff := self.find_hparam(["prefix_dense_intermediate_size", "intermediate_size", "n_inner", "hidden_dim"], optional=True)) is not None:
+        if (n_ff := self.find_hparam(["prefix_dense_intermediate_size", "dense_intermediate_size", "intermediate_size", "n_inner", "hidden_dim"], optional=True)) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
             logger.info(f"gguf: feed forward length = {n_ff}")
 

conversion/chatglm.py

@@ -81,7 +81,7 @@ class ChatGLMModel(TextModel):
 
     @staticmethod
     def token_bytes_to_string(b):
-        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode  # ty: ignore[unresolved-import]
+        from transformers.convert_slow_tokenizer import bytes_to_unicode
         byte_encoder = bytes_to_unicode()
         return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])
 

conversion/deepseek.py

@@ -447,12 +447,43 @@ class DeepseekV2Model(TextModel):
 class DeepseekV32Model(DeepseekV2Model):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK32
     skip_mtp = False
+    supports_mtp_export = True
+    _n_main_layers: int | None = None
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        # DeepSeek V3.2 appends the NextN/MTP block past num_hidden_layers
+        # (model.layers.61 -> blk.61 in the 62-block file).
+        assert cls._n_main_layers is not None
+        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
+
+        # --no-mtp: drop the appended NextN block entirely.
+        if is_mtp and cls.no_mtp:
+            return None
+        # --mtp: keep ONLY NextN-block tensors plus the shared embeddings/
+        # norm/lm_head (so the resulting GGUF carries just the draft head).
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return name, gen
+
     def set_vocab(self):
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
@@ -463,7 +494,7 @@ class DeepseekV32Model(DeepseekV2Model):
         super().set_gguf_parameters()
 
         # NextN/MTP prediction layers
-        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        if not self.no_mtp and (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
             self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
 
         # DSA indexer parameters
@@ -475,7 +506,10 @@ class DeepseekV32Model(DeepseekV2Model):
 @ModelBase.register("DeepseekV4ForCausalLM")
 class DeepseekV4Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK4
+    supports_mtp_export = True
     _skipped_mtp_tensors = 0
+    _dsv4_main_layers: int | None = None
+    _dsv4_nextn_layers: int = 0
 
     def __init__(self, *args, **kwargs):
         type(self)._skipped_mtp_tensors = 0
@@ -487,6 +521,8 @@ class DeepseekV4Model(TextModel):
             self.hparams.setdefault(key, value)
 
         self.block_count = self.hparams["num_hidden_layers"]
+        if self.mtp_only:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
         self._dsv4_fp8_dequantized: set[str] = set()
@@ -499,18 +535,71 @@ class DeepseekV4Model(TextModel):
             logger.info("Skipping %d DeepSeek-V4 MTP tensor(s) for conversion v0", type(self)._skipped_mtp_tensors)
 
         # add a default chat template; if the model has a built-in template, it will be overridden later
-        template_path = Path(__file__).parent.parent / "models" / "templates" / "deepseek-ai-DeepSeek-V4.jinja"
+        model_id_hint = self.remote_hf_model_id or self.dir_model.name
+        is_0731 = "0731" in model_id_hint
+        template_name = "deepseek-ai-DeepSeek-V4-Flash-0731.jinja" if is_0731 else "deepseek-ai-DeepSeek-V4.jinja"
+        template_path = Path(__file__).parent.parent / "models" / "templates" / template_name
         if template_path.is_file():
             with open(template_path, "r", encoding="utf-8") as f:
                 self.gguf_writer.add_chat_template(f.read())
 
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        type(self)._dsv4_main_layers = self.hparams["num_hidden_layers"]
+        type(self)._dsv4_nextn_layers = self.hparams.get("num_nextn_predict_layers", 0)
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, _ = item
+        name, gen = item
         if name.startswith("mtp."):
-            cls._skipped_mtp_tensors += 1
-            return None
-        return super().filter_tensors(item)
+            if not cls.mtp_only:
+                cls._skipped_mtp_tensors += 1
+                return None
+
+            assert cls._dsv4_main_layers is not None
+            parts = name.split(".", 2)
+            if len(parts) < 3 or not parts[1].isdecimal():
+                raise ValueError(f"Unexpected DeepSeek-V4 MTP tensor {name!r}")
+
+            mtp_idx = int(parts[1])
+            if mtp_idx >= cls._dsv4_nextn_layers:
+                raise ValueError(f"Unexpected DeepSeek-V4 MTP layer {mtp_idx}")
+
+            bid = cls._dsv4_main_layers + mtp_idx
+            suffix = parts[2]
+            root_hc_head = {
+                "hc_head_fn",
+                "hc_head_base",
+                "hc_head_scale",
+            }
+            if suffix in root_hc_head:
+                name = suffix
+            elif suffix in (
+                "e_proj.weight", "e_proj.scale",
+                "h_proj.weight", "h_proj.scale",
+            ):
+                name = f"layers.{bid}.nextn.{suffix}"
+            elif suffix == "enorm.weight":
+                name = f"layers.{bid}.nextn.enorm.weight"
+            elif suffix == "hnorm.weight":
+                name = f"layers.{bid}.nextn.hnorm.weight"
+            elif suffix == "norm.weight":
+                name = f"layers.{bid}.nextn.shared_head_norm.weight"
+            else:
+                name = f"layers.{bid}.{suffix}"
+            return name, gen
+
+        if cls.mtp_only:
+            keep = name in (
+                "embed.weight",
+                "norm.weight",
+                "head.weight",
+                "head.scale",
+            )
+            if not keep:
+                return None
+
+        return super().filter_tensors((name, gen))
 
     @staticmethod
     def _float8_dtypes() -> tuple[torch.dtype, ...]:
@@ -565,6 +654,10 @@ class DeepseekV4Model(TextModel):
         self.gguf_writer.add_hyper_connection_sinkhorn_iterations(hparams["hc_sinkhorn_iters"])
         self.gguf_writer.add_hyper_connection_epsilon(hparams["hc_eps"])
         self.gguf_writer.add_hash_layer_count(hparams["num_hash_layers"])
+        if self.model_arch == gguf.MODEL_ARCH.DEEPSEEK4:
+            self.gguf_writer.add_embedding_length_out(hparams["hidden_size"] * hparams["hc_mult"])
+        if self.mtp_only and (num_nextn_predict_layers := hparams.get("num_nextn_predict_layers", 0)) > 0:
+            self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
 
     def dequant_model(self):
         fp8_dtypes = self._float8_dtypes()
@@ -669,12 +762,37 @@ class DeepseekV4Model(TextModel):
         if self._dsv4_mxfp4_generated:
             return ()
 
-        consumed: list[str] = self._write_hash_routing_tensors()
+        consumed: list[str] = []
+        main_layers = self.hparams["num_hidden_layers"]
+        if not self.mtp_only:
+            consumed.extend(self._write_hash_routing_tensors())
+        elif self.hparams["num_hash_layers"] > 0:
+            for bid in range(self.hparams["num_hash_layers"]):
+                name = f"layers.{bid}.ffn.gate.tid2eid"
+                if name in self.model_tensors:
+                    consumed.extend(self._write_hash_routing_tensors())
+                    break
+
         for bid in range(self.block_count):
+            if self.mtp_only and bid < main_layers:
+                continue
             consumed.extend(self._write_mxfp4_expert_tensor(bid, "w1", gguf.MODEL_TENSOR.FFN_GATE_EXP))
             consumed.extend(self._write_mxfp4_expert_tensor(bid, "w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP))
             consumed.extend(self._write_mxfp4_expert_tensor(bid, "w3", gguf.MODEL_TENSOR.FFN_UP_EXP))
 
+        for bid in range(main_layers, self.block_count):
+            e_name = f"layers.{bid}.nextn.e_proj.weight"
+            h_name = f"layers.{bid}.nextn.h_proj.weight"
+            if e_name not in self.model_tensors and h_name not in self.model_tensors:
+                continue
+            if e_name not in self.model_tensors or h_name not in self.model_tensors:
+                raise KeyError(f"Missing DeepSeek-V4 MTP e/h projection pair for block {bid}")
+
+            e_proj = LazyTorchTensor.to_eager(self.model_tensors[e_name]())
+            h_proj = LazyTorchTensor.to_eager(self.model_tensors[h_name]())
+            yield (f"layers.{bid}.nextn.eh_proj.weight", torch.cat((e_proj, h_proj), dim=1).contiguous())
+            consumed.extend((e_name, h_name))
+
         for name in consumed:
             del self.model_tensors[name]
 
@@ -737,6 +855,12 @@ class DeepseekV4Model(TextModel):
             "ffn.shared_experts.w1.weight": (gguf.MODEL_TENSOR.FFN_GATE_SHEXP, ".weight"),
             "ffn.shared_experts.w2.weight": (gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, ".weight"),
             "ffn.shared_experts.w3.weight": (gguf.MODEL_TENSOR.FFN_UP_SHEXP, ".weight"),
+            "nextn.eh_proj.weight": (gguf.MODEL_TENSOR.NEXTN_EH_PROJ, ".weight"),
+            "nextn.enorm.weight": (gguf.MODEL_TENSOR.NEXTN_ENORM, ".weight"),
+            "nextn.hnorm.weight": (gguf.MODEL_TENSOR.NEXTN_HNORM, ".weight"),
+            "nextn.shared_head_norm.weight": (gguf.MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM, ".weight"),
+            "nextn.embed_tokens.weight": (gguf.MODEL_TENSOR.NEXTN_EMBED_TOKENS, ".weight"),
+            "nextn.shared_head_head.weight": (gguf.MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD, ".weight"),
         }
 
         tensor_name = match.group(2)
@@ -759,10 +883,12 @@ class DeepseekV4Model(TextModel):
         return [(self._format_dsv4_tensor_name(tensor_key, bid, suffix), data_torch)]
 
     def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
-        del new_name, bid  # unused
+        del bid  # unused
 
         if name in self._dsv4_fp8_dequantized and n_dims >= 2:
             return gguf.GGMLQuantizationType.Q8_0
+        if new_name.endswith(".nextn.eh_proj.weight"):
+            return gguf.GGMLQuantizationType.Q8_0
         if name in self._dsv4_f32_tensors:
             return gguf.GGMLQuantizationType.F32
         if name in self._dsv4_bf16_tensors and n_dims >= 2:
@@ -770,7 +896,122 @@ class DeepseekV4Model(TextModel):
 
         return False
 
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
     def prepare_tensors(self):
         super().prepare_tensors()
         self._is_mxfp4 = True
         self.ftype = gguf.LlamaFileType.MOSTLY_MXFP4_MOE
+
+
+@ModelBase.register("DeepseekV4DSparkModel")
+class DeepseekV4DSparkModel(DeepseekV4Model):
+    model_arch = gguf.MODEL_ARCH.DFLASH
+
+    _DSPARK_ROOT_MAP: dict[str, tuple[gguf.MODEL_TENSOR, str]] = {
+        "main_proj.weight": (gguf.MODEL_TENSOR.FC, ".weight"),
+        "main_norm.weight": (gguf.MODEL_TENSOR.ENC_OUTPUT_NORM, ".weight"),
+        "markov_head.markov_w1.weight": (gguf.MODEL_TENSOR.DSPARK_MARKOV_W1, ".weight"),
+        "markov_head.markov_w2.weight": (gguf.MODEL_TENSOR.DSPARK_MARKOV_W2, ".weight"),
+        "confidence_head.proj.weight": (gguf.MODEL_TENSOR.DSPARK_CONF_PROJ, ".weight"),
+    }
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        self.block_count = 1 + max(
+            int(match.group(1)) for name in self.model_tensors
+            if (match := re.match(r"layers\.(\d+)\.", name))
+        )
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+        self.hparams["compress_ratios"] = [0] * self.block_count
+        self.hparams["num_hash_layers"] = 0
+
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        if remote_hf_model_id is None:
+            return super().index_tensors()
+
+        with open(self.dir_model / "model.safetensors.index.json", "r", encoding="utf-8") as f:
+            weight_map = json.load(f)["weight_map"]
+
+        part_names = sorted({
+            part_name for name, part_name in weight_map.items()
+            if name.startswith("mtp.")
+        })
+        tensors: dict[str, Callable[[], Tensor]] = {}
+
+        for part_name in part_names:
+            from huggingface_hub import hf_hub_download
+
+            logger.info("gguf: caching remote DSpark part '%s'", part_name)
+            part_path = Path(hf_hub_download(repo_id=remote_hf_model_id, filename=part_name))
+            with gguf.utility.SafetensorsLocal(part_path) as model_part:
+                for name in model_part:
+                    data = model_part[name]
+                    data_gen = lambda data=data: LazyTorchTensor.from_local_tensor(data)  # noqa: E731
+                    if titem := self.filter_tensors((name, data_gen)):
+                        tensor_name, tensor_gen = titem
+                        tensors[tensor_name] = tensor_gen
+
+        return tensors
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if not name.startswith("mtp."):
+            return None
+        return super().filter_tensors((cls._rekey_mtp_tensor_name(name), gen))
+
+    @staticmethod
+    def _rekey_mtp_tensor_name(name: str) -> str:
+        match = re.match(r"mtp\.(\d+)\.(.+)$", name)
+        if match is None:
+            raise ValueError(f"Unexpected DSpark tensor {name!r}")
+
+        stage, rest = match.group(1), match.group(2)
+        root_names = (
+            "main_proj.scale",
+            "norm.weight",
+            "hc_head_fn",
+            "hc_head_base",
+            "hc_head_scale",
+        )
+        if rest in DeepseekV4DSparkModel._DSPARK_ROOT_MAP or rest in root_names:
+            return rest
+        return f"layers.{stage}.{rest}"
+
+    def _map_dsv4_tensor_name(self, name: str, bid: int | None) -> tuple[gguf.MODEL_TENSOR, str]:
+        if name in self._DSPARK_ROOT_MAP:
+            return self._DSPARK_ROOT_MAP[name]
+        return super()._map_dsv4_tensor_name(name, bid)
+
+    def set_vocab(self):
+        if self.target_model_dir is None:
+            raise ValueError("DeepSeek-V4 DSpark requires --target-model-dir with the target tokenizer")
+
+        original_dir = self.dir_model
+        try:
+            self.dir_model = self.target_model_dir
+            super().set_vocab()
+        finally:
+            self.dir_model = original_dir
+
+        self.gguf_writer.add_mask_token_id(self.hparams["dspark_noise_token_id"])
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        self.gguf_writer.add_block_size(self.hparams["dspark_block_size"])
+        self.gguf_writer.add_target_layers([layer + 1 for layer in self.hparams["dspark_target_layer_ids"]])

conversion/glm.py

@@ -1,6 +1,8 @@
 from __future__ import annotations
 
-from typing import Iterable, TYPE_CHECKING
+import re
+
+from typing import Callable, Iterable, TYPE_CHECKING
 
 import torch
 
@@ -204,21 +206,116 @@ class Glm4MoeModel(TextModel):
 @ModelBase.register("Glm4MoeLiteForCausalLM")
 class Glm4MoeLiteModel(DeepseekV2Model):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
+    skip_mtp = False
+    supports_mtp_export = True
+    _n_main_layers: int | None = None
 
     def set_vocab(self):
         return self._set_vocab_glm()
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        num_hidden_layers = self.hparams["num_hidden_layers"]
+        self.num_nextn_predict_layers = self.hparams.get("num_nextn_predict_layers", 0)
+        self.skip_mtp = self.no_mtp or self.num_nextn_predict_layers == 0
+
+        if self.skip_mtp:
+            self.block_count = num_hidden_layers
+        else:
+            self.block_count = num_hidden_layers + self.num_nextn_predict_layers
+
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        if self.skip_mtp:
+            return
+
+        self.gguf_writer.add_nextn_predict_layers(self.num_nextn_predict_layers)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item):
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        if cls._n_main_layers is not None:
+            match = re.match(r"model\.layers\.(\d+)\.", name)
+            is_mtp = match is not None and int(match.group(1)) >= cls._n_main_layers
+            if is_mtp and cls.no_mtp:
+                return None
+            if cls.mtp_only and not is_mtp and name not in (
+                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+            ):
+                return None
+
+        return name, gen
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
 
 @ModelBase.register("GlmMoeDsaForCausalLM")
 class GlmMoeDsaModel(DeepseekV2Model):
     model_arch = gguf.MODEL_ARCH.GLM_DSA
     skip_mtp = False
+    supports_mtp_export = True
+
+    # Trunk layer count, stashed before indexing so the classmethod
+    # filter_tensors can identify the appended NextN/MTP block (mirrors
+    # HYV3Model / Step35Model).
+    _n_main_layers: int | None = None
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        # GLM-5.2 appends the NextN/MTP block past num_hidden_layers
+        # (model.layers.78 -> blk.78 in the 79-block file).
+        assert cls._n_main_layers is not None
+        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
+
+        # --no-mtp: drop the appended NextN block entirely.
+        if is_mtp and cls.no_mtp:
+            return None
+        # --mtp: keep ONLY NextN-block tensors plus the shared embeddings/
+        # norm/lm_head (so the resulting GGUF carries just the draft head).
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return name, gen
+
     def set_vocab(self):
         return self._set_vocab_glm()
 
@@ -230,7 +327,7 @@ class GlmMoeDsaModel(DeepseekV2Model):
         self.gguf_writer.add_rope_dimension_count(int(rope_dim * partial_rotary_factor))
 
         # NextN/MTP prediction layers
-        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        if not self.no_mtp and (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
             self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
 
         # DSA indexer parameters

conversion/kimivl.py

@@ -152,3 +152,19 @@ class KimiK25Model(MmprojModel):
             name = name.replace(".proj.2.", ".proj.linear_2.")
 
         yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Glm5vForConditionalGeneration")
+class Glm5vModel(KimiK25Model):
+    """GLM-5.2-Vision MoonViT3d encoder and projector
+
+    Uses the same vision encoder and projector as Kimi-K2.5, so it reuses the
+    kimik25 projector type. The image begin/end tokens differ, but they are
+    resolved at runtime from the text model vocab.
+    """
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("mm_projector.linear_"):
+            name = name.replace("mm_projector.linear_", "mm_projector.proj.linear_", 1)
+
+        yield from super().modify_tensors(data_torch, name, bid)

conversion/llama.py

@@ -69,9 +69,14 @@ class LlamaModel(TextModel):
                 target_config = {**target_config, **target_config["text_config"]}
             self.target_vocab_size = target_config["vocab_size"]
 
-            # target_layers: derived from target model layer count (low/mid/high)
+            # target_layers: use the eagle3 config's explicit aux hidden-state layer ids
+            # if present, else derive from the target layer count.
             target_num_layers = target_config["num_hidden_layers"]
-            target_layers = [2, target_num_layers // 2, target_num_layers - 3]
+            aux_layer_ids = eagle3_raw_config.get("eagle_aux_hidden_state_layer_ids")
+            if aux_layer_ids:
+                target_layers = aux_layer_ids
+            else:
+                target_layers = [2, target_num_layers // 2, target_num_layers - 3]
             logger.info(f"EAGLE-3: target_layers = {target_layers} (target model has {target_num_layers} layers)")
             self.gguf_writer.add_target_layers(target_layers)
 
@@ -90,6 +95,12 @@ class LlamaModel(TextModel):
             logger.info(f"EAGLE-3: norm_before_residual = {norm_before_residual}")
             self.gguf_writer.add_norm_before_residual(norm_before_residual)
 
+            # norm_before_fc: RMSNorm applied to the fused target features before the
+            # fc projection (e.g. nvidia/gpt-oss-120b-Eagle3-v3)
+            norm_before_fc = eagle3_raw_config.get("norm_before_fc", False)
+            logger.info(f"EAGLE-3: norm_before_fc = {norm_before_fc}")
+            self.gguf_writer.add_norm_before_fc(norm_before_fc)
+
     def set_vocab(self):
         # eagle3: use tokenizer from target model if provided
         original_dir_model = None
@@ -108,7 +119,7 @@ class LlamaModel(TextModel):
         path_tekken_json = self.dir_model / "tekken.json"
         path_tokenizer_json = self.dir_model / "tokenizer.json"
         if path_tekken_json.is_file() and not path_tokenizer_json.is_file():
-            self._set_vocab_mistral()
+            return self._set_vocab_mistral()
 
         tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
         if tokenizer_config_file.is_file():
@@ -222,6 +233,9 @@ class LlamaModel(TextModel):
             if name == "fc.weight":
                 yield (name, data_torch)
                 return
+            if name == "input_norm.weight":
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ENC_OUTPUT_NORM), data_torch)
+                return
             if name == "d2t":
                 # store for manual int64 handling in prepare_tensors (avoid F32 conversion)
                 if not hasattr(self, '_eagle3_int_tensors'):

conversion/mimo.py

@@ -1,8 +1,9 @@
 from __future__ import annotations
 
+import json
 import re
 
-from typing import Callable, TYPE_CHECKING
+from typing import Any, Callable, Iterable, TYPE_CHECKING
 
 import torch
 
@@ -229,7 +230,13 @@ class MimoV2Model(TextModel):
 
 
 @ModelBase.register("MiMoV2ForCausalLM")
-class MiMoV2VisionModel(MmprojModel):
+class MiMoV2VisionAudioModel(MmprojModel):
+    has_audio_encoder = True
+
+    _audio_tok_hparams: dict[str, Any] | None = None
+    _rvq_codebook_sizes: list[int] | None = None
+    _code_embd: dict[int, Tensor] | None = None
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         assert self.hparams_vision is not None
@@ -253,10 +260,22 @@ class MiMoV2VisionModel(MmprojModel):
         self.visual_token_window_size = int(hp.get("visual_token_window_size", -1))
         self.use_sink = bool(hp.get("use_sink", False))
 
+    def get_audio_config(self) -> dict[str, Any] | None:
+        if self._audio_tok_hparams is None:
+            path = self.dir_model / "audio_tokenizer" / "config.json"
+            with open(path, "r", encoding="utf-8") as f:
+                cfg = json.load(f)
+            # aliases so MmprojModel.find_aparam() / n_block_keys can resolve them
+            cfg["hidden_size"] = cfg["d_model"]
+            cfg["intermediate_size"] = cfg["encoder_ffn_dim"]
+            cfg["num_attention_heads"] = cfg["encoder_attention_heads"]
+            self._audio_tok_hparams = cfg
+        return self._audio_tok_hparams
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MIMOVL)
+        self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.MIMOVL)
         self.gguf_writer.add_vision_use_silu(True)
         self.gguf_writer.add_vision_head_count_kv(self.num_kv_heads)
         self.gguf_writer.add_vision_spatial_merge_size(self.spatial_merge_size)
@@ -266,19 +285,45 @@ class MiMoV2VisionModel(MmprojModel):
         self.gguf_writer.add_vision_min_pixels(int(self.preprocessor_config["min_pixels"]))
         self.gguf_writer.add_vision_max_pixels(int(self.preprocessor_config["max_pixels"]))
 
+        assert self.hparams_audio is not None
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.MIMO_AUDIO)
+        self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["n_mels"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(self.hparams_audio.get("layer_norm_eps", 1e-5))
+
+        assert self._rvq_codebook_sizes is not None
+        self.gguf_writer.add_audio_rvq_num_quantizers(len(self._rvq_codebook_sizes))
+        self.gguf_writer.add_audio_rvq_codebook_size(self._rvq_codebook_sizes)
+
+        n_layer = self.hparams_audio["encoder_layers"]
+        swa_per_block = self.hparams_audio.get("swa_per_block", 1)
+        if self.hparams_audio.get("hybrid_attention") and swa_per_block > 1:
+            wa_pattern = [0 if i % swa_per_block < swa_per_block - 1 else -1 for i in range(n_layer)]
+        else:
+            wa_pattern = [-1] * n_layer
+        self.gguf_writer.add_audio_wa_pattern_mode(wa_pattern)
+        self.gguf_writer.add_audio_window_size(int(self.hparams_audio["encoder_attn_window_size"][0]))
+
+        audio_cfg = self.global_config["audio_config"]
+        self.gguf_writer.add_audio_local_block_count(int(audio_cfg["input_local_layers"]))
+        self.gguf_writer.add_audio_local_group_size(int(audio_cfg["group_size"]))
+
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        # Sinks must be F32: any sink-style softmax/mask add in ggml requires
-        # F32, and we fold sinks into a host-built F32 mask at encode time.
-        if new_name.endswith(".attn_sinks"):
+        # for audio encoder: keep codebook in F32
+        if new_name in (
+            gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.A_ENC_RVQ_CODEBOOK] + ".weight",
+            gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.A_MM_CODE_EMBD] + ".weight",
+        ):
+            return gguf.GGMLQuantizationType.F32
+        if ("encoder.conv" in name or "encoder.down_sample_layer" in name) and name.endswith(".weight"):
             return gguf.GGMLQuantizationType.F32
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, _ = item
-        if not name.startswith("visual."):
-            return None
-        return super().filter_tensors(item)
+        if name.startswith("visual.") or name.startswith("speech_embeddings.") or name.startswith("audio_encoder."):
+            return super().filter_tensors(item)
+        return None
 
     def modify_tensors(self, data_torch, name, bid):
         # Conv3D patch embed: split along the temporal axis (kt=2) into two Conv2D
@@ -292,4 +337,64 @@ class MiMoV2VisionModel(MmprojModel):
             yield (embd_name + ".weight.1", data_torch[:, :, 1, ...])
             return
 
+        if m := re.match(r"^speech_embeddings\.(\d+)\.weight$", name):
+            if self._code_embd is None:
+                self._code_embd = {}
+            self._code_embd[int(m.group(1))] = data_torch
+
+            n_channels = int(self.global_config["audio_config"]["audio_channels"])
+            if len(self._code_embd) < n_channels:
+                return
+            merged = torch.stack([self._code_embd.pop(i) for i in range(n_channels)], dim=0)
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MM_CODE_EMBD), merged)
+            return
+
+        if "conv1.bias" in name or "conv2.bias" in name:
+            # transpose conv1/conv2 bias so it broadcasts against [n_frames, C_out, 1]
+            data_torch = data_torch.unsqueeze(-1)
+
+        if name == "audio_encoder.projection.mlp.0.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MMPROJ, 1), data_torch)
+            return
+        if name == "audio_encoder.projection.mlp.2.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_MMPROJ, 2), data_torch)
+            return
+
         yield from super().modify_tensors(data_torch, name, bid)
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # note: audio encoder is in its own subdir "audio_tokenizer"
+        from safetensors.torch import load_file
+
+        tok_dir = self.dir_model / "audio_tokenizer"
+        state_dict = load_file(tok_dir / "model.safetensors")
+
+        codebook_re = re.compile(r"^encoder\.quantizer\.vq\.layers\.(\d+)\._codebook\.embed$")
+        codebooks: dict[int, Tensor] = {}
+
+        # EMA/training-only RVQ buffers - not needed for inference (nearest-codebook
+        # lookup only reads "_codebook.embed")
+        skip_suffixes = (
+            "_codebook.cluster_size",
+            "_codebook.embed_avg",
+            "_codebook.inited",
+        )
+        for name, tensor in state_dict.items():
+            if name.endswith(skip_suffixes):
+                continue
+            if m := codebook_re.match(name):
+                codebooks[int(m.group(1))] = tensor
+                continue
+            yield name, tensor
+
+        # gather codebooks and merge into 3D tensor, similar to MoE MLP tensors
+        n_q = len(codebooks)
+        ordered = [codebooks[i] for i in range(n_q)]
+        self._rvq_codebook_sizes = [int(cb.shape[0]) for cb in ordered]
+        max_bins = max(self._rvq_codebook_sizes)
+        dim = ordered[0].shape[1]
+        merged = ordered[0].new_zeros(n_q, max_bins, dim)
+        for i, cb in enumerate(ordered):
+            merged[i, : cb.shape[0], :] = cb
+
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_ENC_RVQ_CODEBOOK), merged)

conversion/minicpm.py

@@ -137,6 +137,15 @@ class MiniCPMV4_6TextModel(Qwen3_5TextModel):
 class MiniCPMV4_6VisionModel(MmprojModel):
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
+        self.downsample_mode = self.preprocessor_config.get("downsample_mode", "16x")
+        if self.downsample_mode not in {"4x", "16x"}:
+            raise ValueError(f"Unsupported downsample mode: {self.downsample_mode}")
+        if self.downsample_mode == "4x":
+            self.model_tensors = {
+                name: tensor for name, tensor in self.model_tensors.items()
+                if ".vit_merger." not in name
+            }
+
         if self.hparams_vision is not None:
             # In MiniCPM-V 4.6 `vision_config.image_size` (980) describes the SigLIP
             # positional embedding bucket grid (70 x 70), while the per-slice processing
@@ -156,8 +165,8 @@ class MiniCPMV4_6VisionModel(MmprojModel):
         # (mapped to PROJECTOR_TYPE_MINICPMV4_6).
         self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MINICPMV4_6)
 
-        # ViT merger 2x2 + final merger 2x2 = 4x spatial merge per dimension; used for slice alignment
-        self.gguf_writer.add_vision_projector_scale_factor(4)
+        self.gguf_writer.add_vision_projector_scale_factor(
+            2 if self.downsample_mode == "4x" else 4)
 
         # borrow wa_layer_indexes for vit_merger insertion point
         insert_layer_id = int(self.global_config.get(

conversion/minimax.py

@@ -7,7 +7,7 @@ import torch
 if TYPE_CHECKING:
     from torch import Tensor
 
-from .base import ModelBase, TextModel, gguf
+from .base import ModelBase, TextModel, MmprojModel, gguf
 
 
 @ModelBase.register("MiniMaxM2ForCausalLM")
@@ -23,7 +23,7 @@ class MiniMaxM2Model(TextModel):
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
         # merge expert weights
-        if 'experts' in name:
+        if "block_sparse_moe.experts." in name:
             n_experts = self.find_hparam(["num_local_experts", "num_experts"])
             assert bid is not None
 
@@ -52,3 +52,118 @@ class MiniMaxM2Model(TextModel):
             return
 
         yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("MiniMaxM3SparseForCausalLM", "MiniMaxM3SparseForConditionalGeneration")
+class MiniMaxM3Model(MiniMaxM2Model):
+    model_arch = gguf.MODEL_ARCH.MINIMAXM3
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        if ".indexer." in new_name:
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        self.gguf_writer.add_expert_shared_count(self.find_hparam(["n_shared_experts"]))
+        self.gguf_writer.add_expert_weights_scale(self.find_hparam(["routed_scaling_factor"]))
+        self.gguf_writer.add_expert_weights_norm(True)
+
+        sac = self.find_hparam(["sparse_attention_config"])
+        self.gguf_writer.add_indexer_head_count(sac["sparse_num_index_heads"])
+        self.gguf_writer.add_indexer_key_length(sac["sparse_index_dim"])
+        self.gguf_writer.add_indexer_top_k(sac["sparse_topk_blocks"])
+        self.gguf_writer.add_indexer_block_size(sac["sparse_block_size"])
+        self.gguf_writer.add_indexer_local_blocks(sac["sparse_local_block"])
+
+        moe_layer_freq = self.find_hparam(["moe_layer_freq"])
+        n_dense = 0
+        for v in moe_layer_freq:
+            if v == 0:
+                n_dense += 1
+            else:
+                break
+        self.gguf_writer.add_leading_dense_block_count(n_dense)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        # Gemma-style (1 + w) RMSNorm: bake the +1 in so llama.cpp can use plain RMSNorm
+        if name.endswith("norm.weight"):
+            data_torch = data_torch + 1.0
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("MiniMaxM3SparseForConditionalGeneration", "MiniMaxM3VLForConditionalGeneration")
+class MiniMaxM3VisionModel(MmprojModel):
+    @classmethod
+    def filter_tensors(cls, item):
+        name, gen = item
+        # keep only the vision-side tensors; text / mtp / sparse-index are dropped
+        if not name.startswith(("vision_tower.", "multi_modal_projector.", "patch_merge_mlp.")):
+            return None
+        return super().filter_tensors((name, gen))
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MINIMAXM3)
+        self.gguf_writer.add_vision_use_gelu(True)
+
+        # the ViT carries its own LayerNorm eps (text tower uses a different one)
+        self.gguf_writer.add_vision_attention_layernorm_eps(
+            self.hparams_vision.get("layer_norm_eps", 1e-5)
+        )
+
+        comp = self.hparams_vision.get("img_token_compression_config", {})
+        merge_size = comp.get("spatial_merge_size", 2)
+        self.gguf_writer.add_vision_spatial_merge_size(int(merge_size))
+
+    def modify_tensors(self, data_torch, name, bid):
+        assert self.hparams_vision is not None
+
+        # Conv3d patch embed -> Conv2d slices
+        if name == "vision_tower.vision_model.embeddings.patch_embedding.weight":
+            if data_torch.ndim != 5:
+                raise ValueError(f"unexpected patch_embedding rank {data_torch.ndim} for {name}")
+            kt = data_torch.shape[2]
+            base = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH]
+            for t in range(kt):
+                suffix = ".weight" if t == 0 else f".weight.{t}"
+                yield (base + suffix, data_torch[:, :, t, ...])
+            return
+
+        # Permute ViT q/k. HF [Ta Ha Wa | Tb Hb Wb | pad] reorder to [Ta Tb | Ha Hb | Wa Wb | pad].
+        for new_name, tensor in super().modify_tensors(data_torch, name, bid):
+            if ".attn_q." in new_name or ".attn_k." in new_name:
+                tensor = self._permute_vit_qk(tensor, new_name)
+            yield new_name, tensor
+
+    def _permute_vit_qk(self, t: "Tensor", new_name: str) -> "Tensor":
+        assert self.hparams_vision is not None
+        n_head = self.hparams_vision["num_attention_heads"]
+        d_head = t.shape[0] // n_head
+        axis_dim = 2 * ((2 * (d_head // 2) // 3) // 2)
+        ah = axis_dim // 2
+        half = 3 * ah
+        perm = []
+        perm += list(range(0, ah))
+        perm += list(range(half, half + ah))
+        perm += list(range(ah, 2 * ah))
+        perm += list(range(half + ah, half + 2 * ah))
+        perm += list(range(2 * ah, 3 * ah))
+        perm += list(range(half + 2 * ah, half + 3 * ah))
+        perm += list(range(2 * half, d_head))
+
+        assert axis_dim % 2 == 0
+        assert 3 * axis_dim <= d_head
+        assert len(perm) == d_head
+        assert sorted(perm) == list(range(d_head)), "perm is not a bijection of d_head"
+        assert t.shape[0] == n_head * d_head, f"{new_name}: {t.shape[0]} != {n_head}*{d_head}"
+        assert d_head == 80
+
+        idx = torch.tensor(perm, dtype=torch.long)
+        if t.ndim == 2:
+            return t.reshape(n_head, d_head, t.shape[1])[:, idx, :].reshape(t.shape)
+        return t.reshape(n_head, d_head)[:, idx].reshape(t.shape)

conversion/nanbeige.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+from .base import ModelBase, gguf, logger
+from .llama import LlamaModel
+
+
+@ModelBase.register("NanbeigeForCausalLM")
+class NanbeigeModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.NANBEIGE
+    undo_permute = True
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+
+        n_loops = int(hparams.get("num_loops", 1) or 1)
+        if n_loops < 1:
+            n_loops = 1
+        self.gguf_writer.add_num_loops(n_loops)
+        logger.info(f"gguf: num_loops = {n_loops}")
+
+        skip_loop_final_norm = bool(hparams.get("skip_loop_final_norm", False))
+        self.gguf_writer.add_skip_loop_final_norm(skip_loop_final_norm)
+        logger.info(f"gguf: skip_loop_final_norm = {skip_loop_final_norm}")

conversion/nemotron.py

@@ -39,28 +39,48 @@ class NemotronNanoV2VLModel(MmprojModel):
         }
         return vision_config
 
+    def get_audio_config(self) -> dict[str, Any] | None:
+        return self.global_config.get("sound_config")
+
     def set_gguf_parameters(self):
         if "image_mean" not in self.preprocessor_config:
             self.preprocessor_config["image_mean"] = [0.485, 0.456, 0.406]
         if "image_std" not in self.preprocessor_config:
             self.preprocessor_config["image_std"] = [0.229, 0.224, 0.225]
 
+        if self.hparams_audio is not None:
+            self.has_vision_encoder = True
+            self.has_audio_encoder = True
+            self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["num_mel_bins"])
+            self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+            self.gguf_writer.add_audio_subsampling_factor(self.hparams_audio["subsampling_factor"])
+            self.gguf_writer.add_audio_conv_kernel_size(self.hparams_audio["conv_kernel_size"])
+            self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.PARAKEET)
+            self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
+        else:
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
+
         super().set_gguf_parameters()
         hparams = self.global_config
-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.NEMOTRON_V2_VL)
         self.gguf_writer.add_vision_attention_layernorm_eps(1e-6)
         self.gguf_writer.add_vision_use_gelu(True)
         downsample_ratio = hparams.get("downsample_ratio", 0.5)
         self.gguf_writer.add_vision_projector_scale_factor(int(1.0 / downsample_ratio))
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
-        if ".position_embd." in new_name or "pos_embed" in new_name:
-            return gguf.GGMLQuantizationType.F32
+        if "sound_encoder" in name or new_name.startswith("mm.a."):
+            if "bias" in new_name or "norm" in new_name:
+                return gguf.GGMLQuantizationType.F32
+            if "conv" in new_name and "weight" in new_name:
+                return gguf.GGMLQuantizationType.F32
+
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
 
         if "input_conditioner" in name:
             return None
@@ -69,14 +89,18 @@ class NemotronNanoV2VLModel(MmprojModel):
         if "radio_model.model.patch_generator.video_embedder" in name:
             return None
 
-        if not name.startswith("vision_model.radio_model.model.") and not name.startswith("mlp1."):
+        if not name.startswith(("vision_model.radio_model.model.", "mlp1.", "sound_encoder.", "sound_projection.")):
             return None
 
         if "patch_generator.pos_embed" in name:
             if not name.endswith(".weight"):
                 name += ".weight"
 
-        return super().filter_tensors((name, gen))
+        # num_batches is only used for training not inference.
+        if "conv.norm" in name and "num_batches" in name:
+            return None
+
+        return name, gen
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # RADIO's pos_embed doesn't have .weight suffix, but clip.cpp expects it
@@ -104,7 +128,26 @@ class NemotronNanoV2VLModel(MmprojModel):
             n_embd = self.hparams["hidden_size"]
             data_torch = data_torch.reshape(n_embd, 3, patch_size, patch_size)
 
-        yield from super().modify_tensors(data_torch, name, bid)
+        if "depthwise_conv.weight" in name:
+            data_torch = data_torch.unsqueeze(-1)
+            data_torch = data_torch.permute(3, 1, 0, 2).contiguous()
+
+        if "pointwise_conv" in name and name.endswith(".weight"):
+            if len(data_torch.shape) == 3 and data_torch.shape[2] == 1:
+                data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[1])
+
+        if "subsampling.layers" in name and name.endswith(".bias"):
+            if len(data_torch.shape) == 1:
+                data_torch = data_torch.reshape(1, -1, 1, 1)
+
+        if "pointwise_conv" in name and name.endswith(".bias"):
+            if len(data_torch.shape) == 1:
+                data_torch = data_torch.reshape(1, -1, 1, 1)
+
+        for mapped_name, tensor in super().modify_tensors(data_torch, name, bid):
+            if name.startswith("sound_projection.") and mapped_name.startswith("mm.model.mlp."):
+                mapped_name = mapped_name.replace("mm.model.mlp.", "mm.a.mlp.")
+            yield mapped_name, tensor
 
 
 @ModelBase.register("NemotronForCausalLM")

conversion/qwen.py

@@ -18,7 +18,7 @@ class QwenModel(TextModel):
 
     @staticmethod
     def token_bytes_to_string(b):
-        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode  # ty: ignore[unresolved-import]
+        from transformers.convert_slow_tokenizer import bytes_to_unicode
         byte_encoder = bytes_to_unicode()
         return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])
 
@@ -268,8 +268,101 @@ class Qwen3MoeModel(Qwen2MoeModel):
         super().set_vocab()
 
 
+class _QwenMtpMixin:
+    """Shared MTP wiring for Qwen3-Next and Qwen3.5/3.6 text variants. The HF
+    config carries the MTP block under `mtp_num_hidden_layers` (computed from
+    the checkpoint when absent, e.g. Qwen3-Next) and the tensors under
+    `mtp.*`; we extend block_count, emit the nextn metadata key, and remap
+    `mtp.*` to the standard layer-indexed nextn naming so the existing
+    tensor_map handles them."""
+
+    supports_mtp_export = True
+    hparams: dict[str, Any]
+    model_arch: gguf.MODEL_ARCH
+    gguf_writer: gguf.GGUFWriter
+    block_count: int
+    tensor_map: gguf.TensorNameMap
+    no_mtp: bool
+    mtp_only: bool
+    _original_block_count: int | None = None
+    opt_num_mtp_layers: int = 0
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            n_mtp = self.hparams.get("mtp_num_hidden_layers", 0)
+            # Qwen-3-Next doesn't include `mtp_num_hidden_layers` in config.
+            if n_mtp == 0:
+                assert self.opt_num_mtp_layers != 0
+                n_mtp = self.opt_num_mtp_layers
+            self.block_count += n_mtp
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
+        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
+        type(self)._original_block_count = hparams.get(key)
+        type(self).opt_num_mtp_layers = 0
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[unresolved-attribute]
+
+    @classmethod
+    def filter_tensors(cls, item):
+        assert cls._original_block_count is not None
+        # TODO: change TextModel to super()
+        if (titem := TextModel.filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+        if name.startswith("model.mtp."):
+            name = name.replace("model.", "", 1)
+        if name.startswith("mtp."):
+            if cls.no_mtp:
+                return None
+            remapper = {
+                "fc":                    "eh_proj",
+                "pre_fc_norm_embedding": "enorm",
+                "pre_fc_norm_hidden":    "hnorm",
+                "norm":                  "shared_head.norm",
+            }
+            parts = name.split(".", 3)
+            if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal():
+                mtp_idx = int(parts[2])
+                name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}"
+                cls.opt_num_mtp_layers = max(cls.opt_num_mtp_layers, mtp_idx + 1)
+            elif len(parts) == 3 and parts[1] in remapper:
+                name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}"
+        elif cls.mtp_only:
+            keep = name in (
+                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+                "embed_tokens.weight", "norm.weight",
+            )
+            if not keep:
+                return None
+        return name, gen
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
+        if self.no_mtp:
+            return
+        if (n := self.block_count - self.hparams["num_hidden_layers"]) > 0:
+            self.gguf_writer.add_nextn_predict_layers(n)
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)  # ty: ignore[unresolved-attribute]
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,                  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
+
 @ModelBase.register("Qwen3NextForCausalLM")
-class Qwen3NextModel(Qwen2MoeModel):
+class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):
     model_arch = gguf.MODEL_ARCH.QWEN3NEXT
 
     def set_gguf_parameters(self):
@@ -284,16 +377,6 @@ class Qwen3NextModel(Qwen2MoeModel):
             rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.rope_parameters.get("partial_rotary_factor", 0.25)))
 
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
-
-        if name.startswith("mtp"):
-            # ignore MTP layers for now
-            return None
-
-        return super().filter_tensors(item)
-
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name.endswith(".A_log"):
             data_torch = -torch.exp(data_torch)
@@ -536,97 +619,13 @@ class _Qwen35MRopeMixin:
             self.gguf_writer.add_rope_dimension_sections(self._QWEN35_DEFAULT_MROPE_SECTION)
 
 
-class _Qwen35MtpMixin:
-    """Shared MTP wiring for Qwen3.5/3.6 text variants. The HF config carries
-    the MTP block under `mtp_num_hidden_layers` and the tensors under
-    `mtp.*`; we extend block_count, emit the nextn metadata key, and remap
-    `mtp.*` to the standard layer-indexed nextn naming so the existing
-    tensor_map handles them."""
-
-    supports_mtp_export = True
-    hparams: dict[str, Any]
-    model_arch: gguf.MODEL_ARCH
-    gguf_writer: gguf.GGUFWriter
-    block_count: int
-    tensor_map: gguf.TensorNameMap
-    no_mtp: bool
-    mtp_only: bool
-    _original_block_count: int | None = None
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"]
-        if not self.no_mtp:
-            self.block_count += self.hparams.get("mtp_num_hidden_layers", 0)
-        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
-
-    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
-        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
-        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
-        type(self)._original_block_count = hparams.get(key)
-        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[unresolved-attribute]
-
-    @classmethod
-    def filter_tensors(cls, item):
-        assert cls._original_block_count is not None
-        # TODO: change TextModel to super()
-        if (titem := TextModel.filter_tensors(item)) is None:
-            return None
-        name, gen = titem
-        if name.startswith("model.mtp."):
-            name = name.replace("model.", "", 1)
-        if name.startswith("mtp."):
-            if cls.no_mtp:
-                return None
-            remapper = {
-                "fc":                    "eh_proj",
-                "pre_fc_norm_embedding": "enorm",
-                "pre_fc_norm_hidden":    "hnorm",
-                "norm":                  "shared_head.norm",
-            }
-            parts = name.split(".", 3)
-            if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal():
-                mtp_idx = int(parts[2])
-                name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}"
-            elif len(parts) == 3 and parts[1] in remapper:
-                name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}"
-        elif cls.mtp_only:
-            keep = name in (
-                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
-                "embed_tokens.weight", "norm.weight",
-            )
-            if not keep:
-                return None
-        return name, gen
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
-        if self.no_mtp:
-            return
-        if (n := self.hparams.get("mtp_num_hidden_layers", 0)) > 0:
-            self.gguf_writer.add_nextn_predict_layers(n)
-
-    def prepare_metadata(self, vocab_only: bool):
-        from_dir = self.fname_out.is_dir()
-        super().prepare_metadata(vocab_only=vocab_only)  # ty: ignore[unresolved-attribute]
-
-        if not self.mtp_only or not from_dir:
-            return
-
-        output_type: str = self.ftype.name.partition("_")[2]  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-        fname_default: str = gguf.naming_convention(
-            self.metadata.name, self.metadata.basename, self.metadata.finetune,                  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-            self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
-
-
 @ModelBase.register("Qwen3_5ForConditionalGeneration", "Qwen3_5ForCausalLM")
-class Qwen3_5TextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase):
+class Qwen3_5TextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35
 
 
 @ModelBase.register("Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForCausalLM")
-class Qwen3_5MoeTextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase):
+class Qwen3_5MoeTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35MOE
 
 
@@ -688,3 +687,23 @@ class DFlashModel(Qwen3Model):
         if not name.startswith("model."):
             name = "model." + name
         return super().filter_tensors((name, gen))
+
+
+@ModelBase.register("Qwen3DSparkModel")
+class DSparkModel(DFlashModel):
+    # DSpark = DFlash + a semi-autoregressive Markov head
+    model_arch = gguf.MODEL_ARCH.DFLASH
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # normalize the flat DeepSpec schema to DFlash's nested dflash_config
+        self.hparams.setdefault("dflash_config", {
+            k: self.hparams[k] for k in ("target_layer_ids", "mask_token_id") if k in self.hparams
+        })
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if name.endswith(("embed_tokens.weight", "lm_head.weight")):
+            return None
+        return super().filter_tensors((name, gen))

conversion/qwenvl.py

@@ -179,12 +179,12 @@ class Qwen25OmniModel(Qwen2VLVisionModel, Qwen25AudioModel):
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, gen = item
 
-        if not name.startswith("visual.") and not name.startswith("audio_tower."):
-            return None
-
         if name.startswith("thinker."):
             name = name.replace("thinker.", "")
 
+        if not name.startswith("visual.") and not name.startswith("audio_tower."):
+            return None
+
         if "audio_bos_eos_token" in name:
             # this tensor is left unused in transformers code
             # https://github.com/huggingface/transformers/blob/6e3063422c4b1c014aa60c32b9254fd2902f0f28/src/transformers/models/qwen2_5_omni/modular_qwen2_5_omni.py#L1809

convert_hf_to_gguf.py

@@ -122,8 +122,12 @@ def parse_args() -> argparse.Namespace:
         help="Export only the multi-token prediction (MTP) head as a separate GGUF, suitable for use as a speculative draft. An 'mtp-' prefix will be added to the output file name.",
     )
     parser.add_argument(
-        "--no-mtp", action="store_true",
-        help="Exclude the multi-token prediction (MTP) head from the converted GGUF. Pair with --mtp on a second run to publish trunk and MTP as two files. Note: the split form duplicates embeddings, but even though the bundled default is more space-efficient overall, this allows differing quantization which may be more performant.",
+        "--no-nextn", "--no-mtp", dest="no_mtp", action="store_true",
+        help="Exclude NextN speculative draft tensors from the converted GGUF. Pair with --mtp or --dspark on a second run to publish target and draft as two files.",
+    )
+    parser.add_argument(
+        "--dspark", action="store_true",
+        help="Export only the DeepSeek-V4 DSpark draft tensors as a separate GGUF.",
     )
     parser.add_argument(
         "--mistral-format", action="store_true",
@@ -254,13 +258,20 @@ def main() -> None:
             from conversion.mistral import MistralModel
             model_class = MistralModel
 
-        if args.mtp and args.no_mtp:
-            logger.error("--mtp and --no-mtp are mutually exclusive")
+        if sum((args.mtp, args.no_mtp, args.dspark)) > 1:
+            logger.error("--mtp, --no-nextn, and --dspark are mutually exclusive")
             sys.exit(1)
 
+        if args.dspark:
+            if is_mistral_format or model_architecture != "DeepseekV4ForCausalLM":
+                logger.error("--dspark is only supported for DeepseekV4ForCausalLM")
+                sys.exit(1)
+            from conversion.deepseek import DeepseekV4DSparkModel
+            model_class = DeepseekV4DSparkModel
+
         if args.mtp or args.no_mtp:
             if not model_class.supports_mtp_export:
-                logger.error("--mtp / --no-mtp are not supported for %s", model_architecture)
+                logger.error("--mtp / --no-nextn are not supported for %s", model_architecture)
                 sys.exit(1)
             if args.no_mtp:
                 model_class.no_mtp = True

docs/backend/SYCL.md

@@ -788,13 +788,18 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | Name              | Value            | Function                                                                                                                  |
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
-| GGML_SYCL_DEV2DEV_MEMCPY | 0 (default) or 1 | Choose the SYCL or L0 API in dev2dev memory copy.<br>Value: <br>*  0: SYCL API (default)<br>* 1: L0 API -- L0 API is found to lead to abnormal crash in some case. This debug flag is used to check the issue.|
+| GGML_SYCL_DEV2DEV_MEMCPY | 0 (default), 1, 2 | Choose the method of dev2dev memory copy.<br>Value: <br>*  0: SYCL API (default), only support dGPUs.<br>* 1: L0 API -- Better performance, only support dGPUs, found to lead to abnormal crash in some case. <br>* 2: Host Forward -- Most stable method for all cases (including iGPU + dGPU*N), but with lower performance (-2% to -5%).<br>SYCL & L0 API are easy to be impacted by Intel GPU driver issue. When you meet the garbled output or crash issues in multiple GPUs case, try with this debug flag to work around or check the issue.|
 | GGML_SYCL_ENABLE_FLASH_ATTN | 1 (default) or 0| Enable Flash-Attention. It can reduce memory usage. The performance impact depends on the LLM.|
 | GGML_SYCL_ENABLE_OPT | 0 or 1 (default)| Enable optimize features for Intel GPUs. (Recommended to 0 for Intel devices older than Gen 10) |
 | GGML_SYCL_ENABLE_GRAPH | 0 (default) or 1 | Enable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
 | GGML_SYCL_USE_LEVEL_ZERO_API | 1 (default) or 0 | Use Level Zero API for device memory allocation instead of SYCL. Reduces system RAM usage on Intel dGPUs by avoiding DMA-buf/TTM host memory staging. Requires GGML_SYCL_SUPPORT_LEVEL_ZERO_API=ON at build time. SYCL backend always runs on Level Zero running time even if it's set as OFF (The SYCL api will be usage for memory allocation).|
 | GGML_SYCL_ENABLE_DNN | 0 or 1 (default)| Enable running computations through oneDNN and always use oneMKL. |
+| GGML_SYCL_FA_ONEDNN | 1 (default) or 0 | Enable the oneDNN fused SDPA (flash-attention) path on supported GPUs. Set to 0 to always use the native SYCL flash-attention kernel. |
+| GGML_SYCL_FA_ONEDNN_MAX_KV | 0 (default, disabled) or positive integer | By default (0), all sequences are handled by the oneDNN fused SDPA path, regardless of KV length; a positive value caps that length, past which sequences fall back to the native kernel. If GPU driver watchdog resets (DEVICE_LOST) occur during long-context inference, set this near the context depth where they start, e.g. 24576. |
 | GGML_SYCL_ENABLE_VMM | 0 or 1 (default) | Enable the virtual-memory device pool. |
+| GGML_SYCL_ENABLE_MKL_FA | 1 (default) or 0 | Enable oneMKL GEMM flash attention for XMX-accelerated prompt processing with quantized KV cache. Automatically activates during prefill (prompt processing) when all conditions are met: (1) flash-attn enabled (`-fa` or `--flash-attn on`), (2) KV cache quantized (`--cache-type-k q8_0 --cache-type-v q8_0` or other `*_0/*_1` types), (3) batch size ≥ 1024 (`--batch-size 1024`), (4) prompt length ≥ 1024 tokens. Set to 0 to force the TILE kernel for A/B testing. Example minimum command: `llama-cli -m model.gguf -fa -ngl 99 --cache-type-k q8_0 --cache-type-v q8_0 --batch-size 1024 -p "your prompt"` |
+| GGML_SYCL_MKL_FA_DEBUG | 0 (default) or 1 | Enable per-call diagnostic logging for MKL flash attention: GEMM/softmax timings, interleaved-head detection, and buffer memory usage. |
+| GGML_SYCL_MKL_FA_DIAG | 0 (default) or 1 | Enable output fingerprinting for MKL flash attention. Dumps the first 64 float output values for the first 6 FA calls with n_kv ≥ 1024, labeled with kernel type (MKL/TILE/VEC) for cross-kernel comparison. |
 | GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute (currently top-k MoE gating). |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Allow SYCL/Unified Runtime Level Zero device allocations larger than 4 GiB. llama.cpp's direct Level Zero allocation path requests the relaxed maximum-size limit itself when GGML_SYCL_ENABLE_LEVEL_ZERO=1. |

docs/completions.md

@@ -0,0 +1,17 @@
+# Completions
+
+Command-line completion is available for some environments.
+
+## Bash Completion
+
+```bash
+$ build/bin/llama-cli --completion-bash > ~/.llama-completion.bash
+$ source ~/.llama-completion.bash
+```
+
+Optionally this can be added to your `.bashrc` or `.bash_profile` to load it
+automatically. For example:
+
+```console
+$ echo "source ~/.llama-completion.bash" >> ~/.bashrc
+```

docs/development/HOWTO-add-model.md

@@ -144,6 +144,8 @@ Examples:
 - Gemma 3 folds the `1 +` of its `norm(1 + weight)` normalization into the weights at conversion time, so the graph just does a plain RMS norm.
 - Qwen3-Next applies its tensor permutation during conversion (in `modify_tensors`), so the graph can consume the already-permuted weights directly.
 
+Exception: a plain `weight * scale` with a constant scale is usually better left to inference time rather than folded into the weight at conversion. The scale conceptually applies to the activation, not the weight, so folding it into the weight can hurt numerical stability, and it shifts the weight's value range in a way that can make quantization worse. In this case, write the scale to GGUF as its own metadata key (e.g. `%s.attention.output_scale`, `%s.attention.value_scale`, `%s.embedding_scale`) and apply it in the graph, instead of pre-multiplying the weight tensor during conversion.
+
 ### Working with ggml_rope_ext
 
 PyTorch implementations usually prefer explicitly calculating `freq_cis`/`sin`/`cos` components. However, in llama.cpp, most RoPE operations can be handled via `ggml_rope_ext`, which does not require a sin/cos matrix. This saves memory while allowing the GGML RoPE kernel to be fused with other ops.

docs/install.md

@@ -16,22 +16,22 @@ conda-forge provides builds for:
  - Apple Metal (macOS)
 
 ```sh
-conda install -c conda-forge llama-cpp
+conda install -c conda-forge llama.cpp
 ```
 
 ```sh
-mamba install -c conda-forge llama-cpp
+mamba install -c conda-forge llama.cpp
 ```
 
 ```sh
 # Project-local installation
-pixi add llama-cpp
+pixi add llama.cpp
 
 # Global installation
-pixi global install llama-cpp
+pixi global install llama.cpp
 ```
 
-This distribution is managed on [`conda-forge/llama-cpp-feedstock`](https://github.com/conda-forge/llama.cpp-feedstock/).
+This distribution is managed on [`conda-forge/llama.cpp-feedstock`](https://github.com/conda-forge/llama.cpp-feedstock/).
 
 Shall you have any problems, please open an issue on [its issue tracker](https://github.com/conda-forge/llama.cpp-feedstock/issues).
 

docs/models.md

@@ -0,0 +1,26 @@
+# Obtaining and quantizing models
+
+The [Hugging Face](https://huggingface.co) platform hosts [thousands of models](https://huggingface.co/models?library=gguf&sort=trending) compatible with `llama.cpp`:
+
+- [Trending](https://huggingface.co/models?library=gguf&sort=trending)
+
+You can use any `llama.cpp`-compatible model from [Hugging Face](https://huggingface.co/) using this CLI argument: `-hf <user>/<model>[:quant]`. For example:
+
+```sh
+llama cli -hf ggml-org/gemma-3-1b-it-GGUF
+```
+
+You can use the same CLI invocation to download from other sites, by pointing the `MODEL_ENDPOINT` environment variable to an endpoint compatible with the Hugging Face API.
+`llama.cpp` can also run models you have downloaded locally to your filesystem.
+
+After downloading a model, use the CLI tools to run it locally - see below.
+
+`llama.cpp` requires the model to be stored in the [GGUF](https://github.com/ggml-org/ggml/blob/master/docs/gguf.md) file format. Models in other data formats can be converted to GGUF using the `convert_*.py` Python scripts in this repo.
+To learn more about model quantization, [read this documentation](../tools/quantize/README.md)
+
+The Hugging Face platform provides a variety of online tools for converting, quantizing and hosting models with `llama.cpp`:
+
+- Use the [GGUF-my-repo space](https://huggingface.co/spaces/ggml-org/gguf-my-repo) to convert to GGUF format and quantize model weights to smaller sizes
+- Use the [GGUF-my-LoRA space](https://huggingface.co/spaces/ggml-org/gguf-my-lora) to convert LoRA adapters to GGUF format (more info: https://github.com/ggml-org/llama.cpp/discussions/10123)
+- Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggml-org/llama.cpp/discussions/9268)
+- Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggml-org/llama.cpp/discussions/9669)

docs/ops.md

@@ -23,16 +23,16 @@ Legend:
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
@@ -51,8 +51,8 @@ Legend:
 |                             FILL | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |                            FLOOR | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -60,14 +60,14 @@ Legend:
 |                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -76,13 +76,13 @@ Legend:
 |                 MUL_MAT_HADAMARD | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ❌ |
+|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | 🟡 |
+|                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 |
 |                              PAD | ❌ | 🟡 | ✅ | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -103,13 +103,13 @@ Legend:
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |

docs/speculative.md

@@ -78,6 +78,38 @@ See:
 
 - #22105
 
+### DSpark (`draft-dspark`)
+
+DSpark extends DFlash with a semi-autoregressive _Markov head_: the draft still emits a whole
+block per forward pass, but each block position's logits are biased by a low-rank term keyed on
+the previous token, chained in-graph across the block. This keeps drafting at one decode per
+block while recovering some of the left-to-right signal that pure block diffusion loses.
+
+The draft is a small DeepSpec checkpoint trained for a specific target (for example
+[`deepseek-ai/dspark_qwen3_4b_block7`](https://huggingface.co/deepseek-ai/dspark_qwen3_4b_block7)
+for `Qwen/Qwen3-4B`). Convert it with `--target-model-dir` so it inherits the target's tokenizer
+and token embeddings:
+
+```bash
+python convert_hf_to_gguf.py deepseek-ai/dspark_qwen3_4b_block7 \
+    --target-model-dir Qwen/Qwen3-4B --outtype bf16 --outfile Qwen3-4B-DSpark.gguf
+
+llama-server -m Qwen3-4B.gguf -md Qwen3-4B-DSpark.gguf \
+    --spec-type draft-dspark --spec-draft-n-max 7 -fa on --jinja
+```
+
+`--spec-draft-n-max` is clamped to the draft model's trained block size.
+
+`--spec-draft-conf-min P` truncates each drafted block at the first position whose predicted
+acceptance (from the draft's confidence head, if present) falls below `P` (default 0 = disabled).
+
+Currently only drafts with a Qwen3 backbone are supported; support for other backbones
+(e.g. Gemma4) is planned.
+
+See:
+
+- #25173
+
 ### n-gram Cache (`ngram-cache`)
 
 An n-gram is a sequence of n tokens. The n-gram cache implementation maintains statistics about short n-gram sequences.
@@ -173,7 +205,7 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
 ### General Speculative Parameters
 
 ```
---spec-type [none|draft-simple|draft-eagle3|draft-dflash|draft-mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
+--spec-type [none|draft-simple|draft-eagle3|draft-dflash|draft-dspark|draft-mtp|ngram-cache|ngram-simple|ngram-map-k|ngram-map-k4v|ngram-mod]
                                         comma-separated list of types of speculative decoding to use
                                         (default: none)
                                         (env: LLAMA_ARG_SPEC_TYPE)
@@ -314,6 +346,7 @@ Specifies a comma-separated list of speculative decoding types to use.
 | `draft-simple` | Use a simple draft model for speculation |
 | `draft-eagle3` | Use an EAGLE-3 draft model that reads the target's hidden states |
 | `draft-dflash` | Use a DFlash block-diffusion draft model that emits a block per step |
+| `draft-dspark` | Use a DSpark draft model (DFlash backbone + semi-autoregressive Markov head) |
 | `draft-mtp` | Use Multi Token Prediction (MTP) heads from the main model |
 | `ngram-cache` | Use n-gram cache lookup |
 | `ngram-simple` | Use simple n-gram pattern matching |

docs/xcframework.md

@@ -0,0 +1,31 @@
+# XCFramework
+
+The XCFramework is a precompiled version of the library for iOS, visionOS, tvOS,
+and macOS. It can be used in Swift projects without the need to compile the
+library from source. For example:
+
+```swift
+// swift-tools-version: 5.10
+// The swift-tools-version declares the minimum version of Swift required to build this package.
+
+import PackageDescription
+
+let package = Package(
+    name: "MyLlamaPackage",
+    targets: [
+        .executableTarget(
+            name: "MyLlamaPackage",
+            dependencies: [
+                "LlamaFramework"
+            ]),
+        .binaryTarget(
+            name: "LlamaFramework",
+            url: "https://github.com/ggml-org/llama.cpp/releases/download/b5046/llama-b5046-xcframework.zip",
+            checksum: "c19be78b5f00d8d29a25da41042cb7afa094cbf6280a225abe614b03b20029ab"
+        )
+    ]
+)
+```
+
+The above example is using an intermediate build `b5046` of the library. This can be modified
+to use a different version by changing the URL and checksum.

examples/gen-docs/gen-docs.cpp

@@ -70,6 +70,8 @@ static void write_table(std::ostringstream & ss, std::vector<common_arg *> & opt
 
 static void write_help(std::ostringstream & ss, const md_file & md) {
     common_params params;
+    params.is_gen_docs = true;
+
     auto ctx_arg = common_params_parser_init(params, md.ex);
 
     std::vector<common_arg *> common_options;

ggml/CMakeLists.txt

@@ -4,8 +4,8 @@ project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 17)
-set(GGML_VERSION_PATCH 0)
+set(GGML_VERSION_MINOR 18)
+set(GGML_VERSION_PATCH 1)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -341,9 +341,6 @@ set(GGML_PUBLIC_HEADERS
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
-#if (GGML_METAL)
-#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
-#endif()
 install(TARGETS ggml LIBRARY PUBLIC_HEADER)
 install(TARGETS ggml-base LIBRARY)
 

ggml/include/ggml-rpc.h

@@ -6,9 +6,9 @@
 extern "C" {
 #endif
 
-#define RPC_PROTO_MAJOR_VERSION    4
+#define RPC_PROTO_MAJOR_VERSION    5
 #define RPC_PROTO_MINOR_VERSION    0
-#define RPC_PROTO_PATCH_VERSION    3
+#define RPC_PROTO_PATCH_VERSION    0
 
 #ifdef  __cplusplus
 static_assert(GGML_OP_COUNT == 101, "GGML_OP_COUNT has changed - update RPC_PROTO_PATCH_VERSION");

ggml/src/ggml-backend.cpp

@@ -765,8 +765,9 @@ struct ggml_backend_sched_split {
     int backend_id;
     int i_start;
     int i_end;
-    struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    struct ggml_tensor ** inputs;
     int n_inputs;
+    int inputs_capacity;
     // graph view of this split
     struct ggml_cgraph graph;
 };
@@ -805,8 +806,9 @@ struct ggml_backend_sched {
     int cur_copy;
     int next_copy;
     ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
-    struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    struct ggml_tensor ** graph_inputs;
     int n_graph_inputs;
+    int graph_inputs_capacity;
 
     struct ggml_context * ctx;
 
@@ -832,6 +834,36 @@ struct ggml_backend_sched {
 #define tensor_id_copy(id, backend_id, copy_id) sched->hv_tensor_copies[(id) * sched->n_backends * sched->n_copies + (backend_id) * sched->n_copies + (copy_id)]
 #define tensor_copy(tensor, backend_id, copy_id) tensor_id_copy(hash_id(tensor), backend_id, copy_id)
 
+static void ggml_backend_sched_split_inputs_grow(struct ggml_backend_sched_split * split) {
+    int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
+    if (split->inputs_capacity > 0) {
+        new_cap = 2*split->inputs_capacity;
+        GGML_LOG_WARN("%s: increasing split inputs capacity from %d to %d\n", __func__, split->inputs_capacity, new_cap);
+    }
+    auto * pnew = (struct ggml_tensor **) realloc((void *) split->inputs, new_cap * sizeof(struct ggml_tensor *));
+    if (pnew == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate %zu bytes\n", __func__, new_cap * sizeof(struct ggml_tensor *));
+        GGML_ABORT("failed to grow split inputs container");
+    }
+    split->inputs = pnew;
+    split->inputs_capacity = new_cap;
+}
+
+static void ggml_backend_sched_graph_inputs_grow(ggml_backend_sched_t sched) {
+    int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
+    if (sched->graph_inputs_capacity > 0) {
+        new_cap = 2*sched->graph_inputs_capacity;
+        GGML_LOG_WARN("%s: increasing graph inputs capacity from %d to %d\n", __func__, sched->graph_inputs_capacity, new_cap);
+    }
+    auto * pnew = (struct ggml_tensor **) realloc((void *) sched->graph_inputs, new_cap * sizeof(struct ggml_tensor *));
+    if (pnew == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate %zu bytes\n", __func__, new_cap * sizeof(struct ggml_tensor *));
+        GGML_ABORT("failed to grow graph inputs container");
+    }
+    sched->graph_inputs = pnew;
+    sched->graph_inputs_capacity = new_cap;
+}
+
 // returns the priority of the backend, lower id is higher priority
 static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
     for (int i = 0; i < sched->n_backends; i++) {
@@ -906,26 +938,35 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
     }
 
     // operations with weights are preferably run on the same backend as the weights
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        const struct ggml_tensor * src = tensor->src[i];
-        if (src == NULL) {
-            continue;
-        }
-        // skip ROPE since the rope freqs tensor is too small to choose a backend based on it
-        // not an ideal solution
-        if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
-            int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
-            // check if a backend with higher prio wants to offload the op
-            if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
-                for (int b = 0; b < src_backend_id; b++) {
-                    if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
-                        SET_CAUSE(tensor, "1.off");
-                        return b;
+    // TODO: there are exceptions (see below) - not an ideal solution
+    bool allow = true;
+
+    // skip ROPE since the rope freqs tensor is too small to choose a backend based on it
+    allow = allow && tensor->op != GGML_OP_ROPE;
+
+    // skip FLASH_ATTN_EXT since the sinks tensor is too small to choose a based based on it
+    allow = allow && tensor->op != GGML_OP_FLASH_ATTN_EXT;
+
+    if (allow) {
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            const struct ggml_tensor * src = tensor->src[i];
+            if (src == NULL) {
+                continue;
+            }
+            if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+                int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
+                // check if a backend with higher prio wants to offload the op
+                if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
+                    for (int b = 0; b < src_backend_id; b++) {
+                        if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
+                            SET_CAUSE(tensor, "1.off");
+                            return b;
+                        }
                     }
                 }
+                SET_CAUSE(tensor, "1.wgt%d", i);
+                return src_backend_id;
             }
-            SET_CAUSE(tensor, "1.wgt%d", i);
-            return src_backend_id;
         }
     }
 
@@ -1288,7 +1329,7 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                     }
                     // check if the split has too many inputs
                     // FIXME: count the number of inputs instead of only checking when full
-                    if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
+                    if (split->n_inputs >= split->inputs_capacity) {
                         const size_t id = hash_id(src);
                         int src_backend_id = sched->hv_tensor_backend_ids[id];
                         bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
@@ -1304,10 +1345,14 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                 split->i_end = i;
                 i_split++;
                 if (i_split >= sched->splits_capacity) {
+                    int old_cap = sched->splits_capacity;
                     sched->splits_capacity *= 2;
                     sched->splits = (ggml_backend_sched_split *)
                         realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
                     GGML_ASSERT(sched->splits != NULL);
+                    for (int k = old_cap; k < sched->splits_capacity; k++) {
+                        memset(&sched->splits[k], 0, sizeof(struct ggml_backend_sched_split));
+                    }
                 }
                 split = &sched->splits[i_split];
                 split->backend_id = node_backend_id;
@@ -1344,7 +1389,9 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_graph_inputs = sched->n_graph_inputs++;
-                        GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        if (n_graph_inputs >= sched->graph_inputs_capacity) {
+                            ggml_backend_sched_graph_inputs_grow(sched);
+                        }
                         sched->graph_inputs[n_graph_inputs] = src;
                     }
                 }
@@ -1364,7 +1411,9 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_inputs = split->n_inputs++;
-                        GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        if (n_inputs >= split->inputs_capacity) {
+                            ggml_backend_sched_split_inputs_grow(split);
+                        }
                         split->inputs[n_inputs] = src;
                     }
                     node->src[j] = tensor_id_copy(src_id, cur_backend_id, sched->cur_copy);
@@ -1390,7 +1439,11 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         sched->prev_leaf_backend_ids = tmp;
     }
 
-    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies;
+    int total_inputs = sched->n_graph_inputs;
+    for (int i = 0; i < sched->n_splits; i++) {
+        total_inputs += sched->splits[i].n_inputs;
+    }
+    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + total_inputs * 2 * sched->n_copies;
 
     // remember the actual graph_size for performing reallocation checks later [GGML_SCHED_DEBUG_REALLOC]
     sched->debug_prev_graph_size = sched->debug_graph_size;
@@ -1773,6 +1826,9 @@ ggml_backend_sched_t ggml_backend_sched_new(
     sched->splits = (ggml_backend_sched_split *) calloc(initial_splits_capacity, sizeof(sched->splits[0]));
     sched->splits_capacity = initial_splits_capacity;
 
+    sched->graph_inputs_capacity = GGML_SCHED_MAX_SPLIT_INPUTS;
+    sched->graph_inputs = (struct ggml_tensor **) calloc(sched->graph_inputs_capacity, sizeof(struct ggml_tensor *));
+
     for (int b = 0; b < n_backends; b++) {
         sched->backends[b] = backends[b];
         sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
@@ -1805,7 +1861,11 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
     ggml_gallocr_free(sched->galloc);
     ggml_free(sched->ctx);
     ggml_hash_set_free(&sched->hash_set);
+    for (int i = 0; i < sched->splits_capacity; i++) {
+        free(sched->splits[i].inputs);
+    }
     free(sched->splits);
+    free(sched->graph_inputs);
     free(sched->hv_tensor_backend_ids);
     free(sched->hv_tensor_copies);
     free(sched->node_backend_ids);

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -469,6 +469,8 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
             return (src0->type == GGML_TYPE_F32 ||
                     ((src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && src0->ne[2] == src1->ne[2] && src0->ne[3] == src1->ne[3])) &&
                 src1->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_CONV_2D:
+            return ggml_is_contiguous(op->src[0]);
         default:
             return true;
     }

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -1797,14 +1797,6 @@ class tinyBLAS_Q0_AVX {
 //PPC Implementation
 #if defined(__MMA__)
 
-#define SAVE_ACC(ACC, ii, jj) \
-   __builtin_mma_disassemble_acc(vec_C, ACC); \
-   for (int I = 0; I < 4; I++) { \
-      for (int J = 0; J < 4; J++) { \
-         *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J); \
-      } \
-   } \
-
 template<typename T>
 struct mma_instr;
 
@@ -1834,10 +1826,49 @@ class tinyBLAS_HP16_PPC {
     }
 
     void matmul(int64_t m, int64_t n) {
-        mnpack(0, m, 0, n);
+        int64_t mc = 256;
+        int64_t nc = 256;
+        int64_t kc = 256;
+    #if defined(_AIX) || defined(__BIG_ENDIAN__)
+        mc = 128;
+        nc = 128;
+        kc = 128;
+    #endif
+        if (k < kc) {
+            kc = k;
+        }
+        bool can_use_tiled = (m % mc == 0) && (n % nc == 0) && (k % kc == 0);
+        if (can_use_tiled) {
+            matmul_tiled(m, n, mc, nc, kc);
+        } else {
+            mnpack(0, m, 0, n);
+        }
     }
 
   private:
+    __attribute__((always_inline))
+    inline void save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                *((float *)(C+ii+((jj+J)*ldc)+I)) = *((float *)&vec_C[I]+J);
+            }
+        }
+    }
+
+    __attribute__((always_inline))
+    inline void add_save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                float * c_ptr = (float *)(C+ii+((jj+J)*ldc)+I);
+                *c_ptr += *((float *)&vec_C[I]+J);
+            }
+        }
+    }
+
     void vector_permute_store(vec_t *c, int numVec, unsigned char *vecOffset) {
         vec_t t[8], s[8];
         vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
@@ -1896,6 +1927,7 @@ class tinyBLAS_HP16_PPC {
         j = (rows >> 3);
         if (j > 0) {
             do {
+                aoffsets[0] = aoffset;
                 if (cols == 4) {
                     aoffsets[0] = aoffset;
                     for (int it = 1; it < 4; ++it)
@@ -1910,17 +1942,17 @@ class tinyBLAS_HP16_PPC {
                 }
                 i = (cols >> 3);
                 if (i > 0) {
-                    aoffsets[0] = aoffset;
                     for (int it = 1; it < 8; ++it) {
                         aoffsets[it] = aoffsets[it-1] + lda;
                     }
                     aoffset += 8 * lda;
+
                     do {
                         for (int it = 0; it < 8; ++it)
                             c_arr[it] = vec_xl(0, (vector unsigned char*)aoffsets[it]);
                         vector_permute_store(c_arr, 8, vecOffset);
                         for (int it = 0; it < 8; ++it)
-                            aoffsets[it] = aoffsets[it] + 8*lda;
+                            aoffsets[it] = aoffsets[it] + 8;
                         vecOffset += 128;
                         i--;
                     } while(i > 0);
@@ -2147,8 +2179,8 @@ class tinyBLAS_HP16_PPC {
                 mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
             }
         }
-        SAVE_ACC(&acc_0, ii, jj);
-        SAVE_ACC(&acc_1, ii, jj+4);
+        save_acc(&acc_0, ii, jj);
+        save_acc(&acc_1, ii, jj+4);
     }
 
     void KERNEL_8x4(int64_t ii, int64_t jj) {
@@ -2164,8 +2196,8 @@ class tinyBLAS_HP16_PPC {
                 mma_instr<TA>::outer_product(&acc_1, vec_A[x+4], vec_B[x]);
             }
         }
-        SAVE_ACC(&acc_0, ii, jj);
-        SAVE_ACC(&acc_1, ii+4, jj);
+        save_acc(&acc_0, ii, jj);
+        save_acc(&acc_1, ii+4, jj);
     }
 
 
@@ -2186,13 +2218,64 @@ class tinyBLAS_HP16_PPC {
                 mma_instr<TA>::outer_product(&acc_3, vec_A[x+4], vec_B[x+4]);
             }
         }
-
-        SAVE_ACC(&acc_0, ii, jj);
-        SAVE_ACC(&acc_1, ii, jj+4);
-        SAVE_ACC(&acc_2, ii+4, jj);
-        SAVE_ACC(&acc_3, ii+4, jj+4);
+        save_acc(&acc_0, ii, jj);
+        save_acc(&acc_1, ii, jj+4);
+        save_acc(&acc_2, ii+4, jj);
+        save_acc(&acc_3, ii+4, jj+4);
     }
 
+    inline void MMA_16x8(vec_t * vec_A0, vec_t * vec_A1, vec_t * vec_B, acc_t * acc) {
+        for (int x = 0; x < 4; x ++) {
+             mma_instr<TA>::outer_product(&acc[0], vec_A0[x], vec_B[x]);
+             mma_instr<TA>::outer_product(&acc[1], vec_A0[x], vec_B[x+4]);
+             mma_instr<TA>::outer_product(&acc[2], vec_A0[x+4], vec_B[x]);
+             mma_instr<TA>::outer_product(&acc[3], vec_A0[x+4], vec_B[x+4]);
+             mma_instr<TA>::outer_product(&acc[4], vec_A1[x], vec_B[x]);
+             mma_instr<TA>::outer_product(&acc[5], vec_A1[x], vec_B[x+4]);
+             mma_instr<TA>::outer_product(&acc[6], vec_A1[x+4], vec_B[x]);
+             mma_instr<TA>::outer_product(&acc[7], vec_A1[x+4], vec_B[x+4]);
+        }
+    }
+    void KERNEL(int64_t ii, int64_t jj, int64_t mc, int64_t nc, int64_t kc, vec_t * vec_A, vec_t * vec_B, int64_t kk) {
+        for (int64_t i = 0; i < mc; i += 16) {
+            int A_base_addr = (mc / 8) * (i / 8) * 8;
+            for (int64_t j = 0; j < nc; j += 8) {
+                 int B_base_addr = (nc / 8) * (j / 8) * 8;
+                 acc_t acc[8];
+                 vec_t A0_block[8]; vec_t A1_block[8];
+                 for (int x = 0; x < 8; x++)
+                     __builtin_mma_xxsetaccz(&acc[x]);
+                 for (int64_t l = 0; l < kc; l += 8) {
+                     int A0_block_idx = A_base_addr + (l / 8) * 8;
+                     int A1_block_idx = A0_block_idx + (mc / 8) * 8;
+                     int B_block_idx = B_base_addr + (l / 8) * 8;
+                     vec_t* A0_block = &vec_A[A0_block_idx];
+                     vec_t* A1_block = &vec_A[A1_block_idx];
+                     vec_t* B_block = &vec_B[B_block_idx];
+                     MMA_16x8(A0_block, A1_block, B_block, acc);
+                 }
+                 if (kk == 0) {
+                     save_acc(&acc[0], ii + i, jj + j);
+                     save_acc(&acc[1], ii + i, jj + j + 4);
+                     save_acc(&acc[2], ii + i + 4, jj + j);
+                     save_acc(&acc[3], ii + i + 4, jj + j + 4);
+                     save_acc(&acc[4], ii + i + 8, jj + j);
+                     save_acc(&acc[5], ii + i + 8, jj + j + 4);
+                     save_acc(&acc[6], ii + i + 12, jj + j);
+                     save_acc(&acc[7], ii + i + 12, jj + j + 4);
+                 } else {
+                     add_save_acc(&acc[0], ii + i, jj + j);
+                     add_save_acc(&acc[1], ii + i, jj + j + 4);
+                     add_save_acc(&acc[2], ii + i + 4, jj + j);
+                     add_save_acc(&acc[3], ii + i + 4, jj + j + 4);
+                     add_save_acc(&acc[4], ii + i + 8, jj + j);
+                     add_save_acc(&acc[5], ii + i + 8, jj + j + 4);
+                     add_save_acc(&acc[6], ii + i + 12, jj + j);
+                     add_save_acc(&acc[7], ii + i + 12, jj + j + 4);
+                 }
+            }
+        }
+    }
     template<int RM, int RN>
     void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int64_t ytiles = (m - m0) / RM;
@@ -2281,6 +2364,29 @@ class tinyBLAS_HP16_PPC {
        }
     }
 
+    void matmul_tiled(int64_t m, int64_t n, int64_t mc, int64_t nc, int64_t kc) {
+        int64_t ytiles = m / mc;
+        int64_t xtiles = n / nc;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles) {
+            end = tiles;
+        }
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = (job / xtiles) * mc;
+            int64_t jj = (job % xtiles) * nc;
+            for (int64_t kk = 0; kk < k; kk += kc) {
+                 vec_t A_pack[kc * mc / 8];
+                 vec_t B_pack[kc * nc / 8];
+                 packNormal(A + (ii * lda) + kk, lda, kc, mc, (uint8_t *)A_pack);
+                 packNormal(B + (jj * ldb) + kk, ldb, kc, nc, (uint8_t *)B_pack);
+                 KERNEL(ii, jj, mc, nc, kc, A_pack, B_pack, kk);
+            }
+        }
+    }
+
     template <int RM, int RN>
     NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int64_t ytiles = (m - m0) / RM;

ggml/src/ggml-cpu/repack.cpp

@@ -2739,7 +2739,7 @@ static block_q8_0x4 make_block_q8_0x4(block_q8_0 * in, unsigned int blck_size_in
     return out;
 }
 
-static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, int blck_size_interleave) {
     block_q4_0x4 out;
 
     for (int i = 0; i < 4; i++) {

ggml/src/ggml-cuda/common.cuh

@@ -627,7 +627,8 @@ template <typename T> struct block_reduce_policy<block_reduce_method::MAX, T> {
 };
 
 template <block_reduce_method reduce_method_t, const unsigned int block_size_template = 0, typename T>
-static __device__ T block_reduce(T val, T * shared_vals) {
+static __device__ T block_reduce(T val, [[maybe_unused]] T * shared_vals) {
+    // for multi-warp reductions, callers must not reuse shared_vals until all reads from this invocation have completed
     val                           = block_reduce_policy<reduce_method_t, T>::reduce(val);
     const unsigned int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
     if (block_size > WARP_SIZE) {
@@ -977,6 +978,13 @@ struct ggml_cuda_type_traits<GGML_TYPE_Q1_0> {
     static constexpr int qi = QI1_0;
 };
 
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q2_0> {
+    static constexpr int qk = QK2_0;
+    static constexpr int qr = QR2_0;
+    static constexpr int qi = QI2_0;
+};
+
 template<>
 struct ggml_cuda_type_traits<GGML_TYPE_Q4_0> {
     static constexpr int qk = QK4_0;

ggml/src/ggml-cuda/conv2d.cu

@@ -126,6 +126,7 @@ void ggml_cuda_op_conv2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const float *       X_D    = (const float *) input->data;
     float *             Y_D    = (float *) dst->data;
 
+    GGML_ASSERT(ggml_is_contiguous(input));
     GGML_ASSERT(ggml_is_contiguous(kernel));
     GGML_ASSERT(kernel->type == GGML_TYPE_F16 || kernel->type == GGML_TYPE_F32);
 

ggml/src/ggml-cuda/convert.cu

@@ -459,6 +459,8 @@ to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q1_0:
             return dequantize_block_cont_cuda<QK1_0, QR1_0, dequantize_q1_0>;
+        case GGML_TYPE_Q2_0:
+            return dequantize_block_cont_cuda<QK2_0, QR2_0, dequantize_q2_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_row_q4_0_cuda;
         case GGML_TYPE_Q4_1:
@@ -514,6 +516,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q1_0:
             return dequantize_block_cont_cuda<QK1_0, QR1_0, dequantize_q1_0>;
+        case GGML_TYPE_Q2_0:
+            return dequantize_block_cont_cuda<QK2_0, QR2_0, dequantize_q2_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_row_q4_0_cuda;
         case GGML_TYPE_Q4_1:
@@ -572,6 +576,8 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q1_0:
             return dequantize_block_cont_cuda<QK1_0, QR1_0, dequantize_q1_0>;
+        case GGML_TYPE_Q2_0:
+            return dequantize_block_cont_cuda<QK2_0, QR2_0, dequantize_q2_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_row_q4_0_cuda;
         case GGML_TYPE_Q4_1:
@@ -629,6 +635,8 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
             return convert_unary_cuda<float>;
         case GGML_TYPE_Q1_0:
             return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
+        case GGML_TYPE_Q2_0:
+            return dequantize_block_cuda<QK2_0, QR2_0, dequantize_q2_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:
@@ -652,6 +660,8 @@ to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
             return convert_unary_cuda<float, nv_bfloat16>;
         case GGML_TYPE_Q1_0:
             return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
+        case GGML_TYPE_Q2_0:
+            return dequantize_block_cuda<QK2_0, QR2_0, dequantize_q2_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:
@@ -675,6 +685,8 @@ to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
             return convert_unary_cuda<half, float>;
         case GGML_TYPE_Q1_0:
             return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
+        case GGML_TYPE_Q2_0:
+            return dequantize_block_cuda<QK2_0, QR2_0, dequantize_q2_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:

ggml/src/ggml-cuda/dequantize.cuh

@@ -23,6 +23,26 @@ static __device__ __forceinline__ void dequantize_q1_0(const void * vx, const in
     v.y = (2*bit_1 - 1) * d;
 }
 
+static __device__ __forceinline__ void dequantize_q2_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
+    const block_q2_0 * x = (const block_q2_0 *) vx;
+
+    const float d = x[ib].d;
+
+    // Q2_0: 2 bits per element, 4 elements per byte.
+    // Stored code c in {0,1,2,3} maps to symbol s = c - 1 in {-1, 0, +1, +2}.
+    const int byte_index_0 = iqs / 4;
+    const int bit_offset_0 = (iqs % 4) * 2;
+
+    const int byte_index_1 = (iqs + 1) / 4;
+    const int bit_offset_1 = ((iqs + 1) % 4) * 2;
+
+    const int c0 = (x[ib].qs[byte_index_0] >> bit_offset_0) & 0x3;
+    const int c1 = (x[ib].qs[byte_index_1] >> bit_offset_1) & 0x3;
+
+    v.x = (c0 - 1) * d;
+    v.y = (c1 - 1) * d;
+}
+
 static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q4_0 * x = (const block_q4_0 *) vx;
 

ggml/src/ggml-cuda/getrows.cu

@@ -320,6 +320,10 @@ static void ggml_cuda_get_rows_switch_src0_type(
             get_rows_cuda_q<QK1_0, QR1_0, dequantize_q1_0>(src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
             break;
+        case GGML_TYPE_Q2_0:
+            get_rows_cuda_q<QK2_0, QR2_0, dequantize_q2_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
         case GGML_TYPE_Q4_0:
             get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -1836,6 +1836,20 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         ggml_cuda_mul_mat_vec_f(ctx, src0, src1, nullptr, dst);
         return;
     }
+    // A transposed vector can still use MMVQ (i.e. ne01 == 1)
+    if (ne01 == 1 && ne11 > MMVF_MAX_BATCH_SIZE && ne2 == 1 && ne3 == 1
+            && src0->type == GGML_TYPE_F32
+            && ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)
+            && ggml_cuda_should_use_mmvf(src1->type, cc, src1->ne, src1->nb, /*ne11 =*/ 1)) {
+        ggml_tensor dst_vec = *dst;
+        dst_vec.ne[0] = ne11;
+        dst_vec.ne[1] = 1;
+        dst_vec.nb[1] = dst_vec.nb[0]*ne11;
+        dst_vec.nb[2] = dst_vec.nb[1];
+        dst_vec.nb[3] = dst_vec.nb[1];
+        ggml_cuda_mul_mat_vec_f(ctx, src1, src0, nullptr, &dst_vec);
+        return;
+    }
     if (ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, ne11, /*mul_mat_id =*/ false)) {
         ggml_cuda_mul_mat_f(ctx, src0, src1, nullptr, dst);
         return;
@@ -4802,6 +4816,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
                     case GGML_TYPE_Q1_0:
+                    case GGML_TYPE_Q2_0:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -4840,6 +4855,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_BF16:
                     case GGML_TYPE_I32:
                     case GGML_TYPE_Q1_0:
+                    case GGML_TYPE_Q2_0:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -5089,7 +5105,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_3D:
         case GGML_OP_CONV_2D:
-            return true;
+            return (ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]));
         case GGML_OP_CONV_2D_DW:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_TRANSPOSE_2D:

ggml/src/ggml-cuda/mmq-config-ampere.cuh

@@ -16,6 +16,23 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_Q1_0, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
     CASE(GGML_TYPE_Q1_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
 
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 256, 1, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+
     CASE(GGML_TYPE_Q4_0, 256, 1, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
     CASE(GGML_TYPE_Q4_0, 256, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
     CASE(GGML_TYPE_Q4_0, 256, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);

ggml/src/ggml-cuda/mmq-config-cdna.cuh

@@ -7,6 +7,14 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_Q1_0, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
     CASE(GGML_TYPE_Q1_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
 
+    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
+    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+    CASE(GGML_TYPE_Q2_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, false);
+
     CASE(GGML_TYPE_Q4_0, 512, 1, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
     CASE(GGML_TYPE_Q4_0, 512, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);
     CASE(GGML_TYPE_Q4_0, 512, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, true, true);

ggml/src/ggml-cuda/mmq-config-pascal.cuh

@@ -11,6 +11,18 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
     CASE(GGML_TYPE_Q4_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);

ggml/src/ggml-cuda/mmq-config-rdna2.cuh

@@ -11,6 +11,18 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
     CASE(GGML_TYPE_Q1_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
     CASE(GGML_TYPE_Q4_0, 256, 2, 128,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);

ggml/src/ggml-cuda/mmq-config-rdna3-5.cuh

@@ -0,0 +1,290 @@
+static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna3_5(ggml_type type, int J, bool fallback) {
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+// ---------------------------------------------------------------------------------------------
+
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+
+// ---------------------------------------------------------------------------------------------
+
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+// ---------------------------------------------------------------------------------------------
+
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+}

ggml/src/ggml-cuda/mmq-config-rdna3.cuh

@@ -0,0 +1,290 @@
+static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna3(ggml_type type, int J, bool fallback) {
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+// ---------------------------------------------------------------------------------------------
+
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+
+// ---------------------------------------------------------------------------------------------
+
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+
+// ---------------------------------------------------------------------------------------------
+
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+
+    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
+}

ggml/src/ggml-cuda/mmq-config-rdna4.cuh

@@ -1,77 +1,89 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna4(ggml_type type, int J, bool fallback) {
-    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_1, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_1, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_1, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_1, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q8_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q8_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q8_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
@@ -79,66 +91,62 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
 
 // ---------------------------------------------------------------------------------------------
 
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q2_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q3_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q3_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q3_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q6_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q6_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q6_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
@@ -146,105 +154,105 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
 
 // ---------------------------------------------------------------------------------------------
 
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
@@ -252,27 +260,27 @@ static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_conf
 
 // ---------------------------------------------------------------------------------------------
 
-    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_MXFP4, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_MXFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_MXFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
 
-    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);

ggml/src/ggml-cuda/mmq-load-tiles.cuh

@@ -95,6 +95,87 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
     }
 }
 
+template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q2_0(
+        const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
+    constexpr int nwarps      = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
+    constexpr int I           = ggml_cuda_mmq_get_I(type, J, fallback);
+    constexpr int sram_stride = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, I);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+
+    constexpr int blocks_per_iter = MMQ_ITER_K / QK2_0;
+    constexpr int threads_per_row = blocks_per_iter * QI2_0;
+    constexpr int nrows = warp_size / threads_per_row;
+    constexpr int scale_entries_per_block = QK2_0 / QK8_1;
+    constexpr int scale_entries_per_row = blocks_per_iter * scale_entries_per_block;
+
+    const int txi  = threadIdx.x % threads_per_row;
+    const int kbx  = txi / QI2_0;
+    const int kqsx = txi % QI2_0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < I; i0 += nrows*nwarps) {
+        int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
+
+        if (fallback) {
+            i = min(i, i_max);
+        }
+
+        const block_q2_0 * bxi = (const block_q2_0 *) x + kbx0 + i*stride + kbx;
+        const int16_t    * qxi = (const int16_t *) bxi->qs + kqsx * 4;
+
+        const int dst_offset = kbx*(scale_entries_per_block*QI8_0) + kqsx*QI8_0;
+
+#pragma unroll
+        for (int j = 0; j < 4; ++j) {
+            const int q  = qxi[j];
+
+            // unpack even and odd crumbs into byte values
+            const int qe = __byte_perm(0x020100FF, 0x020100FF, q >> 0);
+            const int qo = __byte_perm(0x020100FF, 0x020100FF, q >> 2);
+            // unshuffle values
+            const int qx = __byte_perm(qe, qo, 0x5140);
+            const int qy = __byte_perm(qe, qo, 0x7362);
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+            x_qs[i*sram_stride           + dst_offset + j*2+0] = qx;
+            x_qs[i*sram_stride           + dst_offset + j*2+1] = qy;
+#else
+            x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j*2+0] = qx;
+            x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j*2+1] = qy;
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        }
+    }
+
+    const int ksx = threadIdx.x % scale_entries_per_row;
+    const int scale_block = ksx / scale_entries_per_block;
+
+#pragma unroll
+    for (int i0 = 0; i0 < I; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (fallback) {
+            i = min(i, i_max);
+        }
+
+        const block_q2_0 * bxi = (const block_q2_0 *) x + kbx0 + i*stride + scale_block;
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        x_df[i*sram_stride                           + ksx] = bxi->d;
+#else
+        x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + ksx] = bxi->d;
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+    }
+}
+
 template <ggml_type type, int J, bool fallback> static __device__ __forceinline__ void ggml_cuda_mmq_load_tiles_q4_0(
         const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
     constexpr int warp_size   = ggml_cuda_get_physical_warp_size();

ggml/src/ggml-cuda/mmq.cu

@@ -10,6 +10,9 @@ static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, con
         case GGML_TYPE_Q1_0:
             mul_mat_q_case<GGML_TYPE_Q1_0>(ctx, args, stream);
             break;
+        case GGML_TYPE_Q2_0:
+            mul_mat_q_case<GGML_TYPE_Q2_0>(ctx, args, stream);
+            break;
         case GGML_TYPE_Q4_0:
             mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
             break;
@@ -262,6 +265,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
 
     switch (type) {
         case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -296,6 +300,15 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
         return false;
     }
 
+    // MMQ tiles require at least 48 KiB per-block shared memory; fall back to BLAS otherwise.
+    {
+        const int    id    = ggml_cuda_get_device();
+        const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+        if (smpbo < 48 * 1024) {
+            return false;
+        }
+    }
+
     if (turing_mma_available(cc)) {
         return true;
     }

ggml/src/ggml-cuda/mmq.cuh

@@ -60,6 +60,7 @@ static_assert(sizeof(block_fp4_mmq)  == sizeof(block_q8_1_mmq),    "Unexpected b
 static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
     switch (type_x) {
         case GGML_TYPE_Q1_0:
+        case GGML_TYPE_Q2_0:
             return MMQ_Q8_1_DS_LAYOUT_D4;
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
@@ -218,6 +219,8 @@ struct ggml_cuda_mmq_config {
 
 #include "mmq-config-cdna.cuh"
 #include "mmq-config-rdna2.cuh"
+#include "mmq-config-rdna3.cuh"
+#include "mmq-config-rdna3-5.cuh"
 #include "mmq-config-rdna4.cuh"
 
 #undef CASE
@@ -227,9 +230,15 @@ static __host__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(const ggml_type ty
         if (GGML_CUDA_CC_IS_CDNA(cc)) {
             return ggml_cuda_mmq_get_config_cdna(type, J, fallback);
         }
-        if (amd_wmma_available(cc)) {
+        if (GGML_CUDA_CC_IS_RDNA4(cc)) {
             return ggml_cuda_mmq_get_config_rdna4(type, J, fallback);
         }
+        if (GGML_CUDA_CC_IS_RDNA3_5(cc)) {
+            return ggml_cuda_mmq_get_config_rdna3_5(type, J, fallback);
+        }
+        if (GGML_CUDA_CC_IS_RDNA3(cc)) {  // covers RDNA 3.0
+            return ggml_cuda_mmq_get_config_rdna3(type, J, fallback);
+        }
         return ggml_cuda_mmq_get_config_rdna2(type, J, fallback);
     }
     if (blackwell_mma_available(cc)) {
@@ -245,8 +254,12 @@ static constexpr __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(ggml_t
 #ifdef GGML_USE_HIP
 #ifdef CDNA
     return ggml_cuda_mmq_get_config_cdna(type, J, fallback);
-#elif defined(AMD_WMMA_AVAILABLE)
+#elif defined(RDNA4)
     return ggml_cuda_mmq_get_config_rdna4(type, J, fallback);
+#elif defined(RDNA3_5)
+    return ggml_cuda_mmq_get_config_rdna3_5(type, J, fallback);
+#elif defined(RDNA3)
+    return ggml_cuda_mmq_get_config_rdna3(type, J, fallback);
 #else
     return ggml_cuda_mmq_get_config_rdna2(type, J, fallback);
 #endif // CDNA
@@ -373,6 +386,7 @@ static constexpr __device__ int ggml_cuda_mmq_get_rows_per_warp(ggml_type type,
 static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int I) {
     switch (type) {
         case GGML_TYPE_Q1_0:    return MMQ_DP4A_TXS_Q8_0;
+        case GGML_TYPE_Q2_0:    return MMQ_DP4A_TXS_Q8_0;
         case GGML_TYPE_Q4_0:    return MMQ_DP4A_TXS_Q4_0;
         case GGML_TYPE_Q4_1:    return MMQ_DP4A_TXS_Q4_1;
         case GGML_TYPE_Q5_0:    return MMQ_DP4A_TXS_Q8_0;
@@ -530,6 +544,12 @@ static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_func
                     ggml_cuda_mmq_load_tiles_q1_0<type, J, fallback>,
                     ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a<type, J, fallback>,
                     ggml_cuda_mmq_write_back_dp4a<type, J, fallback>);
+            case GGML_TYPE_Q2_0:
+                return ggml_cuda_mmq_util_funcs(
+                    VDR_Q2_0_Q8_1_MMQ,
+                    ggml_cuda_mmq_load_tiles_q2_0<type, J, fallback>,
+                    ggml_cuda_mmq_vec_dot_q8_0_q8_1_dp4a<type, J, fallback>,
+                    ggml_cuda_mmq_write_back_dp4a<type, J, fallback>);
             case GGML_TYPE_Q4_0:
                 return ggml_cuda_mmq_util_funcs(
                     VDR_Q4_0_Q8_1_MMQ,
@@ -688,6 +708,12 @@ static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_func
                 ggml_cuda_mmq_load_tiles_q1_0<type, J, fallback>,
                 ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
                 ggml_cuda_mmq_write_back_mma<type, J, fallback>);
+        case GGML_TYPE_Q2_0:
+            return ggml_cuda_mmq_util_funcs(
+                -1,
+                ggml_cuda_mmq_load_tiles_q2_0<type, J, fallback>,
+                ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma<type, J, fallback, MMQ_Q8_1_DS_LAYOUT_D4>,
+                ggml_cuda_mmq_write_back_mma<type, J, fallback>);
         case GGML_TYPE_Q4_0:
             return ggml_cuda_mmq_util_funcs(
                 -1,
@@ -1538,6 +1564,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
     template void mul_mat_q_case<type>(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) \
 
 extern DECL_MMQ_CASE(GGML_TYPE_Q1_0);
+extern DECL_MMQ_CASE(GGML_TYPE_Q2_0);
 extern DECL_MMQ_CASE(GGML_TYPE_Q4_0);
 extern DECL_MMQ_CASE(GGML_TYPE_Q4_1);
 extern DECL_MMQ_CASE(GGML_TYPE_Q5_0);

ggml/src/ggml-cuda/mmvq.cu

@@ -10,6 +10,7 @@ typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_
 static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q1_0:    return vec_dot_q1_0_q8_1;
+        case GGML_TYPE_Q2_0:    return vec_dot_q2_0_q8_1;
         case GGML_TYPE_Q4_0:    return vec_dot_q4_0_q8_1;
         case GGML_TYPE_Q4_1:    return vec_dot_q4_1_q8_1;
         case GGML_TYPE_Q5_0:    return vec_dot_q5_0_q8_1;
@@ -38,6 +39,7 @@ static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type)
 static constexpr __host__ __device__ int get_vdr_mmvq(ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q1_0:    return VDR_Q1_0_Q8_1_MMVQ;
+        case GGML_TYPE_Q2_0:    return VDR_Q2_0_Q8_1_MMVQ;
         case GGML_TYPE_Q4_0:    return VDR_Q4_0_Q8_1_MMVQ;
         case GGML_TYPE_Q4_1:    return VDR_Q4_1_Q8_1_MMVQ;
         case GGML_TYPE_Q5_0:    return VDR_Q5_0_Q8_1_MMVQ;
@@ -1010,6 +1012,12 @@ static void mul_mat_vec_q_switch_type(
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                  nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
+        case GGML_TYPE_Q2_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_0>
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
+            break;
         case GGML_TYPE_Q4_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,

ggml/src/ggml-cuda/norm.cu

@@ -64,7 +64,7 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
         tmp += xi * xi;
     }
 
-    tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum);
+    tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum + 32);
 
     const float variance = tmp / group_size;
     const float scale = rsqrtf(variance + eps);
@@ -297,7 +297,7 @@ static void group_norm_f32_cuda(
         group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
     } else {
         const dim3 block_dims(1024, 1, 1);
-        group_norm_f32<1024><<<num_groups, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(x, dst, group_size, ne_elements, eps);
+        group_norm_f32<1024><<<num_groups, block_dims, block_dims.x > WARP_SIZE ? 2 * 32 * sizeof(float): 0, stream>>>(x, dst, group_size, ne_elements, eps);
     }
 }
 

ggml/src/ggml-cuda/softmax.cu

@@ -116,6 +116,11 @@ static __global__ void soft_max_f32(
         vals[col] = val;
     }
 
+    if (block_size > WARP_SIZE) {
+        // sync is needed as we reuse buf_iw across block_reduce invocations, see #26385
+        // for block_size <= WARP_SIZE, block_reduce does not access buf_iw
+        __syncthreads();
+    }
     // find the sum of exps in the block
     tmp = block_reduce<block_reduce_method::SUM, block_size_template>(tmp, buf_iw);
 
@@ -142,6 +147,8 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
                                                                 float * __restrict__ dst,
                                                                 float * __restrict__ tmp_maxs,
                                                                 float * __restrict__ tmp_sums,
+                                                                float * shared_vals_max,
+                                                                float * shared_vals_sum,
                                                                 const soft_max_params p) {
     namespace cg = cooperative_groups;
 
@@ -154,7 +161,6 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     float     local_vals[n_elem_per_thread] = { -INFINITY, -INFINITY, -INFINITY, -INFINITY };
     float     local_max                     = -INFINITY;
     const int step_size                     = gridDim.x * blockDim.x;
-    __shared__ float shared_vals[32];
 
     // Compute thread-local max
     for (int col = col_start; col < p.ncols;) {
@@ -171,7 +177,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     }
 
     // Compute CTA-level max
-    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals);
+    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals_max);
 
     // Store CTA-level max to GMEM
     if (tid == 0) {
@@ -186,7 +192,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     } else {
         local_max = -INFINITY;
     }
-    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals);
+    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals_max);
 
     // Compute softmax dividends, accumulate divisor
     float tmp_expf = 0.0f;
@@ -209,7 +215,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     }
 
     // Reduce divisor within CTA
-    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals);
+    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals_sum);
 
     // Store CTA-level sum to GMEM
     if (tid == 0) {
@@ -223,7 +229,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     } else {
         tmp_expf = 0.0f;
     }
-    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals);
+    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals_sum);
 
     // Divide dividend by global sum + store data
     for (int col = col_start; col < p.ncols;) {
@@ -310,9 +316,11 @@ __launch_bounds__(8*WARP_SIZE, 1) static __global__ void soft_max_f32_paralleliz
 // https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/device-callable-apis.html#grid-synchronization
 // https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/device-callable-apis.html#class-cluster-group
 {
+    __shared__ float shared_vals[2][32];
+
     for (int rowx = 0; rowx < p.ne01 * p.ne02 * p.ne03; rowx++) {
         soft_max_f32_parallelize_cols_single_row(x + int64_t(rowx) * p.ncols, dst + int64_t(rowx) * p.ncols, tmp_maxs,
-                                                 tmp_sums, p);
+                                                 tmp_sums, shared_vals[0], shared_vals[1], p);
     }
 }
 

ggml/src/ggml-cuda/ssm-scan.cu

@@ -9,6 +9,21 @@ using namespace cub;
 
 #include "ssm-scan.cuh"
 
+
+// Minimum number of tokens to use SSD (State Space Duality) matmul path instead of scan path.
+// For n_tok <= this threshold, the scan kernel is used (lower overhead for short sequences).
+#define SSM_SSD_MIN_TOKENS 128
+
+// prepare_dt kernel dimensions: one block per (head, seq), each block handles DT_MAX_ITEMS items.
+#define SSM_SSD_DT_BLOCK     256
+#define SSM_SSD_DT_MAX_ITEMS  32
+
+// Maximum tokens the SSD path supports, derived from the prepare_dt kernel block capacity.
+#define SSM_SSD_MAX_TOKENS (SSM_SSD_DT_BLOCK * SSM_SSD_DT_MAX_ITEMS)
+
+// Chunk size for chunked SSD. Caps matmul cost at O(chunk^2) per chunk.
+#define SSM_SSD_CHUNK_SIZE 256
+
 // We would like to keep pragma unroll for cases where L_template is not 0,
 // so we suppress the clang transformation warning.
 #ifdef __clang__
@@ -316,6 +331,429 @@ static void ssm_scan_f32_cuda(const float * src0, const float * src1, const floa
     }
 }
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+// ============================================================================
+// SSD (State Space Duality) kernels for Mamba-2 prefill (n_tok > SSM_SSD_MIN_TOKENS)
+//
+// Instead of a sequential scan, SSD reformulates the output as:
+//   Y = (L (.) (C @ B^T)) @ (X * dt)  +  decay * C @ s_init
+// where L is a causal decay mask derived from A and dt.
+//
+// This converts the O(T*N) sequential scan into parallel matmuls.
+// ============================================================================
+// Softplus(dt) and inclusive prefix sum per head using CUB BlockScan.
+// Grid: (n_head, n_seqs)
+template <int BLOCK_SIZE, int MAX_ITEMS>
+__global__ void ssm_ssd_prepare_dt_kernel(
+        const float * __restrict__ dt_raw,
+        float * __restrict__ dt_sp_out,
+        float * __restrict__ cs_out,
+        const int n_head, const int n_tok,
+        const int dt_stride_tok,   // elements between tokens in dt
+        const int dt_stride_seq) { // elements between sequences in dt
+
+    const int h = blockIdx.x;
+    const int s = blockIdx.y;
+
+    const float * dt_seq = dt_raw + s * dt_stride_seq;
+
+    float * dt_sp_seq = dt_sp_out + s * n_tok * n_head;
+    float * cs_seq    = cs_out    + s * n_tok * n_head;
+
+    const int items_per_thread = (n_tok + BLOCK_SIZE - 1) / BLOCK_SIZE;
+
+    // Phase 1: softplus with interleaved distribution (t = i*BLOCK_SIZE + threadIdx.x).
+    // Each warp reads BLOCK_SIZE consecutive tokens, giving coalesced dt_raw loads
+    // (stride n_head between threads vs. items_per_thread*n_head in blocked layout).
+    float local_vals[MAX_ITEMS];
+    for (int i = 0; i < items_per_thread; i++) {
+        const int t = i * BLOCK_SIZE + threadIdx.x;
+        if (t < n_tok) {
+            float val = dt_seq[h + t * dt_stride_tok];
+            float sp = (val <= 20.0f) ? log1pf(expf(val)) : val;
+            local_vals[i] = sp;
+            dt_sp_seq[t * n_head + h] = sp;
+        } else {
+            local_vals[i] = 0.0f;
+        }
+    }
+
+    // Phase 2+3: per-step inclusive scan to build cs[] in token order.
+    // With interleaved distribution the per-thread total scan would not give token-order
+    // prefix sums, so we scan one BLOCK_SIZE slab at a time and carry a running total.
+#ifdef USE_CUB
+    using BlockScan = cub::BlockScan<float, BLOCK_SIZE>;
+    __shared__ typename BlockScan::TempStorage scan_temp;
+    __shared__ float step_total;
+
+    float running = 0.0f;
+    for (int i = 0; i < items_per_thread; i++) {
+        float inclusive;
+        BlockScan(scan_temp).InclusiveSum(local_vals[i], inclusive);
+        const int t = i * BLOCK_SIZE + threadIdx.x;
+        if (t < n_tok) {
+            cs_seq[t * n_head + h] = running + inclusive;
+        }
+        if (threadIdx.x == BLOCK_SIZE - 1) {
+            step_total = inclusive;
+        }
+        __syncthreads();
+        running += step_total;
+    }
+#else
+    // Fallback: sequential prefix scan in shared memory, one slab at a time.
+    __shared__ float sdata[BLOCK_SIZE];
+    float running = 0.0f;
+    for (int i = 0; i < items_per_thread; i++) {
+        const int t = i * BLOCK_SIZE + threadIdx.x;
+        sdata[threadIdx.x] = local_vals[i];
+        __syncthreads();
+        if (threadIdx.x == 0) {
+            for (int j = 1; j < BLOCK_SIZE; j++) {
+                sdata[j] += sdata[j - 1];
+            }
+        }
+        __syncthreads();
+        if (t < n_tok) {
+            cs_seq[t * n_head + h] = running + sdata[threadIdx.x];
+        }
+        running += sdata[BLOCK_SIZE - 1];
+        __syncthreads();
+    }
+#endif
+}
+
+// Prepare SSD matmul inputs for one chunk: X_dt, B_weighted, C_scaled.
+// T_matmul controls precision for X_dt, B_weighted (float or half).
+// C_scaled is always float (pairs with float s_cur in step 3c).
+// Computation is always FP32; only the final store converts to T_matmul.
+// Also materializes the causal M matrix = exp(A*(cs_out - cs_in)) * CB (fused with prep to save a launch).
+// Grid: (ceil(max(C*head_dim, d_state*C, chunk_len^2) / BLOCK), n_head, n_seqs)
+template <int BLOCK_SIZE, typename T_matmul>
+__global__ void ssm_ssd_pre_matmul_kernel(
+        const float * __restrict__ cs,         // {n_tok, n_head} cumulative dt sums
+        const float * __restrict__ dt_sp,      // {n_tok, n_head} softplus(dt)
+        const float * __restrict__ A,          // {1, n_head}
+        const float * __restrict__ x,          // {head_dim, n_head, n_tok, n_seqs}
+        const float * __restrict__ B,          // {d_state, n_group, n_tok, n_seqs}
+        const float * __restrict__ C_src,      // {d_state, n_group, n_tok, n_seqs}
+        T_matmul * __restrict__ X_dt,          // {head_dim, C, n_head} x * dt, d-fastest
+        T_matmul * __restrict__ B_weighted,    // {d_state, C, n_head} B * decay_from_end
+        float * __restrict__ C_scaled,         // {d_state, C, n_head} C * decay_to_pos (always float)
+        const float * __restrict__ CB,         // {chunk_len, chunk_len, n_group, n_seqs}
+        half * __restrict__ M_out,             // {chunk_len, chunk_len, n_head, n_seqs}
+        const int chunk_len, const int head_dim, const int n_head, const int n_group,
+        const int d_state, const int A_stride,
+        const int x_stride_tok, const int x_stride_seq,
+        const int B_stride_tok, const int B_stride_seq,
+        const int C_stride_tok, const int C_stride_seq,
+        const int chunk_offset,
+        const int n_tok_total) {
+
+    const int h = blockIdx.y;
+    const int s = blockIdx.z;
+    const int g = h / (n_head / n_group);
+
+    const float A_h = A[h * A_stride];
+    const int idx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
+
+    const int cs_seq_off = s * n_tok_total * n_head;
+    const float cs_base = (chunk_offset > 0) ? cs[cs_seq_off + (chunk_offset - 1) * n_head + h] : 0.0f;
+    const float cs_last = cs[cs_seq_off + (chunk_offset + chunk_len - 1) * n_head + h] - cs_base;
+
+    // Prepare X_dt = x * dt, stored d-fastest for coalesced reads and writes.
+    const int n_xdt = chunk_len * head_dim;
+    if (idx < n_xdt) {
+        const int d = idx % head_dim;
+        const int t = idx / head_dim;
+
+        const float x_val = x[s * x_stride_seq + (chunk_offset + t) * x_stride_tok + d + h * head_dim];
+        const float dt_val = dt_sp[cs_seq_off + (chunk_offset + t) * n_head + h];
+
+        X_dt[d + t * head_dim + h * n_xdt + s * n_xdt * n_head] = (T_matmul)(x_val * dt_val);
+    }
+
+    // Prepare B_weighted and C_scaled together: both share the same index space (d_state * chunk_len)
+    // and the same cs_t load, so merging halves the cs[] global memory traffic.
+    const int n_bw = d_state * chunk_len;
+    if (idx < n_bw) {
+        const int n = idx % d_state;
+        const int t = idx / d_state;
+
+        const float cs_t = cs[cs_seq_off + (chunk_offset + t) * n_head + h] - cs_base;
+
+        const float B_val = B[s * B_stride_seq + (chunk_offset + t) * B_stride_tok + g * d_state + n];
+        B_weighted[n + t * d_state + h * n_bw + s * n_bw * n_head] = (T_matmul)(B_val * __expf(A_h * (cs_last - cs_t)));
+
+        const float C_val = C_src[s * C_stride_seq + (chunk_offset + t) * C_stride_tok + g * d_state + n];
+        C_scaled[n + t * d_state + h * n_bw + s * n_bw * n_head] = C_val * __expf(A_h * cs_t);
+    }
+
+    // Materialize M = exp(A*(cs_out - cs_in)) * CB with causal mask.
+    const int n_M = chunk_len * chunk_len;
+    if (idx < n_M) {
+        const int t_out = idx % chunk_len;
+        const int t_in  = idx / chunk_len;
+
+        half val;
+        if (t_in <= t_out) {
+            const float cs_out = cs[cs_seq_off + (chunk_offset + t_out) * n_head + h] - cs_base;
+            const float cs_in  = cs[cs_seq_off + (chunk_offset + t_in)  * n_head + h] - cs_base;
+            const float decay  = __expf(A_h * (cs_out - cs_in));
+            const float * CB_g = CB + (int64_t)s * chunk_len * chunk_len * n_group
+                                   + (int64_t)g * chunk_len * chunk_len;
+            const float cb_val = CB_g[t_out + t_in * chunk_len];
+            val = __float2half(decay * cb_val);
+        } else {
+            val = __float2half(0.0f);
+        }
+
+        M_out[(int64_t)s * n_M * n_head + (int64_t)h * n_M + t_in * chunk_len + t_out] = val;
+    }
+}
+
+// Scale running state in-place: s_cur *= decay_total(chunk).
+// Called BEFORE cuBLAS state update (beta=1) to fuse inter-chunk decay.
+// Eliminates the s_old buffer and D2D memcpy vs the old approach of:
+//   memcpy(s_old, s_cur) -> cuBLAS(beta=0) -> s_cur += decay * s_old
+// Grid: (ceil(d_state * head_dim / BLOCK), n_head, n_seqs)
+template <int BLOCK_SIZE>
+__global__ void ssm_ssd_scale_state_kernel(
+        float * __restrict__ s_cur,            // {d_state, head_dim, n_head, n_seqs}
+        const float * __restrict__ cs,         // {n_tok, n_head} cumulative dt sums
+        const float * __restrict__ A,          // {1, n_head}
+        const int d_state, const int head_dim, const int n_head,
+        const int chunk_offset, const int chunk_len,
+        const int n_tok_total, const int A_stride) {
+
+    const int h = blockIdx.y;
+    const int s = blockIdx.z;
+    const int idx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
+    const int state_per_head = d_state * head_dim;
+    if (idx >= state_per_head) return;
+
+    const float A_h = A[h * A_stride];
+    const int cs_seq_off = s * n_tok_total * n_head;
+    const float cs_base = (chunk_offset > 0) ? cs[cs_seq_off + (chunk_offset - 1) * n_head + h] : 0.0f;
+    const float cs_last = cs[cs_seq_off + (chunk_offset + chunk_len - 1) * n_head + h] - cs_base;
+    const float decay_total = __expf(A_h * cs_last);
+
+    const int off = s * state_per_head * n_head + h * state_per_head + idx;
+    s_cur[off] *= decay_total;
+}
+
+// Copy initial state from src0[ids[s]] into s_cur for each sequence.
+// Grid: (ceil(d_state * head_dim * n_head / BLOCK), n_seqs)
+template <int BLOCK_SIZE>
+__global__ void ssm_ssd_init_state_kernel(
+        const float * __restrict__ src0,       // {d_state, head_dim, n_head, n_rs}
+        const int32_t * __restrict__ ids,      // {n_seqs}
+        float * __restrict__ s_cur,            // {d_state, head_dim, n_head, n_seqs}
+        const int state_size,                  // d_state * head_dim * n_head
+        const int64_t s0_stride_seq) {         // elements between state rows
+    const int s = blockIdx.y;
+    const int idx = blockIdx.x * BLOCK_SIZE + threadIdx.x;
+    if (idx >= state_size) return;
+
+    const float * s_src = src0 + (int64_t)ids[s] * s0_stride_seq;
+    s_cur[s * state_size + idx] = s_src[idx];
+}
+
+// SSD (State Space Duality) dispatch for Mamba-2 prefill.
+// Chunked matmuls: CB, materialize M + cuBLAS Y, S@C, B@X_dt.
+// All strides are in elements (floats), not bytes.
+static void ssm_scan_ssd_f32_cuda(
+        ggml_backend_cuda_context & ctx,
+        const float * src0_d, const float * src1_d, const float * src2_d, const float * src3_d,
+        const float * src4_d, const float * src5_d, const int32_t * src6_d, float * dst_d,
+        const int64_t s0_stride_seq,                                   // state (src0) stride between seqs
+        const int x_stride_tok,  const int x_stride_seq,               // x (src1) strides
+        const int dt_stride_tok, const int dt_stride_seq,              // dt (src2) strides
+        const int A_stride,                                            // A (src3) stride between heads
+        const int B_stride_tok,  const int B_stride_seq,               // B (src4) strides
+        const int C_stride_tok,  const int C_stride_seq,               // C (src5) strides
+        const int64_t s_off, const int64_t d_state, const int64_t head_dim,
+        const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq) {
+
+    cudaStream_t stream = ctx.stream();
+    const int64_t d_inner = head_dim * n_head;
+
+    const int64_t chunk_size = SSM_SSD_CHUNK_SIZE;
+    const int64_t n_chunks = (n_tok + chunk_size - 1) / chunk_size;
+
+    const int64_t state_per_head = d_state * head_dim;
+
+    using matmul_t = half;
+    static constexpr cudaDataType_t matmul_dtype = CUDA_R_16F;
+
+    ggml_cuda_pool_alloc<float>    dt_sp_buf(ctx.pool(), n_tok * n_head * n_seq);
+    ggml_cuda_pool_alloc<float>    cs_buf(ctx.pool(), n_tok * n_head * n_seq);
+    ggml_cuda_pool_alloc<float>    CB_buf(ctx.pool(), chunk_size * chunk_size * n_group * n_seq);
+    ggml_cuda_pool_alloc<matmul_t> X_dt_buf(ctx.pool(), chunk_size * head_dim * n_head * n_seq);
+    ggml_cuda_pool_alloc<matmul_t> B_w_buf(ctx.pool(), d_state * chunk_size * n_head * n_seq);
+    ggml_cuda_pool_alloc<float>    C_s_buf(ctx.pool(), d_state * chunk_size * n_head * n_seq);
+    float    * dt_sp      = dt_sp_buf.get();
+    float    * cs         = cs_buf.get();
+    float    * CB         = CB_buf.get();
+    matmul_t * X_dt       = X_dt_buf.get();
+    matmul_t * B_weighted = B_w_buf.get();
+    float    * C_scaled   = C_s_buf.get();
+    float    * s_cur      = (float *)((char *)dst_d + s_off); // write state directly to dst
+
+    // Step 1: softplus(dt) and parallel prefix sum over full sequence
+    {
+        dim3 grid(n_head, n_seq);
+        ssm_ssd_prepare_dt_kernel<SSM_SSD_DT_BLOCK, SSM_SSD_DT_MAX_ITEMS><<<grid, SSM_SSD_DT_BLOCK, 0, stream>>>(
+            src2_d, dt_sp, cs, n_head, n_tok, dt_stride_tok, dt_stride_seq);
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    // Step 2: initialize running state from src0[ids[s]]
+    {
+        constexpr int BLOCK = 256;
+        const int64_t state_size = d_state * head_dim * n_head;
+        dim3 grid((state_size + BLOCK - 1) / BLOCK, n_seq);
+        ssm_ssd_init_state_kernel<BLOCK><<<grid, BLOCK, 0, stream>>>(
+            src0_d, src6_d, s_cur, state_size, s0_stride_seq);
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    // Step 3: chunked SSD loop
+    // Per chunk: pre_matmul (incl. M) + 4 cuBLAS (CB, Y, S@C, state update) + scale_state
+    cublasHandle_t handle = ctx.cublas_handle();
+    CUBLAS_CHECK(cublasSetStream(handle, stream));
+    const float alpha_one  = 1.0f;
+    const float beta_zero  = 0.0f;
+    const float beta_one   = 1.0f;
+    const int lda_C_src = C_stride_tok;  // leading dim for C in CB = C^T @ B
+    const int ldb_B_src = B_stride_tok;  // leading dim for B in CB = C^T @ B
+
+    // Scratch buffer for causal M matrix, reused across chunks (max size at chunk_size)
+    const int64_t n_M_max = chunk_size * chunk_size;
+    ggml_cuda_pool_alloc<half> M_buf(ctx.pool(), n_M_max * n_head * n_seq);
+    half * M_mat = M_buf.get();
+
+    for (int64_t k = 0; k < n_chunks; k++) {
+        const int64_t chunk_offset = k * chunk_size;
+        const int64_t chunk_len = (chunk_offset + chunk_size <= n_tok) ? chunk_size : (n_tok - chunk_offset);
+
+        // 3a: CB = C^T @ B per group
+        for (int64_t s = 0; s < n_seq; s++) {
+            const float * C_s = src5_d + s * C_stride_seq + chunk_offset * C_stride_tok;
+            const float * B_s = src4_d + s * B_stride_seq + chunk_offset * B_stride_tok;
+            float *      CB_s = CB + s * chunk_len * chunk_len * n_group;
+
+            if (n_group == 1) {
+                CUBLAS_CHECK(cublasSgemm(handle, CUBLAS_OP_T, CUBLAS_OP_N,
+                    chunk_len, chunk_len, d_state,
+                    &alpha_one, C_s, lda_C_src, B_s, ldb_B_src,
+                    &beta_zero, CB_s, (int)chunk_len));
+            } else {
+                CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_T, CUBLAS_OP_N,
+                    chunk_len, chunk_len, d_state,
+                    &alpha_one,
+                    C_s, CUDA_R_32F, lda_C_src, d_state,
+                    B_s, CUDA_R_32F, ldb_B_src, d_state,
+                    &beta_zero,
+                    CB_s, CUDA_R_32F, (int)chunk_len, (long long)(chunk_len * chunk_len),
+                    n_group,
+                    CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
+            }
+        }
+
+        // 3b: prepare X_dt, B_weighted, C_scaled + materialize causal M matrix
+        const int64_t n_M = chunk_len * chunk_len;
+        {
+            constexpr int BLOCK = 256;
+            const int64_t n_xdt   = chunk_len * head_dim;
+            const int64_t n_bw    = d_state * chunk_len;
+            int64_t max_work = n_xdt;
+            if (n_bw  > max_work) max_work = n_bw;
+            if (n_M   > max_work) max_work = n_M;
+            dim3 grid((max_work + BLOCK - 1) / BLOCK, n_head, n_seq);
+            ssm_ssd_pre_matmul_kernel<BLOCK, matmul_t><<<grid, BLOCK, 0, stream>>>(
+                cs, dt_sp, src3_d, src1_d, src4_d, src5_d,
+                X_dt, B_weighted, C_scaled,
+                CB, M_mat,
+                chunk_len, head_dim, n_head, n_group, d_state, A_stride,
+                x_stride_tok, x_stride_seq, B_stride_tok, B_stride_seq, C_stride_tok, C_stride_seq,
+                chunk_offset, n_tok);
+            CUDA_CHECK(cudaGetLastError());
+        }
+
+        // 3c: dst = S_cur^T @ C_scaled (state contribution)
+        {
+            const int64_t stride_S  = state_per_head;
+            const int64_t stride_Cs = d_state * chunk_len;
+
+            for (int64_t s = 0; s < n_seq; s++) {
+                float * dst_chunk = dst_d + s * d_inner * n_tok + chunk_offset * d_inner;
+
+                CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_T, CUBLAS_OP_N,
+                    head_dim, chunk_len, d_state,
+                    &alpha_one,
+                    s_cur    + s * stride_S  * n_head, CUDA_R_32F, d_state, stride_S,
+                    C_scaled + s * stride_Cs * n_head, CUDA_R_32F, d_state, stride_Cs,
+                    &beta_zero,
+                    dst_chunk, CUDA_R_32F, d_inner, head_dim,
+                    n_head,
+                    CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
+            }
+        }
+
+        // 3d: dst += X_dt @ M^T (intra-chunk contribution, adds to 3c result)
+        // M is stored as M[t_out, t_in] (lower-triangular), transpose needed for Y = X @ M^T.
+        {
+            const int64_t stride_M = n_M;
+            const int64_t stride_X_h = (int64_t)chunk_len * head_dim;
+
+            for (int64_t s = 0; s < n_seq; s++) {
+                float * dst_chunk = dst_d + s * d_inner * n_tok + chunk_offset * d_inner;
+                CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_N, CUBLAS_OP_T,
+                    head_dim, chunk_len, chunk_len,
+                    &alpha_one,
+                    X_dt       + s * stride_X_h * n_head, matmul_dtype, head_dim, stride_X_h,
+                    M_mat      + s * stride_M   * n_head, matmul_dtype, chunk_len, stride_M,
+                    &beta_one,
+                    dst_chunk, CUDA_R_32F, d_inner, head_dim,
+                    n_head,
+                    CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
+            }
+        }
+
+        // 3e: s_cur = B_weighted @ X_dt^T + decay_total * s_cur_old (state update)
+        {
+            // Scale s_cur in-place by per-head decay_total BEFORE cuBLAS overwrites it
+            constexpr int BLOCK = 256;
+            dim3 grid((state_per_head + BLOCK - 1) / BLOCK, n_head, n_seq);
+            ssm_ssd_scale_state_kernel<BLOCK><<<grid, BLOCK, 0, stream>>>(
+                s_cur, cs, src3_d,
+                d_state, head_dim, n_head,
+                chunk_offset, chunk_len, n_tok, A_stride);
+            CUDA_CHECK(cudaGetLastError());
+
+            // cuBLAS with beta=1: s_cur = B_weighted @ X_dt^T + 1.0 * s_cur (already scaled)
+            const int64_t stride_Bw = d_state * chunk_len;
+            const int64_t stride_X  = chunk_len * head_dim;
+            const int64_t stride_S  = state_per_head;
+
+            for (int64_t s = 0; s < n_seq; s++) {
+                // X_dt is d-fastest {hd, C}, read as OP_T to get {C, hd}
+                CUBLAS_CHECK(cublasGemmStridedBatchedEx(handle, CUBLAS_OP_N, CUBLAS_OP_T,
+                    d_state, head_dim, chunk_len,
+                    &alpha_one,
+                    B_weighted + s * stride_Bw * n_head, matmul_dtype, d_state, stride_Bw,
+                    X_dt       + s * stride_X  * n_head, matmul_dtype, head_dim, stride_X,
+                    &beta_one,
+                    s_cur      + s * stride_S  * n_head, CUDA_R_32F, d_state, stride_S,
+                    n_head,
+                    CUBLAS_COMPUTE_32F, CUBLAS_GEMM_DEFAULT));
+            }
+        }
+    }
+}
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
 void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const struct ggml_tensor * src0 = dst->src[0];  // s
     const struct ggml_tensor * src1 = dst->src[1];  // x
@@ -357,6 +795,49 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(src6->type == GGML_TYPE_I32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
+    // Byte strides are narrowed to int for both scan and SSD paths.
+    GGML_ASSERT(src0->nb[2] <= (size_t)INT_MAX);
+    GGML_ASSERT(src0->nb[3] <= (size_t)INT_MAX);
+    GGML_ASSERT(src1->nb[2] <= (size_t)INT_MAX);
+    GGML_ASSERT(src1->nb[3] <= (size_t)INT_MAX);
+    GGML_ASSERT(src2->nb[1] <= (size_t)INT_MAX);
+    GGML_ASSERT(src2->nb[2] <= (size_t)INT_MAX);
+    GGML_ASSERT(src3->nb[1] <= (size_t)INT_MAX);
+    GGML_ASSERT(src4->nb[2] <= (size_t)INT_MAX);
+    GGML_ASSERT(src4->nb[3] <= (size_t)INT_MAX);
+    GGML_ASSERT(src5->nb[2] <= (size_t)INT_MAX);
+    GGML_ASSERT(src5->nb[3] <= (size_t)INT_MAX);
+
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    // Mamba-2 with scalar A per head: use SSD matmul path for long sequences.
+    // Requires NVIDIA Turing+ otherwise fallback to scan.
+    const bool is_mamba2 = (src3->nb[1] == sizeof(float));
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const bool use_ssd = is_mamba2 && n_t > SSM_SSD_MIN_TOKENS
+                      && n_t <= SSM_SSD_MAX_TOKENS
+                      && GGML_CUDA_CC_IS_NVIDIA(cc)
+                      && cc >= GGML_CUDA_CC_TURING
+                      && nr % 8 == 0;  // cuBLAS requires 8-element (16-byte) alignment
+
+    if (use_ssd) {
+        // ssm_ssd_init_state_kernel uses flat linear indexing within each sequence,
+        // so src0 must be fully contiguous across all inner dimensions.
+        // The scan path handles non-contiguous nb[2] via src0_nb2 but does not handle nb[1].
+        GGML_ASSERT(src0->nb[1] == nc         * sizeof(float));
+        GGML_ASSERT(src0->nb[2] == nc * nr    * sizeof(float));
+
+        ssm_scan_ssd_f32_cuda(ctx,
+            src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
+            (int64_t)(src0->nb[3] / sizeof(float)),
+            (int)(src1->nb[2] / sizeof(float)), (int)(src1->nb[3] / sizeof(float)),
+            (int)(src2->nb[1] / sizeof(float)), (int)(src2->nb[2] / sizeof(float)),
+            (int)(src3->nb[1] / sizeof(float)),
+            (int)(src4->nb[2] / sizeof(float)), (int)(src4->nb[3] / sizeof(float)),
+            (int)(src5->nb[2] / sizeof(float)), (int)(src5->nb[3] / sizeof(float)),
+            s_off, nc, nr, nh, ng, n_t, n_s);
+        return;
+    }
+#endif
     ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
                       src0->nb[2], src0->nb[3], src1->nb[2], src1->nb[3], src2->nb[1], src2->nb[2],
                       src3->nb[1], src4->nb[2], src4->nb[3], src5->nb[2], src5->nb[3],

ggml/src/ggml-cuda/template-instances/generate_cu_files.py

@@ -36,6 +36,7 @@ SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size_kq}, {head_size_v},
 
 TYPES_MMQ = [
     "GGML_TYPE_Q1_0",
+    "GGML_TYPE_Q2_0",
     "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0",
     "GGML_TYPE_Q2_K", "GGML_TYPE_Q3_K", "GGML_TYPE_Q4_K", "GGML_TYPE_Q5_K", "GGML_TYPE_Q6_K",
     "GGML_TYPE_IQ2_XXS", "GGML_TYPE_IQ2_XS", "GGML_TYPE_IQ2_S", "GGML_TYPE_IQ3_XXS", "GGML_TYPE_IQ3_S",

ggml/src/ggml-cuda/template-instances/mmq-instance-q2_0.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q2_0);

ggml/src/ggml-cuda/vecdotq.cuh

@@ -109,6 +109,9 @@ static __device__ __forceinline__ uint32_t unpack_ksigns(const uint8_t v) {
 #define VDR_Q1_0_Q8_1_MMVQ 1  // Process one 32-element chunk at a time for parallelism
 #define VDR_Q1_0_Q8_1_MMQ  4  // Q1_0 has 128 bits (4 ints) per block
 
+#define VDR_Q2_0_Q8_1_MMVQ 1  // Process one 32-element chunk at a time for parallelism
+#define VDR_Q2_0_Q8_1_MMQ  2  // Q2_0 group 64: 128 bits (4 ints) per block, 2 32-element chunks
+
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 
@@ -722,6 +725,44 @@ static __device__ __forceinline__ float vec_dot_q1_0_q8_1(
     return d1 * d8 * sumi;
 }
 
+static __device__ __forceinline__ float vec_dot_q2_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q2_0 * bq2_0 = (const block_q2_0 *) vbq + kbx;
+
+    // Q2_0 (group 64): 64 elements with ONE scale, 2 bits per element (4 elements per byte)
+    // Q8_1: 32 elements per block with individual scales
+    // iqs selects which of the 2 chunks of 32 elements to process (0-1)
+
+    const float     d2 = bq2_0->d;
+    const int16_t * qs = (const int16_t *) bq2_0->qs + iqs * 4;
+
+    // Process only the chunk specified by iqs
+    const block_q8_1 * bq8_1_chunk = bq8_1 + iqs;
+
+    int sumi = 0;
+#pragma unroll
+    for (int j = 0; j < 4; ++j) {
+        const int q  = qs[j];
+        const int u  = get_int_b4(bq8_1_chunk->qs, j*2+0);
+        const int v  = get_int_b4(bq8_1_chunk->qs, j*2+1);
+
+        // unpack even and odd crumbs into byte values
+        const int qe = __byte_perm(0x020100FF, 0x020100FF, q >> 0);
+        const int qo = __byte_perm(0x020100FF, 0x020100FF, q >> 2);
+        // unshuffle values
+        const int qx = __byte_perm(qe, qo, 0x5140);
+        const int qy = __byte_perm(qe, qo, 0x7362);
+
+        sumi = ggml_cuda_dp4a(u, qx, sumi);
+        sumi = ggml_cuda_dp4a(v, qy, sumi);
+    }
+
+    // Apply Q2_0's single scale and this chunk's Q8_1 scale
+    const float d8 = __low2float(bq8_1_chunk->ds);
+    return d2 * d8 * sumi;
+}
+
 static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
 

ggml/src/ggml-hip/CMakeLists.txt

@@ -154,5 +154,3 @@ if (GGML_HIP_RCCL)
 endif()
 
 target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas)
-
-target_compile_options(ggml-hip PRIVATE "$<$<COMPILE_LANGUAGE:HIP>:-ffast-math;-fno-finite-math-only>")

ggml/src/ggml-metal/CMakeLists.txt

@@ -24,62 +24,119 @@ if (GGML_METAL_NDEBUG)
 endif()
 
 set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
+set(METALLIB_KERNELS_COMMON     "${CMAKE_CURRENT_SOURCE_DIR}/kernels/common.h")
+set(METALLIB_KERNELS_DEQUANTIZE "${CMAKE_CURRENT_SOURCE_DIR}/kernels/dequantize.h")
+set(METALLIB_KERNELS_QUANTIZE   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantize.h")
+
+set(METALLIB_KERNEL_SOURCES
+    kernels/fa.metal
+    kernels/mul_mv.metal
+    kernels/mul_mm.metal
+    kernels/quantize.metal
+    kernels/softmax.metal
+    kernels/norm.metal
+    kernels/unary.metal
+    kernels/binbcast.metal
+    kernels/reduce.metal
+    kernels/tri.metal
+    kernels/ssm.metal
+    kernels/wkv.metal
+    kernels/gated_delta_net.metal
+    kernels/solve_tri.metal
+    kernels/rope.metal
+    kernels/conv.metal
+    kernels/upscale.metal
+    kernels/argsort.metal
+    kernels/pool.metal
+    kernels/misc.metal
+)
+
 if (GGML_METAL_EMBED_LIBRARY)
     enable_language(ASM)
 
     add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
 
-    set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
+    set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
 
     file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
 
-    # merge ggml-common.h and ggml-metal.metal into a single file
-    set(METALLIB_EMBED_ASM        "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-    set(METALLIB_SOURCE_EMBED     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
+    set(METALLIB_EMBED_ASM_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(kind ${src} NAME_WE)
+        # symbol names must be valid C identifiers ('-' is not allowed)
+        string(REPLACE "-" "_" kind_sym ${kind})
 
-    add_custom_command(
-        OUTPUT "${METALLIB_EMBED_ASM}"
-        COMMAND echo "Embedding Metal library"
-        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
-        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
-        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
-        COMMENT "Generate assembly for embedded Metal library"
-        VERBATIM
-    )
+        set(SRC   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${kind}.metal")
+        set(EMBED "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.metal")
+        set(ASM   "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.s")
 
-    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
+        # only prepend headers that this source actually includes
+        set(HEADERS_FOR_SRC ${METALLIB_KERNELS_COMMON})
+        file(STRINGS ${SRC} _has_dequantize REGEX "#include \"dequantize\\.h\"")
+        file(STRINGS ${SRC} _has_quantize   REGEX "#include \"quantize\\.h\"")
+        if(_has_dequantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_DEQUANTIZE})
+        endif()
+        if(_has_quantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_QUANTIZE})
+        endif()
+
+        add_custom_command(
+            OUTPUT "${ASM}"
+            # Step 1: concatenate shared headers + this kernel source
+            COMMAND cat ${HEADERS_FOR_SRC} ${SRC} > "${EMBED}.tmp1"
+            # Step 2: remove internal #include and #pragma once
+            COMMAND sed -e "/\#include \"common.h\"/d" -e "/\#include \"dequantize.h\"/d" -e "/\#include \"quantize.h\"/d" -e "/\#pragma once/d" < "${EMBED}.tmp1" > "${EMBED}.tmp2"
+            # Step 3: inline ggml-common.h (replacing __embed_ggml-common.h__ sentinel)
+            COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}" -e "/__embed_ggml-common.h__/d" < "${EMBED}.tmp2" > "${EMBED}.tmp3"
+            # Step 4: inline ggml-metal-impl.h
+            COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${EMBED}.tmp3" > "${EMBED}"
+            # Step 5: emit an asm chunk with kind-specific start/end symbols
+            #   note: '-' is illegal in C symbols, so we use kind_sym; the macOS
+            #   section name is limited to 16 chars so we keep it shared
+            #   across kinds (__ggml_metallib) and only vary the global symbols.
+            COMMAND echo ".section __DATA,__ggml_metallib"                       >  "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_start"               >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_start:"                     >> "${ASM}"
+            COMMAND echo .incbin "\"${EMBED}\""                                  >> "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_end"                 >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_end:"                       >> "${ASM}"
+            DEPENDS ../ggml-common.h ggml-metal-impl.h
+                    kernels/common.h kernels/dequantize.h kernels/quantize.h
+                    kernels/${kind}.metal
+            COMMENT "Generate embedded Metal library for ${kind}"
+            VERBATIM
+        )
+
+        list(APPEND METALLIB_EMBED_ASM_FILES "${ASM}")
+    endforeach()
+
+    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM_FILES})
 else()
-    # copy metal files to bin directory
+    # copy header files to bin directory
     configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
-    configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
     configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
 
+    file(MAKE_DIRECTORY "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels")
+    configure_file(kernels/common.h     ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/common.h     COPYONLY)
+    configure_file(kernels/dequantize.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/dequantize.h COPYONLY)
+    configure_file(kernels/quantize.h   ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/quantize.h   COPYONLY)
+
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        configure_file(${src} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} COPYONLY)
+    endforeach()
+
     if (GGML_METAL_SHADER_DEBUG)
-        # custom command to do the following:
-        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: disabling fast math is needed in order to pass tests/test-backend-ops
         # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
         # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
         #       ref: https://github.com/ggml-org/whisper.cpp/issues/1720
         # note: adding -g causes segmentation fault during compile
-        #set(XC_FLAGS -fno-fast-math -fno-inline -g)
         set(XC_FLAGS -fno-fast-math -fno-inline)
     else()
         set(XC_FLAGS -O3)
     endif()
 
-    # Append macOS metal versioning flags
     if (GGML_METAL_MACOSX_VERSION_MIN)
         message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
         list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
@@ -90,35 +147,46 @@ else()
         list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
     endif()
 
+    # Compile each kernel source to .air, then link into default.metallib
+    set(AIR_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(name ${src} NAME_WE)
+        set(AIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${name}.air")
+        list(APPEND AIR_FILES ${AIR})
+        add_custom_command(
+            OUTPUT ${AIR}
+            COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} -o ${AIR}
+            DEPENDS ${src} kernels/common.h kernels/dequantize.h kernels/quantize.h ${METALLIB_COMMON} ggml-metal-impl.h
+            COMMENT "Compiling ${src}"
+            VERBATIM
+        )
+    endforeach()
+
     add_custom_command(
         OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
-                xcrun -sdk macosx metallib        - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
         COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-        DEPENDS ggml-metal.metal ${METALLIB_COMMON}
-        COMMENT "Compiling Metal kernels"
-        )
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h
+        COMMAND rm -rf ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels
+        DEPENDS ${AIR_FILES}
+        COMMENT "Linking Metal kernels into default.metallib"
+    )
 
-    # FIXME: only add to the ggml-metal target?
     add_custom_target(
         ggml-metal-lib ALL
         DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
+    )
 endif() # GGML_METAL_EMBED_LIBRARY
 
 if (NOT GGML_METAL_EMBED_LIBRARY)
     install(
-        FILES src/ggml-metal/ggml-metal.metal
-        PERMISSIONS
-            OWNER_READ
-            OWNER_WRITE
-            GROUP_READ
-            WORLD_READ
-        DESTINATION ${CMAKE_INSTALL_BINDIR})
+        DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/kernels/
+        DESTINATION ${CMAKE_INSTALL_BINDIR}/kernels
+        FILES_MATCHING PATTERN "*.metal" PATTERN "*.h"
+    )
 
-        install(
-            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DESTINATION ${CMAKE_INSTALL_BINDIR}
-        )
+    install(
+        FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        DESTINATION ${CMAKE_INSTALL_BINDIR}
+    )
 endif()

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -477,6 +477,41 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max(ggml_me
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_lightning_indexer(
+        ggml_metal_library_t lib,
+        const ggml_tensor * op) {
+    GGML_ASSERT(op->op == GGML_OP_LIGHTNING_INDEXER);
+
+    char name[256];
+
+    snprintf(name, 256, "kernel_lightning_indexer_%s", ggml_type_name(op->src[1]->type));
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, name, name, nullptr);
+    }
+
+    return res;
+}
+
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_dsv4_hc(ggml_metal_library_t lib, ggml_op op) {
+    const char * name = nullptr;
+
+    switch (op) {
+        case GGML_OP_DSV4_HC_COMB: name = "kernel_dsv4_hc_comb_f32"; break;
+        case GGML_OP_DSV4_HC_PRE:  name = "kernel_dsv4_hc_pre_f32";  break;
+        case GGML_OP_DSV4_HC_POST: name = "kernel_dsv4_hc_post_f32"; break;
+        default: GGML_ABORT("fatal error");
+    }
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, name, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv(ggml_metal_library_t lib, const ggml_tensor * op) {
     GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
@@ -1252,6 +1287,21 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort_merge(gg
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_fwht(ggml_metal_library_t lib, int n) {
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_fwht_f32_%d", n);
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 // note: reuse the argsort kernel for top_k
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_TOP_K);
@@ -2102,6 +2152,23 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_opt_step_sgd(ggm
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_silu_back(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_SILU_BACK);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_silu_back_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_memset(ggml_metal_library_t lib, const ggml_tensor *  op) {
     GGML_ASSERT(op->type == GGML_TYPE_I64);
 

ggml/src/ggml-metal/ggml-metal-device.h

@@ -117,6 +117,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_concat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary             (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_silu_back         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum_rows          (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -124,6 +125,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_bl
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_add        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_tri               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_lightning_indexer (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_dsv4_hc           (ggml_metal_library_t lib, enum ggml_op op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -139,6 +142,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_id
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argmax            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort_merge     (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_fwht              (ggml_metal_library_t lib, int n);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k             (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse );
@@ -212,7 +216,7 @@ typedef void * ggml_metal_rset_t;
 // a collection of residency sets (non-owning)
 typedef struct ggml_metal_rsets * ggml_metal_rsets_t;
 
-ggml_metal_rsets_t ggml_metal_rsets_init(void);
+ggml_metal_rsets_t ggml_metal_rsets_init(ggml_metal_device_t dev);
 void ggml_metal_rsets_free(ggml_metal_rsets_t rsets);
 
 //

ggml/src/ggml-metal/ggml-metal-device.m

@@ -2,6 +2,7 @@
 
 #import "ggml-impl.h"
 #import "ggml-backend-impl.h"
+#import "ggml-metal-impl.h"
 
 #include <Foundation/Foundation.h>
 
@@ -94,8 +95,63 @@ int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_wi
     return pipeline.pipeline->obj.maxTotalThreadsPerThreadgroup;
 }
 
+//
+// MTLLibrary collection (one library per op-source, compiled separately)
+//
+
+// Single source of truth for the per-kind metal libraries. The order here
+// defines the enum values and every per-kind table below, so adding a library
+// is a one-line change here (plus adding its source to CMakeLists.txt).
+//   X(suffix, name): name is both the kernels/<name>.metal basename and the
+//   ggml_metallib_<name>_{start,end} embed-symbol stem.
+#define GGML_METAL_LIBS \
+    X(FA,              fa)             \
+    X(MUL_MV,          mul_mv)         \
+    X(MUL_MM,          mul_mm)         \
+    X(QUANTIZE,        quantize)       \
+    X(SOFTMAX,         softmax)        \
+    X(NORM,            norm)           \
+    X(UNARY,           unary)          \
+    X(BINBCAST,        binbcast)       \
+    X(REDUCE,          reduce)         \
+    X(TRI,             tri)            \
+    X(SSM,             ssm)            \
+    X(WKV,             wkv)            \
+    X(GATED_DELTA_NET, gated_delta_net)\
+    X(SOLVE_TRI,       solve_tri)      \
+    X(ROPE,            rope)           \
+    X(CONV,            conv)           \
+    X(UPSCALE,         upscale)        \
+    X(ARGSORT,         argsort)        \
+    X(POOL,            pool)           \
+    X(MISC,            misc)
+
+enum ggml_metal_lib_kind {
+#define X(e, s) GGML_METAL_LIB_##e,
+    GGML_METAL_LIBS
+#undef X
+    GGML_METAL_LIB_COUNT,
+};
+
+static const char * const k_lib_names[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = #s,
+    GGML_METAL_LIBS
+#undef X
+};
+
 struct ggml_metal_library {
-    id<MTLLibrary> obj;
+    // Per-kind compiled libraries. When single_library is true, the whole library
+    // (e.g. a pre-compiled default.metallib or a from-source build) lives at
+    // objs[0] and the remaining slots are nil.
+    id<MTLLibrary> objs[GGML_METAL_LIB_COUNT];
+    bool single_library; // true: combined library at objs[0]; false: per-kind libs in objs[*]
+
+    // Routing table: kernel function name -> objs[] index, populated from each
+    // compiled library's -[MTLLibrary functionNames]. The actual compiled
+    // libraries are the single source of truth for which library owns a kernel,
+    // so adding kernels later requires no manual routing maintenance.
+    // nil in single_library mode (everything resolves to objs[0]).
+    NSMutableDictionary<NSString *, NSNumber *> * fn_to_lib;
 
     ggml_metal_device_t dev;
     ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
@@ -103,160 +159,376 @@ struct ggml_metal_library {
     NSLock * lock;
 };
 
-ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
-    id<MTLLibrary> library = nil;
-    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+// Build the fn_to_lib routing table by querying each compiled library's public
+// function names. Call once after all per-kind libraries have been compiled.
+static void ggml_metal_library_build_index(ggml_metal_library_t lib) {
+    @autoreleasepool {
+        NSMutableDictionary<NSString *, NSNumber *> * index = [[NSMutableDictionary alloc] init];
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            for (NSString * fname in [lib->objs[kind] functionNames]) {
+                index[fname] = @(kind);
+            }
+        }
+        lib->fn_to_lib = index;
+    }
+}
 
-    // load library
-    //
-    // - first check if the library is embedded
-    // - then check if the library is in the bundle
-    // - if not found, load the source and compile it
-    // - if that fails, return NULL
-    //
-    // TODO: move to a function
-    {
-        const int64_t t_start = ggml_time_us();
+// Parse a `#include "name"` line. Returns the quoted name in *include_name on
+// success. Whitespace-tolerant; ignores `#include <...>` (system headers).
+static bool ggml_metal_library_parse_quoted_include(NSString * line, NSString ** include_name) {
+    NSScanner * scanner = [NSScanner scannerWithString:line];
+    scanner.charactersToBeSkipped = [NSCharacterSet whitespaceCharacterSet];
 
-        NSError * error = nil;
-        NSString * src = nil;
+    if (![scanner scanString:@"#" intoString:NULL] ||
+        ![scanner scanString:@"include" intoString:NULL] ||
+        ![scanner scanString:@"\"" intoString:NULL]) {
+        return false;
+    }
 
-#if GGML_METAL_EMBED_LIBRARY
-        GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+    NSString * name = nil;
+    if (![scanner scanUpToString:@"\"" intoString:&name]) {
+        return false;
+    }
 
-        extern const char ggml_metallib_start[];
-        extern const char ggml_metallib_end[];
+    if (include_name) {
+        *include_name = name;
+    }
+    return true;
+}
 
-        src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
+// Recursively inline `#include "name"` directives. System includes (<...>),
+// `#if/#else/#endif`, and other preprocessor lines are passed through to the
+// Metal compiler unchanged. `#pragma once` is dropped since `seen` already
+// guards against double-inclusion.
+static bool ggml_metal_library_flatten_file(NSMutableString * dst, NSString * path,
+                                            NSArray<NSString *> * search_paths,
+                                            NSMutableSet<NSString *> * seen, NSError ** error) {
+    NSString * key = [path stringByStandardizingPath];
+    if ([seen containsObject:key]) {
+        return true;
+    }
+    [seen addObject:key];
 
-#ifdef SWIFT_PACKAGE
-        NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
-#else
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-#endif
+    NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:error];
+    if (!src) {
+        return false;
+    }
 
-        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (path_lib == nil) {
-            // Try to find the resource in the directory where the current binary located.
-            NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
-            NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+    NSFileManager * fm = [NSFileManager defaultManager];
+    for (NSString * line in [src componentsSeparatedByString:@"\n"]) {
+        NSString * trimmed = [line stringByTrimmingCharactersInSet:[NSCharacterSet whitespaceCharacterSet]];
+        if ([trimmed isEqualToString:@"#pragma once"]) {
+            continue;
+        }
 
-            NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
-            if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
-
-                NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
-                if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
-                    // Optionally, if this is a symlink, try to resolve it.
-                    path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
-                    if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
-                        // It is a relative path, adding the binary directory as directory prefix.
-                        path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
-                    }
-                    if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                        // Link to the resource could not be resolved.
-                        path_lib_default = nil;
-                    } else {
-                        GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
-                    }
+        NSString * include_name = nil;
+        if (ggml_metal_library_parse_quoted_include(line, &include_name)) {
+            NSString * resolved = nil;
+            for (NSString * dir in search_paths) {
+                NSString * candidate = [dir stringByAppendingPathComponent:include_name];
+                if ([fm isReadableFileAtPath:candidate]) {
+                    resolved = candidate;
+                    break;
                 }
-            } else {
-                // The resource couldn't be found in the binary's directory.
-                path_lib_default = nil;
             }
-
-            path_lib = path_lib_default;
+            if (!resolved) {
+                if (error) {
+                    NSString * msg = [NSString stringWithFormat:@"could not resolve include \"%@\" from '%@'", include_name, path];
+                    *error = [NSError errorWithDomain:@"ggml-metal-source-flatten" code:1
+                                             userInfo:@{NSLocalizedDescriptionKey: msg}];
+                }
+                return false;
+            }
+            if (!ggml_metal_library_flatten_file(dst, resolved, search_paths, seen, error)) {
+                return false;
+            }
+            continue;
         }
 
-        if (path_lib != nil) {
-            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+        [dst appendString:line];
+        [dst appendString:@"\n"];
+    }
 
-            library = [device newLibraryWithURL:libURL error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        } else {
-            GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+    return true;
+}
 
-            NSString * path_source;
-            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+static NSString * ggml_metal_library_flatten_source(NSString * path_source, NSError ** error) {
+    // Search paths cover both runtime layout (build/bin/kernels + build/bin)
+    // and source-tree layout (ggml/src/ggml-metal/kernels + ggml/src/ggml-metal + ggml/src).
+    NSString * path_kernels = [path_source stringByDeletingLastPathComponent];
+    NSString * path_base    = [path_kernels stringByDeletingLastPathComponent];
+    NSArray<NSString *> * search_paths = @[
+        path_kernels,
+        path_base,
+        [path_base stringByDeletingLastPathComponent],
+    ];
 
-            GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
+    NSMutableString * src = [[NSMutableString alloc] init];
+    NSMutableSet<NSString *> * seen = [NSMutableSet set];
 
-            if (path_resource) {
-                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
-            } else {
-                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+    if (!ggml_metal_library_flatten_file(src, path_source, search_paths, seen, error)) {
+        [src release];
+        return nil;
+    }
+    return src;
+}
+
+// Compile all per-kind libraries in parallel. `source_for_kind` returns the MSL
+// source for a kind (the helper takes ownership and releases it), or nil with
+// *err set on failure. On success the objs[] slots are populated and the routing
+// index is built; on any failure every error is logged and false is returned
+// (the caller is responsible for freeing `res`).
+static bool ggml_metal_library_compile_all(
+        ggml_metal_library_t res,
+        id<MTLDevice> device,
+        NSDictionary * prep,
+        NSString * (^source_for_kind)(int kind, NSError ** err),
+        const char * origin) {
+    const int64_t t_start = ggml_time_us();
+
+    int64_t  * t_per_lib   = calloc(GGML_METAL_LIB_COUNT, sizeof(int64_t));
+    NSError ** err_per_lib = calloc(GGML_METAL_LIB_COUNT, sizeof(NSError *));
+    __block atomic_bool any_failure = false;
+
+    dispatch_group_t group = dispatch_group_create();
+    dispatch_queue_t queue = dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0);
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        dispatch_group_async(group, queue, ^{
+
+            const int64_t t0 = ggml_time_us();
+
+            NSError * error = nil;
+
+            NSString * src = source_for_kind(kind, &error);
+            if (!src) {
+                err_per_lib[kind] = [error retain];
+                atomic_store(&any_failure, true);
+                return;
             }
 
-            if (path_source == nil) {
-                GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                path_source = @"ggml-metal.metal";
-            }
+            id<MTLLibrary> lib = nil;
 
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
-
-            src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        }
-#endif
-
-        if (!library) {
             @autoreleasepool {
-                // dictionary of preprocessor macros
-                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
-
-                if (ggml_metal_device_get_props(dev)->has_bfloat) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
-                }
-
-                if (ggml_metal_device_get_props(dev)->has_tensor) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
-                }
-
-#if GGML_METAL_EMBED_LIBRARY
-                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
-#endif
-
                 MTLCompileOptions * options = [MTLCompileOptions new];
                 options.preprocessorMacros = prep;
 
-                //[options setFastMathEnabled:false];
+                lib = [device newLibraryWithSource:src options:options error:&error];
 
-                library = [device newLibraryWithSource:src options:options error:&error];
-                if (error) {
-                    GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                    return nil;
-                }
-
-#if !__has_feature(objc_arc)
                 [options release];
-#endif
+
+                // retain the error before the autorelease pool drains it
+                if (!lib) {
+                    err_per_lib[kind] = [error retain];
+                }
+            }
+
+            [src release];
+
+            t_per_lib[kind] = ggml_time_us() - t0;
+
+            if (!lib) {
+                atomic_store(&any_failure, true);
+                return;
+            }
+
+            res->objs[kind] = lib;
+        });
+    }
+    dispatch_group_wait(group, DISPATCH_TIME_FOREVER);
+    dispatch_release(group);
+
+    const bool ok = !atomic_load(&any_failure);
+
+    if (ok) {
+        const int64_t t_total = ggml_time_us() - t_start;
+        int64_t t_max = 0;
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            GGML_LOG_DEBUG("%s: compiled '%s' library in %.3f sec\n",
+                           __func__, k_lib_names[kind], t_per_lib[kind] / 1e6);
+            if (t_per_lib[kind] > t_max) t_max = t_per_lib[kind];
+        }
+        GGML_LOG_INFO("%s: loaded %d libraries from %s in %.3f sec (max single = %.3f sec)\n",
+                      __func__, GGML_METAL_LIB_COUNT, origin, t_total / 1e6, t_max / 1e6);
+
+        ggml_metal_library_build_index(res);
+    } else {
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            if (err_per_lib[kind]) {
+                GGML_LOG_ERROR("%s: failed to build '%s' library: %s\n", __func__,
+                               k_lib_names[kind], [[err_per_lib[kind] description] UTF8String]);
+                [err_per_lib[kind] release];
             }
         }
-
-#if GGML_METAL_EMBED_LIBRARY
-        [src release];
-#endif // GGML_METAL_EMBED_LIBRARY
-
-        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
     }
 
-    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
+    free(err_per_lib);
+    free(t_per_lib);
 
-    res->obj       = library;
+    return ok;
+}
+
+ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+
+    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
     res->dev       = dev;
     res->pipelines = ggml_metal_pipelines_init();
     res->lock      = [NSLock new];
 
+    // shared MTLCompileOptions preprocessor macros (matches the build-time defines)
+    NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+    if (ggml_metal_device_get_props(dev)->has_bfloat) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
+    }
+    if (ggml_metal_device_get_props(dev)->has_tensor) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
+    }
+#if GGML_METAL_EMBED_LIBRARY
+    [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
+#endif
+
+#if GGML_METAL_EMBED_LIBRARY
+    GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+
+    // start/end symbols emitted by CMake (see CMakeLists.txt), one pair per kind
+#define X(e, s) extern const char ggml_metallib_##s##_start[]; extern const char ggml_metallib_##s##_end[];
+    GGML_METAL_LIBS
+#undef X
+
+    static const char * const lib_start[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_start,
+    GGML_METAL_LIBS
+#undef X
+    };
+    static const char * const lib_end[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_end,
+    GGML_METAL_LIBS
+#undef X
+    };
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            (void) err;
+            return [[NSString alloc] initWithBytes:lib_start[kind]
+                                            length:(lib_end[kind] - lib_start[kind])
+                                          encoding:NSUTF8StringEncoding];
+        }, "embedded data");
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
     return res;
+#else
+#ifdef SWIFT_PACKAGE
+    NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
+#else
+    NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
+
+    const int64_t t_start = ggml_time_us();
+
+    NSError * error = nil;
+    NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+    if (path_lib == nil) {
+        // Try to find the resource in the directory where the current binary located.
+        NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
+        NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+
+        NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
+        if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+            GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
+
+            NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
+            if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
+                // Optionally, if this is a symlink, try to resolve it.
+                path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
+                if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
+                    // It is a relative path, adding the binary directory as directory prefix.
+                    path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
+                }
+                if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+                    // Link to the resource could not be resolved.
+                    path_lib_default = nil;
+                } else {
+                    GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
+                }
+            }
+        } else {
+            // The resource couldn't be found in the binary's directory.
+            path_lib_default = nil;
+        }
+
+        path_lib = path_lib_default;
+    }
+
+    if (path_lib != nil) {
+        // pre-compiled library found: a single combined default.metallib
+        NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+        GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
+        res->objs[0]        = [device newLibraryWithURL:libURL error:&error];
+        res->single_library = true;
+        if (!res->objs[0]) {
+            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            ggml_metal_library_free(res);
+            return NULL;
+        }
+
+        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
+        return res;
+    }
+
+    // no pre-compiled metallib: fall back to compiling each kernel source separately
+    GGML_LOG_INFO("%s: default.metallib not found, loading kernel sources\n", __func__);
+
+    NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+    if (path_resource) {
+        GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, [path_resource UTF8String]);
+    }
+
+    // resolve each kind's source path up front (file lookup/logging stays on the calling thread)
+    NSString ** path_per_kind = calloc(GGML_METAL_LIB_COUNT, sizeof(NSString *));
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        NSString * rel = [NSString stringWithFormat:@"kernels/%s.metal", k_lib_names[kind]];
+
+        NSString * path_source = nil;
+        if (path_resource) {
+            path_source = [path_resource stringByAppendingPathComponent:rel];
+        } else {
+            NSString * stem = [NSString stringWithFormat:@"kernels/%s", k_lib_names[kind]];
+            path_source = [bundle pathForResource:stem ofType:@"metal"];
+        }
+
+        if (path_source == nil || ![[NSFileManager defaultManager] isReadableFileAtPath:path_source]) {
+            GGML_LOG_WARN("%s: could not locate %s in bundle, falling back to cwd\n", __func__, [rel UTF8String]);
+            path_source = rel;
+        }
+
+        GGML_LOG_DEBUG("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+        path_per_kind[kind] = [path_source retain];
+    }
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            return ggml_metal_library_flatten_source(path_per_kind[kind], err);
+        }, "source");
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        [path_per_kind[kind] release];
+    }
+    free(path_per_kind);
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
+    return res;
+#endif
 }
 
 ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose) {
@@ -318,10 +590,11 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
         return NULL;
     }
 
-    res->obj       = library;
-    res->dev       = dev;
-    res->pipelines = ggml_metal_pipelines_init();
-    res->lock      = [NSLock new];
+    res->objs[0]        = library;
+    res->single_library = true;
+    res->dev            = dev;
+    res->pipelines      = ggml_metal_pipelines_init();
+    res->lock           = [NSLock new];
 
     return res;
 }
@@ -331,8 +604,14 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
         return;
     }
 
-    if (lib->obj) {
-        [lib->obj release];
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        if (lib->objs[kind]) {
+            [lib->objs[kind] release];
+        }
+    }
+
+    if (lib->fn_to_lib) {
+        [lib->fn_to_lib release];
     }
 
     ggml_metal_pipelines_free(lib->pipelines);
@@ -393,11 +672,28 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_
 
         GGML_LOG_DEBUG("%s: compiling pipeline: base = '%s', name = '%s'\n", __func__, base, name);
 
+        // route to the library that actually defines this kernel; fn_to_lib is
+        // built from -[MTLLibrary functionNames] so it's always in sync
+        int lib_idx = 0;
+        if (!lib->single_library) {
+            NSNumber * idx = lib->fn_to_lib[base_func];
+            if (!idx) {
+                [lib->lock unlock];
+
+                GGML_LOG_ERROR("%s: kernel not found in any metal library: base = '%s', name = '%s'\n", __func__, base, name);
+
+                return res;
+            }
+            lib_idx = [idx intValue];
+        }
+
+        id<MTLLibrary> mtl_lib = lib->objs[lib_idx];
+
         id<MTLFunction> mtl_function;
         if (!cv) {
-            mtl_function = [lib->obj newFunctionWithName:base_func];
+            mtl_function = [mtl_lib newFunctionWithName:base_func];
         } else {
-            mtl_function = [lib->obj newFunctionWithName:base_func constantValues:cv->obj error:&error];
+            mtl_function = [mtl_lib newFunctionWithName:base_func constantValues:cv->obj error:&error];
         }
         if (!mtl_function) {
             [lib->lock unlock];
@@ -557,7 +853,32 @@ struct ggml_metal_rsets {
     dispatch_group_t d_group;
 };
 
-ggml_metal_rsets_t ggml_metal_rsets_init(void) {
+#if defined(GGML_METAL_HAS_RESIDENCY_SETS)
+static void ggml_metal_dummy_work(ggml_metal_device_t dev) {
+    if (dev->mtl_queue == nil) {
+        return;
+    }
+
+    @autoreleasepool {
+        // perform a minimal dummy operation on the GPU
+        id<MTLBuffer> buf = [dev->mtl_device newBufferWithLength:1 options:MTLResourceStorageModePrivate];
+        id<MTLCommandBuffer> cmd_buf = [dev->mtl_queue commandBuffer];
+
+        {
+            id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
+
+            [encoder fillBuffer:buf range:NSMakeRange(0, 1) value:0];
+
+            [encoder endEncoding];
+        }
+
+        [cmd_buf commit];
+        [buf release];
+    }
+}
+#endif
+
+ggml_metal_rsets_t ggml_metal_rsets_init(ggml_metal_device_t dev) {
     ggml_metal_rsets_t res = calloc(1, sizeof(struct ggml_metal_rsets));
 
     res->lock = [[NSLock alloc] init];
@@ -610,6 +931,15 @@ ggml_metal_rsets_t ggml_metal_rsets_init(void) {
 #endif
     });
 
+#if defined(GGML_METAL_HAS_RESIDENCY_SETS)
+    if (@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, *)) {
+        // workaround for residency set memory not being released if no GPU operation occurs
+        // https://developer.apple.com/forums/thread/839089
+        // https://github.com/ggml-org/llama.cpp/issues/25937
+        ggml_metal_dummy_work(dev);
+    }
+#endif
+
     return res;
 }
 
@@ -864,7 +1194,7 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
             }
 
             if (dev->props.use_residency_sets) {
-                dev->rsets = ggml_metal_rsets_init();
+                dev->rsets = ggml_metal_rsets_init(dev);
             } else {
                 dev->rsets = nil;
             }
@@ -1103,6 +1433,14 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 default:
                     return false;
             }
+        case GGML_OP_SILU_BACK:
+            return (op->src[0]->type == GGML_TYPE_F32) &&
+                (op->src[1]->type == GGML_TYPE_F32) &&
+                (op->type == GGML_TYPE_F32) &&
+                ggml_is_contiguous(op->src[0]) &&
+                ggml_is_contiguous(op->src[1]) &&
+                ggml_is_contiguous(op) &&
+                ggml_are_same_shape(op->src[0], op->src[1]);
         case GGML_OP_GLU:
             switch (ggml_get_glu_op(op)) {
                 case GGML_GLU_OP_REGLU:
@@ -1147,6 +1485,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_ADD_ID:
+            return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) && (op->src[0]->type == op->src[1]->type);
         case GGML_OP_ACC:
             return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_REPEAT:
@@ -1265,6 +1604,72 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                     return false;
             }
             return has_simdgroup_mm; // TODO: over-restricted for vec-kernels
+        case GGML_OP_LIGHTNING_INDEXER:
+            if (op->src[0]->ne[0] != OP_LIGHTNING_INDEXER_DK ||
+                op->src[0]->ne[1] != OP_LIGHTNING_INDEXER_NH) {
+                return false;
+            }
+            if (!has_simdgroup_mm ||
+                op->src[0]->type != GGML_TYPE_F32 ||
+                op->src[2]->type != GGML_TYPE_F32 ||
+                op->src[3]->type != GGML_TYPE_F16 ||
+                op->type         != GGML_TYPE_F32 ||
+                !ggml_is_contiguous_rows(op->src[0]) ||
+                !ggml_is_contiguous_rows(op->src[1]) ||
+                !ggml_is_contiguous_rows(op->src[2]) ||
+                !ggml_is_contiguous_rows(op->src[3])) {
+                return false;
+            }
+            switch (op->src[1]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                    return true;
+                case GGML_TYPE_BF16:
+                    return has_bfloat;
+                default:
+                    return false;
+            }
+        case GGML_OP_DSV4_HC_COMB:
+            return has_simdgroup_reduction &&
+                op->src[0]->type == GGML_TYPE_F32 &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->src[2]->type == GGML_TYPE_F32 &&
+                op->type         == GGML_TYPE_F32 &&
+                op->src[0]->ne[0] == 24 &&
+                op->src[1]->ne[0] >= 3 &&
+                op->src[2]->ne[0] == 24 &&
+                ggml_is_contiguous_rows(op->src[0]) &&
+                ggml_is_contiguous_rows(op->src[1]) &&
+                ggml_is_contiguous_rows(op->src[2]);
+        case GGML_OP_DSV4_HC_PRE:
+            return has_simdgroup_reduction &&
+                op->src[0]->type == GGML_TYPE_F32 &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->type         == GGML_TYPE_F32 &&
+                op->src[0]->ne[1] == 4 &&
+                op->src[1]->ne[0] == 4 &&
+                ggml_is_contiguous_rows(op->src[0]) &&
+                ggml_is_contiguous_rows(op->src[1]);
+        case GGML_OP_DSV4_HC_POST:
+            return has_simdgroup_reduction &&
+                op->src[0]->type == GGML_TYPE_F32 &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->src[2]->type == GGML_TYPE_F32 &&
+                op->src[3]->type == GGML_TYPE_F32 &&
+                op->type         == GGML_TYPE_F32 &&
+                op->src[1]->ne[1] == 4 &&
+                op->src[2]->ne[0] == 4 &&
+                op->src[3]->ne[0] == 4 &&
+                op->src[3]->ne[1] == 4 &&
+                ggml_is_contiguous_rows(op->src[0]) &&
+                ggml_is_contiguous_rows(op->src[1]) &&
+                ggml_is_contiguous_rows(op->src[2]) &&
+                ggml_is_contiguous_rows(op->src[3]);
         case GGML_OP_SSM_CONV:
         case GGML_OP_SSM_SCAN:
             return has_simdgroup_reduction;
@@ -1484,6 +1889,7 @@ static void ggml_metal_buffer_rset_free(ggml_metal_buffer_t buf) {
         if (buf->rset) {
             [buf->rset endResidency];
             [buf->rset removeAllAllocations];
+            [buf->rset commit];
             [buf->rset release];
         }
     }

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -112,6 +112,13 @@
 #define OP_FLASH_ATTN_EXT_VEC_NQPSG 1
 #define OP_FLASH_ATTN_EXT_VEC_NCPSG 32
 
+#define OP_LIGHTNING_INDEXER_DK    128
+#define OP_LIGHTNING_INDEXER_NH     64
+#define OP_LIGHTNING_INDEXER_NHPTG   8
+#define OP_LIGHTNING_INDEXER_NKPSG   8
+#define OP_LIGHTNING_INDEXER_NSG     8
+#define OP_LIGHTNING_INDEXER_NBPTG   8
+
 #define OP_UNARY_NUM_SCALE      10
 #define OP_UNARY_NUM_FILL       11
 #define OP_UNARY_NUM_CLAMP      12
@@ -1157,6 +1164,10 @@ typedef struct {
     int32_t  len;
 } ggml_metal_kargs_argsort_merge;
 
+typedef struct {
+    int32_t nrows;
+} ggml_metal_kargs_fwht;
+
 typedef struct {
     int64_t  ne0;
     float    start;
@@ -1167,6 +1178,66 @@ typedef struct {
     int64_t val;
 } ggml_metal_kargs_memset;
 
+typedef struct {
+    int32_t  n_kv;
+    int32_t  n_batch;
+    int32_t  mask_ne3;
+    uint64_t nb1;
+    uint64_t nb3;
+    uint64_t nbq1;
+    uint64_t nbq2;
+    uint64_t nbq3;
+    uint64_t nbk2;
+    uint64_t nbk3;
+    uint64_t nbw1;
+    uint64_t nbw3;
+    uint64_t nbm1;
+    uint64_t nbm3;
+} ggml_metal_kargs_lightning_indexer;
+
+typedef struct {
+    int32_t  n_tokens;
+    int32_t  n_iter;
+    uint64_t nb_m0;
+    uint64_t nb_m1;
+    uint64_t nb_s0;
+    uint64_t nb_b0;
+    uint64_t nb_d0;
+    uint64_t nb_d1;
+    uint64_t nb_d2;
+    float    eps;
+} ggml_metal_kargs_dsv4_hc_comb;
+
+typedef struct {
+    int32_t  n_embd;
+    int32_t  n_tokens;
+    uint64_t nb_x0;
+    uint64_t nb_x1;
+    uint64_t nb_x2;
+    uint64_t nb_w0;
+    uint64_t nb_w1;
+    uint64_t nb_d0;
+    uint64_t nb_d1;
+} ggml_metal_kargs_dsv4_hc_pre;
+
+typedef struct {
+    int32_t  n_embd;
+    int32_t  n_tokens;
+    uint64_t nb_x0;
+    uint64_t nb_x1;
+    uint64_t nb_r0;
+    uint64_t nb_r1;
+    uint64_t nb_r2;
+    uint64_t nb_p0;
+    uint64_t nb_p1;
+    uint64_t nb_c0;
+    uint64_t nb_c1;
+    uint64_t nb_c2;
+    uint64_t nb_d0;
+    uint64_t nb_d1;
+    uint64_t nb_d2;
+} ggml_metal_kargs_dsv4_hc_post;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -1218,4 +1289,8 @@ typedef struct {
     int64_t  np;
 } ggml_metal_kargs_opt_step_sgd;
 
+typedef struct {
+    int64_t ne;
+} ggml_metal_kargs_silu_back;
+
 #endif // GGML_METAL_IMPL

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -299,6 +299,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_unary(ctx, idx);
             } break;
+        case GGML_OP_SILU_BACK:
+            {
+                n_fuse = ggml_metal_op_silu_back(ctx, idx);
+            } break;
         case GGML_OP_GLU:
             {
                 n_fuse = ggml_metal_op_glu(ctx, idx);
@@ -316,6 +320,16 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_cumsum(ctx, idx);
             } break;
+        case GGML_OP_LIGHTNING_INDEXER:
+            {
+                n_fuse = ggml_metal_op_lightning_indexer(ctx, idx);
+            } break;
+        case GGML_OP_DSV4_HC_COMB:
+        case GGML_OP_DSV4_HC_PRE:
+        case GGML_OP_DSV4_HC_POST:
+            {
+                n_fuse = ggml_metal_op_dsv4_hc(ctx, idx);
+            } break;
         case GGML_OP_SOFT_MAX:
             {
                 n_fuse = ggml_metal_op_soft_max(ctx, idx);
@@ -1297,6 +1311,203 @@ int ggml_metal_op_diag(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_lightning_indexer(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_ASSERT(op->op == GGML_OP_LIGHTNING_INDEXER);
+
+    const ggml_tensor * q = op->src[0];
+    const ggml_tensor * k = op->src[1];
+    const ggml_tensor * w = op->src[2];
+    const ggml_tensor * m = op->src[3];
+
+    GGML_ASSERT(q->type == GGML_TYPE_F32);
+    GGML_ASSERT(k->type == GGML_TYPE_F32  ||
+                k->type == GGML_TYPE_F16  ||
+                k->type == GGML_TYPE_BF16 ||
+                k->type == GGML_TYPE_Q4_0 ||
+                k->type == GGML_TYPE_Q4_1 ||
+                k->type == GGML_TYPE_Q5_0 ||
+                k->type == GGML_TYPE_Q5_1 ||
+                k->type == GGML_TYPE_Q8_0);
+    GGML_ASSERT(w->type == GGML_TYPE_F32);
+    GGML_ASSERT(m->type == GGML_TYPE_F16);
+    GGML_ASSERT(op->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(q->ne[0] == OP_LIGHTNING_INDEXER_DK);
+    GGML_ASSERT(q->ne[1] == OP_LIGHTNING_INDEXER_NH);
+
+    ggml_metal_kargs_lightning_indexer args = {
+        /*.n_kv      =*/ (int32_t) k->ne[2],
+        /*.n_batch   =*/ (int32_t) q->ne[2],
+        /*.mask_ne3  =*/ (int32_t) m->ne[3],
+        /*.nb1       =*/ op->nb[1],
+        /*.nb3       =*/ op->nb[3],
+        /*.nbq1      =*/ q->nb[1],
+        /*.nbq2      =*/ q->nb[2],
+        /*.nbq3      =*/ q->nb[3],
+        /*.nbk2      =*/ k->nb[2],
+        /*.nbk3      =*/ k->nb[3],
+        /*.nbw1      =*/ w->nb[1],
+        /*.nbw3      =*/ w->nb[3],
+        /*.nbm1      =*/ m->nb[1],
+        /*.nbm3      =*/ m->nb[3],
+    };
+
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(q),  1);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(k),  2);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(w),  3);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(m),  4);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 5);
+
+    const int nsg   = OP_LIGHTNING_INDEXER_NSG;
+    const int nkptg = OP_LIGHTNING_INDEXER_NKPSG*nsg;
+    const int nbptg = OP_LIGHTNING_INDEXER_NBPTG;
+
+    auto pipeline = ggml_metal_library_get_pipeline_lightning_indexer(ctx->lib, op);
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_dispatch_threadgroups(enc,
+            (k->ne[2] + nkptg - 1)/nkptg,
+            (q->ne[2] + nbptg - 1)/nbptg,
+            q->ne[3], 32, nsg, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_dsv4_hc(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_encoder_t enc = ctx->enc;
+    auto pipeline = ggml_metal_library_get_pipeline_dsv4_hc(ctx->lib, op->op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+
+    switch (op->op) {
+        case GGML_OP_DSV4_HC_COMB:
+            {
+                const ggml_tensor * mixes = op->src[0];
+                const ggml_tensor * scale = op->src[1];
+                const ggml_tensor * base  = op->src[2];
+
+                GGML_ASSERT(mixes->type == GGML_TYPE_F32);
+                GGML_ASSERT(scale->type == GGML_TYPE_F32);
+                GGML_ASSERT(base->type  == GGML_TYPE_F32);
+                GGML_ASSERT(op->type    == GGML_TYPE_F32);
+                GGML_ASSERT(mixes->ne[0] == 24);
+                GGML_ASSERT(op->ne[0] == 4 && op->ne[1] == 4);
+
+                ggml_metal_kargs_dsv4_hc_comb args = {
+                    /*.n_tokens =*/ (int32_t) mixes->ne[1],
+                    /*.n_iter   =*/ ggml_get_op_params_i32(op, 1),
+                    /*.nb_m0    =*/ mixes->nb[0],
+                    /*.nb_m1    =*/ mixes->nb[1],
+                    /*.nb_s0    =*/ scale->nb[0],
+                    /*.nb_b0    =*/ base->nb[0],
+                    /*.nb_d0    =*/ op->nb[0],
+                    /*.nb_d1    =*/ op->nb[1],
+                    /*.nb_d2    =*/ op->nb[2],
+                    /*.eps      =*/ ggml_get_op_params_f32(op, 0),
+                };
+
+                ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(mixes), 1);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(scale), 2);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(base),  3);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),    4);
+
+                // One SIMDgroup owns one 4x4 Sinkhorn matrix. Packing up to four
+                // independent tokens per threadgroup keeps both decode and prompt
+                // dispatches compact without any threadgroup-memory synchronization.
+                const int nsg = std::min(4, args.n_tokens);
+                ggml_metal_encoder_dispatch_threadgroups(
+                        enc, (args.n_tokens + nsg - 1)/nsg, 1, 1, 32, nsg, 1);
+            } break;
+        case GGML_OP_DSV4_HC_PRE:
+            {
+                const ggml_tensor * x       = op->src[0];
+                const ggml_tensor * weights = op->src[1];
+
+                GGML_ASSERT(x->type       == GGML_TYPE_F32);
+                GGML_ASSERT(weights->type == GGML_TYPE_F32);
+                GGML_ASSERT(op->type      == GGML_TYPE_F32);
+                GGML_ASSERT(x->ne[1] == 4);
+
+                ggml_metal_kargs_dsv4_hc_pre args = {
+                    /*.n_embd   =*/ (int32_t) x->ne[0],
+                    /*.n_tokens =*/ (int32_t) x->ne[2],
+                    /*.nb_x0    =*/ x->nb[0],
+                    /*.nb_x1    =*/ x->nb[1],
+                    /*.nb_x2    =*/ x->nb[2],
+                    /*.nb_w0    =*/ weights->nb[0],
+                    /*.nb_w1    =*/ weights->nb[1],
+                    /*.nb_d0    =*/ op->nb[0],
+                    /*.nb_d1    =*/ op->nb[1],
+                };
+
+                ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(x),       1);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(weights), 2);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),      3);
+
+                const int n_tiles = (args.n_embd + 31)/32;
+                const int nsg = std::min(4, n_tiles);
+                ggml_metal_encoder_dispatch_threadgroups(
+                        enc, (n_tiles + nsg - 1)/nsg, args.n_tokens, 1, 32, nsg, 1);
+            } break;
+        case GGML_OP_DSV4_HC_POST:
+            {
+                const ggml_tensor * x        = op->src[0];
+                const ggml_tensor * residual = op->src[1];
+                const ggml_tensor * post     = op->src[2];
+                const ggml_tensor * comb     = op->src[3];
+
+                GGML_ASSERT(x->type        == GGML_TYPE_F32);
+                GGML_ASSERT(residual->type == GGML_TYPE_F32);
+                GGML_ASSERT(post->type     == GGML_TYPE_F32);
+                GGML_ASSERT(comb->type     == GGML_TYPE_F32);
+                GGML_ASSERT(op->type       == GGML_TYPE_F32);
+                GGML_ASSERT(residual->ne[1] == 4);
+
+                ggml_metal_kargs_dsv4_hc_post args = {
+                    /*.n_embd   =*/ (int32_t) x->ne[0],
+                    /*.n_tokens =*/ (int32_t) x->ne[1],
+                    /*.nb_x0    =*/ x->nb[0],
+                    /*.nb_x1    =*/ x->nb[1],
+                    /*.nb_r0    =*/ residual->nb[0],
+                    /*.nb_r1    =*/ residual->nb[1],
+                    /*.nb_r2    =*/ residual->nb[2],
+                    /*.nb_p0    =*/ post->nb[0],
+                    /*.nb_p1    =*/ post->nb[1],
+                    /*.nb_c0    =*/ comb->nb[0],
+                    /*.nb_c1    =*/ comb->nb[1],
+                    /*.nb_c2    =*/ comb->nb[2],
+                    /*.nb_d0    =*/ op->nb[0],
+                    /*.nb_d1    =*/ op->nb[1],
+                    /*.nb_d2    =*/ op->nb[2],
+                };
+
+                ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(x),        1);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(residual), 2);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(post),     3);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(comb),     4);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),       5);
+
+                const int n_tiles = (args.n_embd + 31)/32;
+                const int nsg = std::min(4, n_tiles);
+                ggml_metal_encoder_dispatch_threadgroups(
+                        enc, (n_tiles + nsg - 1)/nsg, args.n_tokens, 1, 32, nsg, 1);
+            } break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+
+    return 1;
+}
+
 int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -1979,6 +2190,46 @@ int ggml_metal_op_pool_1d(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+// supported FWHT sizes, must stay in sync with the
+// kernel_fwht_f32_<N> templates in ggml-metal.metal
+static bool ggml_metal_fwht_supported_size(int64_t n) {
+    return n == 64 || n == 128 || n == 256 || n == 512;
+}
+
+int ggml_metal_op_fwht(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    ggml_tensor * src1 = op->src[1];
+
+    const int64_t n = src1->ne[0];
+    const int64_t nrows = ggml_nrows(src1);
+
+    ggml_metal_kargs_fwht args = {
+        /*.nrows = */ (int32_t) nrows,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_fwht(lib, n);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(src1), 1);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 2);
+
+    const int th_max = ggml_metal_pipeline_max_theads_per_threadgroup(pipeline);
+    const int simd_size = 32;
+
+    int sg_per_tg = 2;
+    sg_per_tg = std::min(sg_per_tg, th_max/simd_size);
+    sg_per_tg = std::max(sg_per_tg, 1);
+
+    const int64_t n_tg = (nrows + sg_per_tg - 1) / sg_per_tg;
+    ggml_metal_encoder_dispatch_threadgroups(enc, n_tg, 1, 1, 32*sg_per_tg, 1, 1);
+
+    return 1;
+}
 
 int ggml_metal_op_pool_2d(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
@@ -2046,6 +2297,18 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
     ggml_metal_library_t lib = ctx->lib;
     ggml_metal_encoder_t enc = ctx->enc;
 
+    const int32_t hint = ggml_get_op_params_i32(op, 1);
+
+    if (hint == GGML_HINT_SRC0_IS_HADAMARD) {
+        if (op->src[1]->type == GGML_TYPE_F32 &&
+            op->type == GGML_TYPE_F32 &&
+            ggml_is_contiguous(op->src[1]) &&
+            ggml_is_contiguous(op) &&
+            ggml_are_same_shape(op->src[1], op) &&
+            ggml_metal_fwht_supported_size(op->src[1]->ne[0])) {
+            return ggml_metal_op_fwht(ctx, idx);
+        }
+    }
     const ggml_metal_device_props * props_dev = ggml_metal_device_get_props(ctx->dev);
 
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
@@ -3145,9 +3408,6 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
-
     GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
     GGML_ASSERT(ggml_is_contiguous_rows(op->src[1]));
 
@@ -3287,6 +3547,36 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     return n_fuse;
 }
 
+int ggml_metal_op_silu_back(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    auto pipeline = ggml_metal_library_get_pipeline_silu_back(lib, op);
+
+    const int64_t ne = ggml_nelements(op);
+
+    ggml_metal_kargs_silu_back args = {
+        /*.ne =*/ ne,
+    };
+
+    int arg_idx{0};
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op), arg_idx++);
+
+    const int nth = std::min<int64_t>(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne);
+    const int64_t n = (ne + nth - 1) / nth;
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -54,6 +54,8 @@ int ggml_metal_op_cumsum            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_get_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_set_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_diag              (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_lightning_indexer (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_dsv4_hc           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_soft_max          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_conv          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_scan          (ggml_metal_op_t ctx, int idx);
@@ -64,11 +66,13 @@ int ggml_metal_op_set               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_cpy               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_1d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_2d           (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_fwht              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_mul_mat           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_mul_mat_id        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_add_id            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_flash_attn_ext    (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_bin               (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_silu_back         (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_l2_norm           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_group_norm        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_norm              (ggml_metal_op_t ctx, int idx);

ggml/src/ggml-metal/kernels/argsort.metal

@@ -0,0 +1,232 @@
+#include "common.h"
+
+// bitonic sort implementation following the CUDA kernels as reference
+typedef void (argsort_t)(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_f32_i32(
+        constant   ggml_metal_kargs_argsort & args,
+        device   const char * src0,
+        device      int32_t * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    // bitonic sort
+    const int col = tpitg[0];
+    const int ib  = tgpig[0] / args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
+
+    // initialize indices
+    shmem_i32[col] = i00 + col;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (int k = 2; k <= ntg.x; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (shmem_i32[col] >= args.ne00 ||
+                       (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                } else {
+                    if (shmem_i32[ixj] >= args.ne00 ||
+                       (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
+                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
+                    ) {
+                        SWAP(shmem_i32[col], shmem_i32[ixj]);
+                    }
+                }
+            }
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+        }
+    }
+
+    const int64_t i0 = ib*args.top_k;
+
+    // copy the result to dst without the padding
+    if (i0 + col < args.ne0 && col < args.top_k) {
+        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
+
+        dst[col] = shmem_i32[col];
+    }
+}
+
+template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
+
+typedef void (argsort_merge_t)(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_merge_f32_i32(
+        constant   ggml_metal_kargs_argsort_merge & args,
+        device const char    * src0,
+        device const int32_t * tmp,
+        device       int32_t * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int im  = tgpig[0] / args.ne01;
+    const int i01 = tgpig[0] % args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    const int start = im * (2 * args.len);
+
+    const int len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
+    const int len1 = MIN(args.len, MAX(0, args.ne0 - (int)(start + args.len)));
+
+    const int total = len0 + len1;
+
+    device const int32_t * tmp0 = tmp + start
+        + i01*args.ne0
+        + i02*args.ne0*args.ne01
+        + i03*args.ne0*args.ne01*args.ne02;
+
+    device const int32_t * tmp1 = tmp0 + args.len;
+
+    dst += start
+        + i01*args.top_k
+        + i02*args.top_k*args.ne01
+        + i03*args.top_k*args.ne01*args.ne02;
+
+    device const float * src0_row = (device const float *)(src0
+        + args.nb01*i01
+        + args.nb02*i02
+        + args.nb03*i03);
+
+    if (total == 0) {
+        return;
+    }
+
+    const int chunk = (total + ntg.x - 1) / ntg.x;
+
+    const int k0 = tpitg.x * chunk;
+    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
+
+    if (k0 >= args.top_k) {
+        return;
+    }
+
+    if (k0 >= total) {
+        return;
+    }
+
+    int low  = k0 > len1 ? k0 - len1 : 0;
+    int high = MIN(k0, len0);
+
+    // binary-search partition (i, j) such that i + j = k
+    while (low < high) {
+        const int mid = (low + high) >> 1;
+
+        const int32_t idx0 = tmp0[mid];
+        const int32_t idx1 = tmp1[k0 - mid - 1];
+
+        const float val0 = src0_row[idx0];
+        const float val1 = src0_row[idx1];
+
+        bool take_left;
+        if (order == GGML_SORT_ORDER_ASC) {
+            take_left = (val0 <= val1);
+        } else {
+            take_left = (val0 >= val1);
+        }
+
+        if (take_left) {
+            low = mid + 1;
+        } else {
+            high = mid;
+        }
+    }
+
+    int i = low;
+    int j = k0 - i;
+
+    // keep the merge fronts into registers
+    int32_t idx0 = 0;
+    float   val0 = 0.0f;
+    if (i < len0) {
+        idx0 = tmp0[i];
+        val0 = src0_row[idx0];
+    }
+
+    int32_t idx1 = 0;
+    float   val1 = 0.0f;
+    if (j < len1) {
+        idx1 = tmp1[j];
+        val1 = src0_row[idx1];
+    }
+
+    for (int k = k0; k < k1; ++k) {
+        int32_t out_idx;
+
+        if (i >= len0) {
+            while (k < k1) {
+                dst[k++] = tmp1[j++];
+            }
+            break;
+        } else if (j >= len1) {
+            while (k < k1) {
+                dst[k++] = tmp0[i++];
+            }
+            break;
+        } else {
+            bool take_left;
+
+            if (order == GGML_SORT_ORDER_ASC) {
+                take_left = (val0 <= val1);
+            } else {
+                take_left = (val0 >= val1);
+            }
+
+            if (take_left) {
+                out_idx = idx0;
+                ++i;
+                if (i < len0) {
+                    idx0 = tmp0[i];
+                    val0 = src0_row[idx0];
+                }
+            } else {
+                out_idx = idx1;
+                ++j;
+                if (j < len1) {
+                    idx1 = tmp1[j];
+                    val1 = src0_row[idx1];
+                }
+            }
+        }
+
+        dst[k] = out_idx;
+    }
+}
+
+template [[host_name("kernel_argsort_merge_f32_i32_asc")]]  kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_merge_f32_i32_desc")]] kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_DESC>;

ggml/src/ggml-metal/kernels/binbcast.metal

@@ -0,0 +1,228 @@
+#include "common.h"
+
+// OP: 0 - add, 1 - sub, 2 - mul, 3 - div
+constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
+constant short FC_bin_f  [[function_constant(FC_BIN + 1)]];
+constant bool  FC_bin_rb [[function_constant(FC_BIN + 2)]];
+constant bool  FC_bin_cb [[function_constant(FC_BIN + 3)]];
+
+template <typename T0, typename T1, typename T>
+kernel void kernel_bin_fuse_impl(
+        constant ggml_metal_kargs_bin & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP FC_bin_op
+#define FC_F  FC_bin_f
+#define FC_RB FC_bin_rb
+#define FC_CB FC_bin_cb
+
+    if (FC_RB) {
+        // row broadcast
+        const uint i0 = tgpig.y*args.ne00 + tgpig.x;
+        const uint i1 = FC_CB ? tgpig.x%args.ne10 : tgpig.x;
+
+        device const T0 * src0_row = (device const T0 *) (src0);
+        device       T  * dst_row  = (device       T  *) (dst);
+
+        if (FC_F == 1) {
+            device const T1 * src1_row = (device const T1 *) (src1 + args.o1[0]);
+
+            if (FC_OP == 0) {
+                dst_row[i0] = src0_row[i0] + src1_row[i1];
+            }
+
+            if (FC_OP == 1) {
+                dst_row[i0] = src0_row[i0] - src1_row[i1];
+            }
+
+            if (FC_OP == 2) {
+                dst_row[i0] = src0_row[i0] * src1_row[i1];
+            }
+
+            if (FC_OP == 3) {
+                dst_row[i0] = src0_row[i0] / src1_row[i1];
+            }
+        } else {
+            T0 res = src0_row[i0];
+
+            if (FC_OP == 0) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res += ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 1) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res -= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 2) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res *= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 3) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res /= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            dst_row[i0] = res;
+        }
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int i01 = tgpig.x;
+
+        if (i01 >= args.ne01) {
+            return;
+        }
+
+        const int i13 = i03%args.ne13;
+        const int i12 = i02%args.ne12;
+        const int i11 = i01%args.ne11;
+
+        device const T0 * src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs);
+        device       T  * dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs);
+
+        if (FC_F == 1) {
+            device const T1 * src1_ptr = (device const T1 *) (src1 + args.o1[0] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                if (FC_OP == 0) {
+                    dst_ptr[i0] = src0_ptr[i0] + src1_ptr[i10];
+                }
+
+                if (FC_OP == 1) {
+                    dst_ptr[i0] = src0_ptr[i0] - src1_ptr[i10];
+                }
+
+                if (FC_OP == 2) {
+                    dst_ptr[i0] = src0_ptr[i0] * src1_ptr[i10];
+                }
+
+                if (FC_OP == 3) {
+                    dst_ptr[i0] = src0_ptr[i0] / src1_ptr[i10];
+                }
+            }
+        } else {
+            device const T1 * src1_ptr[8];
+            FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                src1_ptr[j] = (device const T1 *) (src1 + args.o1[j] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+            }
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                T res = src0_ptr[i0];
+
+                if (FC_OP == 0) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res += src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 1) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res -= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 2) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res *= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 3) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res /= src1_ptr[j][i10];
+                    }
+                }
+
+                dst_ptr[i0] = res;
+            }
+        }
+    }
+
+#undef FC_OP
+#undef FC_F
+#undef FC_RB
+#undef FC_CB
+}
+
+typedef decltype(kernel_bin_fuse_impl<float, float, float>) kernel_bin_fuse_t;
+
+template [[host_name("kernel_bin_fuse_f32_f32_f32")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float,  float,  float>;
+template [[host_name("kernel_bin_fuse_f32_f32_f32_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float4, float4, float4>;
+template [[host_name("kernel_bin_fuse_f16_f16_f16")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<half,   half,   half>;
+template [[host_name("kernel_bin_fuse_f16_f16_f16_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<half4,  half4,  half4>;
+
+kernel void kernel_add_id(
+        constant ggml_metal_kargs_add_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i1 = tgpig.x;
+    const int i2 = tgpig.y;
+
+    const int i11 = *((device const int32_t *) (src2 + i1*sizeof(int32_t) + i2*args.nb21));
+
+    const size_t nb1 = args.ne0 * sizeof(float);
+    const size_t nb2 = args.ne1 * nb1;
+
+    device       float * dst_row  = (device       float *)((device char *)dst  +  i1*nb1       + i2*nb2);
+    device const float * src0_row = (device const float *)((device char *)src0 +  i1*args.nb01 + i2*args.nb02);
+    device const float * src1_row = (device const float *)((device char *)src1 + i11*args.nb11);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
+
+template<typename T>
+kernel void kernel_repeat(
+        constant ggml_metal_kargs_repeat & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    const int i03 = i3%args.ne03;
+    const int i02 = i2%args.ne02;
+    const int i01 = i1%args.ne01;
+
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device       char * dst_ptr  = dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i00 = i0%args.ne00;
+        *((device T *)(dst_ptr + i0*args.nb0)) = *((device T *)(src0_ptr + i00*args.nb00));
+    }
+}
+
+typedef decltype(kernel_repeat<float>) kernel_repeat_t;
+
+template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
+template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_repeat_bf16")]] kernel kernel_repeat_t kernel_repeat<bfloat>;
+#endif
+template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
+template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;

ggml/src/ggml-metal/kernels/common.h

@@ -0,0 +1,126 @@
+#pragma once
+
+#include "ggml-metal-impl.h"
+
+#include <metal_stdlib>
+
+#ifdef GGML_METAL_HAS_TENSOR
+#include <metal_tensor>
+
+#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>
+#endif
+
+using namespace metal;
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
+
+#define PAD2(x, n) (((x) + (n) - 1) & ~((n) - 1))
+
+#define FOR_UNROLL(x) _Pragma("clang loop unroll(full)") for (x)
+
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+//
+// cmd:
+//   .../usr/bin/metal -dM -E -c                             ggml/src/ggml-metal/kernels/<src>.metal
+//   .../usr/bin/metal -dM -E -c -target air64-apple-ios14.0 ggml/src/ggml-metal/kernels/<src>.metal
+//
+#if __METAL_VERSION__ < 310 && defined(GGML_METAL_HAS_BF16)
+#undef GGML_METAL_HAS_BF16
+#endif
+
+#if defined(GGML_METAL_HAS_BF16)
+typedef matrix<bfloat, 4, 4> bfloat4x4;
+typedef matrix<bfloat, 2, 4> bfloat2x4;
+#endif
+
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+constexpr constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline int best_index_int8(int n, constant float * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static inline float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    return as_type<float>(bits);
+}
+
+static inline float dot(float x, float y) {
+    return x*y;
+}
+
+static inline float sum(float x) {
+    return x;
+}
+
+static inline float sum(float4 x) {
+    return x[0] + x[1] + x[2] + x[3];
+}
+
+enum ggml_sort_order {
+    GGML_SORT_ORDER_ASC,
+    GGML_SORT_ORDER_DESC,
+};
+
+constant float GELU_COEF_A     = 0.044715f;
+constant float GELU_QUICK_COEF = -1.702f;
+constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+constant float SQRT_2_INV      = 0.70710678118654752440084436210484f;
+
+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+constant float p_erf  = 0.3275911f;
+constant float a1_erf = 0.254829592f;
+constant float a2_erf = -0.284496736f;
+constant float a3_erf = 1.421413741f;
+constant float a4_erf = -1.453152027f;
+constant float a5_erf = 1.061405429f;
+
+template<typename T>
+inline T erf_approx(T x) {
+    T sign_x = sign(x);
+    x = fabs(x);
+    T t = 1.0f / (1.0f + p_erf * x);
+    T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    return sign_x * y;
+}
+
+template<typename T> T elu_approx(T x);
+
+template<> inline float elu_approx<float>(float x) {
+    return (x > 0.f) ? x : (exp(x) - 1);
+}
+
+template<> inline float4 elu_approx<float4>(float4 x) {
+    float4 res;
+
+    res[0] = (x[0] > 0.0f) ? x[0] : (exp(x[0]) - 1.0f);
+    res[1] = (x[1] > 0.0f) ? x[1] : (exp(x[1]) - 1.0f);
+    res[2] = (x[2] > 0.0f) ? x[2] : (exp(x[2]) - 1.0f);
+    res[3] = (x[3] > 0.0f) ? x[3] : (exp(x[3]) - 1.0f);
+
+    return res;
+}

ggml/src/ggml-metal/kernels/conv.metal

@@ -0,0 +1,723 @@
+#include "common.h"
+
+typedef void (im2col_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+//    const int64_t IC = tgpg[0];
+    const int64_t OH = tgpg[1];
+    const int64_t OW = tgpg[2];
+
+    const int64_t KH = ntg[1];
+    const int64_t KW = ntg[2];
+
+          int64_t in  = tpitg[0];
+    const int64_t ikh = tpitg[1];
+    const int64_t ikw = tpitg[2];
+
+    const int64_t iic = tgpig[0];
+    const int64_t ioh = tgpig[1];
+    const int64_t iow = tgpig[2];
+
+    const int64_t iiw = iow*args.s0 + ikw*args.d0 - args.p0;
+    const int64_t iih = ioh*args.s1 + ikh*args.d1 - args.p1;
+
+    int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*args.CHW + (iic*(KH*KW) + ikh*KW + ikw);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        while (in < args.N) {
+            pdst[offset_dst] = 0.0f;
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+
+            in += ntg[0];
+        }
+    } else {
+        int64_t offset_src = in*args.ofs0 + iic*args.ofs1 + iih*args.IW + iiw;
+
+        while (in < args.N) {
+            pdst[offset_dst] = x[offset_src];
+
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+            offset_src += ntg[0]*args.ofs0;
+
+            in += ntg[0];
+        }
+    }
+}
+
+template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
+template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
+
+// TODO: optimize
+typedef void (im2col_ext_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col_ext(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
+    const int64_t KHW = (int64_t)args.KHW;
+
+    const int64_t d   = tgpig[0] / args.CHW;
+    const int64_t chw = tgpig[0] % args.CHW;
+    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int64_t HW = tgpig[0] % KHW;
+
+    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    if (tpitg_0 >= args.N) {
+        return;
+    }
+
+    const int64_t tpitg_1 = HW / args.KW;
+    const int64_t tpitg_2 = HW % args.KW;
+
+    const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0;
+    const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1;
+
+    const int64_t offset_dst =
+        (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW +
+        (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        pdst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1;
+        pdst[offset_dst] = x[offset_src + iih * args.IW + iiw];
+    }
+}
+
+template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
+template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
+
+template <typename T>
+kernel void kernel_col2im_1d(
+        constant ggml_metal_kargs_col2im_1d & args,
+        device const T * col,
+        device       T * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T_out * args.OC) {
+        return;
+    }
+
+    const int t_out = idx % args.T_out;
+    const int oc    = idx / args.T_out;
+    const int t_abs = t_out + args.p0;  // absolute position in uncropped signal
+
+    int t_in_min = (t_abs - args.K + args.s0) / args.s0;  // ceil((t_abs - K + 1) / s0)
+    if (t_in_min < 0) {
+        t_in_min = 0;
+    }
+    int t_in_max = t_abs / args.s0;
+    if (t_in_max >= args.T_in) {
+        t_in_max = args.T_in - 1;
+    }
+
+    float sum = 0.0f;
+    for (int t_in = t_in_min; t_in <= t_in_max; t_in++) {
+        const int k = t_abs - t_in * args.s0;
+        sum += float(col[(oc * args.K + k) + t_in * args.K_OC]);
+    }
+
+    dst[t_out + oc * args.T_out] = T(sum);
+}
+
+template [[host_name("kernel_col2im_1d_f32")]]  kernel void kernel_col2im_1d<float>(constant ggml_metal_kargs_col2im_1d &, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_col2im_1d_f16")]]  kernel void kernel_col2im_1d<half>(constant ggml_metal_kargs_col2im_1d &, device const half *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_col2im_1d_bf16")]] kernel void kernel_col2im_1d<bfloat>(constant ggml_metal_kargs_col2im_1d &, device const bfloat *, device bfloat *, uint, uint, uint);
+#endif
+
+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+typedef void (conv_transpose_1d_t)(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_1d(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const     T * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tgpg[[threadgroups_per_grid]]) {
+
+    // For output position j on the time axis, only input positions
+    //   i such that i*s0 <= j < i*s0 + K
+    // contribute -- i.e. i in [ceil((j - K + 1)/s0), floor(j/s0)]
+    // intersected with [0, IL-1]. That's at most ceil(K/s0) values
+    // (typically 2 for stride==K/2 transposed convs).
+    const int32_t j  = tgpig[0];
+    const int32_t s0 = args.s0;
+    const int32_t K  = args.K;
+    const int32_t IL = args.IL;
+
+    int32_t i_min;
+    {
+        int32_t a = j - K + 1;
+        i_min = a <= 0 ? 0 : (a + s0 - 1) / s0; // ceil(a/s0) for a>0
+    }
+    int32_t i_max = j / s0;
+    if (i_max > IL - 1) i_max = IL - 1;
+
+    float v = 0.0f;
+    if (i_min <= i_max) {
+        for (int64_t c = 0; c < args.IC; c++) {
+            const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1];
+            const int32_t input_offset  = c * IL;
+
+            for (int32_t i = i_min; i <= i_max; i++) {
+                v += float(src0[kernel_offset + j - i * s0]) * src1[input_offset + i];
+            }
+        }
+    }
+
+    device float * dst_ptr = (device float *) (dst + tgpig[0] * args.nb0 + tgpig[1] * args.nb1);
+
+    dst_ptr[0] = v;
+}
+
+template [[host_name("kernel_conv_transpose_1d_f32_f32")]]
+kernel void kernel_conv_transpose_1d<float>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+template [[host_name("kernel_conv_transpose_1d_f16_f32")]]
+kernel void kernel_conv_transpose_1d<half>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+
+typedef void (conv_transpose_2d_t)(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_2d(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const T * src0,
+        device const float * src1,
+        device        char * dst,
+        threadgroup float * shared_sum [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int64_t out_x = tgpig[0];
+    const int64_t out_y = tgpig[1];
+    const int64_t out_c = tgpig[2];
+
+    const int64_t kw = tpitg[0];
+    const int64_t kh = tpitg[1];
+
+    float v = 0.0f;
+
+    for (int64_t in_c = 0; in_c < args.IC; in_c++) {
+        int64_t in_y = out_y - kh;
+
+        if (in_y < 0 || in_y % args.s0) continue;
+
+        in_y /= args.s0;
+
+        if (in_y >= args.IH) continue;
+
+        int64_t in_x = out_x - kw;
+
+        if (in_x < 0 || in_x % args.s0) continue;
+
+        in_x /= args.s0;
+
+        if (in_x >= args.IW) continue;
+
+        const int64_t input_idx = (args.IW * args.IH) * in_c + (args.IW) * in_y + in_x;
+        const int64_t kernel_idx = (args.KH * args.KW * args.OC) * in_c + (args.KH * args.KW) * out_c + (args.KW) * kh + kw;
+
+        v += (float)src0[kernel_idx] * src1[input_idx];
+    }
+
+    const uint tid = tpitg.y * ntg.x + tpitg.x;
+    shared_sum[tid] = v;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tid == 0) {
+        float total = 0.0f;
+        const uint num_threads = ntg.x * ntg.y;
+        for (uint i = 0; i < num_threads; i++) {
+            total += shared_sum[i];
+        }
+
+        device float * dst_ptr = (device float *) (dst + out_x*args.nb0 + out_y * args.nb1 + out_c*args.nb2);
+        dst_ptr[0] = total;
+    }
+}
+
+template [[host_name("kernel_conv_transpose_2d_f32_f32")]]
+kernel void kernel_conv_transpose_2d<float>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_transpose_2d_f16_f32")]]
+kernel void kernel_conv_transpose_2d<half>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+// grid: x = C tile, y = OH, z = OW * N (for channel-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw_tiled(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t c = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (c >= args.C) {
+        return;
+    }
+
+    const int32_t oh = tgpig.y;
+    const int32_t own = tgpig.z;
+    const int32_t ow = own % args.OW;
+    const int32_t n  = own / args.OW;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    float acc = 0.0f;
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        const uint64_t w_base   = (uint64_t) c * args.nb02;
+        const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+// grid: x = OW tile, y = OH, z = C * N (for spatially-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t oh = tgpig.y;
+    const int32_t cn = tgpig.z;
+    const int32_t c  = cn % args.C;
+    const int32_t n  = cn / args.C;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const uint64_t w_base   = (uint64_t) c * args.nb02;
+    const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+    const int32_t ow = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (ow >= args.OW) {
+        return;
+    }
+
+    float acc = 0.0f;
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+template [[host_name("kernel_conv_2d_dw_f32_f32")]]
+kernel void kernel_conv_2d_dw<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_f16_f32")]]
+kernel void kernel_conv_2d_dw<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f32_f32")]]
+kernel void kernel_conv_2d_dw_tiled<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f16_f32")]]
+kernel void kernel_conv_2d_dw_tiled<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_conv_3d(
+        constant ggml_metal_kargs_conv_3d & args,
+        device const  char * src0, // Weights [IC * OC, KD, KH, KW]
+        device const  char * src1, // Inputs  [IC * N,  ID, IH, IW]
+        device       char  * dst,  // Outputs [OC * N,  OD, OH, OW]
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
+
+    // 1. Un-flatten the spatial dimension from Grid X
+    int64_t spatial_idx = tgpig.x * 32 + tpitg.x;
+
+    if (spatial_idx >= args.OW * args.OH * args.OD) {
+        return; // Thread falls outside the spatial volume
+    }
+
+    int64_t od = spatial_idx / (args.OW * args.OH);
+    int64_t oh = (spatial_idx / args.OW) % args.OH;
+    int64_t ow = spatial_idx % args.OW;
+
+    // 2. Map Y to Channels, Z to Batch
+    int64_t oc = tgpig.y;
+    int64_t batch_idx = tgpig.z;
+
+    // 3. Calculate anchor coordinates in the Input volume
+    int64_t i_w_base = ow * args.s0 - args.p0;
+    int64_t i_h_base = oh * args.s1 - args.p1;
+    int64_t i_d_base = od * args.s2 - args.p2;
+
+    float sum = 0.0f;
+
+    // 4. Gather Loop (Iterate over Input Channels -> Depth -> Height -> Width)
+    for (int64_t ic = 0; ic < args.IC; ++ic) {
+
+        // ggml packs batch and channel together in the 4th dimension
+        int64_t src_cn_idx = batch_idx * args.IC + ic;
+        int64_t w_cn_idx   = oc * args.IC + ic;
+
+        for (int64_t kz = 0; kz < args.KD; ++kz) {
+            int64_t id = i_d_base + kz * args.d2;
+            if (id < 0 || id >= args.ID) continue; // Boundary check (Padding)
+
+            for (int64_t ky = 0; ky < args.KH; ++ky) {
+                int64_t ih = i_h_base + ky * args.d1;
+                if (ih < 0 || ih >= args.IH) continue;
+
+                for (int64_t kx = 0; kx < args.KW; ++kx) {
+                    int64_t iw = i_w_base + kx * args.d0;
+                    if (iw < 0 || iw >= args.IW) continue;
+
+                    // Convert multi-dimensional coordinates to flat byte offsets
+                    int64_t w_idx = kx*args.nb00 + ky*args.nb01 + kz*args.nb02 + w_cn_idx*args.nb03;
+                    int64_t i_idx = iw*args.nb10 + ih*args.nb11 + id*args.nb12 + src_cn_idx*args.nb13;
+
+                    // Dereference memory and cast weights to f32 if they were f16
+                    float w_val = (float)*(device const T*)((device const char*)src0 + w_idx);
+                    float i_val = *(device const float*)((device const char*)src1 + i_idx);
+
+                    sum += w_val * i_val;
+                }
+            }
+        }
+    }
+
+    // 5. Write the accumulated value out to RAM
+    int64_t dst_cn_idx = batch_idx * args.OC + oc;
+    int64_t d_idx = ow*args.nb0 + oh*args.nb1 + od*args.nb2 + dst_cn_idx*args.nb3;
+
+    *(device float*)(dst + d_idx) = sum;
+}
+
+// Explicit instantiations so the JIT compiler can find them by name
+template [[host_name("kernel_conv_3d_f32_f32")]]
+kernel void kernel_conv_3d<float>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
+
+// Explicit instantiation for f16 weights
+template [[host_name("kernel_conv_3d_f16_f32")]]
+kernel void kernel_conv_3d<half>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char  * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);

ggml/src/ggml-metal/kernels/dequantize.h

@@ -0,0 +1,719 @@
+#pragma once
+
+#include "common.h"
+
+#define GGML_COMMON_DECL_METAL
+#define GGML_COMMON_IMPL_METAL
+#if defined(GGML_METAL_EMBED_LIBRARY)
+__embed_ggml-common.h__
+#else
+#include "ggml-common.h"
+#endif
+
+#define QK_NL 16 // shared by mul_mm and get_rows_q instantiations
+
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f32_t4(device const float4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*src);
+}
+
+template <typename type4x4>
+void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+
+#if defined(GGML_METAL_HAS_BF16)
+template <typename type4x4>
+void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_bf16_t4(device const bfloat4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+#endif
+
+template <typename type4x4>
+void dequantize_q1_0(device const block_q1_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+    const float neg_d = -d;
+
+    const int byte_offset = il * 2;  // il*16 bits = il*2 bytes
+    const uint8_t b0 = qs[byte_offset];
+    const uint8_t b1 = qs[byte_offset + 1];
+
+    float4x4 reg_f;
+
+    reg_f[0][0] = select(neg_d, d, bool(b0 & 0x01));
+    reg_f[0][1] = select(neg_d, d, bool(b0 & 0x02));
+    reg_f[0][2] = select(neg_d, d, bool(b0 & 0x04));
+    reg_f[0][3] = select(neg_d, d, bool(b0 & 0x08));
+    reg_f[1][0] = select(neg_d, d, bool(b0 & 0x10));
+    reg_f[1][1] = select(neg_d, d, bool(b0 & 0x20));
+    reg_f[1][2] = select(neg_d, d, bool(b0 & 0x40));
+    reg_f[1][3] = select(neg_d, d, bool(b0 & 0x80));
+
+    reg_f[2][0] = select(neg_d, d, bool(b1 & 0x01));
+    reg_f[2][1] = select(neg_d, d, bool(b1 & 0x02));
+    reg_f[2][2] = select(neg_d, d, bool(b1 & 0x04));
+    reg_f[2][3] = select(neg_d, d, bool(b1 & 0x08));
+    reg_f[3][0] = select(neg_d, d, bool(b1 & 0x10));
+    reg_f[3][1] = select(neg_d, d, bool(b1 & 0x20));
+    reg_f[3][2] = select(neg_d, d, bool(b1 & 0x40));
+    reg_f[3][3] = select(neg_d, d, bool(b1 & 0x80));
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q1_0_t4(device const block_q1_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const float neg_d = -d;
+    const int base = il * 4;
+    const uint8_t byte = xb->qs[base / 8];
+    const int s = base % 8;
+
+    float4 reg_f;
+    reg_f[0] = select(neg_d, d, bool((byte >> (s    )) & 1));
+    reg_f[1] = select(neg_d, d, bool((byte >> (s + 1)) & 1));
+    reg_f[2] = select(neg_d, d, bool((byte >> (s + 2)) & 1));
+    reg_f[3] = select(neg_d, d, bool((byte >> (s + 3)) & 1));
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q2_0(device const block_q2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    const int byte_offset = il * 4;  // il*16 elements = il*4 bytes (4 elements per byte)
+    float4x4 reg_f;
+
+    for (int i = 0; i < 4; i++) {
+        const uint8_t b = qs[byte_offset + i];
+        reg_f[i][0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+        reg_f[i][1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+        reg_f[i][2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+        reg_f[i][3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q2_0_t4(device const block_q2_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const uint8_t b = xb->qs[il];
+
+    float4 reg_f;
+    reg_f[0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+    reg_f[1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+    reg_f[2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+    reg_f[3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = d1 * (qs[i] & mask0) + md;
+        reg_f[i/2][2*(i%2) + 1] = d2 * (qs[i] & mask1) + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + md;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + md;
+    }
+}
+
+
+
+template <typename type4x4>
+void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = ((qs[i] & mask0) * d1) + m;
+        reg_f[i/2][2*(i%2) + 1] = ((qs[i] & mask1) * d2) + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_1_t4(device const block_q4_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + m;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + md;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_0_t4(device const block_q5_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + md;
+        reg[2*ii + 1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + m;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + m;
+        reg[2*ii + 1] = d * x1 + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 16; i++) {
+        reg_f[i/4][i%4] = (qs[i + 16*il] * d);
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    for (int i = 0; i < 4; i++) {
+        reg[i] = (qs[4*(il%4) + i + 16*(il/4)] * d);
+    }
+}
+
+template <typename type4x4>
+void dequantize_mxfp4(device const block_mxfp4 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const uint8_t shr = il >= 1 ? 4 : 0;
+
+    for (int i = 0; i < 4; ++i) {
+        reg[i][0] = d * kvalues_mxfp4_f[(q2[4*i + 0] >> shr) & 0x0F];
+        reg[i][1] = d * kvalues_mxfp4_f[(q2[4*i + 1] >> shr) & 0x0F];
+        reg[i][2] = d * kvalues_mxfp4_f[(q2[4*i + 2] >> shr) & 0x0F];
+        reg[i][3] = d * kvalues_mxfp4_f[(q2[4*i + 3] >> shr) & 0x0F];
+    }
+}
+
+template <typename type4>
+void dequantize_mxfp4_t4(device const block_mxfp4 * xb, short il, thread type4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const short il4 = il%4;
+
+    const uint8_t shr = il >= 4 ? 4 : 0;
+
+    reg[0] = d * kvalues_mxfp4_f[(q2[4*il4 + 0] >> shr) & 0x0F];
+    reg[1] = d * kvalues_mxfp4_f[(q2[4*il4 + 1] >> shr) & 0x0F];
+    reg[2] = d * kvalues_mxfp4_f[(q2[4*il4 + 2] >> shr) & 0x0F];
+    reg[3] = d * kvalues_mxfp4_f[(q2[4*il4 + 3] >> shr) & 0x0F];
+}
+
+template <typename type4x4>
+void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
+    const float d = xb->d;
+    const float min = xb->dmin;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    float dl, ml;
+    uint8_t sc = xb->scales[il];
+
+    q = q + 32*(il/8) + 16*(il&1);
+    il = (il/2)%4;
+
+    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    device const uint8_t * h = (device const uint8_t *)xb->hmask;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    q = q + 32 * (il/8) + 16 * (il&1);
+    h = h + 16 * (il&1);
+    uint8_t m = 1 << (il/2);
+    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
+                                 ((il/4)>0 ? 12  : 3);
+    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
+    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
+
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
+
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
+    }
+}
+
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
+template <typename type4x4>
+void dequantize_q4_K(device const block_q4_K * xb, short il, thread type4x4 & reg) {
+    device const uchar * q = xb->qs;
+
+    short is = (il/4) * 2;
+    q = q + (il/4) * 32 + 16 * (il&1);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d   = il < 2 ? xb->d : xb->d / 16.h;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask = il < 2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q  = xb->qs;
+    device const uint8_t * qh = xb->qh;
+
+    short is = (il/4) * 2;
+    q  = q + 32 * (il/4) + 16 * (il&1);
+    qh = qh + 16 * (il&1);
+    uint8_t ul = 1 << (il/2);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint16_t * ql = (device const uint16_t *)xb->ql;
+    device const uint16_t * qh = (device const uint16_t *)xb->qh;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    ql = ql + 32*(il/8) + 16*((il/2)&1) + 8*(il&1);
+    qh = qh + 16*(il/8) + 8*(il&1);
+    float sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint32_t kmask1 = il>1 ? (il>2 ? 0xC0C0C0C0 : 0x30303030) : (il>0 ? 0x0C0C0C0C : 0x03030303);
+    const uint32_t kmask2 = il>1 ? 0xF0F0F0F0                       : 0x0F0F0F0F;
+    const float ml = d_all * sc * 32.f;
+    const float dl0 = d_all * sc;
+    const float dl1 = dl0 / 256.f;
+    const float dl2 = dl0 / (256.f * 256.f);
+    const float dl3 = dl0 / (256.f * 256.f * 256.f);
+    const uint8_t shr_h = il>2 ? 2 : 0;
+    const uint8_t shl_h = il>1 ? 0 : (il>0 ? 2 : 4);
+    const uint8_t shr_l = il>1 ? 4 : 0;
+    for (int i = 0; i < 4; ++i) {
+        const uint32_t  low = (ql[2*i] | (uint32_t)(ql[2*i+1] << 16)) & kmask2;
+        const uint32_t high = (qh[2*i] | (uint32_t)(qh[2*i+1] << 16)) & kmask1;
+        const uint32_t q = ((high << shl_h) >> shr_h) | (low >> shr_l);
+        reg[i][0] = dl0 *  ((half)(q & 0xFF))       - ml;
+        reg[i][1] = dl1 * ((float)(q & 0xFF00))     - ml;
+        reg[i][2] = dl2 * ((float)(q & 0xFF0000))   - ml;
+        reg[i][3] = dl3 * ((float)(q & 0xFF000000)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * signs = qs + QK_K/8;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
+        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    const float d = xb->d;
+    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint16_t * qh = xb->qh;
+    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
+    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
+    const uint16_t h = qh[ib32] >> 6*il;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    device const uint16_t * sc = (device const uint16_t *)xb->scales;
+
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = scale.f16;
+
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * qh = xb->qh + 2*ib32 + il;
+
+    const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
+    const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml1;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml2;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4>
+void dequantize_iq4_nl_t4(device const block_iq4_nl * xb, short il, thread type4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    aux32 = ((q4[2*(il%4)] | (q4[2*(il%4)+1] << 16)) >> 4*(il/4)) & 0x0f0f0f0f;
+    reg[0] = d * kvalues_iq4nl_f[q8[0]];
+    reg[1] = d * kvalues_iq4nl_f[q8[1]];
+    reg[2] = d * kvalues_iq4nl_f[q8[2]];
+    reg[3] = d * kvalues_iq4nl_f[q8[3]];
+}
+
+template <typename type4x4>
+void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
+    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)xb->d * (ls - 32);
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+

ggml/src/ggml-metal/kernels/gated_delta_net.metal

@@ -0,0 +1,250 @@
+#include "common.h"
+
+constant short FC_gated_delta_net_ne20 [[function_constant(FC_GATED_DELTA_NET + 0)]];
+constant short FC_gated_delta_net_ne30 [[function_constant(FC_GATED_DELTA_NET + 1)]];
+constant short FC_gated_delta_net_K    [[function_constant(FC_GATED_DELTA_NET + 2)]];
+
+#if 1
+template<short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+#define K   FC_gated_delta_net_K
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B (n_seqs)
+    const uint i21 = tgpig.y; // H (head)
+    const uint i20 = tgpig.x*NSG + ty; // row within S_v
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    // input state layout [S_v, S_v, H, n_seqs] (s0 only): per-seq stride is H*D.
+    // state is stored transposed: M[i20][is] = S[is][i20], so row i20 is contiguous
+    const uint state_in_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+    device const float * s_ptr = (device const float *) (s) + state_in_base;
+
+    float ls[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        ls[j] = s_ptr[is];
+    }
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    // snapshot slot mapping: slot 0 = most recent state, slot s = s tokens back.
+    // When n_tokens < K, only slots 0..n_tokens-1 are written; older slots are caller-owned.
+
+    // output state base offset: after attention scores
+    const uint attn_size = args.ne22 * args.ne21 * S_v * args.ne23;
+    // output state per-slot size: S_v * S_v * H * n_seqs
+    const uint state_size_per_snap = S_v * S_v * args.ne21 * args.ne23;
+    // per-(seq,head) offset within a slot
+    const uint state_out_base = (i23*args.ne21 + i21)*S_v*S_v + i20*S_v;
+
+    for (short t = 0; t < args.ne22; t++) {
+        float s_k = 0.0f;
+
+        if (G == 1) {
+            const float g_exp = exp(g_ptr[0]);
+
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= g_exp;
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        } else {
+            // KDA
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= exp(g_ptr[is]);
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        }
+
+        s_k = simd_sum(s_k);
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        float y = 0.0f;
+
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            ls[j] += k_ptr[is]*d;
+
+            y += ls[j]*q_ptr[is];
+        }
+
+        y = simd_sum(y);
+
+        if (tx == 0) {
+            dst_attn[t*args.ne21*S_v] = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        if (K > 1) {
+            const int target_slot = (int)args.ne22 - 1 - (int)t;
+            if (target_slot >= 0 && target_slot < (int)K) {
+                device float * dst_state = (device float *) (dst) + attn_size + (uint)target_slot * state_size_per_snap + state_out_base;
+                FOR_UNROLL (short j = 0; j < NSG; j++) {
+                    const short is = tx*NSG + j;
+                    dst_state[is] = ls[j];
+                }
+            }
+        }
+    }
+
+    if (K == 1) {
+        device float * dst_state = (device float *) (dst) + attn_size + state_out_base;
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            dst_state[is] = ls[j];
+        }
+    }
+
+#undef S_v
+#undef G
+#undef K
+}
+
+typedef decltype(kernel_gated_delta_net_impl<4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<4>;
+
+#else
+// a simplified version of the above
+// no performance improvement, so keep the above version for now
+
+template<typename T, short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B
+    const uint i21 = tgpig.y; // H
+    const uint i20 = tgpig.x*NSG + ty;
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    device const float * s_ptr = (device const float *) (s) + (i23*args.ne21 + i21)*S_v*S_v + i20;
+
+    float lsf[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        lsf[j] = s_ptr[is*S_v];
+    }
+
+    thread T * ls = (thread T *) (lsf);
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr  = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr  = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    for (short t = 0; t < args.ne22; t++) {
+        device const T * qt_ptr = (device const T *) (q_ptr);
+        device const T * kt_ptr = (device const T *) (k_ptr);
+        device const T * gt_ptr = (device const T *) (g_ptr);
+
+        if (G == 1) {
+            *ls *= exp(g_ptr[0]);
+        } else {
+            // KDA
+            *ls *= exp(gt_ptr[tx]);
+        }
+
+        const float s_k = simd_sum(dot(*ls, kt_ptr[tx]));
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        *ls += kt_ptr[tx]*d;
+
+        const float y = simd_sum(dot(*ls, qt_ptr[tx]));
+
+        if (tx == 0) {
+            *dst_attn = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        dst_attn += args.ne21*S_v;
+    }
+
+    device float * dst_state  = (device float *) (dst) + args.ne23*args.ne22*args.ne21*S_v + (i23*args.ne21 + i21)*S_v*S_v + i20;
+    device T     * dstt_state = (device T     *) (dst_state);
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        dst_state[is*S_v] = lsf[j];
+    }
+
+#undef S_v
+#undef G
+}
+
+typedef decltype(kernel_gated_delta_net_impl<float4, 4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float,  1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float2, 2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float4, 4>;
+#endif

ggml/src/ggml-metal/kernels/misc.metal

@@ -0,0 +1,595 @@
+#include "common.h"
+
+kernel void kernel_argmax_f32(
+        constant ggml_metal_kargs_argmax & args,
+        device   const char * src0,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    device const float * x_row = (device const float *) ((device const char *) src0 + tgpig * args.nb01);
+
+    float   lmax = -INFINITY;
+    int32_t larg = -1;
+
+    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+        if (x_row[i00] > lmax) {
+            lmax = x_row[i00];
+            larg = i00;
+        }
+    }
+
+    // find the argmax value in the block
+    float max_val = simd_max(lmax);
+    int32_t arg_val = simd_max(select(-1, larg, lmax == max_val));
+
+    device int32_t * dst_i32 = (device int32_t *) dst;
+
+    threadgroup   float * shared_maxval = (threadgroup   float *) shmem;
+    threadgroup int32_t * shared_argmax = (threadgroup int32_t *) shmem + N_SIMDWIDTH;
+
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            shared_maxval[tiisg] = -INFINITY;
+            shared_argmax[tiisg] = -1;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            shared_maxval[sgitg] = max_val;
+            shared_argmax[sgitg] = arg_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = shared_maxval[tiisg];
+        arg_val = shared_argmax[tiisg];
+
+        float max_val_reduced   = simd_max(max_val);
+        int32_t arg_val_reduced = simd_max(select(-1, arg_val, max_val == max_val_reduced));
+
+        dst_i32[tgpig] = arg_val_reduced;
+
+        return;
+    }
+
+    dst_i32[tgpig] = arg_val;
+}
+
+kernel void kernel_diag_f32(
+        constant ggml_metal_kargs_diag & args,
+        device   const char * src0,
+        device         char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *)(src0 +                i2*args.nb02 + i3*args.nb03);
+    device       float * dst_ptr  = (device       float *)(dst  + i1*args.nb01 + i2*args.nb2  + i3*args.nb3);
+
+    for (int i0 = tiitg; i0 < args.ne0; i0 += NW) {
+        dst_ptr[i0] = i0 == i1 ? src0_ptr[i0] : 0.0f;
+    }
+}
+
+kernel void kernel_roll_f32(
+    constant ggml_metal_kargs_roll & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *) src0;
+    device       float * dst_ptr  = (device       float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        // apply shifts and wrap around
+        int64_t i00 = i0 - args.s0;
+        int64_t i01 = i1 - args.s1;
+        int64_t i02 = i2 - args.s2;
+        int64_t i03 = i3 - args.s3;
+
+        if (i00 < 0) { i00 += args.ne00; } else if (i00 >= args.ne00) { i00 -= args.ne00; }
+        if (i01 < 0) { i01 += args.ne01; } else if (i01 >= args.ne01) { i01 -= args.ne01; }
+        if (i02 < 0) { i02 += args.ne02; } else if (i02 >= args.ne02) { i02 -= args.ne02; }
+        if (i03 < 0) { i03 += args.ne03; } else if (i03 >= args.ne03) { i03 -= args.ne03; }
+
+        int64_t src_idx = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00 + i00;
+        int64_t dst_idx = i3 *args.ne2 *args.ne1 *args.ne0  + i2 *args.ne1 *args.ne0  + i1 *args.ne0  + i0;
+
+        dst_ptr[dst_idx] = src0_ptr[src_idx];
+    }
+}
+
+template <typename T>
+kernel void kernel_pad_impl(
+    constant ggml_metal_kargs_pad & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i3 = tgpig.z;
+    const int32_t i2 = tgpig.y;
+    const int32_t k0 = tgpig.x/args.ne1;
+    const int32_t i1 = tgpig.x - k0*args.ne1;
+
+    const int32_t i03 = i3;
+    const int32_t i02 = i2;
+    const int32_t i01 = i1;
+
+    device const T * src0_ptr = (device const T *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T * dst_ptr  = (device       T *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    for (int32_t l0 = 0; l0 < 1024; l0 += ntg.x) {
+        const int32_t i0 = k0*1024 + tpitg.x + l0;
+        if (i0 >= args.ne0) {
+            break;
+        }
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            dst_ptr[i0] = src0_ptr[i0];
+        } else {
+            dst_ptr[i0] = 0.0f;
+        }
+    }
+}
+
+typedef decltype(kernel_pad_impl<float>) kernel_pad_t;
+
+template [[host_name("kernel_pad_f32")]]   kernel kernel_pad_t kernel_pad_impl<float>;
+template [[host_name("kernel_pad_f32_4")]] kernel kernel_pad_t kernel_pad_impl<float4>;
+
+// TODO: this is slow - optimize
+kernel void kernel_pad_reflect_1d_f32(
+    constant   ggml_metal_kargs_pad_reflect_1d & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3  tgpg[[threadgroups_per_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1);
+
+    if (i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            if (i0 < args.p0) {
+                dst_ptr[i0] = src0_ptr[args.p0 - i0];
+            } else if (i0 < args.ne0 - args.p1) {
+                dst_ptr[i0] = src0_ptr[i0 - args.p0];
+            } else {
+                dst_ptr[i0] = src0_ptr[(args.ne0 - args.p1 - args.p0) - (args.p1 + 1 - (args.ne0 - i0)) - 1];
+            }
+        }
+    }
+}
+
+kernel void kernel_arange_f32(
+    constant   ggml_metal_kargs_arange & args,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    device float * dst_ptr = (device float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_ptr[i0] = args.start + args.step * i0;
+    }
+}
+
+kernel void kernel_timestep_embedding_f32(
+    constant  ggml_metal_kargs_timestep_embedding & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    int i = tgpig.x;
+    device float * embed_data = (device float *)(dst + i*args.nb1);
+
+    int half_ = args.dim / 2;
+    for (int j = tpitg.x; j < half_; j += ntg.x) {
+        float timestep = ((device float *)src0)[i];
+        float freq = (float)exp(-log((float)args.max_period) * j / half_);
+        float arg = timestep * freq;
+        embed_data[j        ] = cos(arg);
+        embed_data[j + half_] = sin(arg);
+    }
+
+    if (args.dim % 2 != 0 && tpitg.x == 0) {
+        embed_data[2 * half_] = 0.f;
+    }
+}
+
+kernel void kernel_opt_step_adamw_f32(
+        constant    ggml_metal_kargs_opt_step_adamw & args,
+        device       float * x,
+        device const float * g,
+        device       float * g_m,
+        device       float * g_v,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const float alpha  = pars[0];
+    const float beta1  = pars[1];
+    const float beta2  = pars[2];
+    const float eps    = pars[3];
+    const float wd     = pars[4];
+    const float beta1h = pars[5];
+    const float beta2h = pars[6];
+
+    const float gi = g[gid];
+    const float gmi = g_m[gid] * beta1 +      gi * (1.0f - beta1);
+    const float gvi = g_v[gid] * beta2 + gi * gi * (1.0f - beta2);
+
+    g_m[gid] = gmi;
+    g_v[gid] = gvi;
+
+    const float mh =      gmi * beta1h;
+    const float vh = sqrt(gvi * beta2h) + eps;
+
+    x[gid] = x[gid] * (1.0f - alpha * wd) - alpha * mh / vh;
+}
+
+kernel void kernel_opt_step_sgd_f32(
+        constant    ggml_metal_kargs_opt_step_sgd & args,
+        device       float * x,
+        device const float * g,
+        device const float * pars,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    x[gid] = x[gid] * (1.0f - pars[0] * pars[1]) - pars[0] * g[gid];
+}
+
+template<typename T>
+kernel void kernel_memset(
+        constant ggml_metal_kargs_memset & args,
+        device T * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = args.val;
+}
+
+typedef decltype(kernel_memset<int64_t>) kernel_memset_t;
+
+template [[host_name("kernel_memset_i64")]] kernel kernel_memset_t kernel_memset<int64_t>;
+
+constant short FC_count_equal_nsg [[function_constant(FC_COUNT_EQUAL + 0)]];
+
+template<typename T>
+kernel void kernel_count_equal(
+        constant ggml_metal_kargs_count_equal & args,
+        device   const char * src0,
+        device   const char * src1,
+        device   atomic_int * dst,
+        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const short NSG = FC_count_equal_nsg;
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    int sum = 0;
+
+    device const char * base0 = src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03;
+    device const char * base1 = src1 + i1*args.nb11 + i2*args.nb12 + i3*args.nb13;
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        const T v0 = *(device const T *)(base0 + i0*args.nb00);
+        const T v1 = *(device const T *)(base1 + i0*args.nb10);
+        sum += (v0 == v1);
+    }
+
+    sum = simd_sum(sum);
+
+    if (tiisg == 0) {
+        shmem_i32[sgitg] = sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        float v = 0.0f;
+        if (tpitg.x < NSG) {
+            v = shmem_i32[tpitg.x];
+        }
+
+        float total = simd_sum(v);
+        if (tpitg.x == 0) {
+            atomic_fetch_add_explicit(dst, (int32_t) total, memory_order_relaxed);
+        }
+    }
+}
+
+typedef decltype(kernel_count_equal<int32_t>) kernel_count_equal_t;
+
+template [[host_name("kernel_count_equal_i32")]] kernel kernel_count_equal_t kernel_count_equal<int32_t>;
+
+template <typename T>
+kernel void kernel_snake(
+        constant ggml_metal_kargs_snake & args,
+        device const T     * x,
+        device const float * a,
+        device const float * inv_b,
+        device       T     * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T * args.C) {
+        return;
+    }
+
+    const int   c  = idx / args.T;  // x is [T, C], a / inv_b collapse to [1, C]
+    const float xi = float(x[idx]);
+    const float si = sin(a[c] * xi);
+    dst[idx] = T(xi + si * si * inv_b[c]);
+}
+
+template [[host_name("kernel_snake_f32")]]  kernel void kernel_snake<float>(constant ggml_metal_kargs_snake &, device const float *, device const float *, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_snake_f16")]]  kernel void kernel_snake<half>(constant ggml_metal_kargs_snake &, device const half *, device const float *, device const float *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_snake_bf16")]] kernel void kernel_snake<bfloat>(constant ggml_metal_kargs_snake &, device const bfloat *, device const float *, device const float *, device bfloat *, uint, uint, uint);
+#endif
+
+template<int N>
+kernel void kernel_fwht_f32(
+        constant ggml_metal_kargs_fwht & args,
+        device const float * src,
+        device float * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort3  ntg[[threads_per_threadgroup]]) {
+
+    constexpr int NW = N_SIMDWIDTH;
+    constexpr int NE = N / NW;
+
+    const float scale = 1.0f / sqrt((float) N);
+
+    const int sg_per_tg = ntg.x / NW;
+    const int64_t r = tgpig.x * sg_per_tg + sgitg;
+    if (r >= args.nrows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    const int lane = tiisg;
+
+    float reg[NE];
+    for (int i = 0; i < NE; i++) {
+        reg[i] = src[i*NW + lane]*scale;
+    }
+    for (int i = 1; i < NW; i *= 2) {
+        for (int j = 0; j < NE; j++) {
+            const float val = reg[j];
+            const float val2 = simd_shuffle_xor(val, i);
+            reg[j] = (lane & i) == 0 ? val2 + val : val2 - val;
+        }
+    }
+
+    for (int i = NW; i < N; i *= 2) {
+        const int step = i / NW;
+        for (int j = 0; j < NE; j += (2 * step)) {
+            for (int k = 0; k < step; k++) {
+                const float x = reg[j + k ];
+                const float y = reg[j + k + step];
+                reg[j + k]        = x + y;
+                reg[j + k + step] = x - y;
+            }
+        }
+    }
+
+    for (int i = 0; i < NE; i++) {
+        dst[i*NW + lane] = reg[i];
+    }
+}
+
+typedef decltype(kernel_fwht_f32<64>) kernel_fwht_t;
+
+template [[host_name("kernel_fwht_f32_64")]]  kernel kernel_fwht_t kernel_fwht_f32<64>;
+template [[host_name("kernel_fwht_f32_128")]] kernel kernel_fwht_t kernel_fwht_f32<128>;
+template [[host_name("kernel_fwht_f32_256")]] kernel kernel_fwht_t kernel_fwht_f32<256>;
+template [[host_name("kernel_fwht_f32_512")]] kernel kernel_fwht_t kernel_fwht_f32<512>;
+
+kernel void kernel_dsv4_hc_comb_f32(
+        constant ggml_metal_kargs_dsv4_hc_comb & args,
+        device const char * mixes,
+        device const char * scale,
+        device const char * base,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+    constexpr ushort comb_offset = 2*hc;
+
+    const int it = tgpig.x*ntg.y + sgitg;
+    if (it >= args.n_tokens) {
+        return;
+    }
+
+    float scale_lane = 0.0f;
+    if (tiisg == 0) {
+        scale_lane = *(device const float *) (scale + 2*args.nb_s0);
+    }
+    const float scale_comb = simd_shuffle(scale_lane, 0);
+
+    float v = 0.0f;
+    if (tiisg < hc*hc) {
+        v = *(device const float *) (mixes + (comb_offset + tiisg)*args.nb_m0 + it*args.nb_m1)*scale_comb
+          + *(device const float *) (base   + (comb_offset + tiisg)*args.nb_b0);
+    }
+
+    // Softmax across destinations (the four contiguous lanes for each source).
+    float vmax = max(v, simd_shuffle_xor(v, 1));
+    vmax = max(vmax, simd_shuffle_xor(vmax, 2));
+    v = exp(v - vmax);
+
+    float sum = v + simd_shuffle_xor(v, 1);
+    sum += simd_shuffle_xor(sum, 2);
+    v = v/sum + args.eps;
+
+    // Normalize columns: equal destination indices are four lanes apart.
+    sum = v + simd_shuffle_xor(v, 4);
+    sum += simd_shuffle_xor(sum, 8);
+    v /= sum + args.eps;
+
+    for (int i = 1; i < args.n_iter; ++i) {
+        sum = v + simd_shuffle_xor(v, 1);
+        sum += simd_shuffle_xor(sum, 2);
+        v /= sum + args.eps;
+
+        sum = v + simd_shuffle_xor(v, 4);
+        sum += simd_shuffle_xor(sum, 8);
+        v /= sum + args.eps;
+    }
+
+    if (tiisg < hc*hc) {
+        const ushort idst = tiisg & 3;
+        const ushort isrc = tiisg >> 2;
+        *(device float *) (dst + idst*args.nb_d0 + isrc*args.nb_d1 + it*args.nb_d2) = v;
+    }
+}
+
+kernel void kernel_dsv4_hc_pre_f32(
+        constant ggml_metal_kargs_dsv4_hc_pre & args,
+        device const char * x,
+        device const char * weights,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+
+    const int it = tgpig.y;
+    const int i0 = ((int) tgpig.x*ntg.y + sgitg)*32 + tiisg;
+
+    float weight_lane = 0.0f;
+    if (tiisg < hc) {
+        weight_lane = *(device const float *) (weights + tiisg*args.nb_w0 + it*args.nb_w1);
+    }
+
+    float w[hc];
+    FOR_UNROLL (ushort ih = 0; ih < hc; ++ih) {
+        w[ih] = simd_shuffle(weight_lane, ih);
+    }
+
+    if (i0 >= args.n_embd) {
+        return;
+    }
+
+    device const char * xb = x + i0*args.nb_x0 + it*args.nb_x2;
+    float result = 0.0f;
+    FOR_UNROLL (ushort ih = 0; ih < hc; ++ih) {
+        result = fma(*(device const float *) (xb + ih*args.nb_x1), w[ih], result);
+    }
+
+    *(device float *) (dst + i0*args.nb_d0 + it*args.nb_d1) = result;
+}
+
+kernel void kernel_dsv4_hc_post_f32(
+        constant ggml_metal_kargs_dsv4_hc_post & args,
+        device const char * x,
+        device const char * residual,
+        device const char * post,
+        device const char * comb,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+
+    const int it = tgpig.y;
+    const int i0 = ((int) tgpig.x*ntg.y + sgitg)*32 + tiisg;
+
+    float coeff_lane = 0.0f;
+    if (tiisg < hc) {
+        coeff_lane = *(device const float *) (post + tiisg*args.nb_p0 + it*args.nb_p1);
+    } else if (tiisg < hc + hc*hc) {
+        const ushort idx  = tiisg - hc;
+        const ushort idst = idx & 3;
+        const ushort isrc = idx >> 2;
+        coeff_lane = *(device const float *) (comb + idst*args.nb_c0 + isrc*args.nb_c1 + it*args.nb_c2);
+    }
+
+    float post_reg[hc];
+    float comb_reg[hc][hc];
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        post_reg[idst] = simd_shuffle(coeff_lane, idst);
+    }
+    FOR_UNROLL (ushort isrc = 0; isrc < hc; ++isrc) {
+        FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+            comb_reg[isrc][idst] = simd_shuffle(coeff_lane, hc + idst + hc*isrc);
+        }
+    }
+
+    if (i0 >= args.n_embd) {
+        return;
+    }
+
+    const float xv = *(device const float *) (x + i0*args.nb_x0 + it*args.nb_x1);
+    float result[hc];
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        result[idst] = xv*post_reg[idst];
+    }
+
+    device const char * rb = residual + i0*args.nb_r0 + it*args.nb_r2;
+    FOR_UNROLL (ushort isrc = 0; isrc < hc; ++isrc) {
+        const float rv = *(device const float *) (rb + isrc*args.nb_r1);
+        FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+            result[idst] = fma(rv, comb_reg[isrc][idst], result[idst]);
+        }
+    }
+
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        *(device float *) (dst + i0*args.nb_d0 + idst*args.nb_d1 + it*args.nb_d2) = result[idst];
+    }
+}

ggml/src/ggml-metal/kernels/mul_mm.metal

@@ -0,0 +1,842 @@
+#include "common.h"
+#include "dequantize.h"
+
+constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
+constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
+constant short FC_mul_mm_ne12  [[function_constant(FC_MUL_MM + 2)]];
+constant short FC_mul_mm_ne13  [[function_constant(FC_MUL_MM + 3)]];
+constant short FC_mul_mm_r2    [[function_constant(FC_MUL_MM + 4)]];
+constant short FC_mul_mm_r3    [[function_constant(FC_MUL_MM + 5)]];
+
+// each block_q contains 16*nl weights
+#ifdef GGML_METAL_HAS_TENSOR
+template<
+    typename SA, typename SA_4x4, typename SA_8x8,
+    typename SB, typename SB_2x4, typename SB_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread SA_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * srcA,
+        device const char * srcB,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig [[threadgroup_position_in_grid]],
+        ushort tiitg [[thread_index_in_threadgroup]],
+        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
+    (void) sgitg;
+
+    // Matrix dimensions: A(M,K) x B(K,N) -> C(M,N)
+    const int K = args.ne00;
+    const int M = args.ne0;
+    const int N = args.ne1;
+
+    // Batch dimension handling
+    const int im = tgpig.z;
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    // Batch offsets for srcA and srcB
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+
+    // Tile dimensions
+    constexpr int NRB = SZ_SIMDGROUP * N_MM_BLOCK_X * N_MM_SIMD_GROUP_X;
+    constexpr int NRA = SZ_SIMDGROUP * N_MM_BLOCK_Y * N_MM_SIMD_GROUP_Y;
+
+    // Tile offsets in output matrix
+    const int ra = tgpig.y * NRA;
+    const int rb = tgpig.x * NRB;
+
+    // Threadgroup memory for dequantized A tile only
+    threadgroup SA * sa = (threadgroup SA *)(shmem);
+
+    // Work-item count for A loading
+    constexpr int A_WORK_ITEMS = NRA * N_MM_NK;
+    constexpr int NUM_THREADS = N_SIMDWIDTH * N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y;
+
+    // tA wraps threadgroup memory
+    auto tA = tensor(sa, dextents<int32_t, 2>(N_MM_NK_TOTAL, NRA));
+
+    // tB wraps device memory directly
+    device T1 * ptrB = (device T1 *)(srcB + args.nb12*i12 + args.nb13*i13);
+    const int strideB = args.nb11 / sizeof(T1);
+    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));
+
+    // Configure matmul operation
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(
+            NRB, NRA, N_MM_NK_TOTAL, false, true, true,
+            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tB), decltype(tA), float>();
+
+    // Accumulate partial results over K dimension
+    for (int loop_k = 0; loop_k < K; loop_k += N_MM_NK_TOTAL) {
+        // === PHASE 1: Dequantization of A into threadgroup memory ===
+        for (int work = tiitg; work < A_WORK_ITEMS; work += NUM_THREADS) {
+            const int row = work / N_MM_NK;
+            const int k_chunk = work % N_MM_NK;
+            const int k_pos = loop_k + k_chunk * 16;
+            const short k_base = k_chunk * 16;
+
+            // Bounds check: skip device read if row is out of matrix bounds
+            if (ra + row < M) {
+                if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+                    // Element-wise reads when K is not aligned (nb01 not aligned for half4x4/float4x4).
+                    // MSL spec Table 2.5: half4x4 requires 8-byte alignment. When K is odd,
+                    // nb01 = K*2 is not 8-byte aligned, so odd-row pointers are misaligned.
+                    // Mirrors the legacy kernel's existing guard.
+                    device const T0 * row_ptr = (device const T0 *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? (SA) row_ptr[k_pos + i] : (SA)0;
+                    }
+                } else {
+                    const int block_idx = k_pos / (16 * nl);
+                    const short il = (k_pos / 16) % nl;
+
+                    device const block_q * row_ptr = (device const block_q *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    SA_4x4 temp_a;
+                    dequantize_func(row_ptr + block_idx, il, temp_a);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        // Zero-pad A for K positions beyond valid range (handles partial K iterations)
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? temp_a[i/4][i%4] : (SA)0;
+                    }
+                }
+            } else {
+                // Zero-pad rows beyond matrix bounds
+                FOR_UNROLL (short i = 0; i < 16; i++) {
+                    sa[row * N_MM_NK_TOTAL + (k_base + i)] = (SA)0;
+                }
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // === PHASE 2: Tensor matmul ===
+        auto mA = tA.slice(0, 0);
+        auto mB = tB.slice(loop_k, rb);
+
+        mm.run(mB, mA, cT);
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // Store result tile to output matrix (with batch offset)
+    // cT.store handles bounds checking via tD's extents (M, N)
+    device float * dstBatch = (device float *)dst + im * N * M;
+
+    auto tD = tensor(dstBatch, dextents<int32_t, 2>(M, N), array<int, 2>({1, M}));
+    cT.store(tD.slice(ra, rb));
+}
+
+#else
+
+template<
+    typename S0, typename S0_4x4, typename S0_8x8,
+    typename S1, typename S1_2x4, typename S1_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z;
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (args.ne1 - r1 < NR1) ? (args.ne1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 + lr1)
+        + args.nb10*iy);
+
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+    }
+
+    if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
+        // if no bounds checks on the output are needed, we can directly write to device memory
+        device float * C = (device float *) dst +
+            (r0 + 32*(sgitg &  1)) + \
+            (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
+        }
+    } else {
+        // block is smaller than 64x32, we should avoid writing data outside of the matrix
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (sgitg == 0) {
+            for (int j = tiitg; j < nr1; j += NR1) {
+                device float  * D  = (device float  *) dst + r0 + (r1 + j)*args.ne0 + im*args.ne1*args.ne0;
+                device float4 * D4 = (device float4 *) D;
+
+                threadgroup float  * C  = temp_str + (j*NR0);
+                threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+                int i = 0;
+                for (; i < nr0/4; i++) {
+                    *(D4 + i) = *(C4 + i);
+                }
+
+                i *= 4;
+                for (; i < nr0; i++) {
+                    *(D + i) = *(C + i);
+                }
+            }
+        }
+    }
+}
+
+#endif // GGML_METAL_HAS_TENSOR
+
+template<short ne20> // n_expert_used
+kernel void kernel_mul_mm_id_map0(
+        constant ggml_metal_kargs_mul_mm_id_map0 & args,
+        device  const char * src2,
+        device        char * htpe,
+        device        char * hids,
+        threadgroup   char * shmem [[threadgroup(0)]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    const short ide = tpitg; // expert id
+
+    uint32_t n_all = 0;
+
+    device int32_t * ids_i32 = (device int32_t *) hids + ide*args.ne21;
+
+    for (int i21 = 0; i21 < args.ne21; i21 += ntg) { // n_tokens
+        if (i21 + tpitg < args.ne21) {
+            device const int32_t * src2_i32 = (device const int32_t *) (src2 + (i21 + tpitg)*args.nb21);
+
+            threadgroup uint16_t * sids = (threadgroup uint16_t *) shmem + tpitg*ne20;
+
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sids[i20] = src2_i32[i20];
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (short t = 0; t < ntg; t++) {
+            if (i21 + t >= args.ne21) {
+                break;
+            }
+
+            threadgroup const uint16_t * sids = (threadgroup const uint16_t *) shmem + t*ne20;
+
+            short sel = 0;
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sel += (sids[i20] == ide)*(i20 + 1);
+            }
+
+            ids_i32[n_all] = (i21 + t)*ne20 + sel - 1;
+
+            n_all += sel > 0;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    device uint32_t * tpe_u32 = (device uint32_t *) (htpe);
+    tpe_u32[ide] = n_all;
+}
+
+typedef decltype(kernel_mul_mm_id_map0<1>) kernel_mul_mm_id_map0_t;
+
+template [[host_name("kernel_mul_mm_id_map0_ne20_1" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<1>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_2" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<2>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_4" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<4>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_5" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<5>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_6" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<6>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_22")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<22>;
+
+template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm_id(
+        constant ggml_metal_kargs_mul_mm_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * htpe,
+        device const char * hids,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    threadgroup float * sc = (threadgroup float *)(shmem);
+#endif
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z; // expert
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
+    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
+
+    const int32_t neh1 = tpe_u32[im];
+
+    if (r1 >= neh1) {
+        return;
+    }
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int id = ids_i32[im*args.ne21 + r1 + lr1];
+
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
+
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? (S0) *((device T0 *) x + i) : (S0) 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifndef GGML_METAL_HAS_TENSOR
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
+    }
+
+    // block is smaller than 64x32, we should avoid writing data outside of the matrix
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+    cT.store(tC);
+#else
+    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+    for (short i = 0; i < 8; i++) {
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+    }
+#endif
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (short j = sgitg; j < nr1; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1 + j];
+
+        const short ide = id % args.ne20;
+        const short idt = id / args.ne20;
+
+        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float4 * D4 = (device float4 *) D;
+
+        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
+        threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+        int i = tiisg;
+        for (; i < nr0/4; i += 32) {
+            *(D4 + i) = *(C4 + i);
+        }
+
+        i = (4*(nr0/4)) + tiisg;
+        for (; i < nr0; i += 32) {
+            *(D + i) = *(C + i);
+        }
+    }
+}
+
+//
+// matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_t;
+
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_bf16_f32")]]    kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_q1_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_f32_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_f16_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_q1_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+
+//
+// indirect matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_id;
+
+template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_id_bf16_f32")]]    kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_id_q1_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_id_f32_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q1_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;

ggml/src/ggml-metal/kernels/norm.metal

@@ -0,0 +1,308 @@
+#include "common.h"
+
+// F == 1 : norm (no fuse)
+// F == 2 : norm + mul
+// F == 3 : norm + mul + add
+template <typename T, short F>
+kernel void kernel_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    T sumft(0.0f);
+
+    float sumf = 0.0f;
+
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumft += x[i00];
+    }
+    sumf = dot(sumft, T(1.0f));
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean = sumf/args.ne00;
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+
+    sumf = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        y[i00] = x[i00] - mean;
+        sumf += dot(y[i00], y[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float variance = sumf/args.ne00;
+
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (y[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (y[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (y[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_norm_fuse_impl<float4, 1>) kernel_norm_fuse_t;
+
+template [[host_name("kernel_norm_f32")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_norm_mul_f32")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_norm_mul_add_f32")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_norm_f32_4")]]         kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_norm_mul_f32_4")]]     kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_norm_mul_add_f32_4")]] kernel kernel_norm_fuse_t kernel_norm_fuse_impl<float4, 3>;
+
+// F == 1 : rms_norm (no fuse)
+// F == 2 : rms_norm + mul
+// F == 3 : rms_norm + mul + add
+template <typename T, short F>
+kernel void kernel_rms_norm_fuse_impl(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device const char * src1_0,
+        device const char * src1_1,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    const int i01 = tgpig.x;
+    const int i02 = tgpig.y;
+    const int i03 = tgpig.z;
+
+    device const T * x = (device const T *) (src0 + i03*args.nbf3[0] + i02*args.nbf2[0] + i01*args.nbf1[0]);
+
+    device const T * f0 = (device const T *) (src1_0 + (i03%args.nef3[1])*args.nbf3[1] + (i02%args.nef2[1])*args.nbf2[1] + (i01%args.nef1[1])*args.nbf1[1]);
+    device const T * f1 = (device const T *) (src1_1 + (i03%args.nef3[2])*args.nbf3[2] + (i02%args.nef2[2])*args.nbf2[2] + (i01%args.nef1[2])*args.nbf1[2]);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean  = sumf/args.ne00;
+    const float scale = 1.0f/sqrt(mean + args.eps);
+
+    device T * y = (device T *) (dst + i03*args.nb3 + i02*args.nb2 + i01*args.nb1);
+    for (int i00 = tpitg.x; i00 < args.ne00_t; i00 += ntg.x) {
+        if (F == 1) {
+            y[i00] = (x[i00]*scale);
+        }
+        if (F == 2) {
+            y[i00] = (x[i00]*scale)*f0[i00];
+        }
+        if (F == 3) {
+            y[i00] = (x[i00]*scale)*f0[i00] + f1[i00];
+        }
+    }
+}
+
+typedef decltype(kernel_rms_norm_fuse_impl<float4, 1>) kernel_rms_norm_fuse_t;
+
+template [[host_name("kernel_rms_norm_f32")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 1>;
+template [[host_name("kernel_rms_norm_mul_f32")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float, 3>;
+
+template [[host_name("kernel_rms_norm_f32_4")]]         kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 1>;
+template [[host_name("kernel_rms_norm_mul_f32_4")]]     kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 2>;
+template [[host_name("kernel_rms_norm_mul_add_f32_4")]] kernel kernel_rms_norm_fuse_t kernel_rms_norm_fuse_impl<float4, 3>;
+
+template <typename T0, typename T>
+kernel void kernel_l2_norm_impl(
+        constant ggml_metal_kargs_l2_norm & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    device const T0 * x = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T  * y = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float scale = 1.0f/max(sqrt(sumf), args.eps);
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        y[i00] = x[i00] * scale;
+    }
+}
+
+typedef decltype(kernel_l2_norm_impl<float, float>) kernel_l2_norm_t;
+
+template [[host_name("kernel_l2_norm_f32_f32")]]   kernel kernel_l2_norm_t kernel_l2_norm_impl<float,  float>;
+template [[host_name("kernel_l2_norm_f32_f32_4")]] kernel kernel_l2_norm_t kernel_l2_norm_impl<float4, float4>;
+
+kernel void kernel_group_norm_f32(
+        constant ggml_metal_kargs_group_norm & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    const int64_t ne = args.ne00*args.ne01*args.ne02;
+    const int64_t gs = args.ne00*args.ne01*((args.ne02 + args.ngrp - 1) / args.ngrp);
+
+    int start = tgpig * gs;
+    int end   = start + gs;
+
+    start += tpitg;
+
+    if (end >= ne) {
+        end = ne;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += ntg) {
+        tmp += src0[j];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float mean = tmp / gs;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += ntg) {
+        float xi = src0[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float variance = tmp / gs;
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int j = start; j < end; j += ntg) {
+        dst[j] *= scale;
+    }
+}

ggml/src/ggml-metal/kernels/pool.metal

@@ -0,0 +1,148 @@
+#include "common.h"
+
+kernel void kernel_pool_2d_max_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+
+    float res = -INFINITY;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            res = MAX(res, i_ptr[i * args.IW + j]);
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+kernel void kernel_pool_2d_avg_f32(
+        constant    ggml_metal_kargs_pool_2d & args,
+        device  const float * src0,
+        device        float * dst,
+        uint        gid[[thread_position_in_grid]]) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int idx = gid;
+    const int I_HW = args.IH * args.IW;
+    const int O_HW = args.OH * args.OW;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / args.OW;
+    const int cur_ow = idx % O_HW % args.OW;
+
+    device const float * i_ptr = src0 + nc * I_HW;
+    device       float * o_ptr = dst  + nc * O_HW;
+
+    const int start_h = cur_oh * args.s1 - args.p1;
+    const int bh = MAX(0,  start_h);
+    const int eh = MIN(args.IH, start_h + args.k1);
+    const int start_w = cur_ow * args.s0 - args.p0;
+    const int bw = MAX(0,  start_w);
+    const int ew = MIN(args.IW, start_w + args.k0);
+    // const float scale = 1. / ((eh - bh) * (ew - bw));
+    const float scale = 1. / (args.k0 * args.k1);
+
+    float res = 0;
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+            float cur = i_ptr[i * args.IW + j];
+            res += cur * scale;
+        }
+    }
+
+    o_ptr[cur_oh * args.OW + cur_ow] = res;
+}
+
+
+kernel void kernel_pool_1d_max_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = -INFINITY;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        int j = base + ki;
+        if (j < 0 || j >= args.IW){
+            continue;
+        }
+        float v = src[src_off + j];
+        acc = max(acc, v);
+    }
+
+    dst[dst_off + ow] = acc;
+}
+
+kernel void kernel_pool_1d_avg_f32(
+        constant        ggml_metal_kargs_pool_1d & args,
+        device  const   float * src,
+        device          float * dst,
+        uint            gid [[thread_position_in_grid]]
+) {
+
+    if (gid >= args.np) {
+        return;
+    }
+
+    const int ow  = (int)gid % args.OW;
+    const int row = (int)gid / args.OW;
+
+    const int base = ow * args.s0 - args.p0;
+
+    float acc = 0.0f;
+    int   cnt = 0;
+
+    const int src_off = row * args.IW;
+    const int dst_off = row * args.OW;
+
+    for (int ki = 0; ki < args.k0; ++ki) {
+        const int j = base + ki;
+        if (j < 0 || j >= args.IW) {
+            continue;
+        }
+        acc += src[src_off + j];
+        cnt += 1;
+    }
+
+    dst[dst_off + ow] = (cnt > 0) ? (acc / (float)cnt) : 0.0f;
+}