update release requirements (#11897)

* vulkan: support memset_tensor

* vulkan: support GGML_OP_SUM

* vulkan: implement GGML_OP_ARGMAX

* vulkan: implement GGML_OP_SUB

* vulkan: implement GGML_OP_COUNT_EQUAL

* vulkan: implement GGML_OP_OPT_STEP_ADAMW

* vulkan: fix check_results RWKV_WKV6 crash and memory leaks

* vulkan: implement GGML_OP_REPEAT_BACK

* tests: remove invalid test-backend-ops REPEAT_BACK tests

* vulkan: fix COUNT_EQUAL memset using a fillBuffer command

.devops/cuda.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=12.6.0
+ARG CUDA_VERSION=12.4.0
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 

.devops/llama-cpp-cuda.srpm.spec

@@ -17,10 +17,10 @@ Version:        %( date "+%%Y%%m%%d" )
 Release:        1%{?dist}
 Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
 License:        MIT
-Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+Source0:        https://github.com/ggml-org/llama.cpp/archive/refs/heads/master.tar.gz
 BuildRequires:  coreutils make gcc-c++ git cuda-toolkit
 Requires:       cuda-toolkit
-URL:            https://github.com/ggerganov/llama.cpp
+URL:            https://github.com/ggml-org/llama.cpp
 
 %define debug_package %{nil}
 %define source_date_epoch_from_changelog 0

.devops/llama-cpp.srpm.spec

@@ -18,10 +18,10 @@ Version:        %( date "+%%Y%%m%%d" )
 Release:        1%{?dist}
 Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
 License:        MIT
-Source0:        https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz
+Source0:        https://github.com/ggml-org/llama.cpp/archive/refs/heads/master.tar.gz
 BuildRequires:  coreutils make gcc-c++ git libstdc++-devel
 Requires:       libstdc++
-URL:            https://github.com/ggerganov/llama.cpp
+URL:            https://github.com/ggml-org/llama.cpp
 
 %define debug_package %{nil}
 %define source_date_epoch_from_changelog 0

.devops/musa.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG MUSA_VERSION=rc3.1.0
+ARG MUSA_VERSION=rc3.1.1
 # Target the MUSA build image
 ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 

.devops/nix/package.nix

@@ -133,12 +133,12 @@ effectiveStdenv.mkDerivation (finalAttrs: {
       --replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
   '';
 
-  # With PR#6015 https://github.com/ggerganov/llama.cpp/pull/6015,
+  # With PR#6015 https://github.com/ggml-org/llama.cpp/pull/6015,
   # `default.metallib` may be compiled with Metal compiler from XCode
   # and we need to escape sandbox on MacOS to access Metal compiler.
   # `xcrun` is used find the path of the Metal compiler, which is varible
   # and not on $PATH
-  # see https://github.com/ggerganov/llama.cpp/pull/6118 for discussion
+  # see https://github.com/ggml-org/llama.cpp/pull/6118 for discussion
   __noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
 
   nativeBuildInputs =
@@ -220,7 +220,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
     broken = (useMetalKit && !effectiveStdenv.isDarwin);
 
     description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
-    homepage = "https://github.com/ggerganov/llama.cpp/";
+    homepage = "https://github.com/ggml-org/llama.cpp/";
     license = lib.licenses.mit;
 
     # Accommodates `nix run` and `lib.getExe`

.devops/rocm.Dockerfile

@@ -11,7 +11,7 @@ ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-co
 FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 
 # Unless otherwise specified, we make a fat build.
-# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
+# List from https://github.com/ggml-org/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
 # gfx803, gfx900, gfx1032, gfx1101, gfx1102,not officialy supported
 # gfx906 is deprecated

.github/ISSUE_TEMPLATE/020-enhancement.yml

@@ -6,7 +6,7 @@ body:
   - type: markdown
     attributes:
       value: |
-        [Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggerganov/llama.cpp/discussions/categories/ideas)
+        [Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggml-org/llama.cpp/discussions/categories/ideas)
 
   - type: checkboxes
     id: prerequisites
@@ -16,11 +16,11 @@ body:
       options:
         - label: I am running the latest code. Mention the version if possible as well.
           required: true
-        - label: I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
+        - label: I carefully followed the [README.md](https://github.com/ggml-org/llama.cpp/blob/master/README.md).
           required: true
         - label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
           required: true
-        - label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new and useful enhancement to share.
+        - label: I reviewed the [Discussions](https://github.com/ggml-org/llama.cpp/discussions), and have a new and useful enhancement to share.
           required: true
 
   - type: textarea

.github/ISSUE_TEMPLATE/030-research.yml

@@ -6,7 +6,7 @@ body:
   - type: markdown
     attributes:
       value: |
-        Don't forget to check for any [duplicate research issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22)
+        Don't forget to check for any [duplicate research issue tickets](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22)
 
   - type: checkboxes
     id: research-stage

.github/ISSUE_TEMPLATE/040-refactor.yml

@@ -6,8 +6,8 @@ body:
   - type: markdown
     attributes:
       value: |
-        Don't forget to [check for existing refactor issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered.
-        Also you may want to check [Pull request refactor label as well](https://github.com/ggerganov/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too.
+        Don't forget to [check for existing refactor issue tickets](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered.
+        Also you may want to check [Pull request refactor label as well](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too.
 
   - type: textarea
     id: background-description

.github/ISSUE_TEMPLATE/config.yml

@@ -1,11 +1,11 @@
 blank_issues_enabled: true
 contact_links:
   - name: Got an idea?
-    url: https://github.com/ggerganov/llama.cpp/discussions/categories/ideas
+    url: https://github.com/ggml-org/llama.cpp/discussions/categories/ideas
     about: Pop it there. It may then become an enhancement ticket.
   - name: Got a question?
-    url: https://github.com/ggerganov/llama.cpp/discussions/categories/q-a
+    url: https://github.com/ggml-org/llama.cpp/discussions/categories/q-a
     about: Ask a question there!
   - name: Want to contribute?
-    url: https://github.com/ggerganov/llama.cpp/wiki/contribute
+    url: https://github.com/ggml-org/llama.cpp/wiki/contribute
     about: Head to the contribution guide page of the wiki for areas you can help with

.github/pull_request_template.md

@@ -1 +1 @@
-*Make sure to read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md) before submitting a PR*
+*Make sure to read the [contributing guidelines](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md) before submitting a PR*

.github/workflows/bench.yml.disabled

@@ -1,5 +1,5 @@
 # TODO: there have been some issues with the workflow, so disabling for now
-#       https://github.com/ggerganov/llama.cpp/issues/7893
+#       https://github.com/ggml-org/llama.cpp/issues/7893
 #
 # Benchmark
 name: Benchmark
@@ -57,17 +57,7 @@ jobs:
 
     if: |
       inputs.gpu-series == 'Standard_NC4as_T4_v3'
-      || (
-        github.event_name == 'schedule'
-        && github.ref_name == 'master'
-        && github.repository_owner == 'ggerganov'
-      )
       || github.event_name == 'pull_request_target'
-      || (
-        github.event_name == 'push'
-        && github.event.ref == 'refs/heads/master'
-        && github.repository_owner == 'ggerganov'
-      )
     steps:
       - name: Clone
         id: checkout

.github/workflows/build.yml

@@ -129,7 +129,7 @@ jobs:
         run: |
           sysctl -a
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
-          # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
+          # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
           cmake -B build \
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
@@ -374,6 +374,8 @@ jobs:
       - name: Clone
         id: checkout
         uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
 
       - name: ccache
         uses: hendrikmuhs/ccache-action@v1.2.16
@@ -401,7 +403,35 @@ jobs:
         run: |
           cd build
           # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 1800
+          ctest -L main --verbose --timeout 2700
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp LICENSE ./build/bin/
+          cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip
+          name: llama-bin-ubuntu-vulkan-x64.zip
 
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
@@ -443,7 +473,7 @@ jobs:
 
   ubuntu-22-cmake-musa:
     runs-on: ubuntu-22.04
-    container: mthreads/musa:rc3.1.0-devel-ubuntu22.04
+    container: mthreads/musa:rc3.1.1-devel-ubuntu22.04
 
     steps:
       - name: Clone
@@ -1345,8 +1375,10 @@ jobs:
 
     needs:
       - ubuntu-cpu-cmake
+      - ubuntu-22-cmake-vulkan
       - windows-latest-cmake
       - windows-2019-cmake-cuda
+      - windows-latest-cmake-sycl
       - windows-latest-cmake-hip-release
       - macOS-latest-cmake-arm64
       - macOS-latest-cmake-x64

.github/workflows/docker.yml

@@ -51,6 +51,8 @@ jobs:
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v3
+        with:
+          image: tonistiigi/binfmt:qemu-v7.0.0-28
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3

.github/workflows/labeler.yml

@@ -11,7 +11,7 @@ jobs:
     steps:
     - uses: actions/checkout@v4
       with:
-        repository: "ggerganov/llama.cpp"
+        repository: "ggml-org/llama.cpp"
     - uses: actions/labeler@v5
       with:
         configuration-path: '.github/labeler.yml'

CONTRIBUTING.md

@@ -12,7 +12,7 @@
 
 - Squash-merge PRs
 - Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
-- Optionally pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
+- Optionally pick a `<module>` from here: https://github.com/ggml-org/llama.cpp/wiki/Modules
 - Consider adding yourself to [CODEOWNERS](CODEOWNERS)
 
 # Coding guidelines
@@ -40,14 +40,14 @@
 - Try to follow the existing patterns in the code (indentation, spaces, etc.). In case of doubt use `clang-format` to format the added code
 - For anything not covered in the current guidelines, refer to the [C++ Core Guidelines](https://isocpp.github.io/CppCoreGuidelines/CppCoreGuidelines)
 - Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
-- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
+- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggml-org/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
 
 ![matmul](media/matmul.png)
 
 # Naming guidelines
 
 - Use `snake_case` for function, variable and type names
-- Naming usually optimizes for longest common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
+- Naming usually optimizes for longest common prefix (see https://github.com/ggml-org/ggml/pull/302#discussion_r1243240963)
 
     ```cpp
     // not OK
@@ -122,4 +122,4 @@
 
 The Github issues, PRs and discussions contain a lot of information that can be useful to get familiar with the codebase. For convenience, some of the more important information is referenced from Github projects:
 
-https://github.com/ggerganov/llama.cpp/projects
+https://github.com/ggml-org/llama.cpp/projects

Makefile

@@ -1,5 +1,5 @@
 ifndef LLAMA_MAKEFILE
-$(error The Makefile build is deprecated. Use the CMake build instead. For more details, see https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md)
+$(error The Makefile build is deprecated. Use the CMake build instead. For more details, see https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md)
 endif
 
 # Define the default target now so that it is always the first target
@@ -463,7 +463,7 @@ endif
 ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
 	# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
 	# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
-	# https://github.com/ggerganov/llama.cpp/issues/2922
+	# https://github.com/ggml-org/llama.cpp/issues/2922
 	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
 	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
 
@@ -1078,8 +1078,8 @@ endif
 ifdef REMOVE_WARNING
 $(info !!! REMOVAL WARNING !!!)
 $(info The following LLAMA_ options have been removed and are no longer supported)
-$(info   - LLAMA_DISABLE_LOGS   (https://github.com/ggerganov/llama.cpp/pull/9418))
-$(info   - LLAMA_SERVER_VERBOSE (https://github.com/ggerganov/llama.cpp/pull/9418))
+$(info   - LLAMA_DISABLE_LOGS   (https://github.com/ggml-org/llama.cpp/pull/9418))
+$(info   - LLAMA_SERVER_VERBOSE (https://github.com/ggml-org/llama.cpp/pull/9418))
 $(info )
 endif
 

README.md

@@ -3,26 +3,33 @@
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
+[![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)
 
-[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
+[Roadmap](https://github.com/users/ggml-org/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
+> [!IMPORTANT]
+> New `llama.cpp` package location: [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp/pkgs/container/llama.cpp)
+>
+> Update your container URLs to: `ghcr.io/ggml-org/llama.cpp`
+>
+> More info: https://github.com/ggml-org/llama.cpp/discussions/11801
+
 ## Recent API changes
 
-- [Changelog for `libllama` API](https://github.com/ggerganov/llama.cpp/issues/9289)
-- [Changelog for `llama-server` REST API](https://github.com/ggerganov/llama.cpp/issues/9291)
+- [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
 
-- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggerganov/llama.cpp/pull/11427
+- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427
 - **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode
-- Universal tool call support in `llama-server`: https://github.com/ggerganov/llama.cpp/pull/9639
+- Universal tool call support in `llama-server`: https://github.com/ggml-org/llama.cpp/pull/9639
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
-- Introducing GGUF-my-LoRA https://github.com/ggerganov/llama.cpp/discussions/10123
-- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669
-- Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
+- Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
+- 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)
 
 ----
 
@@ -39,7 +46,7 @@ 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/ggerganov/ggml) library.
+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>
@@ -59,23 +66,23 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - [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/ggerganov/llama.cpp/pull/5423)
+- [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/ggerganov/llama.cpp/pull/3187)
+- [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/ggerganov/llama.cpp/pull/3417)
-- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
+- [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/ggerganov/llama.cpp/pull/3557)
+- [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/ggerganov/llama.cpp/pull/11003)
+- [x] [PhiMoE](https://github.com/ggml-org/llama.cpp/pull/11003)
 - [x] [GPT-2](https://huggingface.co/gpt2)
-- [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
+- [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)
@@ -146,7 +153,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 - 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/ggerganov/llama.cpp/pull/6326)
+- 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)
@@ -245,7 +252,7 @@ The project also includes many example programs and tools using the `llama` libr
 - Clone this repository and build locally, see [how to build](docs/build.md)
 - On MacOS or Linux, install `llama.cpp` via [brew, flox or nix](docs/install.md)
 - Use a Docker image, see [documentation for Docker](docs/docker.md)
-- Download pre-built binaries from [releases](https://github.com/ggerganov/llama.cpp/releases)
+- Download pre-built binaries from [releases](https://github.com/ggml-org/llama.cpp/releases)
 
 ## Obtaining and quantizing models
 
@@ -258,14 +265,14 @@ You can either manually download the GGUF file or directly use any `llama.cpp`-c
 
 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/ggerganov/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.
+`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/ggerganov/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/ggerganov/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/ggerganov/llama.cpp/discussions/9669)
+- 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](examples/quantize/README.md)
 
@@ -488,9 +495,9 @@ To learn more about model quantization, [read this documentation](examples/quant
 - Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
 - Collaborators will be invited based on contributions
 - Any help with managing issues, PRs and projects is very appreciated!
-- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
+- 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/ggerganov/llama.cpp/discussions/205)
+- 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
@@ -505,7 +512,7 @@ To learn more about model quantization, [read this documentation](examples/quant
 - [Running on Docker](docs/docker.md)
 - [Build on Android](docs/android.md)
 - [Performance troubleshooting](docs/development/token_generation_performance_tips.md)
-- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
+- [GGML tips & tricks](https://github.com/ggml-org/llama.cpp/wiki/GGML-Tips-&-Tricks)
 
 #### Seminal papers and background on the models
 
@@ -519,5 +526,18 @@ If your issue is with model generation quality, then please at least scan the fo
     - [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)
 
-#### References
+## 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
+```
+
+## References

SECURITY.md

@@ -62,6 +62,6 @@ Beware that none of the topics under [Using llama.cpp securely](#using-llamacpp-
 <!-- normal version -->
 However, If you have discovered a security vulnerability in this project, please report it privately. **Do not disclose it as a public issue.** This gives us time to work with you to fix the issue before public exposure, reducing the chance that the exploit will be used before a patch is released.
 
-Please disclose it as a private [security advisory](https://github.com/ggerganov/llama.cpp/security/advisories/new).
+Please disclose it as a private [security advisory](https://github.com/ggml-org/llama.cpp/security/advisories/new).
 
 A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.

ci/README.md

@@ -1,11 +1,11 @@
 # CI
 
-In addition to [Github Actions](https://github.com/ggerganov/llama.cpp/actions) `llama.cpp` uses a custom CI framework:
+In addition to [Github Actions](https://github.com/ggml-org/llama.cpp/actions) `llama.cpp` uses a custom CI framework:
 
 https://github.com/ggml-org/ci
 
 It monitors the `master` branch for new commits and runs the
-[ci/run.sh](https://github.com/ggerganov/llama.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us
+[ci/run.sh](https://github.com/ggml-org/llama.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us
 to execute heavier workloads compared to just using Github Actions. Also with time, the cloud instances will be scaled
 to cover various hardware architectures, including GPU and Apple Silicon instances.
 

common/CMakeLists.txt

@@ -96,6 +96,22 @@ if (LLAMA_LLGUIDANCE)
     include(ExternalProject)
     set(LLGUIDANCE_SRC ${CMAKE_BINARY_DIR}/llguidance/source)
     set(LLGUIDANCE_PATH ${LLGUIDANCE_SRC}/target/release)
+
+    # Set the correct library file extension based on platform
+    if (WIN32)
+        set(LLGUIDANCE_LIB_NAME "llguidance.lib")
+        # Add Windows-specific libraries
+        set(LLGUIDANCE_PLATFORM_LIBS
+            ws2_32    # Windows Sockets API
+            userenv   # For GetUserProfileDirectoryW
+            ntdll     # For NT functions
+            bcrypt    # For BCryptGenRandom
+        )
+    else()
+        set(LLGUIDANCE_LIB_NAME "libllguidance.a")
+        set(LLGUIDANCE_PLATFORM_LIBS "")
+    endif()
+
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
         # v0.6.12:
@@ -106,17 +122,18 @@ if (LLAMA_LLGUIDANCE)
         CONFIGURE_COMMAND ""
         BUILD_COMMAND cargo build --release
         INSTALL_COMMAND ""
-        BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/libllguidance.a ${LLGUIDANCE_PATH}/llguidance.h
+        BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME} ${LLGUIDANCE_PATH}/llguidance.h
         UPDATE_COMMAND ""
     )
     target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_LLGUIDANCE)
 
     add_library(llguidance STATIC IMPORTED)
-    set_target_properties(llguidance PROPERTIES IMPORTED_LOCATION ${LLGUIDANCE_PATH}/libllguidance.a)
+    set_target_properties(llguidance PROPERTIES IMPORTED_LOCATION ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME})
     add_dependencies(llguidance llguidance_ext)
 
     target_include_directories(${TARGET} PRIVATE ${LLGUIDANCE_PATH})
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance)
+    # Add platform libraries to the main target
+    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance ${LLGUIDANCE_PLATFORM_LIBS})
 endif ()
 
 target_include_directories(${TARGET} PUBLIC .)

common/arg.cpp

@@ -365,6 +365,112 @@ static void common_params_print_usage(common_params_context & ctx_arg) {
     print_options(specific_options);
 }
 
+static void common_params_print_completion(common_params_context & ctx_arg) {
+    std::vector<common_arg *> common_options;
+    std::vector<common_arg *> sparam_options;
+    std::vector<common_arg *> specific_options;
+
+    for (auto & opt : ctx_arg.options) {
+        if (opt.is_sparam) {
+            sparam_options.push_back(&opt);
+        } else if (opt.in_example(ctx_arg.ex)) {
+            specific_options.push_back(&opt);
+        } else {
+            common_options.push_back(&opt);
+        }
+    }
+
+    printf("_llama_completions() {\n");
+    printf("    local cur prev opts\n");
+    printf("    COMPREPLY=()\n");
+    printf("    cur=\"${COMP_WORDS[COMP_CWORD]}\"\n");
+    printf("    prev=\"${COMP_WORDS[COMP_CWORD-1]}\"\n\n");
+
+    printf("    opts=\"");
+    auto print_options = [](const std::vector<common_arg *> & options) {
+        for (const common_arg * opt : options) {
+            for (const char * arg : opt->args) {
+                printf("%s ", arg);
+            }
+        }
+    };
+
+    print_options(common_options);
+    print_options(sparam_options);
+    print_options(specific_options);
+    printf("\"\n\n");
+
+    printf("    case \"$prev\" in\n");
+    printf("        --model)\n");
+    printf("            COMPREPLY=( $(compgen -f -X '!*.gguf' -- \"$cur\") $(compgen -d -- \"$cur\") )\n");
+    printf("            return 0\n");
+    printf("            ;;\n");
+    printf("        --grammar-file)\n");
+    printf("            COMPREPLY=( $(compgen -f -X '!*.gbnf' -- \"$cur\") $(compgen -d -- \"$cur\") )\n");
+    printf("            return 0\n");
+    printf("            ;;\n");
+    printf("        --chat-template-file)\n");
+    printf("            COMPREPLY=( $(compgen -f -X '!*.jinja' -- \"$cur\") $(compgen -d -- \"$cur\") )\n");
+    printf("            return 0\n");
+    printf("            ;;\n");
+    printf("        *)\n");
+    printf("            COMPREPLY=( $(compgen -W \"${opts}\" -- \"$cur\") )\n");
+    printf("            return 0\n");
+    printf("            ;;\n");
+    printf("    esac\n");
+    printf("}\n\n");
+
+    std::set<std::string> executables = {
+        "llama-batched",
+        "llama-batched-bench",
+        "llama-bench",
+        "llama-cli",
+        "llama-convert-llama2c-to-ggml",
+        "llama-cvector-generator",
+        "llama-embedding",
+        "llama-eval-callback",
+        "llama-export-lora",
+        "llama-gbnf-validator",
+        "llama-gen-docs",
+        "llama-gguf",
+        "llama-gguf-hash",
+        "llama-gguf-split",
+        "llama-gritlm",
+        "llama-imatrix",
+        "llama-infill",
+        "llama-llava-cli",
+        "llama-llava-clip-quantize-cli",
+        "llama-lookahead",
+        "llama-lookup",
+        "llama-lookup-create",
+        "llama-lookup-merge",
+        "llama-lookup-stats",
+        "llama-minicpmv-cli",
+        "llama-parallel",
+        "llama-passkey",
+        "llama-perplexity",
+        "llama-q8dot",
+        "llama-quantize",
+        "llama-quantize-stats",
+        "llama-qwen2vl-cli",
+        "llama-retrieval",
+        "llama-run",
+        "llama-save-load-state",
+        "llama-server",
+        "llama-simple",
+        "llama-simple-chat",
+        "llama-speculative",
+        "llama-speculative-simple",
+        "llama-tokenize",
+        "llama-tts",
+        "llama-vdot"
+    };
+
+    for (const auto& exe : executables) {
+        printf("complete -F _llama_completions %s\n", exe.c_str());
+    }
+}
+
 static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & value) {
     std::vector<ggml_backend_dev_t> devices;
     auto dev_names = string_split<std::string>(value, ',');
@@ -426,6 +532,10 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
             }
             exit(0);
         }
+        if (ctx_arg.params.completion) {
+            common_params_print_completion(ctx_arg);
+            exit(0);
+        }
     } catch (const std::invalid_argument & ex) {
         fprintf(stderr, "%s\n", ex.what());
         ctx_arg.params = params_org;
@@ -494,6 +604,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             exit(0);
         }
     ));
+    add_opt(common_arg(
+        {"--completion-bash"},
+        "print source-able bash completion script for llama.cpp",
+        [](common_params & params) {
+            params.completion = true;
+        }
+    ));
     add_opt(common_arg(
         {"--verbose-prompt"},
         string_format("print a verbose prompt before generation (default: %s)", params.verbose_prompt ? "true" : "false"),
@@ -946,6 +1063,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sampling.min_p = std::stof(value);
         }
     ).set_sparam());
+    add_opt(common_arg(
+        {"--top-nsigma"}, "N",
+        string_format("top-n-sigma sampling (default: %.1f, -1.0 = disabled)", params.sampling.top_n_sigma),
+        [](common_params & params, const std::string & value) {
+            params.sampling.top_n_sigma = std::stof(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_MAIN}).set_sparam());
     add_opt(common_arg(
         {"--xtc-probability"}, "N",
         string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability),
@@ -1445,7 +1569,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "- isolate: only spawn threads on CPUs on the node that execution started on\n"
         "- numactl: use the CPU map provided by numactl\n"
         "if run without this previously, it is recommended to drop the system page cache before using this\n"
-        "see https://github.com/ggerganov/llama.cpp/issues/1437",
+        "see https://github.com/ggml-org/llama.cpp/issues/1437",
         [](common_params & params, const std::string & value) {
             /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
             else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
@@ -1975,6 +2099,17 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.use_jinja = true;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
+    add_opt(common_arg(
+        {"--reasoning-format"}, "FORMAT",
+        "reasoning format (default: deepseek; allowed values: deepseek, none)\n"
+        "controls whether thought tags are extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only).\n"
+        "only supported for non-streamed responses",
+        [](common_params & params, const std::string & value) {
+            /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
+            else if (value == "none") {     params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
+            else { std::invalid_argument("invalid value"); }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK"));
     add_opt(common_arg(
         {"--chat-template"}, "JINJA_TEMPLATE",
         string_format(
@@ -2112,7 +2247,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_env("LLAMA_LOG_VERBOSITY"));
     add_opt(common_arg(
         {"--log-prefix"},
-        "Enable prefx in log messages",
+        "Enable prefix in log messages",
         [](common_params &) {
             common_log_set_prefix(common_log_main(), true);
         }

common/chat.cpp

@@ -12,11 +12,13 @@ std::string common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_LLAMA_3_X: return "Llama 3.x";
         case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS: return "Llama 3.x with builtin tools";
         case COMMON_CHAT_FORMAT_DEEPSEEK_R1: return "DeepSeek R1";
+        case COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING: return "DeepSeek R1 (extract reasoning)";
         case COMMON_CHAT_FORMAT_FIREFUNCTION_V2: return "FireFunction v2";
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: return "Functionary v3.2";
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1";
         case COMMON_CHAT_FORMAT_HERMES_2_PRO: return "Hermes 2 Pro";
         case COMMON_CHAT_FORMAT_COMMAND_R7B: return "Command R7B";
+        case COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING: return "Command R7B (extract reasoning)";
         default:
             throw std::runtime_error("Unknown chat format");
     }
@@ -105,7 +107,6 @@ static common_chat_msg parse_json_tool_calls(
         std::sregex_iterator rend;
         std::sregex_iterator rit(it, end, function_regex);
         if (rit == rend) {
-            fprintf(stderr, "No more tool calls found\n");
             result.content += std::string(it, end);
             break;
         }
@@ -115,14 +116,21 @@ static common_chat_msg parse_json_tool_calls(
 
         json arguments;
         if (!parse_json(it, end, arguments)) {
-            throw std::runtime_error("Failed to parse json tool call arguments");
+            throw std::runtime_error("Failed to parse json tool call arguments: " + input);
         }
         if (!std::regex_search(it, end, match, close_regex)) {
-            throw std::runtime_error("Malformed input, missing closing pattern");
+            throw std::runtime_error("Malformed input, missing closing pattern: " + input);
         }
         it = match.suffix().first;
         result.tool_calls.push_back({name, arguments.is_string() ? arguments.get<std::string>() : arguments.dump(), /* id= */ ""});
     }
+
+    if (!result.tool_calls.empty()) {
+        if (!string_strip(result.content).empty()) {
+            LOG_WRN("Content found with tool calls: %s\n", result.content.c_str());
+        }
+        result.content = "";
+    }
     return result;
 }
 
@@ -134,11 +142,11 @@ static common_chat_msg parse_prefixed_json_tool_call_array(const std::string& in
     result.role = "assistant";
     const auto process_tool_calls = [&](const json & tool_calls) {
         for (const auto & tool_call : tool_calls) {
-            const auto & arguments = tool_call["arguments"];
+            const auto & arguments = tool_call.at("arguments");
             result.tool_calls.push_back({
-                tool_call["name"],
+                tool_call.at("name"),
                 arguments.is_string() ? arguments.get<std::string>() : arguments.dump(),
-                tool_call.contains("id") ? tool_call["id"] : "",
+                tool_call.contains("id") ? tool_call.at("id") : "",
             });
         }
     };
@@ -155,7 +163,7 @@ static common_chat_msg parse_prefixed_json_tool_call_array(const std::string& in
 
 static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
     for (const auto & tool : tools) {
-        if (!tool.contains("type") || tool["type"] != "function" || !tool.contains("function")) {
+        if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) {
             LOG_INF("Skipping tool without function: %s", tool.dump(2).c_str());
             continue;
         }
@@ -190,27 +198,27 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
 
     auto tool_call_schemas = json::array();
     foreach_function(inputs.tools, [&](const json & tool) {
-        const auto & function = tool["function"];
+        const auto & function = tool.at("function");
         auto tool_schema = json {
             {"type", "object"},
             {"properties", {
                 {"name", {
                     {"type", "string"},
-                    {"const", function["name"]},
+                    {"const", function.at("name")},
                 }},
-                {"arguments", function["parameters"]},
+                {"arguments", function.at("parameters")},
             }},
             {"required", json::array({"name", "arguments"})},
         };
         if (function.contains("description")) {
-            tool_schema["description"] = function["description"];
+            tool_schema["description"] = function.at("description");
         }
         if (inputs.parallel_tool_calls) {
-            tool_schema["properties"]["id"] = {
+            tool_schema.at("properties")["id"] = {
                 {"type", "string"},
                 {"minLength", 4},
             };
-            tool_schema["required"].push_back("id");
+            tool_schema.at("required").push_back("id");
         }
         tool_call_schemas.emplace_back(tool_schema);
     });
@@ -275,21 +283,21 @@ static common_chat_msg common_chat_parse_generic(const std::string & input) {
     common_chat_msg result;
     result.role = "assistant";
     if (data.contains("tool_calls")) {
-        for (const auto & tool_call : data["tool_calls"]) {
+        for (const auto & tool_call : data.at("tool_calls")) {
             result.tool_calls.push_back({
-                tool_call["name"],
-                tool_call["arguments"].dump(),
-                tool_call.contains("id") ? tool_call["id"] : "",
+                tool_call.at("name"),
+                tool_call.at("arguments").dump(),
+                tool_call.contains("id") ? tool_call.at("id") : "",
             });
         }
     } else if (data.contains("tool_call")) {
         result.tool_calls.push_back({
-            data["tool_call"]["name"],
-            data["tool_call"]["arguments"].dump(),
+            data.at("tool_call").at("name"),
+            data.at("tool_call").at("arguments").dump(),
             /* id= */ "",
         });
     } else if (data.contains("response")) {
-        const auto & response = data["response"];
+        const auto & response = data.at("response");
         result.content = response.is_string() ? response.get<std::string>() : response.dump(2);
     }
     return result;
@@ -301,7 +309,7 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     data.grammar = build_grammar([&](const common_grammar_builder & builder) {
         auto schemas = json::array();
         foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool["function"];
+            const auto & function = tool.at("function");
             schemas.push_back({
                 {"type", "object"},
                 {"properties", {
@@ -309,9 +317,9 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
                     // It's hard to constrain that for now (while reusing the JSON schema conversion), so we're just expecting a plain object.
                     {"name", {
                         {"type", "string"},
-                        {"const", function["name"]},
+                        {"const", function.at("name")},
                     }},
-                    {"arguments", function["parameters"]},
+                    {"arguments", function.at("parameters")},
                     {"id", {
                         {"type", "string"},
                         // Nemo's template expects a 9-character alphanumeric ID.
@@ -346,7 +354,7 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
     data.grammar = build_grammar([&](const common_grammar_builder & builder) {
         auto schemas = json::array();
         foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool["function"];
+            const auto & function = tool.at("function");
             schemas.push_back({
                 {"type", "object"},
                 {"properties", {
@@ -357,9 +365,9 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
                     }},
                     {"tool_name", {
                         {"type", "string"},
-                        {"const", function["name"]},
+                        {"const", function.at("name")},
                     }},
-                    {"parameters", function["parameters"]},
+                    {"parameters", function.at("parameters")},
                 }},
                 {"required", json::array({"tool_call_id", "tool_name", "parameters"})},
             });
@@ -382,39 +390,65 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
         "<|END_THINKING|>",
         "<|END_ACTION|>",
     };
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
-    data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
+    auto adjusted_messages = json::array();
+    for (const auto & msg : inputs.messages) {
+        auto has_reasoning_content = msg.contains("reasoning_content") && msg.at("reasoning_content").is_string();
+        auto has_tool_calls = msg.contains("tool_calls") && msg.at("tool_calls").is_array();
+        if (has_reasoning_content && has_tool_calls) {
+            auto adjusted_message = msg;
+            adjusted_message["tool_plan"] = msg.at("reasoning_content");
+            adjusted_message.erase("reasoning_content");
+            adjusted_messages.push_back(adjusted_message);
+        } else {
+            adjusted_messages.push_back(msg);
+        }
+    }
+    data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});
+    data.format = inputs.extract_reasoning ? COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING : COMMON_CHAT_FORMAT_COMMAND_R7B;
     return data;
 }
-static common_chat_msg common_chat_parse_command_r7b(const std::string & input) {
-    static std::regex response_regex("<\\|START_RESPONSE\\|>([\\s\\S\\n\\r]*?)<\\|END_RESPONSE\\|>");
-    static std::regex thought_action_regex("<\\|START_THINKING\\|>([\\s\\S\\n\\r]*?)<\\|END_THINKING\\|><\\|START_ACTION\\|>([\\s\\S\\n\\r]*?)<\\|END_ACTION\\|>");
+static common_chat_msg common_chat_parse_command_r7b(const std::string & input, bool extract_reasoning) {
+    static std::regex thought_regex("(<\\|START_THINKING\\|>([\\s\\S\\n\\r]*?)<\\|END_THINKING\\|>)([\\s\\S\\n\\r]*)");
+    static std::regex action_regex("<\\|START_ACTION\\|>([\\s\\S\\n\\r]*?)<\\|END_ACTION\\|>");
+    static std::regex response_regex("(?:<\\|START_RESPONSE\\|>)?([\\s\\S\\n\\r]*?)<\\|END_RESPONSE\\|>");
+
     std::smatch match;
 
     common_chat_msg result;
     result.role = "assistant";
-    if (std::regex_match(input, match, response_regex)) {
-        result.content = match[1].str();
-    } else if (std::regex_match(input, match, thought_action_regex)) {
-        result.tool_plan = match[1].str();
-        auto actions_str = match[2].str();
+
+    std::string rest = input;
+
+    if (std::regex_match(rest, match, thought_regex)) {
+        if (extract_reasoning) {
+            result.reasoning_content = match[2].str();
+        } else if (!match[2].str().empty()) {
+            // Let the unparsed thinking tags through in content only if their insides aren't empty.
+            result.content = match[1].str();
+        }
+        rest = match[3].str();
+    }
+    if (std::regex_match(rest, match, action_regex)) {
+        auto actions_str = match[1].str();
         auto actions = json::parse(actions_str);
         for (const auto & action : actions) {
             result.tool_calls.push_back({
-                /* .name = */      action["tool_name"],
-                /* .arguments = */ action["parameters"].dump(),
-                /* .id = */        action["tool_call_id"],
+                /* .name = */      action.at("tool_name"),
+                /* .arguments = */ action.at("parameters").dump(),
+                /* .id = */        action.at("tool_call_id"),
             });
         }
+    } else if (std::regex_match(rest, match, response_regex)) {
+        auto response = match[1].str();
+        result.content += response;
     } else {
-        LOG_ERR("Failed to parse command_r output");
-        result.content = input;
+        result.content += rest;
     }
     return result;
 }
 
 static void expect_tool_parameters(const std::string & name, const json & parameters, const std::vector<std::string> & expected_properties) {
-    if (!parameters.is_object() || !parameters.contains("type") || parameters["type"] != "object" || !parameters.contains("properties") || !parameters.contains("required")) {
+    if (!parameters.is_object() || !parameters.contains("type") || parameters.at("type") != "object" || !parameters.contains("properties") || !parameters.contains("required")) {
         throw std::runtime_error("Parameters of tool " + name + " must be an object w/ required properties");
     }
     const auto & parameters_properties = parameters.at("properties");
@@ -468,9 +502,9 @@ static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const com
         };
 
         foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool["function"];
-            std::string name = function["name"];
-            auto parameters = function["parameters"];
+            const auto & function = tool.at("function");
+            std::string name = function.at("name");
+            auto parameters = function.at("parameters");
             builder.resolve_refs(parameters);
 
             // https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/remote/tool_runtime
@@ -546,34 +580,90 @@ static common_chat_msg common_chat_parse_llama_3_1(const std::string & input, bo
 
 static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
     common_chat_params data;
-    data.grammar_lazy = inputs.tool_choice != "required";
-    data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-        std::vector<std::string> tool_rules;
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool["function"];
-            std::string name = function["name"];
-            auto parameters = function["parameters"];
-            auto args_rule = builder.add_schema(name + "-args", parameters);
-            tool_rules.push_back(builder.add_rule(name + "-call",
-                "\"<｜tool▁call▁begin｜>function<｜tool▁sep｜>" + name + "\\n```json\\n\" " + args_rule + " \"```<｜tool▁call▁end｜>\""));
-        });
-        data.grammar_triggers.push_back({"<｜tool▁calls▁begin｜>", /* .at_start = */ false});
-        data.preserved_tokens = {
-            "<｜tool▁sep｜>",
-            "<｜tool▁call▁end｜>",
-        };
-        builder.add_rule("root", "\"<｜tool▁calls▁begin｜>\" (" + string_join(tool_rules, " | ") + ")" + (inputs.parallel_tool_calls ? "*" : "") + " space");
-    }, grammar_options);
+    if (inputs.tools.is_array() && !inputs.tools.empty()) {
+        data.grammar_lazy = inputs.tool_choice != "required" && inputs.json_schema.is_null();
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            std::vector<std::string> tool_rules;
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                auto parameters = function.at("parameters");
+                auto args_rule = builder.add_schema(name + "-args", parameters);
+                tool_rules.push_back(builder.add_rule(name + "-call",
+                    "\"<｜tool▁call▁begin｜>function<｜tool▁sep｜>" + name + "\\n"
+                    "```json\\n\" " + args_rule + " \"```<｜tool▁call▁end｜>\""));
+            });
+            // Distill Qwen 7B & 32B models seem confused re/ syntax of their tool call opening tag,
+            // so we accept common variants (then it's all constrained)
+            builder.add_rule("root",
+                "( \"<｜tool▁calls▁begin｜>\" | \"<｜tool_calls_begin｜>\" | \"<｜tool calls begin｜>\" | \"<｜tool\\\\_calls\\\\_begin｜>\" ) "
+                "(" + string_join(tool_rules, " | ") + ")" + (inputs.parallel_tool_calls ? "*" : "") + " "
+                "\"<｜tool▁calls▁end｜>\""
+                " space");
+            data.grammar_triggers.push_back({"<｜tool▁calls▁begin｜>", /* .at_start = */ false});
+            data.grammar_triggers.push_back({"<｜tool_calls_begin｜>", /* .at_start = */ false});
+            data.grammar_triggers.push_back({"<｜tool calls begin｜>", /* .at_start = */ false});
+            data.grammar_triggers.push_back({"<｜tool\\_calls\\_begin｜>", /* .at_start = */ false});
+            data.preserved_tokens = {
+                "<think>",
+                "</think>",
+                "<｜tool▁sep｜>",
+                "<｜tool▁calls▁end｜",
+                "<｜tool▁call▁end｜>",
+            };
+        }, grammar_options);
+    }
     auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+
+    // Hacks to fix the official (broken) prompt.
+    // It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead,
+    // until the official template is fixed.
+    if (tmpl.source().find("{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}") != std::string::npos) {
+        // Don't leave the chat dangling after tool results
+        if (string_ends_with(prompt, "<｜tool▁outputs▁end｜>")) {
+            prompt += "<｜end▁of▁sentence｜>";
+            if (inputs.add_generation_prompt) {
+                prompt += "<｜Assistant｜>";
+            }
+        }
+        // Fix up tool call delta example added by Minja
+        prompt = std::regex_replace(
+            prompt,
+            std::regex("(<｜tool▁call▁end｜>)[\\s\\r\\n]*(<｜tool▁outputs▁begin｜>|<｜User｜>)"),
+            "$1<｜tool▁calls▁end｜><｜end▁of▁sentence｜>$2");
+    }
     data.prompt = prompt;
-    data.format = COMMON_CHAT_FORMAT_DEEPSEEK_R1;
+    data.format = inputs.extract_reasoning ? COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING : COMMON_CHAT_FORMAT_DEEPSEEK_R1;
     return data;
 }
-static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input) {
-    static std::regex trigger_regex("<｜tool▁calls▁begin｜>");
+static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input, bool extract_reasoning) {
     static std::regex function_regex("<｜tool▁call▁begin｜>function<｜tool▁sep｜>([^\n]+)\n```json\n");
-    static std::regex close_regex("```<｜tool▁call▁end｜>");
-    return parse_json_tool_calls(input, trigger_regex, function_regex, close_regex);
+    static std::regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>");
+    static std::regex reasoning_content_regex("((?:<think>)?([\\s\\S\\r\\n]*?)</think>)?([\\s\\S\\r\\n]*)");
+    static std::regex tool_calls_regex("[\\s\\r\\n]*(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>)([\\s\\S\\r\\n]*?)<｜tool▁calls▁end｜>");
+    common_chat_msg msg;
+    msg.role = "assistant";
+    std::smatch match;
+    if (std::regex_match(input, match, reasoning_content_regex)) {
+        std::string rest;
+        if (extract_reasoning) {
+            msg.reasoning_content = string_strip(match[2].str());
+        } else {
+            msg.content = match[1].str();
+        }
+        rest = match[3].str();
+
+        if (std::regex_search(rest, match, tool_calls_regex)) {
+            auto tool_calls = match[1].str();
+            auto msg2 = parse_json_tool_calls(tool_calls, std::nullopt, function_regex, close_regex);
+            msg.tool_calls = std::move(msg2.tool_calls);
+        } else {
+            msg.content += std::string(rest.begin() + rest.find_first_not_of(" \r\n"), rest.end());
+        }
+    } else {
+        msg.content = input;
+    }
+    return msg;
 }
 
 static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
@@ -583,20 +673,20 @@ static common_chat_params common_chat_params_init_firefunction_v2(const common_c
         {"datetime", "Jan 29 2025 13:00:00 GMT"},
         {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
     });
-    if (!inputs.tools.is_null() && !inputs.tools.empty()) {
+    if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != "required";
         data.grammar = build_grammar([&](const common_grammar_builder & builder) {
             auto schemas = json::array();
             foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool["function"];
+                const auto & function = tool.at("function");
                 schemas.push_back({
                     {"type", "object"},
                     {"properties", {
                         {"name", {
                             {"type", "string"},
-                            {"const", function["name"]},
+                            {"const", function.at("name")},
                         }},
-                        {"arguments", function["parameters"]},
+                        {"arguments", function.at("parameters")},
                     }},
                     {"required", json::array({"name", "arguments", "id"})},
                 });
@@ -628,15 +718,15 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
     common_chat_params data;
     data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2;
-    if (!inputs.tools.is_null() && !inputs.tools.empty()) {
+    if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != "required";
         data.grammar = build_grammar([&](const common_grammar_builder & builder) {
             std::vector<std::string> first_tool_rules;
             std::vector<std::string> subsequent_tool_rules;
             foreach_function(inputs.tools, [&](const json & tool) {
-                const auto & function = tool["function"];
-                std::string name = function["name"];
-                auto parameters = function["parameters"];
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                auto parameters = function.at("parameters");
                 auto args_rule = builder.add_schema(name + "-args", parameters);
                 first_tool_rules.push_back(builder.add_rule(name + "-call", "\"" + name + "\\n\" " + args_rule));
                 subsequent_tool_rules.push_back(builder.add_rule(name + "-call2", "\">>>" + name + "\\n\" " + args_rule));
@@ -716,9 +806,9 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
     data.grammar = build_grammar([&](const common_grammar_builder & builder) {
         std::vector<std::string> tool_rules;
         foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool["function"];
-            const auto & parameters = function["parameters"];
-            std::string name = function["name"];
+            const auto & function = tool.at("function");
+            const auto & parameters = function.at("parameters");
+            std::string name = function.at("name");
             if (name == "python" || name == "ipython") {
                 if (!parameters.contains("type")) {
                     throw std::runtime_error("Missing type in python tool");
@@ -789,9 +879,9 @@ static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat
     data.grammar = build_grammar([&](const common_grammar_builder & builder) {
         std::vector<std::string> tool_rules;
         foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool["function"];
-            std::string name = function["name"];
-            auto parameters = function["parameters"];
+            const auto & function = tool.at("function");
+            std::string name = function.at("name");
+            auto parameters = function.at("parameters");
             builder.resolve_refs(parameters);
             tool_rules.push_back(builder.add_schema(name + "-call", {
                 {"type", "object"},
@@ -839,9 +929,9 @@ static common_chat_msg common_chat_parse_hermes_2_pro(const std::string & input)
             if (!parse_json(it, end, call)) {
                 throw std::runtime_error("Failed to parse json tool call");
             }
-            const auto & arguments = call["arguments"];
+            const auto & arguments = call.at("arguments");
             result.tool_calls.push_back({
-                call["name"],
+                call.at("name"),
                 arguments.dump(),
                 // arguments.is_string() ? arguments.get<std::string>() : arguments.dump(),
                 /* id= */ "",
@@ -878,53 +968,78 @@ static common_chat_params common_chat_params_init_without_tools(const common_cha
         }
         data.grammar = json_schema_to_grammar(inputs.json_schema);
     } else {
-        data.grammar = inputs.grammar.empty();
+        data.grammar = inputs.grammar;
     }
     return data;
 }
 
 common_chat_params common_chat_params_init(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
-    auto has_tools = !inputs.tools.is_null() && inputs.tool_choice != "none";
-    LOG_DBG("[%s] has_tools=%s\n", __func__, has_tools ? "true" : "false");
+    const auto & src = tmpl.source();
+    const auto & caps = tmpl.original_caps();
 
-    if (has_tools && !inputs.grammar.empty()) {
-        throw std::runtime_error("Cannot specify grammar with tools");
+    if (inputs.tools.is_array()) {
+        if (inputs.tool_choice != "none" && !inputs.grammar.empty()) {
+            throw std::runtime_error("Cannot specify grammar with tools");
+        }
+        if (caps.supports_tool_calls && !caps.supports_tools) {
+            LOG_WRN("Template supports tool calls but does not natively describe tools. The fallback behaviour used may produce bad results, inspect prompt w/ --verbose & consider overriding the template.\n");
+        }
     }
 
-    const auto & src = tmpl.source();
+    // DeepSeek R1: use handler in all cases except json schema (thinking / tools).
+    if (src.find("<｜tool▁calls▁begin｜>") != std::string::npos && inputs.json_schema.is_null()) {
+        return common_chat_params_init_deepseek_r1(tmpl, inputs);
+    }
+
+    // Command R7B: : use handler in all cases except json schema (thinking / tools).
+    if (src.find("<|END_THINKING|><|START_ACTION|>") != std::string::npos && inputs.json_schema.is_null()) {
+        return common_chat_params_init_command_r7b(tmpl, inputs);
+    }
+
+    // Use generic handler when mixing tools + JSON schema.
+    // TODO: support that mix in handlers below.
+    if ((!inputs.tools.is_array() && inputs.json_schema.is_object())) {
+        return common_chat_params_init_generic(tmpl, inputs);
+    }
+
+    // Functionary prepends "all\n" to plain content outputs, so we use its handler in all cases.
     if (src.find(">>>all") != std::string::npos) {
-        // Functionary prepends "all\n" to plain content outputs, so we use the parser no matter when
         return common_chat_params_init_functionary_v3_2(tmpl, inputs);
     }
+
+    // Firefunction v2 requires datetime and functions in the context even w/o tools, so we also use its handler in all cases.
     if (src.find(" functools[") != std::string::npos) {
-        // Firefunction v2 requires datetime and functions in the context, even w/o tools.
         return common_chat_params_init_firefunction_v2(tmpl, inputs);
     }
 
-    if (!has_tools) {
+    // Plain handler (no tools)
+    if (inputs.tools.is_null() || inputs.tool_choice == "none") {
         return common_chat_params_init_without_tools(tmpl, inputs);
     }
 
+    // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
     if (src.find("<tool_call>") != std::string::npos) {
         return common_chat_params_init_hermes_2_pro(tmpl, inputs);
     }
+
+    // Functionary v3.1 (w/ tools)
     if (src.find("<|start_header_id|>") != std::string::npos
         && src.find("<function=") != std::string::npos) {
         return common_chat_params_init_functionary_v3_1_llama_3_1(tmpl, inputs);
     }
+
+    // Llama 3.1, 3.2, 3.3 (w/ tools)
     if (src.find("<|start_header_id|>ipython<|end_header_id|>") != std::string::npos) {
         auto allow_python_tag_builtin_tools = src.find("<|python_tag|>") != std::string::npos;
         return common_chat_params_init_llama_3_1_tool_calls(tmpl, inputs, allow_python_tag_builtin_tools);
     }
-    if (src.find("<｜tool▁calls▁begin｜>") != std::string::npos) {
-        return common_chat_params_init_deepseek_r1(tmpl, inputs);
-    }
+
+    // Mistral Nemo (w/ tools)
     if (src.find("[TOOL_CALLS]") != std::string::npos) {
         return common_chat_params_init_mistral_nemo(tmpl, inputs);
     }
-    if (src.find("<|END_THINKING|><|START_ACTION|>") != std::string::npos) {
-        return common_chat_params_init_command_r7b(tmpl, inputs);
-    }
+
+    // Generic fallback
     return common_chat_params_init_generic(tmpl, inputs);
 }
 
@@ -949,7 +1064,9 @@ common_chat_msg common_chat_parse(const std::string & input, common_chat_format
         case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS:
             return common_chat_parse_llama_3_1(input, /* with_builtin_tools= */ true);
         case COMMON_CHAT_FORMAT_DEEPSEEK_R1:
-            return common_chat_parse_deepseek_r1(input);
+            return common_chat_parse_deepseek_r1(input, /* extract_reasoning= */ false);
+        case COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING:
+            return common_chat_parse_deepseek_r1(input, /* extract_reasoning= */ true);
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2:
             return common_chat_parse_functionary_v3_2(input);
         case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1:
@@ -959,7 +1076,9 @@ common_chat_msg common_chat_parse(const std::string & input, common_chat_format
         case COMMON_CHAT_FORMAT_FIREFUNCTION_V2:
             return common_chat_parse_firefunction_v2(input);
         case COMMON_CHAT_FORMAT_COMMAND_R7B:
-            return common_chat_parse_command_r7b(input);
+            return common_chat_parse_command_r7b(input, /* extract_reasoning= */ false);
+        case COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING:
+            return common_chat_parse_command_r7b(input, /* extract_reasoning= */ true);
         default:
             throw std::runtime_error("Unsupported format: " + common_chat_format_name(format));
     }

common/chat.hpp

@@ -19,6 +19,7 @@ struct common_chat_inputs {
     bool stream;
     std::string grammar;
     bool add_generation_prompt = true;
+    bool extract_reasoning     = true;
 };
 
 enum common_chat_format {
@@ -28,11 +29,13 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_LLAMA_3_X,
     COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
     COMMON_CHAT_FORMAT_DEEPSEEK_R1,
+    COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING,
     COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
     COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
     COMMON_CHAT_FORMAT_HERMES_2_PRO,
     COMMON_CHAT_FORMAT_COMMAND_R7B,
+    COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING,
 
     COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
 };

common/common.h

@@ -140,6 +140,7 @@ struct common_params_sampling {
     int32_t dry_allowed_length = 2;     // tokens extending repetitions beyond this receive penalty
     int32_t dry_penalty_last_n = -1;    // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
     int32_t mirostat           = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   top_n_sigma        = -1.00f;// -1.0 = disabled
     float   mirostat_tau       = 5.00f; // target entropy
     float   mirostat_eta       = 0.10f; // learning rate
     bool    ignore_eos         = false;
@@ -202,6 +203,11 @@ struct common_params_vocoder {
     bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
 };
 
+enum common_reasoning_format {
+    COMMON_REASONING_FORMAT_NONE,
+    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
+};
+
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =  4096; // context size
@@ -292,6 +298,7 @@ struct common_params {
     bool   kl_divergence    = false; // compute KL divergence
 
     bool usage             = false; // print usage
+    bool completion        = false; // print source-able completion script
     bool use_color         = false; // use color to distinguish generations and inputs
     bool special           = false; // enable special token output
     bool interactive       = false; // interactive mode
@@ -346,6 +353,7 @@ struct common_params {
     std::string chat_template = "";                                                                         // NOLINT
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
 
     std::vector<std::string> api_keys;
 
@@ -623,7 +631,7 @@ struct common_chat_msg {
     std::string role;
     std::string content;
     std::vector<common_tool_call> tool_calls;
-    std::string tool_plan = "";
+    std::string reasoning_content = "";
 };
 
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid

common/sampling.cpp

@@ -134,11 +134,11 @@ std::string common_params_sampling::print() const {
     snprintf(result, sizeof(result),
             "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
             "\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n"
-            "\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n"
+            "\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, top_n_sigma = %.3f, temp = %.3f\n"
             "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
             penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
             dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
-            top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp,
+            top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, top_n_sigma, temp,
             mirostat, mirostat_eta, mirostat_tau);
 
     return std::string(result);
@@ -151,12 +151,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
 
     lparams.no_perf = params.no_perf;
 
-    std::vector<const char *> trigger_words;
-    trigger_words.reserve(params.grammar_trigger_words.size());
-    for (const auto & str : params.grammar_trigger_words) {
-        trigger_words.push_back(str.word.c_str());
-    }
-
     struct llama_sampler * grmr;
     if (params.grammar.compare(0, 11, "%llguidance") == 0) {
 #ifdef LLAMA_USE_LLGUIDANCE
@@ -165,6 +159,12 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
     } else {
+        std::vector<const char *> trigger_words;
+        trigger_words.reserve(params.grammar_trigger_words.size());
+        for (const auto & str : params.grammar_trigger_words) {
+            trigger_words.push_back(str.word.c_str());
+        }
+
         grmr = params.grammar_lazy
              ? llama_sampler_init_grammar_lazy(vocab, params.grammar.c_str(), "root",
                                                trigger_words.data(), trigger_words.size(),
@@ -188,45 +188,51 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                 params.logit_bias.data()));
 
     if (params.mirostat == 0) {
-        for (const auto & cnstr : params.samplers) {
-            switch (cnstr) {
-                case COMMON_SAMPLER_TYPE_DRY:
-                    {
-                        std::vector<const char *> c_breakers;
-                        c_breakers.reserve(params.dry_sequence_breakers.size());
-                        for (const auto & str : params.dry_sequence_breakers) {
-                            c_breakers.push_back(str.c_str());
-                        }
+        if (params.top_n_sigma >= 0) {
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k        (params.top_k));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_temp         (params.temp));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma  (params.top_n_sigma));
+        } else {
+            for (const auto & cnstr : params.samplers) {
+                switch (cnstr) {
+                    case COMMON_SAMPLER_TYPE_DRY:
+                        {
+                            std::vector<const char *> c_breakers;
+                            c_breakers.reserve(params.dry_sequence_breakers.size());
+                            for (const auto & str : params.dry_sequence_breakers) {
+                                c_breakers.push_back(str.c_str());
+                            }
 
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
-                    }
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_K:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
-                    break;
-                case COMMON_SAMPLER_TYPE_TOP_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_MIN_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_XTC:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
-                    break;
-                case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
-                    break;
-                case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                    break;
-                case COMMON_SAMPLER_TYPE_INFILL:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
-                    break;
-                case COMMON_SAMPLER_TYPE_PENALTIES:
-                    llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
-                    break;
-                default:
-                    GGML_ASSERT(false && "unknown sampler type");
+                            llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                        }
+                        break;
+                    case COMMON_SAMPLER_TYPE_TOP_K:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TOP_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_MIN_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_XTC:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                        break;
+                    case COMMON_SAMPLER_TYPE_INFILL:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (vocab));
+                        break;
+                    case COMMON_SAMPLER_TYPE_PENALTIES:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                        break;
+                    default:
+                        GGML_ASSERT(false && "unknown sampler type");
+                }
             }
         }
         llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));

convert_hf_to_gguf.py

@@ -558,7 +558,7 @@ class Model:
 
     # NOTE: this function is generated by convert_hf_to_gguf_update.py
     #       do not modify it manually!
-    # ref:  https://github.com/ggerganov/llama.cpp/pull/6920
+    # ref:  https://github.com/ggml-org/llama.cpp/pull/6920
     # Marker: Start get_vocab_base_pre
     def get_vocab_base_pre(self, tokenizer) -> str:
         # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that
@@ -708,7 +708,7 @@ class Model:
             logger.warning("**          - the model has not been added to convert_hf_to_gguf_update.py yet")
             logger.warning("**          - the pre-tokenization config has changed upstream")
             logger.warning("**          Check your model files and convert_hf_to_gguf_update.py and update them accordingly.")
-            logger.warning("** ref:     https://github.com/ggerganov/llama.cpp/pull/6920")
+            logger.warning("** ref:     https://github.com/ggml-org/llama.cpp/pull/6920")
             logger.warning("**")
             logger.warning(f"** chkhsh:  {chkhsh}")
             logger.warning("**************************************************************************************")
@@ -2835,7 +2835,7 @@ class InternLM2Model(Model):
         if chat_eos_token_id is not None:
             # For the chat model, we replace the eos with '<|im_end|>'.
             # TODO: this is a hack, should be fixed
-            #       https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048
+            #       https://github.com/ggml-org/llama.cpp/pull/6745#issuecomment-2067687048
             special_vocab.special_token_ids["eos"] = chat_eos_token_id
             logger.warning(f"Replace eos:{old_eos} with a special token:{chat_eos_token_id}"
                            " in chat mode so that the conversation can end normally.")

convert_hf_to_gguf_update.py

@@ -8,7 +8,7 @@
 # provide the necessary information to llama.cpp via the GGUF header in order to implement
 # the same pre-tokenizer.
 #
-# ref: https://github.com/ggerganov/llama.cpp/pull/6920
+# ref: https://github.com/ggml-org/llama.cpp/pull/6920
 #
 # Instructions:
 #
@@ -246,7 +246,7 @@ src_func = f"""
             logger.warning("**          - the model has not been added to convert_hf_to_gguf_update.py yet")
             logger.warning("**          - the pre-tokenization config has changed upstream")
             logger.warning("**          Check your model files and convert_hf_to_gguf_update.py and update them accordingly.")
-            logger.warning("** ref:     https://github.com/ggerganov/llama.cpp/pull/6920")
+            logger.warning("** ref:     https://github.com/ggml-org/llama.cpp/pull/6920")
             logger.warning("**")
             logger.warning(f"** chkhsh:  {{chkhsh}}")
             logger.warning("**************************************************************************************")

convert_lora_to_gguf.py

@@ -395,7 +395,7 @@ if __name__ == '__main__':
                         logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
                         if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
                             logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
-                            logger.error("Please refer to https://github.com/ggerganov/llama.cpp/pull/9948")
+                            logger.error("Please refer to https://github.com/ggml-org/llama.cpp/pull/9948")
                         sys.exit(1)
 
                     if base_name in tensor_map:
@@ -419,7 +419,7 @@ if __name__ == '__main__':
                 # some archs may have the same tensor for lm_head and output (tie word embeddings)
                 # in this case, adapters targeting lm_head will fail when using llama-export-lora
                 # therefore, we ignore them for now
-                # see: https://github.com/ggerganov/llama.cpp/issues/9065
+                # see: https://github.com/ggml-org/llama.cpp/issues/9065
                 if name == "lm_head.weight" and len(dest) == 0:
                     raise ValueError("lm_head is present in adapter, but is ignored in base model")
                 for dest_name, dest_data in dest:

docs/android.md

@@ -12,7 +12,7 @@ $ apt update && apt upgrade -y
 $ apt install git cmake
 ```
 
-Then, follow the [build instructions](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md), specifically for CMake.
+Then, follow the [build instructions](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md), specifically for CMake.
 
 Once the binaries are built, download your model of choice (e.g., from Hugging Face). It's recommended to place it in the `~/` directory for best performance:
 

docs/backend/OPENCL.md

@@ -122,7 +122,7 @@ cp libOpenCL.so ~/android-sdk/ndk/26.3.11579264/toolchains/llvm/prebuilt/linux-x
 ```sh
 cd ~/dev/llm
 
-git clone https://github.com/ggerganov/llama.cpp && \
+git clone https://github.com/ggml-org/llama.cpp && \
 cd llama.cpp && \
 mkdir build-android && cd build-android
 
@@ -182,7 +182,7 @@ cmake --build . --target install
 mkdir -p ~/dev/llm
 cd ~/dev/llm
 
-git clone https://github.com/ggerganov/llama.cpp && cd llama.cpp
+git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
 mkdir build && cd build
 
 cmake .. -G Ninja `

docs/backend/SYCL.md

@@ -36,8 +36,8 @@ The following release is verified with good quality:
 
 |Commit ID|Tag|Release|Verified  Platform| Update date|
 |-|-|-|-|-|
-|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
-|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
+|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
+|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
 
 
 ## News
@@ -58,7 +58,7 @@ The following release is verified with good quality:
 - 2024.3
   - Release binary files of Windows.
   - A blog is published: **Run LLM on all Intel GPUs Using llama.cpp**: [intel.com](https://www.intel.com/content/www/us/en/developer/articles/technical/run-llm-on-all-gpus-using-llama-cpp-artical.html) or [medium.com](https://medium.com/@jianyu_neo/run-llm-on-all-intel-gpus-using-llama-cpp-fd2e2dcbd9bd).
-  - New base line is ready: [tag b2437](https://github.com/ggerganov/llama.cpp/tree/b2437).
+  - New base line is ready: [tag b2437](https://github.com/ggml-org/llama.cpp/tree/b2437).
   - Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing.
   - Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE.
   - Support detecting all GPUs with level-zero and same top **Max compute units**.

docs/build.md

@@ -3,7 +3,7 @@
 **To get the Code:**
 
 ```bash
-git clone https://github.com/ggerganov/llama.cpp
+git clone https://github.com/ggml-org/llama.cpp
 cd llama.cpp
 ```
 
@@ -46,7 +46,7 @@ cmake --build build --config Release
   ```
 
 - Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers:
-    - Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...):
+    - Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...):
     - Tab Workload: Desktop-development with C++
     - Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang)
     - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test

docs/cuda-fedora.md

@@ -248,7 +248,7 @@ You have successfully set up CUDA on Fedora within a toolbox environment using t
 
 - **Building `llama.cpp`:**
 
-  - With CUDA installed, you can follow these [build instructions for `llama.cpp`](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) to compile it with CUDA support.
+  - With CUDA installed, you can follow these [build instructions for `llama.cpp`](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md) to compile it with CUDA support.
   - Ensure that any CUDA-specific build flags or paths are correctly set in your build configuration.
 
 - **Using the Toolbox Environment:**

docs/development/HOWTO-add-model.md

@@ -104,16 +104,16 @@ Note: to debug the inference graph: you can use [llama-eval-callback](/examples/
 
 ## GGUF specification
 
-https://github.com/ggerganov/ggml/blob/master/docs/gguf.md
+https://github.com/ggml-org/ggml/blob/master/docs/gguf.md
 
 ## Resources
 
-- YaRN RoPE scaling https://github.com/ggerganov/llama.cpp/pull/2268
-- support Baichuan serial models https://github.com/ggerganov/llama.cpp/pull/3009
-- support attention bias https://github.com/ggerganov/llama.cpp/pull/4283
-- Mixtral support https://github.com/ggerganov/llama.cpp/pull/4406
-- BERT embeddings https://github.com/ggerganov/llama.cpp/pull/5423
-- Grok-1 support https://github.com/ggerganov/llama.cpp/pull/6204
-- Command R Plus support https://github.com/ggerganov/llama.cpp/pull/6491
-- support arch DBRX https://github.com/ggerganov/llama.cpp/pull/6515
-- How to convert HuggingFace model to GGUF format https://github.com/ggerganov/llama.cpp/discussions/2948
+- YaRN RoPE scaling https://github.com/ggml-org/llama.cpp/pull/2268
+- support Baichuan serial models https://github.com/ggml-org/llama.cpp/pull/3009
+- support attention bias https://github.com/ggml-org/llama.cpp/pull/4283
+- Mixtral support https://github.com/ggml-org/llama.cpp/pull/4406
+- BERT embeddings https://github.com/ggml-org/llama.cpp/pull/5423
+- Grok-1 support https://github.com/ggml-org/llama.cpp/pull/6204
+- Command R Plus support https://github.com/ggml-org/llama.cpp/pull/6491
+- support arch DBRX https://github.com/ggml-org/llama.cpp/pull/6515
+- How to convert HuggingFace model to GGUF format https://github.com/ggml-org/llama.cpp/discussions/2948

docs/docker.md

@@ -7,21 +7,21 @@
 ## Images
 We have three Docker images available for this project:
 
-1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
-2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
+1. `ghcr.io/ggml-org/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
+2. `ghcr.io/ggml-org/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggml-org/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
 
 Additionally, there the following images, similar to the above:
 
-- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggerganov/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
 
 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).
 
@@ -32,25 +32,25 @@ The easiest way to download the models, convert them to ggml and optimize them i
 Replace `/path/to/models` below with the actual path where you downloaded the models.
 
 ```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
+docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:full --all-in-one "/models/" 7B
 ```
 
 On completion, you are ready to play!
 
 ```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
+docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
 ```
 
 or with a light image:
 
 ```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
+docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
 ```
 
 or with a server image:
 
 ```bash
-docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
+docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggml-org/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
 ```
 
 ## Docker With CUDA
@@ -69,7 +69,7 @@ You may want to pass in some different `ARGS`, depending on the CUDA environment
 
 The defaults are:
 
-- `CUDA_VERSION` set to `12.6.0`
+- `CUDA_VERSION` set to `12.4.0`
 - `CUDA_DOCKER_ARCH` set to the cmake build default, which includes all the supported architectures
 
 The resulting images, are essentially the same as the non-CUDA images:
@@ -104,7 +104,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment
 
 The defaults are:
 
-- `MUSA_VERSION` set to `rc3.1.0`
+- `MUSA_VERSION` set to `rc3.1.1`
 
 The resulting images, are essentially the same as the non-MUSA images:
 

docs/install.md

@@ -7,7 +7,7 @@ On Mac and Linux, the homebrew package manager can be used via
 ```sh
 brew install llama.cpp
 ```
-The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
+The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668
 
 ## Nix
 

docs/llguidance.md

@@ -13,13 +13,15 @@ cmake -B build -DLLAMA_LLGUIDANCE=ON
 make -C build -j
 ```
 
+For Windows use `cmake --build build --config Release` instead of `make`.
+
 This requires the Rust compiler and the `cargo` tool to be [installed](https://www.rust-lang.org/tools/install).
 
 ## Interface
 
 There are no new command-line arguments or modifications to `common_params`. When enabled, grammars starting with `%llguidance` are passed to LLGuidance instead of the [current](../grammars/README.md) llama.cpp grammars. Additionally, JSON Schema requests (e.g., using the `-j` argument in `llama-cli`) are also passed to LLGuidance.
 
-For your existing GBNF grammars, you can use [gbnf_to_lark.py script](https://github.com/guidance-ai/llguidance/blob/main/scripts/gbnf_to_lark.py) to convert them to LLGuidance Lark-like format.
+For your existing GBNF grammars, you can use [gbnf_to_lark.py script](https://github.com/guidance-ai/llguidance/blob/main/python/llguidance/gbnf_to_lark.py) to convert them to LLGuidance Lark-like format.
 
 ## Performance
 

examples/cvector-generator/README.md

@@ -3,9 +3,9 @@
 This example demonstrates how to generate a control vector using gguf models.
 
 Related PRs:
-- [Add support for control vectors](https://github.com/ggerganov/llama.cpp/pull/5970)
-- (Issue) [Generate control vector using llama.cpp](https://github.com/ggerganov/llama.cpp/issues/6880)
-- [Add cvector-generator example](https://github.com/ggerganov/llama.cpp/pull/7514)
+- [Add support for control vectors](https://github.com/ggml-org/llama.cpp/pull/5970)
+- (Issue) [Generate control vector using llama.cpp](https://github.com/ggml-org/llama.cpp/issues/6880)
+- [Add cvector-generator example](https://github.com/ggml-org/llama.cpp/pull/7514)
 
 ## Examples
 

examples/imatrix/README.md

@@ -1,7 +1,7 @@
 # llama.cpp/examples/imatrix
 
-Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantized models.
-More information is available here: https://github.com/ggerganov/llama.cpp/pull/4861
+Compute an importance matrix for a model and given text dataset. Can be used during quantization to enhance the quality of the quantized models.
+More information is available here: https://github.com/ggml-org/llama.cpp/pull/4861
 
 ## Usage
 

examples/imatrix/imatrix.cpp

@@ -100,7 +100,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
     const float * data = is_host ? (const float *) src1->data : m_src1_data.data();
 
     // this has been adapted to the new format of storing merged experts in a single 3d tensor
-    // ref: https://github.com/ggerganov/llama.cpp/pull/6387
+    // ref: https://github.com/ggml-org/llama.cpp/pull/6387
     if (t->op == GGML_OP_MUL_MAT_ID) {
         //   ids  -> [n_experts_used, n_tokens]
         //   src1 -> [cols, n_expert_used, n_tokens]

examples/llama-bench/llama-bench.cpp

@@ -876,8 +876,8 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
 struct test {
     static const std::string build_commit;
     static const int         build_number;
-    static const std::string cpu_info;
-    static const std::string gpu_info;
+    const std::string        cpu_info;
+    const std::string        gpu_info;
     std::string              model_filename;
     std::string              model_type;
     uint64_t                 model_size;
@@ -903,7 +903,10 @@ struct test {
     std::string              test_time;
     std::vector<uint64_t>    samples_ns;
 
-    test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) {
+    test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) :
+        cpu_info(get_cpu_info()),
+        gpu_info(get_gpu_info()) {
+
         model_filename = inst.model;
         char buf[128];
         llama_model_desc(lmodel, buf, sizeof(buf));
@@ -1058,8 +1061,6 @@ struct test {
 
 const std::string test::build_commit = LLAMA_COMMIT;
 const int         test::build_number = LLAMA_BUILD_NUMBER;
-const std::string test::cpu_info     = get_cpu_info();
-const std::string test::gpu_info     = get_gpu_info();
 
 struct printer {
     virtual ~printer() {}

examples/llama.android/llama/src/main/cpp/CMakeLists.txt

@@ -14,7 +14,7 @@ project("llama-android")
 #include(FetchContent)
 #FetchContent_Declare(
 #        llama
-#        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
+#        GIT_REPOSITORY https://github.com/ggml-org/llama.cpp
 #        GIT_TAG        master
 #)
 

examples/llama.swiftui/README.md

@@ -3,9 +3,9 @@
 Local inference of llama.cpp on an iPhone. This is a sample app that can be used as a starting
 point for more advanced projects.
 
-For usage instructions and performance stats, check the following discussion: https://github.com/ggerganov/llama.cpp/discussions/4508
+For usage instructions and performance stats, check the following discussion: https://github.com/ggml-org/llama.cpp/discussions/4508
 
-![image](https://github.com/ggerganov/llama.cpp/assets/1991296/2b40284f-8421-47a2-b634-74eece09a299)
+![image](https://github.com/ggml-org/llama.cpp/assets/1991296/2b40284f-8421-47a2-b634-74eece09a299)
 
 Video demonstration:
 

examples/llama.vim

@@ -39,7 +39,7 @@
 "
 "   :call llama#init()
 "
-" more info: https://github.com/ggerganov/llama.cpp/pull/9787
+" more info: https://github.com/ggml-org/llama.cpp/pull/9787
 "
 
 " colors (adjust to your liking)

examples/llava/README-minicpmo2.6.md

@@ -26,7 +26,7 @@ python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
 ```
 
 Build llama.cpp using `CMake`:
-https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md
+https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md
 
 ```bash
 cmake -B build

examples/llava/README-minicpmv2.5.md

@@ -6,7 +6,7 @@ Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-
 
 Clone llama.cpp:
 ```bash
-git clone https://github.com/ggerganov/llama.cpp
+git clone https://github.com/ggml-org/llama.cpp
 cd llama.cpp
 ```
 

examples/lookahead/README.md

@@ -4,4 +4,4 @@ Demonstration of lookahead decoding technique:
 
 https://lmsys.org/blog/2023-11-21-lookahead-decoding/
 
-More info: https://github.com/ggerganov/llama.cpp/pull/4207
+More info: https://github.com/ggml-org/llama.cpp/pull/4207

examples/lookup/README.md

@@ -8,5 +8,5 @@ The key parameters for lookup decoding are `ngram_min`, `ngram_max` and `n_draft
 
 More info:
 
-https://github.com/ggerganov/llama.cpp/pull/4484
-https://github.com/ggerganov/llama.cpp/issues/4226
+https://github.com/ggml-org/llama.cpp/pull/4484
+https://github.com/ggml-org/llama.cpp/issues/4226

examples/main/README.md

@@ -1,6 +1,6 @@
 # llama.cpp/examples/main
 
-This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggerganov/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts.
+This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggml-org/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts.
 
 ## Table of Contents
 
@@ -121,7 +121,7 @@ When --in-prefix or --in-suffix options are enabled the chat template ( --chat-t
 
 ### Chat templates
 
- `--chat-template JINJA_TEMPLATE`: This option sets a custom jinja chat template. It accepts a string, not a file name.  Default: template taken from model's metadata. Llama.cpp only supports [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template). These include llama2, llama3, gemma, monarch, chatml, orion, vicuna, vicuna-orca, deepseek, command-r, zephyr. When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled.
+ `--chat-template JINJA_TEMPLATE`: This option sets a custom jinja chat template. It accepts a string, not a file name.  Default: template taken from model's metadata. Llama.cpp only supports [some pre-defined templates](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template). These include llama2, llama3, gemma, monarch, chatml, orion, vicuna, vicuna-orca, deepseek, command-r, zephyr. When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled.
 
  Example usage: `--chat-template gemma`
 
@@ -265,6 +265,14 @@ Being experimental and unique, XTC is disabled by default. The recommended combi
 
 Example usage: `--xtc-probability 0.5 --xtc-threshold 0.1`
 
+### Top-nσ Sampling
+
+-   `--top-nsigma N`: Limit the next token selection to a subset of tokens with pre-softmax logits that are within n * σ less than the max logit (default: -1, -1 = disabled).
+
+Top-nσ sampling is a text generation method that selects tokens based on a statistical threshold in pre-softmax logits. It works by only sampling from tokens with logits that are within n * σ of the maximum logit. This method helps maintain a stable sampling space regardless of temperature scaling, allowing it to perform well on reasoning tasks even in high temperatures. Without complex probability manipulation, it efficiently filters tokens directly on the pre-softmax logits. A higher value for top-nsigma (e.g., 5) will take more noisy tokens into consideration, while a lower value (e.g., 1) will focous on the more informative region of the sampling space.
+
+Example usage: `--top-nsigma 1`
+
 ### Logit Bias
 
 -   `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion.

examples/passkey/README.md

@@ -5,8 +5,8 @@ models ability to recall information from long contexts.
 
 See the following PRs for more info:
 
-- https://github.com/ggerganov/llama.cpp/pull/3856
-- https://github.com/ggerganov/llama.cpp/pull/4810
+- https://github.com/ggml-org/llama.cpp/pull/3856
+- https://github.com/ggml-org/llama.cpp/pull/4810
 
 ### Usage
 

examples/pydantic_models_to_grammar_examples.py

@@ -23,7 +23,7 @@ def create_completion(host, prompt, gbnf_grammar):
     """Calls the /completion API on llama-server.
 
     See
-    https://github.com/ggerganov/llama.cpp/tree/HEAD/examples/server#api-endpoints
+    https://github.com/ggml-org/llama.cpp/tree/HEAD/examples/server#api-endpoints
     """
     print(f"  Request:\n    Grammar:\n{textwrap.indent(gbnf_grammar, '      ')}\n    Prompt:\n{textwrap.indent(prompt.rstrip(), '      ')}")
     headers = {"Content-Type": "application/json"}

examples/quantize/README.md

@@ -69,22 +69,22 @@ Several quantization methods are supported. They differ in the resulting model d
 |   13B | ms/tok @ 8th |      - |     73 |     82 |     98 |    105 |    128 |
 |   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
 
-- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
+- [k-quants](https://github.com/ggml-org/llama.cpp/pull/1684)
 - recent k-quants improvements and new i-quants
-  - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
-  - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
-  - [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773)
-  - [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856)
-  - [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861)
-  - [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872)
-  - [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897)
-  - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
-  - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
-  - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
-  - [#4996 - k-quants tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
-  - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
-  - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
-  - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)
+  - [#2707](https://github.com/ggml-org/llama.cpp/pull/2707)
+  - [#2807](https://github.com/ggml-org/llama.cpp/pull/2807)
+  - [#4773 - 2-bit i-quants (inference)](https://github.com/ggml-org/llama.cpp/pull/4773)
+  - [#4856 - 2-bit i-quants (inference)](https://github.com/ggml-org/llama.cpp/pull/4856)
+  - [#4861 - importance matrix](https://github.com/ggml-org/llama.cpp/pull/4861)
+  - [#4872 - MoE models](https://github.com/ggml-org/llama.cpp/pull/4872)
+  - [#4897 - 2-bit quantization](https://github.com/ggml-org/llama.cpp/pull/4897)
+  - [#4930 - imatrix for all k-quants](https://github.com/ggml-org/llama.cpp/pull/4930)
+  - [#4951 - imatrix on the GPU](https://github.com/ggml-org/llama.cpp/pull/4957)
+  - [#4969 - imatrix for legacy quants](https://github.com/ggml-org/llama.cpp/pull/4969)
+  - [#4996 - k-quants tuning](https://github.com/ggml-org/llama.cpp/pull/4996)
+  - [#5060 - Q3_K_XS](https://github.com/ggml-org/llama.cpp/pull/5060)
+  - [#5196 - 3-bit i-quants](https://github.com/ggml-org/llama.cpp/pull/5196)
+  - [quantization tuning](https://github.com/ggml-org/llama.cpp/pull/5320), [another one](https://github.com/ggml-org/llama.cpp/pull/5334), and [another one](https://github.com/ggml-org/llama.cpp/pull/5361)
 
 **Llama 2 7B**
 

examples/retrieval/README.md

@@ -3,7 +3,7 @@
 Demonstration of simple retrieval technique based on cosine similarity
 
 More info:
-https://github.com/ggerganov/llama.cpp/pull/6193
+https://github.com/ggml-org/llama.cpp/pull/6193
 
 ### How to use
 

examples/server/CMakeLists.txt

@@ -5,7 +5,7 @@ option(LLAMA_SERVER_SSL "Build SSL support for the server" OFF)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
 
 if (MINGW)
-    # fix: https://github.com/ggerganov/llama.cpp/actions/runs/9651004652/job/26617901362?pr=8006
+    # fix: https://github.com/ggml-org/llama.cpp/actions/runs/9651004652/job/26617901362?pr=8006
     add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
 endif()
 

examples/server/README.md

@@ -7,14 +7,14 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
 **Features:**
  * LLM inference of F16 and quantized models on GPU and CPU
  * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
- * Reranking endoint (WIP: https://github.com/ggerganov/llama.cpp/pull/9510)
+ * Reranking endoint (WIP: https://github.com/ggml-org/llama.cpp/pull/9510)
  * Parallel decoding with multi-user support
  * Continuous batching
  * Multimodal (wip)
  * Monitoring endpoints
  * Schema-constrained JSON response format
 
-The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216).
+The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggml-org/llama.cpp/issues/4216).
 
 ## Usage
 
@@ -65,7 +65,7 @@ The project is under active development, and we are [looking for feedback and co
 | `-np, --parallel N` | number of parallel sequences to decode (default: 1)<br/>(env: LLAMA_ARG_N_PARALLEL) |
 | `--mlock` | force system to keep model in RAM rather than swapping or compressing<br/>(env: LLAMA_ARG_MLOCK) |
 | `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock)<br/>(env: LLAMA_ARG_NO_MMAP) |
-| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggerganov/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
+| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggml-org/llama.cpp/issues/1437<br/>(env: LLAMA_ARG_NUMA) |
 | `-dev, --device <dev1,dev2,..>` | comma-separated list of devices to use for offloading (none = don't offload)<br/>use --list-devices to see a list of available devices<br/>(env: LLAMA_ARG_DEVICE) |
 | `--list-devices` | print list of available devices and exit |
 | `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
@@ -127,6 +127,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--grammar-file FNAME` | file to read grammar from |
 | `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
 | `--jinja` | Enable experimental Jinja templating engine (required for tool use) |
+| `--reasoning-format FORMAT` | Controls extraction of model thinking traces and the format / field in which they are returned (default: `deepseek`; allowed values: `deepseek`, `none`; requires `--jinja`). `none` will leave thinking traces inline in `message.content` in a model-specific format, while `deepseek` will return them separately under `message.reasoning_content` |
 
 **Example-specific params**
 
@@ -177,7 +178,7 @@ Example usage of docker compose with environment variables:
 ```yml
 services:
   llamacpp-server:
-    image: ghcr.io/ggerganov/llama.cpp:server
+    image: ghcr.io/ggml-org/llama.cpp:server
     ports:
       - 8080:8080
     volumes:
@@ -272,10 +273,10 @@ You can consume the endpoints with Postman or NodeJS with axios library. You can
 ### Docker
 
 ```bash
-docker run -p 8080:8080 -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080
+docker run -p 8080:8080 -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080
 
 # or, with CUDA:
-docker run -p 8080:8080 -v /path/to/models:/models --gpus all ghcr.io/ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99
+docker run -p 8080:8080 -v /path/to/models:/models --gpus all ghcr.io/ggml-org/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99
 ```
 
 ## Testing with CURL
@@ -1065,7 +1066,7 @@ print(completion.choices[0].text)
 
 ### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
 
-Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
+Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
 
 *Options:*
 
@@ -1119,7 +1120,7 @@ curl http://localhost:8080/v1/chat/completions \
 
 *Tool call support*
 
-[Function calling](https://platform.openai.com/docs/guides/function-calling) is supported for all models (see https://github.com/ggerganov/llama.cpp/pull/9639):
+[Function calling](https://platform.openai.com/docs/guides/function-calling) is supported for all models (see https://github.com/ggml-org/llama.cpp/pull/9639):
 
 - Requires `--jinja` flag
 - Native tool call formats supported:
@@ -1136,61 +1137,252 @@ curl http://localhost:8080/v1/chat/completions \
 
   | Template | Format |
   |----------|--------|
-  | CohereForAI-c4ai-command-r-plus-default.jinja | generic tool calls |
-  | CohereForAI-c4ai-command-r-plus-rag.jinja | generic tool calls |
-  | CohereForAI-c4ai-command-r-plus-tool_use.jinja | generic tool calls |
-  | MiniMaxAI-MiniMax-Text-01.jinja | generic tool calls |
-  | NexaAIDev-Octopus-v2.jinja | generic tool calls |
-  | NousResearch-Hermes-2-Pro-Llama-3-8B-default.jinja | generic tool calls |
-  | NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja | hermes 2 pro tool calls |
-  | NousResearch-Hermes-2-Pro-Mistral-7B-default.jinja | generic tool calls |
-  | NousResearch-Hermes-2-Pro-Mistral-7B-tool_use.jinja | hermes 2 pro tool calls |
-  | NousResearch-Hermes-3-Llama-3.1-70B-default.jinja | generic tool calls |
-  | NousResearch-Hermes-3-Llama-3.1-70B-tool_use.jinja | hermes 2 pro tool calls |
-  | OrionStarAI-Orion-14B-Chat.jinja | generic tool calls |
-  | Qwen-QwQ-32B-Preview.jinja | hermes 2 pro tool calls |
-  | Qwen-Qwen2-7B-Instruct.jinja | generic tool calls |
-  | Qwen-Qwen2-VL-7B-Instruct.jinja | generic tool calls |
-  | Qwen-Qwen2.5-7B-Instruct.jinja | hermes 2 pro tool calls |
-  | Qwen-Qwen2.5-Math-7B-Instruct.jinja | hermes 2 pro tool calls |
-  | TheBloke-FusionNet_34Bx2_MoE-AWQ.jinja | generic tool calls |
-  | abacusai-Fewshot-Metamath-OrcaVicuna-Mistral.jinja | generic tool calls |
-  | bofenghuang-vigogne-2-70b-chat.jinja | generic tool calls |
-  | databricks-dbrx-instruct.jinja | generic tool calls |
-  | deepseek-ai-DeepSeek-Coder-V2-Instruct.jinja | generic tool calls |
-  | deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja | deepseek r1 tool calls |
-  | deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja | deepseek r1 tool calls |
-  | deepseek-ai-DeepSeek-R1-Distill-Qwen-7B.jinja | deepseek r1 tool calls |
-  | deepseek-ai-DeepSeek-V2.5.jinja | deepseek r1 tool calls |
-  | deepseek-ai-deepseek-coder-33b-instruct.jinja | generic tool calls |
-  | google-gemma-2-2b-it.jinja | generic tool calls |
-  | google-gemma-7b-it.jinja | generic tool calls |
-  | indischepartij-MiniCPM-3B-OpenHermes-2.5-v2.jinja | generic tool calls |
-  | mattshumer-Reflection-Llama-3.1-70B.jinja | generic tool calls |
-  | meetkai-functionary-medium-v3.2.jinja | functionary v3.2 tool calls |
-  | meta-llama-Llama-3.1-8B-Instruct.jinja | llama 3.x tool calls (w/ builtin tools) |
-  | meta-llama-Llama-3.2-3B-Instruct.jinja | llama 3.x tool calls |
-  | meta-llama-Llama-3.3-70B-Instruct.jinja | llama 3.x tool calls (w/ builtin tools) |
-  | meta-llama-Meta-Llama-3.1-8B-Instruct.jinja | llama 3.x tool calls (w/ builtin tools) |
-  | microsoft-Phi-3-medium-4k-instruct.jinja | generic tool calls |
-  | microsoft-Phi-3-mini-4k-instruct.jinja | generic tool calls |
-  | microsoft-Phi-3-small-8k-instruct.jinja | generic tool calls |
-  | microsoft-Phi-3.5-mini-instruct.jinja | generic tool calls |
-  | microsoft-Phi-3.5-vision-instruct.jinja | generic tool calls |
-  | mistralai-Mistral-7B-Instruct-v0.2.jinja | generic tool calls |
-  | mistralai-Mistral-Large-Instruct-2407.jinja | mistral nemo tool calls |
-  | mistralai-Mistral-Large-Instruct-2411.jinja | generic tool calls |
-  | mistralai-Mistral-Nemo-Instruct-2407.jinja | mistral nemo tool calls |
-  | mistralai-Mixtral-8x7B-Instruct-v0.1.jinja | generic tool calls |
-  | mlabonne-AlphaMonarch-7B.jinja | generic tool calls |
-  | nvidia-Llama-3.1-Nemotron-70B-Instruct-HF.jinja | llama 3.x tool calls (w/ builtin tools) |
-  | openchat-openchat-3.5-0106.jinja | generic tool calls |
-  | teknium-OpenHermes-2.5-Mistral-7B.jinja | generic tool calls |
+  | Almawave-Velvet-14B.jinja | Hermes 2 Pro |
+  | AtlaAI-Selene-1-Mini-Llama-3.1-8B.jinja | Llama 3.x |
+  | CohereForAI-aya-expanse-8b.jinja | Generic |
+  | CohereForAI-c4ai-command-r-plus-default.jinja | Generic |
+  | CohereForAI-c4ai-command-r-plus-rag.jinja | Generic |
+  | CohereForAI-c4ai-command-r-plus-tool_use.jinja | Generic |
+  | CohereForAI-c4ai-command-r7b-12-2024-default.jinja | Command R7B (extract reasoning) |
+  | CohereForAI-c4ai-command-r7b-12-2024-rag.jinja | Command R7B (extract reasoning) |
+  | CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja | Command R7B (extract reasoning) |
+  | CohereForAI-c4ai-command-r7b-12-2024.jinja | Generic |
+  | DavieLion-Llama-3.2-1B-SPIN-iter3.jinja | Generic |
+  | Delta-Vector-Rei-12B.jinja | Mistral Nemo |
+  | EpistemeAI-Mistral-Nemo-Instruct-12B-Philosophy-Math.jinja | Mistral Nemo |
+  | FlofloB-83k_continued_pretraining_Qwen2.5-0.5B-Instruct_Unsloth_merged_16bit.jinja | Hermes 2 Pro |
+  | FlofloB-test_continued_pretraining_Phi-3-mini-4k-instruct_Unsloth_merged_16bit.jinja | Generic |
+  | HelpingAI-HAI-SER.jinja | Generic |
+  | HuggingFaceTB-SmolLM2-1.7B-Instruct.jinja | Generic |
+  | HuggingFaceTB-SmolLM2-135M-Instruct.jinja | Generic |
+  | HuggingFaceTB-SmolLM2-360M-Instruct.jinja | Generic |
+  | INSAIT-Institute-BgGPT-Gemma-2-27B-IT-v1.0.jinja | Generic |
+  | Ihor-Text2Graph-R1-Qwen2.5-0.5b.jinja | Hermes 2 Pro |
+  | Infinigence-Megrez-3B-Instruct.jinja | Generic |
+  | Josephgflowers-TinyLlama_v1.1_math_code-world-test-1.jinja | Generic |
+  | LGAI-EXAONE-EXAONE-3.5-2.4B-Instruct.jinja | Generic |
+  | LGAI-EXAONE-EXAONE-3.5-7.8B-Instruct.jinja | Generic |
+  | LatitudeGames-Wayfarer-12B.jinja | Generic |
+  | Magpie-Align-Llama-3-8B-Magpie-Align-v0.1.jinja | Generic |
+  | Magpie-Align-Llama-3.1-8B-Magpie-Align-v0.1.jinja | Generic |
+  | MaziyarPanahi-calme-3.2-instruct-78b.jinja | Generic |
+  | MiniMaxAI-MiniMax-Text-01.jinja | Generic |
+  | MiniMaxAI-MiniMax-VL-01.jinja | Generic |
+  | NaniDAO-deepseek-r1-qwen-2.5-32B-ablated.jinja | DeepSeek R1 (extract reasoning) |
+  | NexaAIDev-Octopus-v2.jinja | Generic |
+  | NousResearch-Hermes-2-Pro-Llama-3-8B-default.jinja | Generic |
+  | NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja | Hermes 2 Pro |
+  | NousResearch-Hermes-2-Pro-Mistral-7B-default.jinja | Generic |
+  | NousResearch-Hermes-2-Pro-Mistral-7B-tool_use.jinja | Hermes 2 Pro |
+  | NousResearch-Hermes-3-Llama-3.1-70B-default.jinja | Generic |
+  | NousResearch-Hermes-3-Llama-3.1-70B-tool_use.jinja | Hermes 2 Pro |
+  | NovaSky-AI-Sky-T1-32B-Flash.jinja | Hermes 2 Pro |
+  | NovaSky-AI-Sky-T1-32B-Preview.jinja | Hermes 2 Pro |
+  | OnlyCheeini-greesychat-turbo.jinja | Generic |
+  | Orenguteng-Llama-3.1-8B-Lexi-Uncensored-V2.jinja | Llama 3.x |
+  | OrionStarAI-Orion-14B-Chat.jinja | Generic |
+  | PowerInfer-SmallThinker-3B-Preview.jinja | Generic |
+  | PrimeIntellect-INTELLECT-1-Instruct.jinja | Generic |
+  | Qwen-QVQ-72B-Preview.jinja | Generic |
+  | Qwen-QwQ-32B-Preview.jinja | Hermes 2 Pro |
+  | Qwen-Qwen1.5-7B-Chat.jinja | Generic |
+  | Qwen-Qwen2-7B-Instruct.jinja | Generic |
+  | Qwen-Qwen2-VL-72B-Instruct.jinja | Generic |
+  | Qwen-Qwen2-VL-7B-Instruct.jinja | Generic |
+  | Qwen-Qwen2.5-0.5B.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-1.5B-Instruct.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-14B-Instruct-1M.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-14B.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-32B-Instruct.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-32B.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-3B-Instruct.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-72B-Instruct.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-7B-Instruct-1M.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-7B-Instruct.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-7B.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-Coder-32B-Instruct.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-Coder-7B-Instruct.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-Math-1.5B.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-Math-7B-Instruct.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-VL-3B-Instruct.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-VL-72B-Instruct.jinja | Hermes 2 Pro |
+  | Qwen-Qwen2.5-VL-7B-Instruct.jinja | Hermes 2 Pro |
+  | RWKV-Red-Team-ARWKV-7B-Preview-0.1.jinja | Hermes 2 Pro |
+  | SakanaAI-TinySwallow-1.5B-Instruct.jinja | Hermes 2 Pro |
+  | SakanaAI-TinySwallow-1.5B.jinja | Hermes 2 Pro |
+  | Sao10K-70B-L3.3-Cirrus-x1.jinja | Llama 3.x |
+  | SentientAGI-Dobby-Mini-Leashed-Llama-3.1-8B.jinja | Llama 3.x |
+  | SentientAGI-Dobby-Mini-Unhinged-Llama-3.1-8B.jinja | Llama 3.x |
+  | Steelskull-L3.3-Damascus-R1.jinja | Llama 3.x |
+  | Steelskull-L3.3-MS-Nevoria-70b.jinja | Llama 3.x |
+  | Steelskull-L3.3-Nevoria-R1-70b.jinja | Llama 3.x |
+  | THUDM-glm-4-9b-chat.jinja | Generic |
+  | THUDM-glm-edge-1.5b-chat.jinja | Generic |
+  | Tarek07-Progenitor-V1.1-LLaMa-70B.jinja | Llama 3.x |
+  | TheBloke-FusionNet_34Bx2_MoE-AWQ.jinja | Generic |
+  | TinyLlama-TinyLlama-1.1B-Chat-v1.0.jinja | Generic |
+  | UCLA-AGI-Mistral7B-PairRM-SPPO-Iter3.jinja | Generic |
+  | ValiantLabs-Llama3.1-8B-Enigma.jinja | Llama 3.x |
+  | abacusai-Fewshot-Metamath-OrcaVicuna-Mistral.jinja | Generic |
+  | ai21labs-AI21-Jamba-1.5-Large.jinja | Generic |
+  | allenai-Llama-3.1-Tulu-3-405B-SFT.jinja | Generic |
+  | allenai-Llama-3.1-Tulu-3-405B.jinja | Generic |
+  | allenai-Llama-3.1-Tulu-3-8B.jinja | Generic |
+  | arcee-ai-Virtuoso-Lite.jinja | Hermes 2 Pro |
+  | arcee-ai-Virtuoso-Medium-v2.jinja | Hermes 2 Pro |
+  | arcee-ai-Virtuoso-Small-v2.jinja | Hermes 2 Pro |
+  | avemio-GRAG-NEMO-12B-ORPO-HESSIAN-AI.jinja | Generic |
+  | bespokelabs-Bespoke-Stratos-7B.jinja | Hermes 2 Pro |
+  | bfuzzy1-acheron-m1a-llama.jinja | Generic |
+  | bofenghuang-vigogne-2-70b-chat.jinja | Generic |
+  | bytedance-research-UI-TARS-72B-DPO.jinja | Generic |
+  | bytedance-research-UI-TARS-7B-DPO.jinja | Generic |
+  | bytedance-research-UI-TARS-7B-SFT.jinja | Generic |
+  | carsenk-phi3.5_mini_exp_825_uncensored.jinja | Generic |
+  | cyberagent-DeepSeek-R1-Distill-Qwen-14B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
+  | cyberagent-DeepSeek-R1-Distill-Qwen-32B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
+  | databricks-dbrx-instruct.jinja | Generic |
+  | deepseek-ai-DeepSeek-Coder-V2-Instruct.jinja | Generic |
+  | deepseek-ai-DeepSeek-Coder-V2-Lite-Base.jinja | Generic |
+  | deepseek-ai-DeepSeek-Coder-V2-Lite-Instruct.jinja | Generic |
+  | deepseek-ai-DeepSeek-R1-Distill-Llama-70B.jinja | DeepSeek R1 (extract reasoning) |
+  | deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja | DeepSeek R1 (extract reasoning) |
+  | deepseek-ai-DeepSeek-R1-Distill-Qwen-1.5B.jinja | DeepSeek R1 (extract reasoning) |
+  | deepseek-ai-DeepSeek-R1-Distill-Qwen-14B.jinja | DeepSeek R1 (extract reasoning) |
+  | deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja | DeepSeek R1 (extract reasoning) |
+  | deepseek-ai-DeepSeek-R1-Distill-Qwen-7B.jinja | DeepSeek R1 (extract reasoning) |
+  | deepseek-ai-DeepSeek-R1-Zero.jinja | DeepSeek R1 (extract reasoning) |
+  | deepseek-ai-DeepSeek-R1.jinja | DeepSeek R1 (extract reasoning) |
+  | deepseek-ai-DeepSeek-V2-Lite.jinja | Generic |
+  | deepseek-ai-DeepSeek-V2.5.jinja | DeepSeek R1 (extract reasoning) |
+  | deepseek-ai-DeepSeek-V3.jinja | DeepSeek R1 (extract reasoning) |
+  | deepseek-ai-deepseek-coder-33b-instruct.jinja | Generic |
+  | deepseek-ai-deepseek-coder-6.7b-instruct.jinja | Generic |
+  | deepseek-ai-deepseek-coder-7b-instruct-v1.5.jinja | Generic |
+  | deepseek-ai-deepseek-llm-67b-chat.jinja | Generic |
+  | deepseek-ai-deepseek-llm-7b-chat.jinja | Generic |
+  | dicta-il-dictalm2.0-instruct.jinja | Generic |
+  | ehristoforu-Falcon3-8B-Franken-Basestruct.jinja | Hermes 2 Pro |
+  | fireworks-ai-llama-3-firefunction-v2.jinja | FireFunction v2 |
+  | godlikehhd-alpaca_data_sampled_ifd_new_5200.jinja | Hermes 2 Pro |
+  | godlikehhd-alpaca_data_score_max_0.7_2600.jinja | Hermes 2 Pro |
+  | google-gemma-2-27b-it.jinja | Generic |
+  | google-gemma-2-2b-it.jinja | Generic |
+  | google-gemma-2-2b-jpn-it.jinja | Generic |
+  | google-gemma-7b-it.jinja | Generic |
+  | huihui-ai-DeepSeek-R1-Distill-Llama-70B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
+  | huihui-ai-DeepSeek-R1-Distill-Llama-8B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
+  | huihui-ai-DeepSeek-R1-Distill-Qwen-14B-abliterated-v2.jinja | DeepSeek R1 (extract reasoning) |
+  | huihui-ai-DeepSeek-R1-Distill-Qwen-32B-abliterated.jinja | DeepSeek R1 (extract reasoning) |
+  | huihui-ai-DeepSeek-R1-Distill-Qwen-7B-abliterated-v2.jinja | DeepSeek R1 (extract reasoning) |
+  | huihui-ai-Qwen2.5-14B-Instruct-1M-abliterated.jinja | Hermes 2 Pro |
+  | ibm-granite-granite-3.1-8b-instruct.jinja | Generic |
+  | indischepartij-MiniCPM-3B-OpenHermes-2.5-v2.jinja | Generic |
+  | inflatebot-MN-12B-Mag-Mell-R1.jinja | Generic |
+  | jinaai-ReaderLM-v2.jinja | Generic |
+  | kms7530-chemeng_qwen-math-7b_24_1_100_1_nonmath.jinja | Hermes 2 Pro |
+  | knifeayumu-Cydonia-v1.3-Magnum-v4-22B.jinja | Mistral Nemo |
+  | langgptai-qwen1.5-7b-chat-sa-v0.1.jinja | Generic |
+  | lightblue-DeepSeek-R1-Distill-Qwen-7B-Japanese.jinja | DeepSeek R1 (extract reasoning) |
+  | mattshumer-Reflection-Llama-3.1-70B.jinja | Generic |
+  | meetkai-functionary-medium-v3.1.jinja | Functionary v3.1 Llama 3.1 |
+  | meetkai-functionary-medium-v3.2.jinja | Functionary v3.2 |
+  | meta-llama-Llama-2-7b-chat-hf.jinja | Generic |
+  | meta-llama-Llama-3.1-8B-Instruct.jinja | Llama 3.x |
+  | meta-llama-Llama-3.2-11B-Vision-Instruct.jinja | Llama 3.x |
+  | meta-llama-Llama-3.2-1B-Instruct.jinja | Llama 3.x |
+  | meta-llama-Llama-3.2-3B-Instruct.jinja | Llama 3.x |
+  | meta-llama-Llama-3.3-70B-Instruct.jinja | Llama 3.x |
+  | meta-llama-Meta-Llama-3-8B-Instruct.jinja | Generic |
+  | meta-llama-Meta-Llama-3.1-8B-Instruct.jinja | Llama 3.x |
+  | microsoft-Phi-3-medium-4k-instruct.jinja | Generic |
+  | microsoft-Phi-3-mini-4k-instruct.jinja | Generic |
+  | microsoft-Phi-3-small-8k-instruct.jinja | Generic |
+  | microsoft-Phi-3.5-mini-instruct.jinja | Generic |
+  | microsoft-Phi-3.5-vision-instruct.jinja | Generic |
+  | microsoft-phi-4.jinja | Generic |
+  | migtissera-Tess-3-Mistral-Nemo-12B.jinja | Generic |
+  | ministral-Ministral-3b-instruct.jinja | Generic |
+  | mistralai-Codestral-22B-v0.1.jinja | Generic |
+  | mistralai-Mistral-7B-Instruct-v0.1.jinja | Generic |
+  | mistralai-Mistral-7B-Instruct-v0.2.jinja | Generic |
+  | mistralai-Mistral-7B-Instruct-v0.3.jinja | Mistral Nemo |
+  | mistralai-Mistral-Large-Instruct-2407.jinja | Mistral Nemo |
+  | mistralai-Mistral-Large-Instruct-2411.jinja | Generic |
+  | mistralai-Mistral-Nemo-Instruct-2407.jinja | Mistral Nemo |
+  | mistralai-Mistral-Small-24B-Instruct-2501.jinja | Generic |
+  | mistralai-Mixtral-8x7B-Instruct-v0.1.jinja | Generic |
+  | mkurman-Qwen2.5-14B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
+  | mlabonne-AlphaMonarch-7B.jinja | Generic |
+  | mlx-community-Josiefied-Qwen2.5-0.5B-Instruct-abliterated-v1-float32.jinja | Hermes 2 Pro |
+  | mlx-community-Qwen2.5-VL-7B-Instruct-8bit.jinja | Hermes 2 Pro |
+  | mobiuslabsgmbh-DeepSeek-R1-ReDistill-Qwen-1.5B-v1.1.jinja | DeepSeek R1 (extract reasoning) |
+  | netcat420-MFANNv0.20.jinja | Generic |
+  | netcat420-MFANNv0.24.jinja | Generic |
+  | netease-youdao-Confucius-o1-14B.jinja | Hermes 2 Pro |
+  | nvidia-AceMath-7B-RM.jinja | Hermes 2 Pro |
+  | nvidia-Eagle2-1B.jinja | Hermes 2 Pro |
+  | nvidia-Eagle2-9B.jinja | Hermes 2 Pro |
+  | nvidia-Llama-3.1-Nemotron-70B-Instruct-HF.jinja | Llama 3.x |
+  | onnx-community-DeepSeek-R1-Distill-Qwen-1.5B-ONNX.jinja | DeepSeek R1 (extract reasoning) |
+  | open-thoughts-OpenThinker-7B.jinja | Hermes 2 Pro |
+  | openchat-openchat-3.5-0106.jinja | Generic |
+  | pankajmathur-orca_mini_v6_8b.jinja | Generic |
+  | princeton-nlp-Mistral-7B-Base-SFT-RDPO.jinja | Generic |
+  | princeton-nlp-Mistral-7B-Instruct-DPO.jinja | Generic |
+  | princeton-nlp-Mistral-7B-Instruct-RDPO.jinja | Generic |
+  | prithivMLmods-Bellatrix-Tiny-1.5B-R1.jinja | Hermes 2 Pro |
+  | prithivMLmods-Bellatrix-Tiny-1B-R1.jinja | Llama 3.x |
+  | prithivMLmods-Bellatrix-Tiny-1B-v3.jinja | Generic |
+  | prithivMLmods-Bellatrix-Tiny-3B-R1.jinja | Llama 3.x |
+  | prithivMLmods-Blaze-14B-xElite.jinja | Generic |
+  | prithivMLmods-Calcium-Opus-14B-Elite2-R1.jinja | Hermes 2 Pro |
+  | prithivMLmods-Calme-Ties-78B.jinja | Generic |
+  | prithivMLmods-Calme-Ties2-78B.jinja | Generic |
+  | prithivMLmods-Calme-Ties3-78B.jinja | Generic |
+  | prithivMLmods-ChemQwen2-vL.jinja | Generic |
+  | prithivMLmods-GWQ2b.jinja | Generic |
+  | prithivMLmods-LatexMind-2B-Codec.jinja | Generic |
+  | prithivMLmods-Llama-3.2-6B-AlgoCode.jinja | Llama 3.x |
+  | prithivMLmods-Megatron-Opus-14B-Exp.jinja | Hermes 2 Pro |
+  | prithivMLmods-Megatron-Opus-14B-Stock.jinja | Hermes 2 Pro |
+  | prithivMLmods-Megatron-Opus-7B-Exp.jinja | Hermes 2 Pro |
+  | prithivMLmods-Omni-Reasoner-Merged.jinja | Hermes 2 Pro |
+  | prithivMLmods-Omni-Reasoner4-Merged.jinja | Hermes 2 Pro |
+  | prithivMLmods-Primal-Opus-14B-Optimus-v1.jinja | Hermes 2 Pro |
+  | prithivMLmods-QwQ-Math-IO-500M.jinja | Hermes 2 Pro |
+  | prithivMLmods-Qwen-7B-Distill-Reasoner.jinja | DeepSeek R1 (extract reasoning) |
+  | prithivMLmods-Qwen2.5-1.5B-DeepSeek-R1-Instruct.jinja | Hermes 2 Pro |
+  | prithivMLmods-Qwen2.5-14B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
+  | prithivMLmods-Qwen2.5-32B-DeepSeek-R1-Instruct.jinja | Hermes 2 Pro |
+  | prithivMLmods-Qwen2.5-7B-DeepSeek-R1-1M.jinja | Hermes 2 Pro |
+  | prithivMLmods-Triangulum-v2-10B.jinja | Hermes 2 Pro |
+  | qingy2024-Falcon3-2x10B-MoE-Instruct.jinja | Hermes 2 Pro |
+  | rubenroy-Zurich-14B-GCv2-5m.jinja | Hermes 2 Pro |
+  | rubenroy-Zurich-7B-GCv2-5m.jinja | Hermes 2 Pro |
+  | silma-ai-SILMA-Kashif-2B-Instruct-v1.0.jinja | Generic |
+  | simplescaling-s1-32B.jinja | Hermes 2 Pro |
+  | sometimesanotion-Lamarck-14B-v0.7.jinja | Hermes 2 Pro |
+  | sonthenguyen-zephyr-sft-bnb-4bit-DPO-mtbr-180steps.jinja | Generic |
+  | sthenno-tempesthenno-icy-0130.jinja | Generic |
+  | sumink-qwft.jinja | Hermes 2 Pro |
+  | teknium-OpenHermes-2.5-Mistral-7B.jinja | Generic |
+  | thirdeyeai-elevate360m.jinja | Generic |
+  | tiiuae-Falcon3-10B-Instruct.jinja | Hermes 2 Pro |
+  | unsloth-DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit.jinja | DeepSeek R1 (extract reasoning) |
+  | unsloth-DeepSeek-R1-Distill-Llama-8B.jinja | DeepSeek R1 (extract reasoning) |
+  | unsloth-DeepSeek-R1.jinja | DeepSeek R1 (extract reasoning) |
+  | unsloth-Mistral-Small-24B-Instruct-2501-unsloth-bnb-4bit.jinja | Generic |
+  | upstage-solar-pro-preview-instruct.jinja | Generic |
+  | whyhow-ai-PatientSeek.jinja | Generic |
+  | xwen-team-Xwen-72B-Chat.jinja | Hermes 2 Pro |
+  | xwen-team-Xwen-7B-Chat.jinja | Hermes 2 Pro |
 
   This table can be generated with:
 
   ```bash
   ./build/bin/test-chat ../minja/build/tests/*.jinja 2>/dev/null
+  ```
 
   </details>
 
@@ -1202,11 +1394,20 @@ curl http://localhost:8080/v1/chat/completions \
 
   ```shell
   # Native support:
+
   llama-server --jinja -fa -hf bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M
   llama-server --jinja -fa -hf bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q6_K_L
   llama-server --jinja -fa -hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M
   llama-server --jinja -fa -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M
 
+  # Native support for DeepSeek R1 works best w/ our own template (official template buggy)
+
+  llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q6_K_L \
+    --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
+
+  llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M \
+    --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja
+
   # Native support requires the right template for these GGUFs:
 
   llama-server --jinja -fa -hf bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M \
@@ -1236,17 +1437,17 @@ curl http://localhost:8080/v1/chat/completions \
       {
         "type":"function",
         "function":{
-          "name":"get_current_weather",
-          "description":"Get the current weather in a given location",
+          "name":"python",
+          "description":"Runs code in an ipython interpreter and returns the result of the execution after 60 seconds.",
           "parameters":{
             "type":"object",
             "properties":{
-              "location":{
+              "code":{
                 "type":"string",
-                "description":"The city and state, e.g. San Francisco, CA"
+                "description":"The code to run in the ipython interpreter."
               }
             },
-            "required":["location"]
+            "required":["code"]
           }
         }
       }
@@ -1254,7 +1455,7 @@ curl http://localhost:8080/v1/chat/completions \
     "messages": [
       {
         "role": "user",
-        "content": "What is the weather like in Istanbul?."
+        "content": "Print a hello world message with python."
       }
     ]
   }'
@@ -1398,7 +1599,7 @@ Apart from error types supported by OAI, we also have custom types that are spec
 
 ### Legacy completion web UI
 
-A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggerganov/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy`
+A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggml-org/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy`
 
 For example:
 

examples/server/server.cpp

@@ -42,7 +42,7 @@ enum stop_type {
     STOP_TYPE_LIMIT,
 };
 
-// state diagram: https://github.com/ggerganov/llama.cpp/pull/9283
+// state diagram: https://github.com/ggml-org/llama.cpp/pull/9283
 enum slot_state {
     SLOT_STATE_IDLE,
     SLOT_STATE_STARTED, // TODO: this state is only used for setting up the initial prompt processing; maybe merge it with launch_slot_with_task in the future
@@ -173,6 +173,7 @@ struct slot_params {
             {"grammar_trigger_words",     grammar_trigger_words},
             {"grammar_trigger_tokens",    sampling.grammar_trigger_tokens},
             {"preserved_tokens",          sampling.preserved_tokens},
+            {"chat_format",               common_chat_format_name(oaicompat_chat_format)},
             {"samplers",                  samplers},
             {"speculative.n_max",         speculative.n_max},
             {"speculative.n_min",         speculative.n_min},
@@ -724,9 +725,19 @@ struct server_task_result_cmpl_final : server_task_result {
             msg.content = content;
         }
 
-        json tool_calls;
+        json message {
+            {"role", "assistant"},
+        };
+        if (!msg.reasoning_content.empty()) {
+            message["reasoning_content"] = msg.reasoning_content;
+        }
+        if (msg.content.empty() && !msg.tool_calls.empty()) {
+            message["content"] = json();
+        } else {
+            message["content"] = msg.content;
+        }
         if (!msg.tool_calls.empty()) {
-            tool_calls = json::array();
+            auto tool_calls = json::array();
             for (const auto & tc : msg.tool_calls) {
                 tool_calls.push_back({
                     {"type", "function"},
@@ -737,15 +748,7 @@ struct server_task_result_cmpl_final : server_task_result {
                     {"id", tc.id},
                 });
             }
-        }
-
-        json message {
-            {"content", msg.content},
-            {"tool_calls", tool_calls},
-            {"role", "assistant"},
-        };
-        if (!msg.tool_plan.empty()) {
-            message["tool_plan"] = msg.tool_plan;
+            message["tool_calls"] = tool_calls;
         }
 
         json choice {
@@ -2073,8 +2076,8 @@ struct server_context {
 
         if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
             // Might be better to reject the request with a 400 ?
+            SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.params.n_predict, slot.n_predict);
             slot.params.n_predict = slot.n_predict;
-            SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.n_predict, slot.n_predict);
         }
 
         if (slot.params.ignore_eos && has_eos_token) {
@@ -3653,7 +3656,7 @@ int main(int argc, char ** argv) {
             }, {
                     {"name",  "n_busy_slots_per_decode"},
                     {"help",  "Average number of busy slots per llama_decode() call"},
-                    {"value",  (float) res_metrics->n_busy_slots_total / (float) res_metrics->n_decode_total}
+                    {"value",  (float) res_metrics->n_busy_slots_total / std::max((float) res_metrics->n_decode_total, 1.f)}
             }}},
             {"gauge", {{
                     {"name",  "prompt_tokens_seconds"},
@@ -4060,7 +4063,7 @@ int main(int argc, char ** argv) {
         }
 
         auto body = json::parse(req.body);
-        json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates);
+        json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates);
 
         return handle_completions_impl(
             SERVER_TASK_TYPE_COMPLETION,
@@ -4073,7 +4076,7 @@ int main(int argc, char ** argv) {
     // same with handle_chat_completions, but without inference part
     const auto handle_apply_template = [&ctx_server, &params, &res_ok](const httplib::Request & req, httplib::Response & res) {
         auto body = json::parse(req.body);
-        json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates);
+        json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates);
         res_ok(res, {{ "prompt", std::move(data.at("prompt")) }});
     };
 

examples/server/tests/unit/test_tool_call.py

@@ -92,6 +92,7 @@ def do_test_completion_with_required_tool_tiny(template_name: str, tool: dict, a
     tool_calls = choice["message"].get("tool_calls")
     assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
     tool_call = tool_calls[0]
+    assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
     expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"]
     assert expected_function_name == tool_call["function"]["name"]
     actual_arguments = tool_call["function"]["arguments"]
@@ -155,11 +156,11 @@ def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict,
 
     (TEST_TOOL,    "success",  "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
     (PYTHON_TOOL,  "code",     "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
-    (PYTHON_TOOL,  "code",     "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"),
+    # (PYTHON_TOOL,  "code",     "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"),
 
     (TEST_TOOL,    "success",  "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",   ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
     (PYTHON_TOOL,  "code",     "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",   ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
-    (PYTHON_TOOL,  "code",     "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",   "chatml"),
+    # (PYTHON_TOOL,  "code",     "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",   "chatml"),
 
     (TEST_TOOL,    "success",  "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
     (PYTHON_TOOL,  "code",     "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None),
@@ -175,7 +176,7 @@ def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict,
 
     (TEST_TOOL,    "success",  "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      ("meta-llama/Llama-3.2-3B-Instruct", None)),
     (PYTHON_TOOL,  "code",     "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      ("meta-llama/Llama-3.2-3B-Instruct", None)),
-    (PYTHON_TOOL,  "code",     "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      "chatml"),
+    # (PYTHON_TOOL,  "code",     "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      "chatml"),
     # TODO: fix these
     # (TEST_TOOL,    "success",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
     # (PYTHON_TOOL,  "code",     "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
@@ -214,6 +215,7 @@ def test_completion_with_required_tool_real_model(tool: dict, argument_key: str
     tool_calls = choice["message"].get("tool_calls")
     assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
     tool_call = tool_calls[0]
+    assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
     expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"]
     assert expected_function_name == tool_call["function"]["name"]
     actual_arguments = tool_call["function"]["arguments"]
@@ -273,7 +275,6 @@ def test_completion_without_tool_call_slow(template_name: str, n_predict: int, t
 
 @pytest.mark.slow
 @pytest.mark.parametrize("hf_repo,template_override", [
-    ("bartowski/c4ai-command-r7b-12-2024-GGUF:Q4_K_M",   ("CohereForAI/c4ai-command-r7b-12-2024", "tool_use")),
     ("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
     ("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
 
@@ -298,13 +299,16 @@ def test_completion_without_tool_call_slow(template_name: str, n_predict: int, t
     ("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M",      ("meta-llama/Llama-3.2-3B-Instruct", None)),
     ("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M",      "chatml"),
 
+    ("bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L",   ("CohereForAI/c4ai-command-r7b-12-2024", "tool_use")),
+
+    ("bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+
     # Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it.
     ("bartowski/gemma-2-2b-it-GGUF:Q4_K_M",              None),
 
     # ("bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)),
-    # ("bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
 ])
-def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None):
+def test_weather(hf_repo: str, template_override: str | Tuple[str, str | None] | None):
     global server
     n_predict = 512
     server.n_slots = 1
@@ -323,6 +327,7 @@ def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None)
     res = server.make_request("POST", "/chat/completions", data={
         "max_tokens": n_predict,
         "messages": [
+            {"role": "system", "content": "You are a chatbot that uses tools/functions. Dont overthink things."},
             {"role": "user", "content": "What is the weather in Istanbul?"},
         ],
         "tools": [WEATHER_TOOL],
@@ -332,6 +337,7 @@ def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None)
     tool_calls = choice["message"].get("tool_calls")
     assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
     tool_call = tool_calls[0]
+    assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
     assert tool_call["function"]["name"] == WEATHER_TOOL["function"]["name"]
     actual_arguments = json.loads(tool_call["function"]["arguments"])
     assert 'location' in actual_arguments, f"location not found in {json.dumps(actual_arguments)}"
@@ -340,22 +346,166 @@ def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None)
     assert re.match('^Istanbul(, (TR|Turkey|Türkiye))?$', location), f'Expected Istanbul for location, got {location}'
 
 
+@pytest.mark.slow
+@pytest.mark.parametrize("result_override,n_predict,hf_repo,template_override", [
+    (None,                                           128,  "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       "chatml"),
+    (None,                                           128,  "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",         None),
+    (None,                                           128,  "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",         "chatml"),
+    (None,                                           128,  "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M",     ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")),
+    (None,                                           128,  "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M",       ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")),
+    (None,                                           128,  "bartowski/functionary-small-v3.2-GGUF:Q8_0",        ("meetkai/functionary-medium-v3.2", None)),
+    (None,                                           128,  "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M",  None),
+    (None,                                           128,  "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M",  None),
+    ("^> 0.56$",                                     128,  "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M",  "chatml"),
+    (None,                                           128,  "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
+
+    # TODO: fix these (wrong results, either didn't respect decimal instruction or got wrong value)
+    ("^The y-coordinate [\\s\\S]*?\\*\\*0.5\\*\\*",  8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+    ("[\\s\\S]*?\\*\\*0\\.5\\*\\*",                  8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
+])
+def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
+    global server
+    # n_predict = 512
+    server.n_slots = 1
+    server.jinja = True
+    server.n_ctx = 8192 * 2
+    server.n_predict = n_predict
+    server.model_hf_repo = hf_repo
+    server.model_hf_file = None
+    if isinstance(template_override, tuple):
+        (template_hf_repo, template_variant) = template_override
+        server.chat_template_file = f"../../../models/templates/{template_hf_repo.replace('/', '-') + ('-' + template_variant if template_variant else '')}.jinja"
+        assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
+    elif isinstance(template_override, str):
+        server.chat_template = template_override
+    server.start(timeout_seconds=TIMEOUT_SERVER_START)
+    res = server.make_request("POST", "/chat/completions", data={
+        "max_tokens": n_predict,
+        "messages": [
+            {"role": "system", "content": "You are a chatbot that uses tools/functions. Dont overthink things, and provide very concise answers. Do not explain your reasoning to the user. Provide any numerical values back to the user with at most two decimals."},
+            {"role": "user", "content": "What's the y coordinate of a point on the unit sphere at angle 30 degrees?"},
+            {
+                "role": "assistant",
+                "content": None,
+                "tool_calls": [
+                    {
+                        "id": "call_6789",
+                        "type": "function",
+                        "function": {
+                            "name": "calculate",
+                            "arguments": "{\"expression\":\"sin(30 * pi / 180)\"}"
+                        }
+                    }
+                ]
+            },
+            {
+                "role": "tool",
+                "name": "calculate",
+                "content": 0.55644242476,
+                "tool_call_id": "call_6789"
+            }
+        ],
+        "tools": [
+            {
+                "type":"function",
+                "function":{
+                    "name":"calculate",
+                    "description":"A calculator function that computes values of arithmetic expressions in the Python syntax",
+                    "parameters":{
+                        "type":"object",
+                        "properties":{
+                            "expression":{
+                            "type":"string",
+                            "description":"An arithmetic expression to compute the value of (Python syntad, assuming all floats)"
+                            }
+                        },
+                        "required":["expression"]
+                    }
+                }
+            }
+        ]
+    }, timeout=TIMEOUT_HTTP_REQUEST)
+    assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
+    choice = res.body["choices"][0]
+    tool_calls = choice["message"].get("tool_calls")
+    assert tool_calls is None, f'Expected no tool call in {choice["message"]}'
+    content = choice["message"].get("content")
+    assert content is not None, f'Expected content in {choice["message"]}'
+    if result_override is not None:
+        assert re.match(result_override, content), f'Expected {result_override}, got {content}'
+    else:
+        assert re.match('^[\\s\\S]*?The (y[ -])?coordinate [\\s\\S]*?is (approximately )?0\\.56\\b|^0\\.56$', content), \
+            f'Expected something like "The y coordinate is 0.56.", got {content}'
+
+
+@pytest.mark.slow
+@pytest.mark.parametrize("n_predict,reasoning_format,expect_content,expect_reasoning_content,hf_repo,template_override", [
+    (128, 'deepseek',  "^The sum of 102 and 7 is 109.*",                        None,                                          "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
+    (128,  None,        "^The sum of 102 and 7 is 109.*",                       None,                                          "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
+
+    (1024, 'deepseek',  "To find the sum of.*",                                 "I need to calculate the sum of 102 and 7.*",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+    (1024, 'none',      "<think>\n?I need[\\s\\S]*?</think>\n?To find.*",       None,                                          "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+
+    (1024, 'deepseek',  "To find the sum of.*",                                 "First, I [\\s\\S]*",                          "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
+])
+def test_thoughts(n_predict: int, reasoning_format: Literal['deepseek', 'none'] | None, expect_content: str | None, expect_reasoning_content: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
+    global server
+    server.n_slots = 1
+    server.reasoning_format = reasoning_format
+    server.jinja = True
+    server.n_ctx = 8192 * 2
+    server.n_predict = n_predict
+    server.model_hf_repo = hf_repo
+    server.model_hf_file = None
+    if isinstance(template_override, tuple):
+        (template_hf_repo, template_variant) = template_override
+        server.chat_template_file = f"../../../models/templates/{template_hf_repo.replace('/', '-') + ('-' + template_variant if template_variant else '')}.jinja"
+        assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template."
+    elif isinstance(template_override, str):
+        server.chat_template = template_override
+    server.start(timeout_seconds=TIMEOUT_SERVER_START)
+    res = server.make_request("POST", "/chat/completions", data={
+        "max_tokens": n_predict,
+        "messages": [
+            {"role": "user", "content": "What's the sum of 102 and 7?"},
+        ]
+    }, timeout=TIMEOUT_HTTP_REQUEST)
+    assert res.status_code == 200, f"Expected status code 200, got {res.status_code}"
+    choice = res.body["choices"][0]
+    assert choice["message"].get("tool_calls") is None, f'Expected no tool call in {choice["message"]}'
+
+    content = choice["message"].get("content")
+    if expect_content is None:
+        assert content is None, f'Expected no content in {choice["message"]}'
+    else:
+        assert re.match(expect_content, content), f'Expected {expect_content}, got {content}'
+
+    reasoning_content = choice["message"].get("reasoning_content")
+    if expect_reasoning_content is None:
+        assert reasoning_content is None, f'Expected no reasoning content in {choice["message"]}'
+    else:
+        assert re.match(expect_reasoning_content, reasoning_content), f'Expected {expect_reasoning_content}, got {reasoning_content}'
+
+
 @pytest.mark.slow
 @pytest.mark.parametrize("expected_arguments_override,hf_repo,template_override", [
+    (None,                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+    # (None,                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", "chatml"),
+
     (None,                 "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      None),
     (None,                 "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",      "chatml"),
 
     (None,                 "bartowski/functionary-small-v3.2-GGUF:Q8_0",       ("meetkai-functionary-medium-v3.2", None)),
     (None,                 "bartowski/functionary-small-v3.2-GGUF:Q8_0",       "chatml"),
 
-    (None,                 "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
-    ('{"code":"print("}',  "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
+    ('{"code":"print("}',  "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None),
+    (None,                 "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"),
 
-    ('{"code":"print("}',  "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      ("meta-llama-Llama-3.2-3B-Instruct", None)),
+    (None,                 "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      ("meta-llama-Llama-3.2-3B-Instruct", None)),
     (None,                 "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M",      "chatml"),
 
     ('{"code":"print("}',  "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M",      ("meta-llama-Llama-3.2-3B-Instruct", None)),
-    ('{"code":"print("}',  "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M",      "chatml"),
+    (None,                 "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M",      "chatml"),
 
     (None,                 "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        None),
     (None,                 "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M",        "chatml"),
@@ -371,15 +521,13 @@ def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None)
 
     # Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it.
     (None,                 "bartowski/gemma-2-2b-it-GGUF:Q4_K_M",              None),
-
-    # (None,                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
 ])
-def test_hello_world_tool_call(expected_arguments_override: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
+def test_hello_world(expected_arguments_override: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None):
     global server
     server.n_slots = 1
     server.jinja = True
     server.n_ctx = 8192
-    server.n_predict = 128
+    server.n_predict = 512 # High because of DeepSeek R1
     server.model_hf_repo = hf_repo
     server.model_hf_file = None
     if isinstance(template_override, tuple):
@@ -406,6 +554,7 @@ def test_hello_world_tool_call(expected_arguments_override: str | None, hf_repo:
     tool_calls = choice["message"].get("tool_calls")
     assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}'
     tool_call = tool_calls[0]
+    assert choice["message"].get("content") is None, f'Expected no content in {choice["message"]}'
     assert tool_call["function"]["name"] == PYTHON_TOOL["function"]["name"]
     actual_arguments = tool_call["function"]["arguments"]
     if expected_arguments_override is not None:

examples/server/tests/utils.py

@@ -78,6 +78,7 @@ class ServerProcess:
     draft_max: int | None = None
     no_webui: bool | None = None
     jinja: bool | None = None
+    reasoning_format: Literal['deepseek', 'none'] | None = None
     chat_template: str | None = None
     chat_template_file: str | None = None
 
@@ -172,6 +173,8 @@ class ServerProcess:
             server_args.append("--no-webui")
         if self.jinja:
             server_args.append("--jinja")
+        if self.reasoning_format is not None:
+            server_args.extend(("--reasoning-format", self.reasoning_format))
         if self.chat_template:
             server_args.extend(["--chat-template", self.chat_template])
         if self.chat_template_file:

examples/server/utils.hpp

@@ -367,10 +367,10 @@ inline std::string format_chat(const common_chat_template & tmpl, const std::vec
                     }
                 }
             } else {
-                throw std::runtime_error("Invalid 'content' type (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
+                throw std::runtime_error("Invalid 'content' type (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");
             }
         } else {
-            throw std::runtime_error("Missing 'content' (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
+            throw std::runtime_error("Missing 'content' (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");
         }
 
         chat.push_back({role, content, /* tool_calls= */ {}});
@@ -578,6 +578,7 @@ static json oaicompat_completion_params_parse(const json & body) {
 static json oaicompat_completion_params_parse(
     const json & body, /* openai api json semantics */
     bool use_jinja,
+    common_reasoning_format reasoning_format,
     const common_chat_templates & chat_templates)
 {
     json llama_params;
@@ -633,9 +634,10 @@ static json oaicompat_completion_params_parse(
             throw std::runtime_error("Cannot use custom grammar constraints with tools.");
         }
         common_chat_inputs inputs;
-        inputs.messages = body.at("messages");
-        inputs.tools = tools;
-        inputs.tool_choice = tool_choice;
+        inputs.extract_reasoning   = reasoning_format != COMMON_REASONING_FORMAT_NONE;
+        inputs.messages            = body.at("messages");
+        inputs.tools               = tools;
+        inputs.tool_choice         = tool_choice;
         inputs.parallel_tool_calls = json_value(body, "parallel_tool_calls", false);
         if (inputs.parallel_tool_calls && !tmpl.original_caps().supports_parallel_tool_calls) {
             LOG_DBG("Disabling parallel_tool_calls because the template does not support it\n");

examples/simple-cmake-pkg/README.md

@@ -1,6 +1,6 @@
 # llama.cpp/example/simple-cmake-pkg
 
-This program builds [simple](../simple) using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree.
+This program builds [simple](../simple) using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggml-org/llama.cpp) in projects which live outside of the source tree.
 
 ## Building
 
@@ -13,7 +13,7 @@ When hardware acceleration libraries are used (e.g. CUDA, Metal, Vulkan, etc.),
 ### Build llama.cpp and install to llama.cpp/inst
 
 ```sh
-git clone https://github.com/ggerganov/llama.cpp
+git clone https://github.com/ggml-org/llama.cpp
 cd llama.cpp
 cmake -S . -B build
 cmake --build build

examples/speculative/README.md

@@ -4,6 +4,6 @@ Demonstration of speculative decoding and tree-based speculative decoding techni
 
 More info:
 
-- https://github.com/ggerganov/llama.cpp/pull/2926
-- https://github.com/ggerganov/llama.cpp/pull/3624
-- https://github.com/ggerganov/llama.cpp/pull/5625
+- https://github.com/ggml-org/llama.cpp/pull/2926
+- https://github.com/ggml-org/llama.cpp/pull/3624
+- https://github.com/ggml-org/llama.cpp/pull/5625

flake.nix

@@ -36,7 +36,7 @@
   # ```
   # nixConfig = {
   #   extra-substituters = [
-  #     # Populated by the CI in ggerganov/llama.cpp
+  #     # Populated by the CI in ggml-org/llama.cpp
   #     "https://llama-cpp.cachix.org"
   #
   #     # A development cache for nixpkgs imported with `config.cudaSupport = true`.
@@ -56,11 +56,11 @@
   # };
   # ```
 
-  # For inspection, use `nix flake show github:ggerganov/llama.cpp` or the nix repl:
+  # For inspection, use `nix flake show github:ggml-org/llama.cpp` or the nix repl:
   #
   # ```bash
   # ❯ nix repl
-  # nix-repl> :lf github:ggerganov/llama.cpp
+  # nix-repl> :lf github:ggml-org/llama.cpp
   # Added 13 variables.
   # nix-repl> outputs.apps.x86_64-linux.quantize
   # { program = "/nix/store/00000000000000000000000000000000-llama.cpp/bin/llama-quantize"; type = "app"; }
@@ -176,7 +176,7 @@
             #
             # We could test all outputs e.g. as `checks = confg.packages`.
             #
-            # TODO: Build more once https://github.com/ggerganov/llama.cpp/issues/6346 has been addressed
+            # TODO: Build more once https://github.com/ggml-org/llama.cpp/issues/6346 has been addressed
             checks = {
               inherit (config.packages) default vulkan;
             };

ggml/include/ggml-cpu.h

@@ -8,7 +8,7 @@ extern "C" {
 #endif
 
     // the compute plan that needs to be prepared for ggml_graph_compute()
-    // since https://github.com/ggerganov/ggml/issues/287
+    // since https://github.com/ggml-org/ggml/issues/287
     struct ggml_cplan {
         size_t    work_size; // size of work buffer, calculated by `ggml_graph_plan()`
         uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`

ggml/include/ggml-metal.h

@@ -45,7 +45,7 @@ GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
 GGML_DEPRECATED(
         GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
-        "obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713");
+        "obsoleted by the new device interface - https://github.com/ggml-org/llama.cpp/pull/9713");
 
 GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 

ggml/src/ggml-cpu/ggml-cpu-quants.c

@@ -562,6 +562,41 @@ static __m256i lasx_packs_h(__m256i a, __m256i b) {
     return __lasx_xvpickev_b(tmp1, tmp);
 }
 
+static inline __m256i lasx_madd_h_b(__m256i a, __m256i b) {
+    __m256i tmp1, tmp2;
+    tmp1 = __lasx_xvmulwev_h_b(a, b);
+    tmp2 = __lasx_xvmulwod_h_b(a, b);
+    return __lasx_xvadd_h(tmp1, tmp2);
+}
+
+static inline __m256i lasx_xvrepl128vei_h(__m256i a, const unsigned int b) {
+    switch (b) {
+        case 0: return __lasx_xvrepl128vei_h(a, 0);
+        case 1: return __lasx_xvrepl128vei_h(a, 1);
+        case 2: return __lasx_xvrepl128vei_h(a, 2);
+        case 3: return __lasx_xvrepl128vei_h(a, 3);
+        case 4: return __lasx_xvrepl128vei_h(a, 4);
+        case 5: return __lasx_xvrepl128vei_h(a, 5);
+        case 6: return __lasx_xvrepl128vei_h(a, 6);
+        case 7: return __lasx_xvrepl128vei_h(a, 7);
+        default: __builtin_unreachable();
+    }
+}
+
+static inline __m256i lasx_xvandi_b_bit(__m256i a, const unsigned int b) {
+    switch (b) {
+        case 0: return __lasx_xvandi_b(a, 1 << 0);
+        case 1: return __lasx_xvandi_b(a, 1 << 1);
+        case 2: return __lasx_xvandi_b(a, 1 << 2);
+        case 3: return __lasx_xvandi_b(a, 1 << 3);
+        case 4: return __lasx_xvandi_b(a, 1 << 4);
+        case 5: return __lasx_xvandi_b(a, 1 << 5);
+        case 6: return __lasx_xvandi_b(a, 1 << 6);
+        case 7: return __lasx_xvandi_b(a, 1 << 7);
+        default: __builtin_unreachable();
+    }
+}
+
 // multiply int8_t, add results pairwise twice
 static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
     // Get absolute values of x vectors
@@ -656,13 +691,8 @@ static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy)
 
 // multiply int8_t, add results pairwise twice and return as float vector
 static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
-
-    // Get absolute values of x vectors
-    const __m256i ax = __lasx_xvsigncov_b(x, x);
-    // Sign the values of the y vectors
-    const __m256i sy = __lasx_xvsigncov_b(x, y);
-
-    return mul_sum_us8_pairs_float(ax, sy);
+    const __m256i dot = lasx_madd_h_b(x, y);
+    return sum_i16_pairs_float(dot);
 }
 
 static inline __m128i packNibbles( __m256i bytes ) {
@@ -4965,9 +4995,6 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined __loongarch_asx
 
-    const __m256i m3 = __lasx_xvreplgr2vr_b(3);
-    const __m128i m4 = __lsx_vreplgr2vr_b(0xF);
-
     __m256 acc = (__m256)__lasx_xvldi(0);
 
     for (int i = 0; i < nb; ++i) {
@@ -4978,18 +5005,15 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         const uint8_t * restrict q2 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
 
-        const __m128i mins_and_scales = __lsx_vld((const __m128i*)x[i].scales, 0);
-        const __m128i scales8 = __lsx_vand_v(mins_and_scales, m4);
-        const __m128i mins8 = __lsx_vand_v(__lsx_vsrli_h(mins_and_scales, 4), m4);
-        const __m256i mins = lasx_ext8_16(mins8);
+        const __m128i mins_and_scales128 = __lsx_vld((const __m128i*)x[i].scales, 0);
+        const __m128i scales128 = __lsx_vandi_b(mins_and_scales128, 0xf);
+        const __m256i mins = lasx_ext8_16(__lsx_vsrli_b(mins_and_scales128, 4));
         const __m256i prod = lasx_madd_h(mins, __lasx_xvld((const __m256i*)y[i].bsums, 0));
 
         acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(dmin), __lasx_xvffint_s_w(prod), acc);
 
-        const __m256i all_scales = lasx_ext8_16(scales8);
-        const __m128i l_scales = lasx_extracti128(all_scales, 0);
-        const __m128i h_scales = lasx_extracti128(all_scales, 1);
-        const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)};
+        const v16i8 shuffle_mask = {0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15};
+        const __m256i scales_shuffled = lasx_ext8_16(__lsx_vshuf_b(scales128, scales128, (__m128i)shuffle_mask));
 
         __m256i sumi = __lasx_xvldi(0);
 
@@ -5002,20 +5026,20 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
 
-            const __m256i q2_0 = __lasx_xvand_v(q2bits, m3);
-            const __m256i q2_1 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 2), m3);
-            const __m256i q2_2 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 4), m3);
-            const __m256i q2_3 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 6), m3);
+            const __m256i q2_0 = __lasx_xvandi_b(q2bits, 3);
+            const __m256i q2_1 = __lasx_xvandi_b(__lasx_xvsrli_b(q2bits, 2), 3);
+            const __m256i q2_2 = __lasx_xvandi_b(__lasx_xvsrli_b(q2bits, 4), 3);
+            const __m256i q2_3 = __lasx_xvsrli_b(q2bits, 6);
 
-            __m256i p0 = lasx_maddubs_h(q2_0, q8_0);
-            __m256i p1 = lasx_maddubs_h(q2_1, q8_1);
-            __m256i p2 = lasx_maddubs_h(q2_2, q8_2);
-            __m256i p3 = lasx_maddubs_h(q2_3, q8_3);
+            __m256i p0 = lasx_madd_h_b(q2_0, q8_0);
+            __m256i p1 = lasx_madd_h_b(q2_1, q8_1);
+            __m256i p2 = lasx_madd_h_b(q2_2, q8_2);
+            __m256i p3 = lasx_madd_h_b(q2_3, q8_3);
 
-            p0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(0)), p0);
-            p1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(1)), p1);
-            p2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(2)), p2);
-            p3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(3)), p3);
+            p0 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 0), p0);
+            p1 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 1), p1);
+            p2 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 2), p2);
+            p3 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 3), p3);
 
             p0 = __lasx_xvadd_w(p0, p1);
             p2 = __lasx_xvadd_w(p2, p3);
@@ -5771,8 +5795,6 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined __loongarch_asx
 
-    const __m256i m3 = __lasx_xvreplgr2vr_b(3);
-    const __m256i mone = __lasx_xvreplgr2vr_b(1);
     const __m128i m32 = __lsx_vreplgr2vr_b(32);
 
     __m256 acc = (__m256)__lasx_xvldi(0);
@@ -5792,10 +5814,9 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
                 (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
                 (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
         scales128 = __lsx_vsub_b(scales128, m32);
-        const __m256i all_scales = lasx_ext8_16(scales128);
-        const __m128i l_scales = lasx_extracti128(all_scales, 0);
-        const __m128i h_scales = lasx_extracti128(all_scales, 1);
-        const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)};
+
+        const v16i8 shuffle_mask = {0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15};
+        const __m256i scales_shuffled = lasx_ext8_16(__lsx_vshuf_b(scales128, scales128, (__m128i)shuffle_mask));
 
         // high bit
         const __m256i hbits = __lasx_xvld((const __m256i*)x[i].hmask, 0);
@@ -5803,35 +5824,23 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         // integer accumulator
         __m256i sumi = __lasx_xvldi(0);
 
-        int bit = 0;
-        int is  = 0;
-        __m256i xvbit;
-
-
         for (int j = 0; j < QK_K/128; ++j) {
             // load low 2 bits
             const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32;
 
-            xvbit = __lasx_xvreplgr2vr_h(bit);
             // prepare low and high bits
-            const __m256i q3l_0 = __lasx_xvand_v(q3bits, m3);
-            const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit);
-            const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 2), m3);
-            const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit);
-            const __m256i q3l_2 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 4), m3);
-            const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit);
-            const __m256i q3l_3 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 6), m3);
-            const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
+            const __m256i q3l_0 = __lasx_xvandi_b(q3bits, 3);
+            const __m256i q3l_1 = __lasx_xvandi_b(__lasx_xvsrli_b(q3bits, 2), 3);
+            const __m256i q3l_2 = __lasx_xvandi_b(__lasx_xvsrli_b(q3bits, 4), 3);
+            const __m256i q3l_3 = __lasx_xvsrli_b(q3bits, 6);
+            const __m256i q3h_0 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 0), 0), 2);
+            const __m256i q3h_1 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 1), 0), 2);
+            const __m256i q3h_2 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 2), 0), 2);
+            const __m256i q3h_3 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 3), 0), 2);
+            const __m256i q3_0 = __lasx_xvor_v(q3h_0, q3l_0);
+            const __m256i q3_1 = __lasx_xvor_v(q3h_1, q3l_1);
+            const __m256i q3_2 = __lasx_xvor_v(q3h_2, q3l_2);
+            const __m256i q3_3 = __lasx_xvor_v(q3h_3, q3l_3);
 
             // load Q8 quants
             const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
@@ -5839,29 +5848,16 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
 
-            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use lasx_maddubs_h,
-            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
-            // and 2 if the high bit was set)
-            __m256i q8s_0 = lasx_maddubs_h(q3h_0, q8_0);
-            __m256i q8s_1 = lasx_maddubs_h(q3h_1, q8_1);
-            __m256i q8s_2 = lasx_maddubs_h(q3h_2, q8_2);
-            __m256i q8s_3 = lasx_maddubs_h(q3h_3, q8_3);
-
-            __m256i p16_0 = lasx_maddubs_h(q3l_0, q8_0);
-            __m256i p16_1 = lasx_maddubs_h(q3l_1, q8_1);
-            __m256i p16_2 = lasx_maddubs_h(q3l_2, q8_2);
-            __m256i p16_3 = lasx_maddubs_h(q3l_3, q8_3);
-
-            p16_0 = __lasx_xvsub_h(p16_0, q8s_0);
-            p16_1 = __lasx_xvsub_h(p16_1, q8s_1);
-            p16_2 = __lasx_xvsub_h(p16_2, q8s_2);
-            p16_3 = __lasx_xvsub_h(p16_3, q8s_3);
+            __m256i p16_0 = lasx_madd_h_b(q8_0, q3_0);
+            __m256i p16_1 = lasx_madd_h_b(q8_1, q3_1);
+            __m256i p16_2 = lasx_madd_h_b(q8_2, q3_2);
+            __m256i p16_3 = lasx_madd_h_b(q8_3, q3_3);
 
             // multiply with scales
-            p16_0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
-            p16_1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
-            p16_2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
-            p16_3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
+            p16_0 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 0), p16_0);
+            p16_1 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 1), p16_1);
+            p16_2 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 2), p16_2);
+            p16_3 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 3), p16_3);
 
             // accumulate
             p16_0 = __lasx_xvadd_w(p16_0, p16_1);
@@ -5869,7 +5865,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
             sumi  = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_2));
         }
         // multiply with block scale and accumulate
-        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);//FIXME
+        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
     }
 
     *s = hsum_float_8(acc);
@@ -6562,11 +6558,6 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     *s = vec_extract(vsumf0, 0);
 
 #elif defined __loongarch_asx
-    GGML_UNUSED(kmask1);
-    GGML_UNUSED(kmask2);
-    GGML_UNUSED(kmask3);
-
-    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
 
     __m256 acc = (__m256)__lasx_xvldi(0);
     __m128 acc_m = (__m128)__lsx_vldi(0);
@@ -6586,33 +6577,34 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         const uint8_t * restrict q4 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
 
-        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
+        const __m128i mins_and_scales128 = lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i mins128 = __lsx_vexth_h_b(mins_and_scales128);
+        const __m128i scales128 = __lsx_vsllwil_h_b(mins_and_scales128, 0);
 
         const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
         const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
-        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
+        const __m128i prod = lsx_madd_h(mins128, q8s);
         acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
 
-        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
-        const __m256i scales = lasx_insertf128(sc128, sc128);
+        const __m256i scales = lasx_insertf128(scales128, scales128);
 
         __m256i sumi = __lasx_xvldi(0);
 
         for (int j = 0; j < QK_K/64; ++j) {
 
-            const __m256i scale_l = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_h = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
+            const __m256i scale_l = lasx_xvrepl128vei_h(scales, 2 * j + 0);
+            const __m256i scale_h = lasx_xvrepl128vei_h(scales, 2 * j + 1);
 
             const __m256i q4bits = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
-            const __m256i q4l = __lasx_xvand_v(q4bits, m4);
-            const __m256i q4h = __lasx_xvand_v(__lasx_xvsrli_h(q4bits, 4), m4);
+            const __m256i q4l = __lasx_xvandi_b(q4bits, 0xf);
+            const __m256i q4h = __lasx_xvsrli_b(q4bits, 4);
 
             const __m256i q8l = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            __m256i p16l = lasx_maddubs_h(q4l, q8l);
+            __m256i p16l = lasx_madd_h_b(q4l, q8l);
             p16l = lasx_madd_h(scale_l, p16l);
 
             const __m256i q8h = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            __m256i p16h = lasx_maddubs_h(q4h, q8h);
+            __m256i p16h = lasx_madd_h_b(q4h, q8h);
             p16h = lasx_madd_h(scale_h, p16h);
             const __m256i sumj = __lasx_xvadd_w(p16l, p16h);
 
@@ -7289,19 +7281,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     *s = vec_extract(vsumf0, 0);
 
 #elif defined __loongarch_asx
-    GGML_UNUSED(kmask1);
-    GGML_UNUSED(kmask2);
-    GGML_UNUSED(kmask3);
-
-    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
-    const __m128i mzero = __lsx_vldi(0);
-    const __m256i mone  = __lasx_xvreplgr2vr_b(1);
 
     __m256 acc = (__m256)__lasx_xvldi(0);
+    __m128 acc_m = (__m128)__lsx_vldi(0);
 
-    float summs = 0.f;
-
-   for (int i = 0; i < nb; ++i) {
+    for (int i = 0; i < nb; ++i) {
 
         const uint8_t * restrict q5 = x[i].qs;
         const int8_t  * restrict q8 = y[i].qs;
@@ -7316,49 +7300,40 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         utmp[2] = uaux;
         utmp[0] &= kmask1;
 
-        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
+        const __m128i mins_and_scales128 = lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i mins128 = __lsx_vexth_h_b(mins_and_scales128);
+        const __m128i scales128 = __lsx_vsllwil_h_b(mins_and_scales128, 0);
 
         const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
         const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
-        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
-        const __m128i hsum = lsx_hadd_w(lsx_hadd_w(prod, mzero), mzero);
-        summs += dmin * __lsx_vpickve2gr_w(hsum, 0);    //TODO check
+        const __m128i prod = lsx_madd_h(mins128, q8s);
+        acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
 
-        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
-        const __m256i scales = lasx_insertf128(sc128, sc128);
+        const __m256i scales = lasx_insertf128(scales128, scales128);
 
         const __m256i hbits = __lasx_xvld((const __m256i*)x[i].qh, 0);
-        __m256i hmask = mone;
 
         __m256i sumi = __lasx_xvldi(0);
 
-        int bit = 0;
-        __m256i xvbit;
-
         for (int j = 0; j < QK_K/64; ++j) {
 
-            const __m256i scale_0 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_1 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
+            const __m256i scale_0 = lasx_xvrepl128vei_h(scales, 2 * j + 0);
+            const __m256i scale_1 = lasx_xvrepl128vei_h(scales, 2 * j + 1);
 
             const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); q5 += 32;
 
-            xvbit = __lasx_xvreplgr2vr_h(bit++);
-            const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4);
-            const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
-            const __m256i q5_0  = __lasx_xvadd_b(q5l_0, q5h_0);
-            hmask = __lasx_xvslli_h(hmask, 1);
-
-            xvbit = __lasx_xvreplgr2vr_h(bit++);
-            const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4);
-            const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
-            const __m256i q5_1  = __lasx_xvadd_b(q5l_1, q5h_1);
-            hmask = __lasx_xvslli_h(hmask, 1);
+            const __m256i q5l_0 = __lasx_xvandi_b(q5bits, 0xf);
+            const __m256i q5l_1 = __lasx_xvsrli_b(q5bits, 4);
+            const __m256i q5h_0 = __lasx_xvnori_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 2 * j + 0), 0), 0xef);
+            const __m256i q5h_1 = __lasx_xvnori_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 2 * j + 1), 0), 0xef);
+            const __m256i q5_0  = __lasx_xvor_v(q5l_0, q5h_0);
+            const __m256i q5_1  = __lasx_xvor_v(q5l_1, q5h_1);
 
             const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
 
-            __m256i p16_0 = lasx_maddubs_h(q5_0, q8_0);
-            __m256i p16_1 = lasx_maddubs_h(q5_1, q8_1);
+            __m256i p16_0 = lasx_madd_h_b(q5_0, q8_0);
+            __m256i p16_1 = lasx_madd_h_b(q5_1, q8_1);
 
             p16_0 = lasx_madd_h(scale_0, p16_0);
             p16_1 = lasx_madd_h(scale_1, p16_1);
@@ -7372,7 +7347,10 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
     }
 
-    *s = hsum_float_8(acc) + summs;
+    acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vbsrl_v(acc_m, 8));
+    acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vbsrl_v(acc_m, 4));
+
+    *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
 
 #else
 
@@ -8033,8 +8011,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #elif defined __loongarch_asx
 
-    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
-    const __m256i m2 = __lasx_xvreplgr2vr_b(3);
     const __m256i m32s = __lasx_xvreplgr2vr_b(32);
 
     __m256 acc = (__m256)__lasx_xvldi(0);
@@ -8047,58 +8023,42 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         const uint8_t * restrict qh = x[i].qh;
         const int8_t  * restrict q8 = y[i].qs;
 
-        const __m128i scales = __lsx_vld((const __m128i*)x[i].scales, 0);
+        const __m128i scales128 = __lsx_vld((const __m128i*)x[i].scales, 0);
+        const v16i8 shuffle_mask = {0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15};
+        const __m256i scales_shuffled = lasx_ext8_16(__lsx_vshuf_b(scales128, scales128, (__m128i)shuffle_mask));
 
         __m256i sumi = __lasx_xvldi(0);
 
-        int is = 0;
-
         for (int j = 0; j < QK_K/128; ++j) {
 
-            const __m128i scale_0 = lsx_shuffle_b(scales, get_scale_shuffle(is + 0));
-            const __m128i scale_1 = lsx_shuffle_b(scales, get_scale_shuffle(is + 1));
-            const __m128i scale_2 = lsx_shuffle_b(scales, get_scale_shuffle(is + 2));
-            const __m128i scale_3 = lsx_shuffle_b(scales, get_scale_shuffle(is + 3));
-            is += 4;
-
             const __m256i q4bits1 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
             const __m256i q4bits2 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
             const __m256i q4bitsH = __lasx_xvld((const __m256i*)qh, 0); qh += 32;
 
-            const __m256i q4h_0 = __lasx_xvslli_h(__lasx_xvand_v(q4bitsH, m2), 4);
-            const __m256i q4h_1 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 2), m2), 4);
-            const __m256i q4h_2 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 4), m2), 4);
-            const __m256i q4h_3 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 6), m2), 4);
+            const __m256i q4h_0 = __lasx_xvslli_b(__lasx_xvandi_b(q4bitsH, 3), 4);
+            const __m256i q4h_1 = __lasx_xvslli_b(__lasx_xvandi_b(q4bitsH, 3 << 2), 2);
+            const __m256i q4h_2 = __lasx_xvandi_b(q4bitsH, 3 << 4);
+            const __m256i q4h_3 = __lasx_xvsrli_b(__lasx_xvandi_b(q4bitsH, 3 << 6), 2);
 
-            const __m256i q4_0 = __lasx_xvor_v(__lasx_xvand_v(q4bits1, m4), q4h_0);
-            const __m256i q4_1 = __lasx_xvor_v(__lasx_xvand_v(q4bits2, m4), q4h_1);
-            const __m256i q4_2 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits1, 4), m4), q4h_2);
-            const __m256i q4_3 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits2, 4), m4), q4h_3);
+            const __m256i q4_0 = __lasx_xvor_v(__lasx_xvandi_b(q4bits1, 0xf), q4h_0);
+            const __m256i q4_1 = __lasx_xvor_v(__lasx_xvandi_b(q4bits2, 0xf), q4h_1);
+            const __m256i q4_2 = __lasx_xvor_v(__lasx_xvsrli_b(q4bits1, 4), q4h_2);
+            const __m256i q4_3 = __lasx_xvor_v(__lasx_xvsrli_b(q4bits2, 4), q4h_3);
 
             const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
             const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
 
-            __m256i q8s_0 = lasx_maddubs_h(m32s, q8_0);
-            __m256i q8s_1 = lasx_maddubs_h(m32s, q8_1);
-            __m256i q8s_2 = lasx_maddubs_h(m32s, q8_2);
-            __m256i q8s_3 = lasx_maddubs_h(m32s, q8_3);
+            __m256i p16_0 = lasx_madd_h_b(__lasx_xvsub_b(q4_0, m32s), q8_0);
+            __m256i p16_1 = lasx_madd_h_b(__lasx_xvsub_b(q4_1, m32s), q8_1);
+            __m256i p16_2 = lasx_madd_h_b(__lasx_xvsub_b(q4_2, m32s), q8_2);
+            __m256i p16_3 = lasx_madd_h_b(__lasx_xvsub_b(q4_3, m32s), q8_3);
 
-            __m256i p16_0 = lasx_maddubs_h(q4_0, q8_0);
-            __m256i p16_1 = lasx_maddubs_h(q4_1, q8_1);
-            __m256i p16_2 = lasx_maddubs_h(q4_2, q8_2);
-            __m256i p16_3 = lasx_maddubs_h(q4_3, q8_3);
-
-            p16_0 = __lasx_xvsub_h(p16_0, q8s_0);
-            p16_1 = __lasx_xvsub_h(p16_1, q8s_1);
-            p16_2 = __lasx_xvsub_h(p16_2, q8s_2);
-            p16_3 = __lasx_xvsub_h(p16_3, q8s_3);
-
-            p16_0 = lasx_madd_h(lasx_ext8_16(scale_0), p16_0);
-            p16_1 = lasx_madd_h(lasx_ext8_16(scale_1), p16_1);
-            p16_2 = lasx_madd_h(lasx_ext8_16(scale_2), p16_2);
-            p16_3 = lasx_madd_h(lasx_ext8_16(scale_3), p16_3);
+            p16_0 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 0), p16_0);
+            p16_1 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 1), p16_1);
+            p16_2 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 2), p16_2);
+            p16_3 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 3), p16_3);
 
             sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
             sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_2, p16_3));
@@ -10423,13 +10383,9 @@ static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
 }
 #elif defined(__loongarch_asx)
 static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
-    const __m256i ax = __lasx_xvsigncov_b(x, x);
-    const __m256i sy = __lasx_xvsigncov_b(x, y);
-    __m256i tmp1, tmp2, tmp3;
-    tmp1 = __lasx_xvmulwev_h_bu_b(ax, sy);
-    tmp2 = __lasx_xvmulwod_h_bu_b(ax, sy);
-    tmp3 = __lasx_xvadd_h(tmp1, tmp2);
-    return __lasx_xvsat_h(tmp3, 15);
+    const __m256i a = __lasx_xvmulwev_h_b(x, y);
+    const __m256i b = __lasx_xvmulwod_h_b(x, y);
+    return __lasx_xvadd_h(a, b);
 }
 #endif
 
@@ -11479,67 +11435,31 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void *
 #elif defined(__loongarch_asx)
 
     const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
-    const __m128i m4b  = __lsx_vreplgr2vr_b(0x0f);
 
     __m256 accum = (__m256)__lasx_xvldi(0);
-    __m256i tmp1;
-    __m128i tmp0, tmp2, tmp3, tmp4, mask_8f, mask;
 
-    mask_8f = __lsx_vreplgr2vr_b(0x8f);
     for (int ibl = 0; ibl < nb; ++ibl) {
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;
         uint16_t sh = x[ibl].scales_h;
         __m256i sumi1 = __lasx_xvldi(0);
         __m256i sumi2 = __lasx_xvldi(0);
-        __m128i zero = __lsx_vldi(0);
         for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q4bits_1 = __lsx_vld((const __m128i*)qs, 0);  qs += 16;
-            const __m128i q4bits_2 = __lsx_vld((const __m128i*)qs, 0);  qs += 16;
+            const __m128i q4bits_1 = __lsx_vld((const __m128i*)qs, 0); qs += 16;
+            const __m128i q4bits_2 = __lsx_vld((const __m128i*)qs, 0); qs += 16;
             const __m256i q8b_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
             const __m256i q8b_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp3 = __lsx_vand_v(tmp0, mask);
-            tmp3 = __lsx_vshuf_b(values128, zero, tmp3);
-
-            tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_1, m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp4 = __lsx_vand_v(tmp0, mask);
-            tmp4 = __lsx_vshuf_b(values128, zero, tmp4);
-
-            const __m256i q4b_1 = lasx_insertf128(tmp3, tmp4);
-
-            tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp3 = __lsx_vand_v(tmp0, mask);
-            tmp3 = __lsx_vshuf_b(values128, zero, tmp3);
-
-            tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_2, m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp4 = __lsx_vand_v(tmp0, mask);
-            tmp4 = __lsx_vshuf_b(values128, zero, tmp4);
-
-            const __m256i q4b_2 = lasx_insertf128(tmp3, tmp4);
-
+            const __m256i q4b_1 = lasx_insertf128(__lsx_vshuf_b(values128, values128, __lsx_vsrli_b(q4bits_1, 4)),
+                                                  __lsx_vshuf_b(values128, values128, __lsx_vandi_b(q4bits_1, 0xf)));
+            const __m256i q4b_2 = lasx_insertf128(__lsx_vshuf_b(values128, values128, __lsx_vsrli_b(q4bits_2, 4)),
+                                                  __lsx_vshuf_b(values128, values128, __lsx_vandi_b(q4bits_2, 0xf)));
             const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
             const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
             const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
             const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
             sh >>= 4;
-            __m256i tmp5, tmp6;
-            tmp1 = __lasx_xvreplgr2vr_h(ls1);
-            tmp5 = __lasx_xvmulwev_w_h(p16_1, tmp1);
-            tmp6 = __lasx_xvmulwod_w_h(p16_1, tmp1);
-            const __m256i p_1 = __lasx_xvadd_w(tmp5, tmp6);
-            tmp1 = __lasx_xvreplgr2vr_h(ls2);
-            tmp5 = __lasx_xvmulwev_w_h(p16_2, tmp1);
-            tmp6 = __lasx_xvmulwod_w_h(p16_2, tmp1);
-            const __m256i p_2 = __lasx_xvadd_w(tmp5, tmp6);
+            const __m256i p_1 = lasx_madd_h(p16_1, __lasx_xvreplgr2vr_h(ls1));
+            const __m256i p_2 = lasx_madd_h(p16_2, __lasx_xvreplgr2vr_h(ls2));
             sumi1 = __lasx_xvadd_w(p_1, sumi1);
             sumi2 = __lasx_xvadd_w(p_2, sumi2);
         }

ggml/src/ggml-cpu/ggml-cpu.c

@@ -1816,7 +1816,7 @@ inline static float ggml_silu_f32(float x) {
 
 #if __FINITE_MATH_ONLY__
 #error "some routines in ggml.c require non-finite math arithmetics -- pass -fno-finite-math-only to the compiler to fix"
-#error "ref: https://github.com/ggerganov/llama.cpp/pull/7154#issuecomment-2143844461"
+#error "ref: https://github.com/ggml-org/llama.cpp/pull/7154#issuecomment-2143844461"
 #endif
 
 #if defined(__ARM_NEON) && defined(__aarch64__)
@@ -7574,7 +7574,7 @@ UseGgmlGemm2:;
     int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;
 
     // If the chunking is poor for the number of threads on this setup, scrap the whole plan.  Re-chunk it by thread.
-    //   Also, chunking by thread was measured to have perform better on NUMA systems.  See https://github.com/ggerganov/llama.cpp/pull/6915
+    //   Also, chunking by thread was measured to have perform better on NUMA systems.  See https://github.com/ggml-org/llama.cpp/pull/6915
     //   In theory, chunking should be just as useful on NUMA and non NUMA systems, but testing disagreed with that.
     if (nchunk0 * nchunk1 < nth * 4 || ggml_is_numa()) {
         // distribute the thread work across the inner or outer loop based on which one is larger

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

@@ -280,14 +280,6 @@ template <> inline __m256bh load(const float *p) {
 }
 #endif
 
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// CONSTANTS
-
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
-#endif
-
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // FLOATING POINT MATRIX MULTIPLICATION
 
@@ -614,6 +606,14 @@ class tinyBLAS_Q0_AVX {
                     TC *C, int64_t ldc,
                     int ith, int nth)
         : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+        const int8_t kvalues_iq4nl[16] = {
+            -127, -104, -83, -65,
+            -49,  -35,  -22, -10,
+              1,   13,   25,  38,
+             53,   69,   89, 113
+        };
+
+        iq4nlt = _mm_loadu_si128((const __m128i *)kvalues_iq4nl);
     }
 
     void matmul(int64_t m, int64_t n) {
@@ -1038,6 +1038,7 @@ class tinyBLAS_Q0_AVX {
     const int64_t ldc;
     const int ith;
     const int nth;
+    __m128i iq4nlt;
 };
 #endif // __AVX__
 

ggml/src/ggml-cuda/CMakeLists.txt

@@ -15,9 +15,9 @@ if (CUDAToolkit_FOUND)
         if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
             set(CMAKE_CUDA_ARCHITECTURES "native")
         elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
-            set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75;80")
         else()
-            set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
+            set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75;80")
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")

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

@@ -1480,12 +1480,7 @@ static void ggml_cuda_op_mul_mat(
             const size_t nbytes_data    = ggml_nbytes(src0);
             const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding);
-        // TODO: remove this for MUSA once the Guilty Lockup issue is resolved
-#ifndef GGML_USE_MUSA
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd, 0, nbytes_data + nbytes_padding, stream));
-#else // GGML_USE_MUSA
-            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream));
-#endif // !GGML_USE_MUSA
         }
 
         // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared:

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

@@ -24,7 +24,7 @@
 #endif
 
 // create residency sets only on macOS >= 15.0
-#if TARGET_OS_OSX && __MAC_OS_X_VERSION_MAX_ALLOWED >= 150000 || \
+#if !TARGET_CPU_X86_64 && TARGET_OS_OSX && __MAC_OS_X_VERSION_MAX_ALLOWED >= 150000 || \
     TARGET_OS_IOS && __IPHONE_OS_VERSION_MAX_ALLOWED >= 180000 || \
     TARGET_OS_TV && __TV_OS_VERSION_MAX_ALLOWED >= 180000 || \
     TARGET_OS_VISION && __VISION_OS_VERSION_MAX_ALLOWED >= 200000
@@ -1983,7 +1983,7 @@ static void ggml_metal_encode_node(
                 const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
                 // TODO: add ggml_metal_kargs struct
-                // TODO: optimize (see https://github.com/ggerganov/llama.cpp/pull/10238/commits/7941b6b9ec29a2866fec6fa6c51612515ca509f6)
+                // TODO: optimize (see https://github.com/ggml-org/llama.cpp/pull/10238/commits/7941b6b9ec29a2866fec6fa6c51612515ca509f6)
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
                 if (id_src1) {

ggml/src/ggml-metal/ggml-metal.metal

@@ -373,24 +373,33 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
 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 uint8_t * ql = (device const uint8_t *)xb->ql;
-    device const uint8_t * qh = (device const uint8_t *)xb->qh;
+    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 + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
-    qh = qh + 32*(il/8) + 16*(il&1);
+    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 uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
-    const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
-    const float       coef = il>1 ? 1.f/16.f          : 1.f;
+    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 dl = d_all * sc * coef;
-    for (int i = 0; i < 16; ++i) {
-        const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
-                            : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
-        reg[i/4][i%4] = dl * q - ml;
+    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;
     }
 }
 
@@ -1058,7 +1067,7 @@ kernel void kernel_soft_max(
     }
 
     // This barrier fixes a failing test
-    // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
     threadgroup_barrier(mem_flags::mem_none);
 
     float sum = simd_sum(lsum);
@@ -1163,7 +1172,7 @@ kernel void kernel_soft_max_4(
     const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
 
     // This barrier fixes a failing test
-    // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
     threadgroup_barrier(mem_flags::mem_none);
 
     float sum = simd_sum(lsum);

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -143,6 +143,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_rms_norm;
     cl_kernel kernel_diag_mask_inf, kernel_diag_mask_inf_8;
     cl_kernel kernel_soft_max, kernel_soft_max_4;
+    cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
     cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
     cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
@@ -614,6 +615,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     CL_CHECK((backend_ctx->kernel_diag_mask_inf_8    = clCreateKernel(backend_ctx->program, "kernel_diag_mask_inf_8", &err), err));
     CL_CHECK((backend_ctx->kernel_soft_max           = clCreateKernel(backend_ctx->program, "kernel_soft_max", &err), err));
     CL_CHECK((backend_ctx->kernel_soft_max_4         = clCreateKernel(backend_ctx->program, "kernel_soft_max_4", &err), err));
+    CL_CHECK((backend_ctx->kernel_soft_max_f16       = clCreateKernel(backend_ctx->program, "kernel_soft_max_f16", &err), err));
+    CL_CHECK((backend_ctx->kernel_soft_max_4_f16     = clCreateKernel(backend_ctx->program, "kernel_soft_max_4_f16", &err), err));
     CL_CHECK((backend_ctx->kernel_rope_norm_f32      = clCreateKernel(backend_ctx->program, "kernel_rope_norm_f32", &err), err));
     CL_CHECK((backend_ctx->kernel_rope_norm_f16      = clCreateKernel(backend_ctx->program, "kernel_rope_norm_f16", &err), err));
     CL_CHECK((backend_ctx->kernel_rope_neox_f32      = clCreateKernel(backend_ctx->program, "kernel_rope_neox_f32", &err), err));
@@ -1044,8 +1047,16 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return true;
         case GGML_OP_DIAG_MASK_INF:
             return op->ne[3] == 1;
-        case GGML_OP_ROPE:
+        case GGML_OP_ROPE: {
+            const int mode = ((const int32_t *) op->op_params)[2];
+            if (mode & GGML_ROPE_TYPE_MROPE) {
+                return false;
+            }
+            if (mode & GGML_ROPE_TYPE_VISION) {
+                return false;
+            }
             return true;
+        }
         default:
             return false;
     }
@@ -3666,6 +3677,8 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
+    const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
+
     // Local size must be wave size. Each workgroup is a wave, working on a row,
     // where a row corresponds to leading dimension.
     int nth = MIN(32, ne00);
@@ -3683,9 +3696,17 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
     cl_kernel kernel;
 
     if (ne00%4 == 0) {
-        kernel = backend_ctx->kernel_soft_max_4;
+        if (use_f16) {
+            kernel = backend_ctx->kernel_soft_max_4_f16;
+        } else {
+            kernel = backend_ctx->kernel_soft_max_4;
+        }
     } else {
-        kernel = backend_ctx->kernel_soft_max;
+        if (use_f16) {
+            kernel = backend_ctx->kernel_soft_max_f16;
+        } else {
+            kernel = backend_ctx->kernel_soft_max;
+        }
     }
 
     CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
@@ -3766,7 +3787,8 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     const int  nb2 = dst ? dst->nb[2] : 0;
     const int  nb3 = dst ? dst->nb[3] : 0;
 
-    GGML_ASSERT(ne10 == ne02);
+    GGML_ASSERT(ne10 % ne02 == 0);
+    GGML_ASSERT(ne10 >= ne02);
 
     int nth = MIN(64, ne00);
 

ggml/src/ggml-opencl/kernels/ggml-opencl.cl

@@ -679,6 +679,9 @@ kernel void kernel_diag_mask_inf_8(
 //------------------------------------------------------------------------------
 // softmax
 //------------------------------------------------------------------------------
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
 kernel void kernel_soft_max(
         global float * src0,
         ulong offset0,
@@ -811,6 +814,141 @@ kernel void kernel_soft_max_4(
     }
 }
 
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_f16(
+        global float * src0,
+        ulong offset0,
+        global half * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    src1 = (global half *)((global char *)src1 + offset1);
+    dst = (global float *)((global char *)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    global half  * pmask = (global char *)src1 != (global char *)src0 ? src1 + i01*ne00 : 0;
+    global float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum += exp_psrc0;
+        // Remember the result of exp here. exp is expensive, so we really do not
+        // wish to compute it twice.
+        pdst[i00] = exp_psrc0;
+    }
+
+    const float sum = sub_group_reduce_add(lsum);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        pdst[i00] /= sum;
+    }
+}
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_4_f16(
+        global float * src0,
+        ulong offset0,
+        global half * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    src1 = (global half *)((global char *)src1 + offset1);
+    dst = (global float *)((global char *)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * psrc4 = (global float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    global half4  * pmask = (global char *)src1 != (global char *)src0 ? (global half4 *)(src1 + i01*ne00) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f));
+    }
+    float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3));
+
+    const float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f)) - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
+
+    const float sum = sub_group_reduce_add(lsum);
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        pdst4[i00] /= sum;
+    }
+}
+
 //------------------------------------------------------------------------------
 // kernel_rope
 //------------------------------------------------------------------------------

ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp

@@ -0,0 +1,51 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+
+shared FLOAT_TYPE tmpmax[BLOCK_SIZE];
+shared uint tmp[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint col = gl_LocalInvocationID.x;
+
+    if (col >= p.KX) {
+        return;
+    }
+    A_TYPE amax = data_a[row*p.KX + col];
+    tmp[col] = col;
+
+    for (uint i = col + BLOCK_SIZE; i < p.KX; i += BLOCK_SIZE) {
+        A_TYPE val = data_a[row*p.KX + i];
+        if (val > amax) {
+            amax = val;
+            tmp[col] = i;
+        }
+    }
+    tmpmax[col] = amax;
+
+    barrier();
+    [[unroll]] for (int s = int(BLOCK_SIZE) / 2; s > 0; s >>= 1) {
+        if (col < s && col + s < p.KX) {
+            if (tmpmax[col] < tmpmax[col + s]) {
+                tmpmax[col] = tmpmax[col + s];
+                tmp[col] = tmp[col + s];
+            }
+        }
+        barrier();
+    }
+
+    if (col == 0) {
+        data_d[row] = D_TYPE(tmp[0]);
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp

@@ -12,7 +12,7 @@ layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
 #endif
 
 void main() {
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
     if (gl_LocalInvocationIndex.x != 0) {
         return;

ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp

@@ -217,7 +217,7 @@ void quantize(uint dst_idx, uint src_idx)
 #endif
 
 void main() {
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
     if (gl_LocalInvocationIndex.x != 0) {
         return;

ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp

@@ -0,0 +1,31 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+
+#include "types.comp"
+#include "generic_head.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
+layout (binding = 2) buffer D {D_TYPE data_d[];};
+
+const uint CHUNK_SIZE = 512;
+
+void main() {
+    const uint base = gl_WorkGroupID.x * CHUNK_SIZE;
+    const uint col = gl_LocalInvocationID.x;
+
+    uint count = 0;
+    [[unroll]]
+    for (uint i = 0; i < CHUNK_SIZE; i += gl_WorkGroupSize.x) {
+        const uint idx = base + i + col;
+        if (idx >= p.KX) {
+            break;
+        }
+        count += uint(data_a[idx] == data_b[idx]);
+    }
+
+    atomicAdd(data_d[0], D_TYPE(count));
+}

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp

@@ -88,6 +88,83 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 }
 #endif
 
+#if defined(DATA_A_IQ1_S)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint ib32 = iqs / 32;
+    const uint ib8 = iqs / 8;
+    const int i8 = int(iqs % 8);
+    const uint qh = data_a[a_offset + ib].qh[ib32];
+    const uint qs = data_a[a_offset + ib].qs[ib8];
+    const float dl = float(2 * bitfieldExtract(qh, 12, 3) + 1);
+    const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+    const uint idxhi = bitfieldExtract(qh, 3 * int(ib8 & 3), 3);
+    const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]);
+    // Signed bitfield extract.
+    const ivec2 gvec = ivec2(
+      bitfieldExtract(grid, 2 * (i8), 2),
+      bitfieldExtract(grid, 2 * (i8 + 1), 2)
+    );
+    return dl * (vec2(gvec) + delta);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint ib32 = iqs / 32;
+    const uint ib8 = iqs / 8;
+    const int i8 = int(iqs % 8);
+    const uint qh = data_a[a_offset + ib].qh[ib32];
+    const uint qs = data_a[a_offset + ib].qs[ib8];
+    const float dl = 2 * bitfieldExtract(qh, 12, 3) + 1;
+    const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+    const int16_t grid = int16_t(iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]);
+    // Signed bitfield extract.
+    const ivec4 gvec = ivec4(
+      bitfieldExtract(grid, 2 * (i8), 2),
+      bitfieldExtract(grid, 2 * (i8 + 1), 2),
+      bitfieldExtract(grid, 2 * (i8 + 2), 2),
+      bitfieldExtract(grid, 2 * (i8 + 3), 2)
+    );
+    return dl * (vec4(gvec) + delta);
+}
+#endif
+
+#if defined(DATA_A_IQ1_M)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint ib8 = iqs / 8;
+    const uint ib16 = iqs / 16;
+    const int i8 = int(iqs % 8);
+    const uint sc = data_a[a_offset + ib].scales[iqs / 64];
+    const uint qs = data_a[a_offset + ib].qs[ib8];
+    const uint qh = data_a[a_offset + ib].qh[ib16] >> (4 * (ib8 & 1));
+    const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1;
+    const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+    const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
+    // Signed bitfield extract.
+    const ivec2 gvec = ivec2(
+      bitfieldExtract(grid, 2 * (i8), 2),
+      bitfieldExtract(grid, 2 * (i8 + 1), 2)
+    );
+    return dl * (vec2(gvec) + delta);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint ib8 = iqs / 8;
+    const uint ib16 = iqs / 16;
+    const int i8 = int(iqs % 8);
+    const uint sc = data_a[a_offset + ib].scales[iqs / 64];
+    const uint qs = data_a[a_offset + ib].qs[ib8];
+    const uint qh = data_a[a_offset + ib].qh[ib16] >> (4 * (ib8 & 1));
+    const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1;
+    const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+    const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
+    // Signed bitfield extract.
+    const ivec4 gvec = ivec4(
+      bitfieldExtract(grid, 2 * (i8), 2),
+      bitfieldExtract(grid, 2 * (i8 + 1), 2),
+      bitfieldExtract(grid, 2 * (i8 + 2), 2),
+      bitfieldExtract(grid, 2 * (i8 + 3), 2)
+    );
+    return dl * (vec4(gvec) + delta);
+}
+#endif
+
 #if defined(DATA_A_IQ2_XXS)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint ib32 = iqs / 32;
@@ -357,7 +434,16 @@ vec2 get_dm(uint ib, uint a_offset) {
 }
 #endif
 
-#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#if defined(DATA_A_IQ1_M)
+vec2 get_dm(uint ib, uint a_offset) {
+    const uint16_t[4] scales = data_a[a_offset + ib].scales;
+    const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
+    const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x);
+    return vec2(d, 0);
+}
+#endif
+
+#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(float(data_a[a_offset + ib].d), 0);
 }

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -301,6 +301,56 @@ float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2
     return ret;
 }
 
+#if defined(DATA_A_IQ1_S)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_S {
+   block_iq1_s block;
+};
+
+float16_t dequantFuncIQ1_S(const in decodeBufIQ1_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float16_t d = bl.block.d;
+    const uint idx = coordInBlock[1];
+
+    const uint ib32 = idx / 32;
+    const uint ib8 = idx / 8;
+
+    const uint qh = bl.block.qh[ib32];
+    const uint qs = bl.block.qs[ib8];
+    const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
+    const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+    const uint grid = iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)];
+
+    float16_t ret = float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * int(idx % 8), 2)) + float16_t(delta));
+    return ret;
+}
+#endif
+
+#if defined(DATA_A_IQ1_M)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_M {
+   block_iq1_m block;
+};
+
+float16_t dequantFuncIQ1_M(const in decodeBufIQ1_M bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const u16vec4 scales = u16vec4(bl.block.scales[0], bl.block.scales[1], bl.block.scales[2], bl.block.scales[3]) >> 12;
+    const float16_t d = uint16BitsToHalf(scales.x | (scales.y << 4) | (scales.z << 8) | (scales.w << 12));
+    const uint idx = coordInBlock[1];
+
+    const uint ib8 = idx / 8;
+    const uint ib16 = idx / 16;
+    const int i8 = int(idx % 8);
+    const uint sc = bl.block.scales[ib8 / 8];
+    const uint qs = bl.block.qs[ib8];
+    const uint qh = bl.block.qh[ib16] >> (4 * (ib8 & 1));
+    const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1;
+    const float delta = ((qh & 8) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+    const uint grid = iq1s_grid[qs | ((qh & 7) << 8)];
+
+    float16_t ret = d * float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * i8, 2)) + float16_t(delta));
+    return ret;
+}
+#endif
+
 #if defined(DATA_A_IQ2_XXS)
 layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS {
    block_iq2_xxs block;
@@ -512,6 +562,10 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor
 #define dequantFuncA dequantFuncQ5_K
 #elif defined(DATA_A_Q6_K)
 #define dequantFuncA dequantFuncQ6_K
+#elif defined(DATA_A_IQ1_S)
+#define dequantFuncA dequantFuncIQ1_S
+#elif defined(DATA_A_IQ1_M)
+#define dequantFuncA dequantFuncIQ1_M
 #elif defined(DATA_A_IQ2_XXS)
 #define dequantFuncA dequantFuncIQ2_XXS
 #elif defined(DATA_A_IQ2_XS)

ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp

@@ -0,0 +1,42 @@
+#version 450
+
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq1_m data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    // Each thread handles 1 subblock (32 values with 2 scales)
+    const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    if (ib >= p.nel / 256) {
+        return;
+    }
+
+    const uint ib32 = gl_LocalInvocationID.x % 8;
+    const uint ib64 = ib32 / 2;
+    const uint b_idx = 256 * ib + 32 * ib32;
+
+    const uint16_t[4] scales = data_a[ib].scales;
+    const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
+    const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x);
+
+    const uint sc = data_a[ib].scales[ib64];
+    [[unroll]] for (int l = 0; l < 4; ++l) {
+        const uint ib16 = 2 * ib32 + l / 2;
+        const float dl = d * (2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1);
+        const uint qh = data_a[ib].qh[ib16] >> (4 * (l & 1));
+        const uint qs = data_a[ib].qs[4 * ib32 + l];
+        const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+        const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
+        [[unroll]] for (int j = 0; j < 8; ++j) {
+            data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta));
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp

@@ -0,0 +1,35 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq1_s data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    // Each thread handles 1 subblock (32 values with 2 scales)
+    const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    if (ib >= p.nel / 256) {
+        return;
+    }
+
+    const uint ib32 = gl_LocalInvocationID.x % 8;
+    const uint b_idx = 256 * ib + 32 * ib32;
+
+    uint qh = data_a[ib].qh[ib32];
+    const float d = float(data_a[ib].d);
+    const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
+    const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+    [[unroll]] for (uint l = 0; l < 4; ++l) {
+        const uint qs = data_a[ib].qs[4 * ib32 + l];
+        const uint hi = bitfieldExtract(qh, 3 * int(l), 3);
+        const int16_t grid = int16_t(iq1s_grid[qs | (hi << 8)]);
+        [[unroll]] for (int j = 0; j < 8; ++j) {
+            data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta));
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -104,7 +104,7 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
 #endif
 
 void main() {
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 

ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp

@@ -12,7 +12,7 @@ void main() {
     const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
     const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
 
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp

@@ -133,7 +133,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 void main() {
     const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
 
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp

@@ -0,0 +1,82 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 32 * ib32;
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint16_t[4] scales = data_a[ibi].scales;
+        const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
+        const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x);
+
+        const uint sc = data_a[ibi].scales[ib32 / 2] >> (6 * (ib32 & 1));
+        [[unroll]] for (uint l = 0; l < 4; ++l) {
+            const uint qh = data_a[ibi].qh[2 * ib32 + l / 2] >> (4 * (l&1));
+            const uint qs = data_a[ibi].qs[4 * ib32 + l];
+            const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+            const float dl = d * (2 * bitfieldExtract(sc, 3 * int(l / 2), 3) + 1);
+
+            const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+                [[unroll]] for (int k = 0; k < 4; ++k) {
+                    sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
+                          fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
+                }
+                temp[j][n] = fma(dl, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 8 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/8;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 8;  // 0...7
+    const uint ix = tid / 8;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp

@@ -0,0 +1,79 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx = i * QUANT_K + 32 * ib32;
+
+    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const float d = float(data_a[ibi].d);
+        const uint qh = data_a[ibi].qh[ib32];
+        const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
+        const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+
+        [[unroll]] for (uint l = 0; l < 4; ++l) {
+            const uint qs = data_a[ibi].qs[4 * ib32 + l];
+            const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3);
+            const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]);
+
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
+                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+
+                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+                [[unroll]] for (int k = 0; k < 4; ++k) {
+                    sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
+                          fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
+                }
+                temp[j][n] = fma(dl, sum, temp[j][n]);
+            }
+        }
+        ibi += num_blocks_per_row;
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+
+    // 8 threads are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/8;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid % 8;  // 0...7
+    const uint ix = tid / 8;
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
+        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
+
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    init_iq_shmem(gl_WorkGroupSize);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -6,6 +6,9 @@
 #ifdef FLOAT16
 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
 #endif
+#if defined(DATA_A_IQ1_M)
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#endif
 
 #ifdef COOPMAT
 #extension GL_KHR_cooperative_matrix : enable
@@ -95,7 +98,7 @@ shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
 #endif
 
 void main() {
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 
@@ -437,6 +440,56 @@ void main() {
 
             buf_a[buf_idx    ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32));
             buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
+#elif defined(DATA_A_IQ1_S)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                  // 2 values per idx
+            const uint ib32 = (idx % 128) / 16;         // 0..7
+            const uint ib8 = (idx % 128) / 4;
+            const int i8 = 2 * int(idx % 4);
+
+            const float d = float(data_a[ib].d);
+            const uint qh = data_a[ib].qh[ib32];
+            const uint qs = data_a[ib].qs[ib8];
+            const float dl = d * (2 * bitfieldExtract(qh, 12, 3) + 1);
+            const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+            const int16_t grid = int16_t(iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]);
+
+            const ivec2 gvec = ivec2(
+              bitfieldExtract(grid, 2 * (i8), 2),
+              bitfieldExtract(grid, 2 * (i8 + 1), 2)
+            );
+            const vec2 v = dl * (vec2(gvec) + delta);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_IQ1_M)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                  // 2 values per idx
+            const uint ib8 = (idx % 128) / 4;
+            const uint ib16 = ib8 / 2;
+            const int i8 = 2 * int(idx % 4);
+
+            const uint16_t[4] scales = data_a[ib].scales;
+            const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
+            const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x);
+            const uint sc = scales[ib8 / 8];
+            const uint qs = data_a[ib].qs[ib8];
+            const uint qh = data_a[ib].qh[ib16] >> (4 * (ib8 & 1));
+            const float dl = d * (2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1);
+            const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
+            const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
+            const ivec2 gvec = ivec2(
+              bitfieldExtract(grid, 2 * (i8), 2),
+              bitfieldExtract(grid, 2 * (i8 + 1), 2)
+            );
+            const vec2 v = dl * (vec2(gvec) + delta);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
 #elif defined(DATA_A_IQ2_XXS)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
             const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -106,7 +106,7 @@ D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem
 #endif
 
 void main() {
-#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
+#ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 

ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp

@@ -0,0 +1,42 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) buffer X {A_TYPE x[];};
+layout (binding = 1) readonly buffer G {A_TYPE grad[];};
+layout (binding = 2) buffer GM {A_TYPE gradm[];};
+layout (binding = 3) buffer GV {A_TYPE gradv[];};
+layout (binding = 4) readonly buffer P {float params[7];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float alpha  = params[0];
+    const float beta1  = params[1];
+    const float beta2  = params[2];
+    const float eps    = params[3];
+    const float wd     = params[4];
+    const float beta1h = params[5];
+    const float beta2h = params[6];
+
+    const float gi = grad[i];
+    const float gmi = gradm[i]*beta1 +    gi*(1.0f - beta1);
+    const float gvi = gradv[i]*beta2 + gi*gi*(1.0f - beta2);
+
+    gradm[i] = gmi;
+    gradv[i] = gvi;
+
+    const float mh =      gmi*beta1h;
+    const float vh = sqrt(gvi*beta2h) + eps;
+
+    x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh;
+}

ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp

@@ -0,0 +1,37 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    // Destination multi-index (inlined dst_idx)
+    const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L);
+    const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
+    const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L);
+    const uint i12_offset = i12*p.ne11*p.ne10;
+    const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L);
+    const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
+    const uint d_idx = i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10;
+
+    // Accumulate from sources
+    A_TYPE acc = A_TYPE(0);
+    for (uint i3 = i13; i3 < p.ne03; i3 += p.ne13) {
+        for (uint i2 = i12; i2 < p.ne02; i2 += p.ne12) {
+            for (uint i1 = i11; i1 < p.ne01; i1 += p.ne11) {
+                for (uint i0 = i10; i0 < p.ne00; i0 += p.ne10) {
+                    acc += data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00];
+                }
+            }
+        }
+    }
+
+    data_d[get_doffset() + d_idx] = D_TYPE(acc);
+}

ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp

@@ -25,6 +25,10 @@ layout (push_constant) uniform parameter {
     float corr_dims[2];
     float theta_scale;
     uint has_ff;
+    uint ne02;
+    uint s1;
+    uint s2;
+    int sections[4];
 } p;
 
 float rope_yarn_ramp(const float low, const float high, const uint i0) {

ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp

@@ -0,0 +1,60 @@
+#version 450
+
+#include "rope_head.comp"
+
+void main() {
+    const uint i0 = 2*gl_GlobalInvocationID.y;
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
+    uint ne2 = p.ne02;
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const uint row_dst = gl_GlobalInvocationID.x;
+
+    if (i0 >= p.n_dims) {
+        const uint i = row_dst*ne0 + i0;
+
+        data_d[i + 0] = data_a[i + 0];
+        data_d[i + 1] = data_a[i + 1];
+
+        return;
+    }
+
+    const uint row_x     = row_dst % ne1;
+    const uint channel_x = row_dst / ne1;
+
+    const uint idst = row_dst*ne0 + i0/2;
+    const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0/2;
+
+    const int sect_dims = p.sections[0] + p.sections[1] + p.sections[2] + p.sections[3];
+    const int sec_w = p.sections[1] + p.sections[0];
+    const uint sector = (i0 / 2) % sect_dims;
+
+    float theta_base = 0.0;
+    if (sector < p.sections[0]) {
+        theta_base = data_pos[channel_x]*pow(p.theta_scale, i0/2.0f);
+    }
+    else if (sector >= p.sections[0] && sector < sec_w) {
+        theta_base = data_pos[channel_x + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+    }
+    else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
+        theta_base = data_pos[channel_x + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+    }
+    else if (sector >= sec_w + p.sections[2]) {
+        theta_base = data_pos[channel_x + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+    }
+
+    const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta);
+
+    const float x0 = float(data_a[ix + 0]);
+    const float x1 = float(data_a[ix + p.n_dims/2]);
+
+    data_d[idst + 0]          = D_TYPE(x0*cos_theta - x1*sin_theta);
+    data_d[idst + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta);
+}

ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp

@@ -3,15 +3,18 @@
 #include "rope_head.comp"
 
 void main() {
-    const uint col = gl_GlobalInvocationID.y * 2;
-    const uint row = gl_GlobalInvocationID.x;
+    const uint i0 = 2*gl_GlobalInvocationID.y;
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
 
-    if (col >= p.ncols) {
+    if (i0 >= ne0) {
         return;
     }
 
-    if (col >= p.n_dims) {
-        const uint i = row*p.ncols + col;
+    const uint row_dst = gl_GlobalInvocationID.x;
+
+    if (i0 >= p.n_dims) {
+        const uint i = row_dst*ne0 + i0;
 
         data_d[i + 0] = data_a[i + 0];
         data_d[i + 1] = data_a[i + 1];
@@ -19,19 +22,22 @@ void main() {
         return;
     }
 
-    const uint i  = row*p.ncols + col/2;
-    const uint i2 = row/p.p_delta_rows;
+    const uint row_x     = row_dst % ne1;
+    const uint channel_x = row_dst / ne1;
 
-    const float theta_base = data_pos[i2] * pow(p.theta_scale, col/2.0f);
+    const uint idst = row_dst*ne0 + i0/2;
+    const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0/2;
 
-    const float freq_factor = p.has_ff != 0 ? data_ff[col/2] : 1.0f;
+    const float theta_base = data_pos[channel_x] * pow(p.theta_scale, i0/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f;
 
     float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, col, cos_theta, sin_theta);
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta);
 
-    const float x0 = float(data_a[i + 0]);
-    const float x1 = float(data_a[i + p.n_dims/2]);
+    const float x0 = float(data_a[ix + 0]);
+    const float x1 = float(data_a[ix + p.n_dims/2]);
 
-    data_d[i + 0]        = D_TYPE(x0*cos_theta - x1*sin_theta);
-    data_d[i + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta);
+    data_d[idst + 0]          = D_TYPE(x0*cos_theta - x1*sin_theta);
+    data_d[idst + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta);
 }

ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp

@@ -3,15 +3,18 @@
 #include "rope_head.comp"
 
 void main() {
-    const uint col = gl_GlobalInvocationID.y * 2;
-    const uint row = gl_GlobalInvocationID.x;
+    const uint i0 = 2*gl_GlobalInvocationID.y;
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
 
-    if (col >= p.ncols) {
+    if (i0 >= ne0) {
         return;
     }
 
-    if (col >= p.n_dims) {
-        const uint i = row*p.ncols + col;
+    const uint row_dst = gl_GlobalInvocationID.x;
+
+    if (i0 >= p.n_dims) {
+        const uint i = row_dst*ne0 + i0;
 
         data_d[i + 0] = data_a[i + 0];
         data_d[i + 1] = data_a[i + 1];
@@ -19,19 +22,22 @@ void main() {
         return;
     }
 
-    const uint i = row*p.ncols + col;
-    const uint i2 = row/p.p_delta_rows;
+    const uint row_x     = row_dst % ne1;
+    const uint channel_x = row_dst / ne1;
 
-    const float theta_base = data_pos[i2] * pow(p.theta_scale, col/2.0f);
+    const uint idst = row_dst*ne0 + i0;
+    const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0;
 
-    const float freq_factor = p.has_ff != 0 ? data_ff[col/2] : 1.0f;
+    const float theta_base = data_pos[channel_x] * pow(p.theta_scale, i0/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f;
 
     float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, col, cos_theta, sin_theta);
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta);
 
-    const float x0 = float(data_a[i + 0]);
-    const float x1 = float(data_a[i + 1]);
+    const float x0 = float(data_a[ix + 0]);
+    const float x1 = float(data_a[ix + 1]);
 
-    data_d[i + 0] = D_TYPE(x0*cos_theta - x1*sin_theta);
-    data_d[i + 1] = D_TYPE(x0*sin_theta + x1*cos_theta);
+    data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta);
+    data_d[idst + 1] = D_TYPE(x0*sin_theta + x1*cos_theta);
 }

ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp

@@ -0,0 +1,47 @@
+#version 450
+
+#include "rope_head.comp"
+
+void main() {
+    const uint i0 = 2*gl_GlobalInvocationID.y;
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
+    uint ne2 = p.ne02;
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const uint row_dst = gl_GlobalInvocationID.x;
+
+    const uint row_x     = row_dst % ne1;
+    const uint channel_x = row_dst / ne1;
+
+    const uint idst = row_dst*ne0 + i0/2;
+    const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0/2;
+
+    const int sect_dims = p.sections[0] + p.sections[1];
+    const int sec_w = p.sections[1] + p.sections[0];
+    const uint sector = (i0 / 2) % sect_dims;
+
+    float theta_base = 0.0;
+    if (sector < p.sections[0]) {
+        const uint p0 = sector;
+        theta_base = data_pos[channel_x]*pow(p.theta_scale, p0);
+    }
+    else if (sector >= p.sections[0] && sector < sec_w) {
+        const uint p0 = sector - p.sections[0];
+        theta_base = data_pos[channel_x + ne2]*pow(p.theta_scale, p0);
+    }
+
+    const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta);
+
+    const float x0 = float(data_a[ix + 0]);
+    const float x1 = float(data_a[ix + p.n_dims]);
+
+    data_d[idst + 0]        = D_TYPE(x0*cos_theta - x1*sin_theta);
+    data_d[idst + p.n_dims] = D_TYPE(x0*sin_theta + x1*cos_theta);
+}

ggml/src/ggml-vulkan/vulkan-shaders/sub.comp

@@ -0,0 +1,29 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+const uint num_threads = 256;
+
+layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    uint idx = get_idx();
+
+    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
+    const uint num_iter = 2;
+
+    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
+        if (idx >= p.ne) {
+            continue;
+        }
+        uint i00, i01, i02, i03;
+        get_indices(idx, i00, i01, i02, i03);
+
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) - FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
+
+        idx += num_threads;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/types.comp

@@ -294,6 +294,187 @@ struct block_q6_K_packed16
 
 // IQuants
 
+#define QUANT_K_IQ1_S 256
+#define QUANT_R_IQ1_S 1
+
+struct block_iq1_s {
+    float16_t d;
+    uint8_t  qs[QUANT_K_IQ1_S/8];
+    uint16_t qh[QUANT_K_IQ1_S/32];
+};
+
+#define QUANT_K_IQ1_M 256
+#define QUANT_R_IQ1_M 1
+
+struct block_iq1_m {
+    uint8_t  qs[QUANT_K_IQ1_M/8];
+    uint8_t  qh[QUANT_K_IQ1_M/16];
+    uint16_t scales[QUANT_K_IQ1_M/64];
+};
+
+#if defined(DATA_A_IQ1_S)
+#define QUANT_K QUANT_K_IQ1_S
+#define QUANT_R QUANT_R_IQ1_S
+#define A_TYPE block_iq1_s
+#endif
+
+#if defined(DATA_A_IQ1_M)
+#define QUANT_K QUANT_K_IQ1_M
+#define QUANT_R QUANT_R_IQ1_M
+#define A_TYPE block_iq1_m
+#endif
+
+#if defined(DATA_A_IQ1_S) || defined(DATA_A_IQ1_M)
+#define IQ1S_DELTA 0.125f
+#define IQ1M_DELTA 0.125f
+
+// Packed IQ1S grid where every 2 vec8 are encoded on 32 bits (2 bits per coordinate).
+const uint[1024] iq1s_grid_const = {
+    0xfffdffff, 0xfff7fff0, 0xffccfff5, 0xffdfffc0, 0xffd7ffdd, 0xff30ffd5, 0xff03ff0c, 0xff10ff01,
+    0xff7dff7f, 0xff75ff77, 0xff5fff40, 0xff57ff5d, 0xfcf3ff55, 0xfcccfcf0, 0xfcc1fcc3, 0xfcc5fcc4,
+    0xfc3cfcd0, 0xfc34fc31, 0xfc00fc0d, 0xfc1cfc05, 0xfc11fc13, 0xfc70fc17, 0xfc43fc4c, 0xfc50fc41,
+    0xfdfdfdff, 0xfdf5fdf7, 0xfddffdc0, 0xfdd7fddd, 0xfd30fdd5, 0xfd04fd0c, 0xfd14fd13, 0xfd7dfd7f,
+    0xfd75fd77, 0xfd40fd4c, 0xfd5ffd44, 0xfd57fd5d, 0xf3ccfd55, 0xf3c1f3c3, 0xf33cf3d0, 0xf300f334,
+    0xf313f305, 0xf34cf310, 0xf350f344, 0xf0f3f0fc, 0xf0f1f0f0, 0xf0c7f0c0, 0xf0d4f0c5, 0xf030f03f,
+    0xf00ff035, 0xf003f00c, 0xf001f000, 0xf01ff004, 0xf010f01d, 0xf015f017, 0xf04cf07c, 0xf047f040,
+    0xf05cf045, 0xf050f053, 0xf054f051, 0xf1c4f1c3, 0xf133f13c, 0xf10df10f, 0xf107f100, 0xf11cf11f,
+    0xf114f111, 0xf14cf170, 0xf144f143, 0xf7fdf7ff, 0xf7f5f7f7, 0xf7dff7c0, 0xf7d7f7dd, 0xf730f7d5,
+    0xf701f70c, 0xf77ff710, 0xf777f77d, 0xf740f775, 0xf75df75f, 0xf755f757, 0xf4ccf4f0, 0xf4c4f4c3,
+    0xf4d0f4d3, 0xf40ff43c, 0xf400f40c, 0xf413f41c, 0xf44cf414, 0xf441f443, 0xf450f444, 0xf5fdf5ff,
+    0xf5f5f5f7, 0xf5dff5c0, 0xf5d7f5dd, 0xf530f5d5, 0xf504f50c, 0xf510f51c, 0xf57df57f, 0xf577f570,
+    0xf540f575, 0xf55df55f, 0xf555f557, 0xcfcccfcf, 0xcfc4cfc3, 0xcfd0cfd3, 0xcf33cf3c, 0xcf00cf0f,
+    0xcf1ccf07, 0xcf10cf13, 0xcf4ccf14, 0xcf41cf43, 0xcf50cf5c, 0xccf3ccfc, 0xccf4ccf1, 0xcccdcccf,
+    0xccc7ccc0, 0xccd3ccdc, 0xcc30ccd4, 0xcc0fcc35, 0xcc0dcc0c, 0xcc00cc03, 0xcc04cc01, 0xcc10cc1f,
+    0xcc4dcc73, 0xcc5ccc40, 0xcdcccc53, 0xcdc1cdc3, 0xcd3fcdd0, 0xcd34cd31, 0xcd00cd0d, 0xcd05cd07,
+    0xcd11cd13, 0xcd4ccd70, 0xcd41cd43, 0xc3fccd50, 0xc3f4c3f1, 0xc3c0c3c3, 0xc3c4c3c7, 0xc3d1c3dc,
+    0xc330c33c, 0xc337c331, 0xc30cc335, 0xc300c303, 0xc304c301, 0xc310c31d, 0xc373c317, 0xc34fc374,
+    0xc340c343, 0xc344c347, 0xc35cc345, 0xc350c353, 0xc0fdc354, 0xc0f5c0f0, 0xc0c3c0cc, 0xc0c1c0c0,
+    0xc0dfc0c4, 0xc0d0c0dd, 0xc0d5c0d7, 0xc033c03c, 0xc031c030, 0xc00dc00c, 0xc000c003, 0xc004c001,
+    0xc01cc005, 0xc010c013, 0xc014c011, 0xc07dc07f, 0xc070c073, 0xc075c077, 0xc04cc04f, 0xc040c043,
+    0xc044c041, 0xc05fc045, 0xc050c05d, 0xc1f3c1fc, 0xc1f1c1f0, 0xc1c1c1c0, 0xc1c5c1c7, 0xc1d1c1dc,
+    0xc13dc13f, 0xc130c133, 0xc135c137, 0xc100c10c, 0xc107c101, 0xc11cc104, 0xc110c113, 0xc114c117,
+    0xc171c115, 0xc14dc175, 0xc153c140, 0xc7ccc154, 0xc7d0c7c1, 0xc733c73c, 0xc734c731, 0xc700c70f,
+    0xc705c707, 0xc71cc71f, 0xc711c713, 0xc770c714, 0xc743c74c, 0xc4cfc750, 0xc4c0c4cd, 0xc4dcc4c5,
+    0xc43dc4d0, 0xc430c433, 0xc40cc437, 0xc400c403, 0xc404c401, 0xc41fc405, 0xc415c410, 0xc44cc474,
+    0xc440c44d, 0xc45cc447, 0xc454c451, 0xc5c1c5f4, 0xc5d1c5d3, 0xc531c533, 0xc50fc534, 0xc500c50d,
+    0xc51cc507, 0xc514c511, 0xc54cc570, 0xc545c541, 0xdffddfff, 0xdff5dff7, 0xdfdfdfc0, 0xdfd0dfdd,
+    0xdfd5dfd7, 0xdf0cdf30, 0xdf1cdf04, 0xdf7fdf10, 0xdf77df7d, 0xdf40df75, 0xdf5ddf5f, 0xdf57df50,
+    0xdcf0df55, 0xdcc3dccc, 0xdcd0dcc4, 0xdc33dc3d, 0xdc00dc34, 0xdc05dc07, 0xdc13dc1c, 0xdc11dc10,
+    0xdc4fdc70, 0xdc44dc41, 0xddfcdc50, 0xddf5ddf7, 0xddc0ddcc, 0xdddddddf, 0xddd5ddd7, 0xdd0cdd30,
+    0xdd04dd01, 0xdd7cdd10, 0xdd75dd77, 0xdd40dd4c, 0xdd5ddd5f, 0xdd55dd57, 0xd3c3d3f0, 0xd3c4d3c1,
+    0xd333d3d0, 0xd331d330, 0xd30dd334, 0xd307d300, 0xd311d305, 0xd34cd370, 0xd344d343, 0xd350d35c,
+    0xd0c0d0f4, 0xd0d4d0dc, 0xd030d03f, 0xd00cd037, 0xd000d003, 0xd01dd004, 0xd017d010, 0xd04fd074,
+    0xd040d043, 0xd045d047, 0xd053d05c, 0xd054d051, 0xd1cfd1f0, 0xd1c4d1cd, 0xd13cd1d0, 0xd100d134,
+    0xd11cd11f, 0xd173d114, 0xd14fd171, 0xd7ffd145, 0xd7f7d7fd, 0xd7c0d7f5, 0xd7ddd7df, 0xd7d5d7d7,
+    0xd70cd730, 0xd710d703, 0xd77dd77f, 0xd775d777, 0xd75dd75f, 0xd755d757, 0xd4ccd4f4, 0xd4c4d4c3,
+    0xd431d4d0, 0xd40dd434, 0xd41cd400, 0xd411d413, 0xd470d414, 0xd441d44f, 0xd453d444, 0xd5ffd450,
+    0xd5f7d5fd, 0xd5dfd5f5, 0xd5d7d5dd, 0xd530d5d5, 0xd501d50c, 0xd510d504, 0xd57dd57f, 0xd575d577,
+    0xd55fd540, 0xd557d55d, 0x3ff0d555, 0x3fc13fcc, 0x3f343fd0, 0x3f003f0d, 0x3f053f07, 0x3f133f1c,
+    0x3f433f11, 0x3f5c3f44, 0x3cff3f51, 0x3cf33cfc, 0x3cf43cf1, 0x3cc03ccd, 0x3cc73cc1, 0x3cdc3cc5,
+    0x3cd43cd1, 0x3c373c30, 0x3c0c3c35, 0x3c003c03, 0x3c043c01, 0x3c103c05, 0x3c153c17, 0x3c733c7c,
+    0x3c4f3c71, 0x3c403c4d, 0x3c5c3c5f, 0x3df03c5d, 0x3dc33dcc, 0x3dd03dc1, 0x3d0d3d3c, 0x3d053d00,
+    0x3d143d13, 0x3d433d74, 0x33fc3d50, 0x33c433c0, 0x333033d4, 0x33353337, 0x3303330c, 0x33013300,
+    0x331d331c, 0x33173310, 0x337c3315, 0x33743371, 0x334d334f, 0x335f3340, 0x3354335c, 0x30fd30fc,
+    0x30f530f0, 0x30c330cc, 0x30c130c0, 0x30df30c4, 0x30d530d0, 0x3033303c, 0x30313030, 0x300f3034,
+    0x3003300c, 0x30013000, 0x30043007, 0x3013301c, 0x30113010, 0x307d3014, 0x30703073, 0x304c3077,
+    0x30403043, 0x30443041, 0x30503045, 0x30553057, 0x31f031fc, 0x31c331f4, 0x31c731c0, 0x31dc31c5,
+    0x31d431d3, 0x313d313f, 0x31373130, 0x310c310f, 0x3100310d, 0x31043101, 0x3110311d, 0x317c3117,
+    0x31753170, 0x31403143, 0x3153315c, 0x37f03151, 0x37c037cc, 0x37d037c5, 0x3734373d, 0x3700370f,
+    0x371c3707, 0x37113713, 0x37703714, 0x3743374c, 0x37443741, 0x34fc3750, 0x34f134f0, 0x34cf34f5,
+    0x34c034c3, 0x34dc34c7, 0x34d134d3, 0x3430343f, 0x340c3435, 0x3403340d, 0x34013400, 0x341f3404,
+    0x3410341d, 0x34153411, 0x34743471, 0x3440344d, 0x34473441, 0x3453345c, 0x34543451, 0x353335c1,
+    0x35343531, 0x35073500, 0x35133505, 0x35433514, 0x0ffc3550, 0x0ff00ff3, 0x0ff40ff1, 0x0fc00fcd,
+    0x0fdc0fc5, 0x0fd40fd3, 0x0f300f3f, 0x0f0c0f37, 0x0f000f03, 0x0f040f01, 0x0f170f10, 0x0f740f71,
+    0x0f470f40, 0x0f5c0f5f, 0x0f540f51, 0x0cf70cf0, 0x0cf50cf4, 0x0cc30ccc, 0x0cc10cc0, 0x0cc40cc7,
+    0x0cd00cdf, 0x0cd70cd1, 0x0c3c0cd5, 0x0c300c33, 0x0c340c31, 0x0c0c0c0f, 0x0c030c0d, 0x0c010c00,
+    0x0c040c07, 0x0c1c0c05, 0x0c100c13, 0x0c140c11, 0x0c700c7d, 0x0c430c4c, 0x0c410c40, 0x0c5f0c44,
+    0x0c550c50, 0x0df10dfc, 0x0dc00dcd, 0x0ddc0dc5, 0x0d3d0dd3, 0x0d350d30, 0x0d030d0c, 0x0d010d00,
+    0x0d1d0d04, 0x0d700d10, 0x0d4d0d4f, 0x0d440d40, 0x0d530d45, 0x03f003f3, 0x03c303cc, 0x03c103c0,
+    0x03c403c7, 0x03d003dc, 0x03d503d7, 0x0333033c, 0x03310330, 0x03350334, 0x030c030f, 0x03000303,
+    0x03070301, 0x03050304, 0x031d031c, 0x03100313, 0x03140311, 0x0377037f, 0x034c0375, 0x03400343,
+    0x03440341, 0x0353035c, 0x03550350, 0x00fd00fc, 0x00f000f3, 0x00f400f1, 0x00cc00cf, 0x00c300cd,
+    0x00c100c0, 0x00c500c4, 0x00d300dc, 0x00d100d0, 0x003f00d4, 0x003d003c, 0x00300033, 0x00370031,
+    0x000f0034, 0x000d000c, 0x00000003, 0x00070001, 0x00050004, 0x001c001f, 0x00100013, 0x00170011,
+    0x00150014, 0x0073007c, 0x00740070, 0x004f0075, 0x0043004c, 0x00410040, 0x00440047, 0x0053005c,
+    0x00510050, 0x01ff0054, 0x01fd01fc, 0x01f101f3, 0x01f401f7, 0x01c301cc, 0x01c701c0, 0x01df01c4,
+    0x01dd01dc, 0x01d001d3, 0x01d701d1, 0x013c01d4, 0x01310130, 0x01340137, 0x010f0135, 0x010d010c,
+    0x01000103, 0x01070101, 0x01050104, 0x0113011c, 0x01140110, 0x0170017d, 0x01770171, 0x01750174,
+    0x0140014c, 0x015d0145, 0x01510150, 0x01540157, 0x07f007f3, 0x07f407f1, 0x07c007cf, 0x07dc07c7,
+    0x073007d5, 0x07350737, 0x0703070c, 0x07010700, 0x07040707, 0x071d071f, 0x07100713, 0x0774077d,
+    0x074d074f, 0x07470740, 0x0754075c, 0x04fd04fc, 0x04f504f0, 0x04c304cc, 0x04c104c0, 0x04d004c4,
+    0x0433043c, 0x04310430, 0x040f0434, 0x040d040c, 0x04000403, 0x04070401, 0x04050404, 0x0413041c,
+    0x04110410, 0x047c0414, 0x04740470, 0x0443044c, 0x04410440, 0x04440447, 0x05f30450, 0x05c005f7,
+    0x05df05c5, 0x05d105d0, 0x053005d4, 0x05340537, 0x0500050c, 0x05070501, 0x051d0504, 0x05170510,
+    0x057c0515, 0x054d0575, 0x05410540, 0x05450547, 0x1ff0055c, 0x1fc11fc3, 0x1fd01fc4, 0x1f0f1f33,
+    0x1f011f00, 0x1f051f07, 0x1f131f1c, 0x1f141f11, 0x1f411f7c, 0x1cfc1f50, 0x1cf11cf3, 0x1ccd1cf4,
+    0x1cdc1cc0, 0x1cd11cdd, 0x1c301cd4, 0x1c0c1c34, 0x1c011c00, 0x1c101c04, 0x1c151c11, 0x1c751c73,
+    0x1c401c4d, 0x1c511c5c, 0x1dcc1c54, 0x1dc41dc1, 0x1d3c1d3f, 0x1d001d31, 0x1d071d01, 0x1d701d1f,
+    0x1d411d4c, 0x13cc1d50, 0x13c013cd, 0x13c513c1, 0x13d113dc, 0x133f13d4, 0x1330133d, 0x13351337,
+    0x1303130c, 0x13011300, 0x13051304, 0x131d131f, 0x13731310, 0x13741370, 0x134d134f, 0x13401343,
+    0x13471341, 0x135c1345, 0x13541353, 0x10f710f0, 0x10cc10f5, 0x10c110c0, 0x103310c4, 0x10311030,
+    0x100f1034, 0x1003100c, 0x10011000, 0x101c1004, 0x10101013, 0x10141011, 0x10741071, 0x104c1075,
+    0x10411040, 0x10451044, 0x1050105d, 0x10571051, 0x11f411fd, 0x11df11c0, 0x11d711d1, 0x113f11d4,
+    0x11371130, 0x110c1135, 0x11001103, 0x11071101, 0x111f1105, 0x11171110, 0x117d117f, 0x11751170,
+    0x11411143, 0x11441147, 0x1153115f, 0x11551151, 0x17c417c1, 0x173c17d0, 0x1700170d, 0x171c1705,
+    0x17701714, 0x1747174c, 0x14fc1751, 0x14cf14f3, 0x14dc14c0, 0x14d114d3, 0x143f14d4, 0x1430143c,
+    0x14371431, 0x1403140c, 0x14011400, 0x141f1404, 0x14151410, 0x1473147d, 0x14401475, 0x1453145c,
+    0x14541450, 0x15c115cc, 0x153c15c7, 0x15341533, 0x1500150f, 0x15051507, 0x15101513, 0x15711514,
+    0x15471543, 0x15511545, 0x7ffd7fff, 0x7ff57ff7, 0x7fdd7fdf, 0x7fd57fd7, 0x7f0f7f30, 0x7f037f0c,
+    0x7f047f01, 0x7f7f7f10, 0x7f777f7d, 0x7f407f75, 0x7f5d7f5f, 0x7f557f57, 0x7ccc7cf0, 0x7cc17cc3,
+    0x7cd07cc4, 0x7c337c3c, 0x7c0f7c34, 0x7c007c0d, 0x7c077c01, 0x7c137c04, 0x7c147c11, 0x7c747c70,
+    0x7c417c43, 0x7c507c44, 0x7dfd7dff, 0x7df57df7, 0x7ddf7dc0, 0x7dd77ddd, 0x7d0c7dd5, 0x7d047d03,
+    0x7d7f7d10, 0x7d777d7d, 0x7d407d75, 0x7d5d7d5f, 0x7d557d57, 0x73c473c3, 0x7333733c, 0x7300730c,
+    0x731c7305, 0x73147313, 0x73447343, 0x70f470fc, 0x70c070cd, 0x70d170c5, 0x703f70d4, 0x7030703c,
+    0x700c7037, 0x70007003, 0x70047001, 0x70107005, 0x70177011, 0x707c7015, 0x70717073, 0x704f7074,
+    0x7040704d, 0x70517047, 0x71c171cc, 0x71d071c4, 0x7133713c, 0x71357134, 0x7100710f, 0x71057104,
+    0x7111711c, 0x71707115, 0x7145714c, 0x77ff7153, 0x77f777fd, 0x77c077f5, 0x77dd77df, 0x77d577d7,
+    0x7730773c, 0x7703770c, 0x77107704, 0x777f7714, 0x7777777d, 0x77407775, 0x775d775f, 0x77557757,
+    0x74f174f0, 0x74c374cc, 0x74d074c1, 0x7433743c, 0x74347431, 0x740d740f, 0x74057400, 0x7413741c,
+    0x74417470, 0x74507444, 0x75fd75ff, 0x75f575f7, 0x75df75c0, 0x75d775dd, 0x753075d5, 0x7503750c,
+    0x757f7501, 0x7577757d, 0x75407575, 0x755d755f, 0x75557557, 0x4fcc4ff0, 0x4fc74fc1, 0x4fd04fc4,
+    0x4f314f3c, 0x4f004f34, 0x4f054f07, 0x4f154f14, 0x4f4c4f70, 0x4f414f43, 0x4f504f44, 0x4cf34cfc,
+    0x4cf44cf1, 0x4cc04ccf, 0x4cc54cc7, 0x4cd34cdc, 0x4cd44cd1, 0x4c304c3f, 0x4c0c4c0f, 0x4c004c03,
+    0x4c044c01, 0x4c104c1d, 0x4c714c73, 0x4c404c4d, 0x4c5c4c47, 0x4c514c53, 0x4df04c54, 0x4dc34dcc,
+    0x4dd04dc4, 0x4d314d33, 0x4d0f4d34, 0x4d004d0d, 0x4d114d07, 0x4d704d14, 0x4d414d43, 0x43fc4d54,
+    0x43f143f3, 0x43c043cf, 0x43d143c7, 0x4335433f, 0x4303430c, 0x43014300, 0x43044307, 0x431c431f,
+    0x4310431d, 0x43714373, 0x4343434d, 0x43474340, 0x4354435c, 0x40f040ff, 0x40f540f7, 0x40cc40cf,
+    0x40c040c3, 0x40c440c1, 0x40d040dc, 0x40d540d4, 0x4033403c, 0x40314030, 0x400f4034, 0x400d400c,
+    0x40004003, 0x40074001, 0x40054004, 0x4013401c, 0x40114010, 0x407c4014, 0x40774070, 0x404d404c,
+    0x40404043, 0x40444041, 0x405f4045, 0x4050405d, 0x40554057, 0x41f341fc, 0x41c041cf, 0x41df41c4,
+    0x41d441d1, 0x41374130, 0x410c4134, 0x4100410d, 0x41044101, 0x41174110, 0x4173417d, 0x41754174,
+    0x4143414d, 0x41534140, 0x41544151, 0x47c147f0, 0x47d047c4, 0x4731473c, 0x470d470f, 0x47014700,
+    0x47134705, 0x47704710, 0x4741474c, 0x47504744, 0x44f144f3, 0x44cf44f4, 0x44c044cd, 0x44c544c7,
+    0x44dc44df, 0x44d144d3, 0x443d443f, 0x44374430, 0x440c4435, 0x44004403, 0x44044401, 0x4410441d,
+    0x44154411, 0x4473447c, 0x444d444f, 0x44454440, 0x4451445c, 0x45c045f0, 0x453345d0, 0x45344531,
+    0x4500450f, 0x451c4507, 0x454c4570, 0x45404543, 0x5fff4541, 0x5ff75ffd, 0x5fc05ff5, 0x5fdd5fdf,
+    0x5fd55fd7, 0x5f0c5f30, 0x5f015f03, 0x5f7f5f04, 0x5f775f7d, 0x5f405f75, 0x5f5d5f5f, 0x5f555f57,
+    0x5cf45cf0, 0x5cc35ccc, 0x5cc45cc1, 0x5c315cc5, 0x5c0c5c34, 0x5c075c00, 0x5c1c5c05, 0x5c705c13,
+    0x5c4d5c4f, 0x5c445c41, 0x5df75dfd, 0x5dcf5df5, 0x5ddd5dc4, 0x5dd55dd7, 0x5d0c5d30, 0x5d045d01,
+    0x5d7f5d10, 0x5d775d7d, 0x5d405d75, 0x5d5d5d5f, 0x5d555d57, 0x53d053c4, 0x5333533c, 0x5303530f,
+    0x53075300, 0x531c5305, 0x53115310, 0x53145317, 0x50f15370, 0x50cf50f4, 0x50c050cd, 0x50d150c7,
+    0x503d50d4, 0x500c5030, 0x50005003, 0x50045001, 0x50155010, 0x5073507c, 0x50715070, 0x504d5074,
+    0x50475040, 0x51cc51f0, 0x51c551c1, 0x51d051dc, 0x51315133, 0x510d5135, 0x51015100, 0x511f5107,
+    0x5171511d, 0x5140514f, 0x51445141, 0x5153515c, 0x57ff5151, 0x57f757fd, 0x57df57f5, 0x57d757dd,
+    0x570c57d5, 0x57015703, 0x577f5704, 0x5777577d, 0x57405775, 0x575d575f, 0x57555757, 0x54c354f0,
+    0x54dc54c4, 0x543c54d0, 0x5400540f, 0x541c5405, 0x54145411, 0x5441544f, 0x55fd55ff, 0x55f555f7,
+    0x55dd55df, 0x55d555d7, 0x5503550c, 0x557f5501, 0x5577557d, 0x55405575, 0x555d555f, 0x55555557
+};
+
+shared uint16_t iq1s_grid[2048];
+
+#define NEEDS_INIT_IQ_SHMEM
+void init_iq_shmem(uvec3 wgsize)
+{
+    // copy the table into shared memory and sync
+    for (uint i = gl_LocalInvocationIndex.x; i < iq1s_grid_const.length(); i += wgsize.x) {
+        u16vec2 g = unpack16(iq1s_grid_const[i]);
+        iq1s_grid[2*i+0] = g.x;
+        iq1s_grid[2*i+1] = g.y;
+    }
+    barrier();
+}
+#endif
+
 #define QUANT_K_IQ2_XXS 256
 #define QUANT_R_IQ2_XXS 1
 
@@ -380,6 +561,7 @@ const uvec2[256] iq2xxs_grid_const = {
 
 shared uvec2 iq2xxs_grid[256];
 
+#define NEEDS_INIT_IQ_SHMEM
 void init_iq_shmem(uvec3 wgsize)
 {
     // copy the table into shared memory and sync
@@ -547,6 +729,7 @@ const uvec2 iq2xs_grid_const[512] = {
 
 shared uvec2 iq2xs_grid[512];
 
+#define NEEDS_INIT_IQ_SHMEM
 void init_iq_shmem(uvec3 wgsize)
 {
     // copy the table into shared memory and sync
@@ -836,6 +1019,7 @@ const uvec2 iq2s_grid_const[1024] = {
 
 shared uvec2 iq2s_grid[1024];
 
+#define NEEDS_INIT_IQ_SHMEM
 void init_iq_shmem(uvec3 wgsize)
 {
     // copy the table into shared memory and sync
@@ -904,6 +1088,7 @@ const uint32_t iq3xxs_grid_const[256] = {
 
 shared uint32_t iq3xxs_grid[256];
 
+#define NEEDS_INIT_IQ_SHMEM
 void init_iq_shmem(uvec3 wgsize)
 {
     // copy the table into shared memory and sync
@@ -1011,6 +1196,7 @@ const uint32_t iq3s_grid_const[512] = {
 
 shared uint32_t iq3s_grid[512];
 
+#define NEEDS_INIT_IQ_SHMEM
 void init_iq_shmem(uvec3 wgsize)
 {
     // copy the table into shared memory and sync
@@ -1073,6 +1259,7 @@ const int8_t kvalues_iq4nl_const[16] = {
 
 shared FLOAT_TYPE kvalues_iq4nl[16];
 
+#define NEEDS_INIT_IQ_SHMEM
 void init_iq_shmem(uvec3 wgsize)
 {
     // copy the table into shared memory and sync