server: fix correct time_ms calculation in prompt_progress (#17093)

* fix: correct time_ms calculation in send_partial_response

The time_ms field was incorrectly calculated. The division was happening
before the subtraction leading to incorrect values.

Before: (ggml_time_us() - slot.t_start_process_prompt / 1000) After:
(ggml_time_us() - slot.t_start_process_prompt) / 1000

* docs : document time_ms field in prompt_progress

.github/workflows/build.yml

@@ -161,15 +161,16 @@ jobs:
       - name: Dawn Dependency
         id: dawn-depends
         run: |
-          DAWN_VERSION="v1.0.0"
+          DAWN_VERSION="v2.0.0"
           DAWN_OWNER="reeselevine"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-a1a6b45cced25a3b7f4fb491e0ae70796cc7f22b-macos-latest-Release.tar.gz"
+          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release.zip"
           echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
-          curl -L -o artifact.tar.gz \
+          curl -L -o artifact.zip \
             "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
           mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
+          unzip artifact.zip
+          tar -xvf Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release.tar.gz -C dawn --strip-components=1
 
       - name: Build
         id: cmake_build
@@ -521,15 +522,16 @@ jobs:
         id: dawn-depends
         run: |
           sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
-          DAWN_VERSION="v1.0.0"
+          DAWN_VERSION="v2.0.0"
           DAWN_OWNER="reeselevine"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-a1a6b45cced25a3b7f4fb491e0ae70796cc7f22b-ubuntu-latest-Release.tar.gz"
+          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-ubuntu-latest-Release.zip"
           echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
-          curl -L -o artifact.tar.gz \
+          curl -L -o artifact.zip \
             "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
           mkdir dawn
-          tar -xvf artifact.tar.gz -C dawn --strip-components=1
+          unzip artifact.zip
+          tar -xvf Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-ubuntu-latest-Release.tar.gz -C dawn --strip-components=1
 
       - name: Build
         id: cmake_build

common/CMakeLists.txt

@@ -56,6 +56,8 @@ add_library(${TARGET} STATIC
     common.h
     console.cpp
     console.h
+    download.cpp
+    download.h
     http.h
     json-partial.cpp
     json-partial.h

common/arg.h

@@ -59,8 +59,8 @@ struct common_arg {
     common_arg & set_sparam();
     bool in_example(enum llama_example ex);
     bool is_exclude(enum llama_example ex);
-    bool get_value_from_env(std::string & output);
-    bool has_value_from_env();
+    bool get_value_from_env(std::string & output) const;
+    bool has_value_from_env() const;
     std::string to_string();
 };
 

common/common.h

@@ -507,6 +507,10 @@ struct common_params {
     // return false from callback to abort model loading or true to continue
     llama_progress_callback load_progress_callback = NULL;
     void *                  load_progress_callback_user_data = NULL;
+
+    bool has_speculative() const {
+        return !speculative.model.path.empty() || !speculative.model.hf_repo.empty();
+    }
 };
 
 // call once at the start of a program if it uses libcommon

common/download.h

@@ -0,0 +1,41 @@
+#pragma once
+
+#include <string>
+
+struct common_params_model;
+
+//
+// download functionalities
+//
+
+struct common_hf_file_res {
+    std::string repo; // repo name with ":tag" removed
+    std::string ggufFile;
+    std::string mmprojFile;
+};
+
+// resolve and download model from Docker registry
+// return local path to downloaded model file
+std::string common_docker_resolve_model(const std::string & docker);
+
+/**
+ * Allow getting the HF file from the HF repo with tag (like ollama), for example:
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
+ * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
+ * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
+ *
+ * Return pair of <repo, file> (with "repo" already having tag removed)
+ *
+ * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
+ */
+common_hf_file_res common_get_hf_file(
+    const std::string & hf_repo_with_tag,
+    const std::string & bearer_token,
+    bool offline);
+
+// returns true if download succeeded
+bool common_download_model(
+    const common_params_model & model,
+    const std::string & bearer_token,
+    bool offline);

convert_hf_to_gguf.py

@@ -7187,6 +7187,42 @@ class MiniMaxM2Model(TextModel):
         return super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("PanguEmbeddedForCausalLM")
+class PanguEmbeddedModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.PANGU_EMBED
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+
+        # PanguEmbedded's hparam loaded from config.json without head_dim
+        if (rope_dim := hparams.get("head_dim")) is None:
+            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
+        self.gguf_writer.add_rope_dimension_count(rope_dim)
+
+        if hparams.get("head_dim") is None:
+            self.gguf_writer.add_key_length(rope_dim)
+            self.gguf_writer.add_value_length(rope_dim)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name == "lm_head.weight":
+            if self.hparams.get("tie_word_embeddings", False):
+                logger.info("Skipping tied output layer 'lm_head.weight'")
+                return []
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @ModelBase.register("Dots1ForCausalLM")
 class Dots1Model(Qwen2MoeModel):
     model_arch = gguf.MODEL_ARCH.DOTS1

docs/backend/OPENCL.md

@@ -39,18 +39,23 @@ The llama.cpp OpenCL backend is designed to enable llama.cpp on **Qualcomm Adren
 | Adreno 830 (Snapdragon 8 Elite)      | Support |
 | Adreno X85 (Snapdragon X Elite)      | Support |
 
+> A6x GPUs with a recent driver and compiler are supported; they are usually found in IoT platforms.
+However, A6x GPUs in phones are likely not supported due to the outdated driver and compiler.
+
 ## DataType Supports
 
 | DataType               | Status                     |
 |:----------------------:|:--------------------------:|
 | Q4_0                   | Support                    |
 | Q6_K                   | Support, but not optimized |
+| Q8_0                   | Support                    |
+| MXFP4                  | Support                    |
 
 ## Model Preparation
 
-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration.
+You can refer to the general [llama-quantize tool](/tools/quantize/README.md) for steps to convert a model in Hugging Face safetensor format to GGUF with quantization.
 
-Currently we support `Q4_0` quantization and have optimize for it. To achieve best performance on Adreno GPU, add `--pure` to `llama-quantize`. For example,
+Currently we support `Q4_0` quantization and have optimized for it. To achieve best performance on Adreno GPU, add `--pure` to `llama-quantize` (i.e., make all weights in `Q4_0`). For example,
 
 ```sh
 ./llama-quantize --pure ggml-model-qwen2.5-3b-f16.gguf ggml-model-qwen-3b-Q4_0.gguf Q4_0
@@ -58,6 +63,17 @@ Currently we support `Q4_0` quantization and have optimize for it. To achieve be
 
 Since `Q6_K` is also supported, `Q4_0` quantization without `--pure` will also work. However, the performance will be worse compared to pure `Q4_0` quantization.
 
+### `MXFP4` MoE Models
+
+OpenAI gpt-oss models are MoE models in `MXFP4`. The quantized model will be in `MXFP4_MOE`, a mixture of `MXFP4` and `Q8_0`.
+For this quantization, there is no need to specify `--pure`.
+For gpt-oss-20b model, you can directly [download](https://huggingface.co/ggml-org/gpt-oss-20b-GGUF) the quantized GGUF file in `MXFP4_MOE` from Hugging Face.
+
+Although it is possible to quantize gpt-oss-20b model in pure `Q4_0` (all weights in `Q4_0`), it is not recommended since `MXFP4` has been optimized for MoE while `Q4_0` is not. In addition, accuracy should degrade with such pure `Q4_0` quantization.
+Hence, using the default `MXFP4_MOE` quantization (see the link above) is recommended for this model.
+
+> Note that the `Q4_0` model found [here](https://huggingface.co/unsloth/gpt-oss-20b-GGUF/blob/main/gpt-oss-20b-Q4_0.gguf) is a mixture of `Q4_0`, `Q8_0` and `MXFP4` and gives better performance than `MXFP4_MOE` quantization.
+
 ## CMake Options
 
 The OpenCL backend has the following CMake options that control the behavior of the backend.
@@ -146,10 +162,13 @@ A Snapdragon X Elite device with Windows 11 Arm64 is used. Make sure the followi
 * Ninja
 * Visual Studio 2022
 * Powershell 7
+* Python
 
 Visual Studio provides necessary headers and libraries although it is not directly used for building.
 Alternatively, Visual Studio Build Tools can be installed instead of the full Visual Studio.
 
+> Note that building using Visual Studio's cl compiler is not supported. Clang must be used. Clang depends on libraries provided by Visual Studio to work. Therefore, Visual Studio must be installed. Alternatively, Visual Studio Build Tools can be installed instead of the full Visual Studio.
+
 Powershell 7 is used for the following commands.
 If an older version of Powershell is used, these commands may not work as they are.
 
@@ -201,9 +220,12 @@ ninja
 
 ## Known Issues
 
-- Currently OpenCL backend does not work on Adreno 6xx GPUs.
+- Flash attention does not always improve performance.
+- Currently OpenCL backend works on A6xx GPUs with recent drivers and compilers (usually found in IoT platforms).
+  However, it does not work on A6xx GPUs found in phones with old drivers and compilers.
 
 ## TODO
 
 - Optimization for Q6_K
 - Support and optimization for Q4_K
+- Improve flash attention

docs/build.md

@@ -178,6 +178,48 @@ GeForce RTX 3070      8.6
 cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES="86;89"
 ```
 
+### Overriding the CUDA Version
+
+If you have multiple CUDA installations on your system and want to compile llama.cpp for a specific one, e.g. for CUDA 11.7 installed under `/opt/cuda-11.7`:
+
+```bash
+cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_COMPILER=/opt/cuda-11.7/bin/nvcc -DCMAKE_INSTALL_RPATH="/opt/cuda-11.7/lib64;\$ORIGIN" -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON
+```
+
+#### Fixing Compatibility Issues with Old CUDA and New glibc
+
+If you try to use an old CUDA version (e.g. v11.7) with a new glibc version you can get errors like this:
+
+```
+/usr/include/bits/mathcalls.h(83): error: exception specification is
+  incompatible with that of previous function "cospi"
+
+
+  /opt/cuda-11.7/bin/../targets/x86_64-linux/include/crt/math_functions.h(5545):
+  here
+```
+
+It seems the least bad solution is to patch the CUDA installation to declare the correct signatures.
+Replace the following lines in `/path/to/your/cuda/installation/targets/x86_64-linux/include/crt/math_functions.h`:
+
+```C++
+// original lines
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 cospi(double x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  cospif(float x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 sinpi(double x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  sinpif(float x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 rsqrt(double x);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  rsqrtf(float x);
+
+// edited lines
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 cospi(double x) noexcept (true);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  cospif(float x) noexcept (true);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 sinpi(double x) noexcept (true);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  sinpif(float x) noexcept (true);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ double                 rsqrt(double x) noexcept (true);
+extern __DEVICE_FUNCTIONS_DECL__ __device_builtin__ float                  rsqrtf(float x) noexcept (true);
+```
+
 ### Runtime CUDA environmental variables
 
 You may set the [cuda environmental variables](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) at runtime.

docs/ops.md

@@ -22,11 +22,11 @@ Legend:
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
-|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ |
+|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                          CONV_3D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
@@ -42,7 +42,7 @@ Legend:
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
@@ -84,7 +84,7 @@ Legend:
 |                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
@@ -111,6 +111,6 @@ Legend:
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                         TOPK_MOE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
 |                            XIELU | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |

docs/ops/SYCL.csv

@@ -9307,37 +9307,37 @@
 "SYCL0","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=24,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=0,inplace=1","support","1","yes","SYCL"
 "SYCL0","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=24,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=0,inplace=1","support","1","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=0","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=0","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=0","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=0","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=1","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=1","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=1","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=1","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=2","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=2","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=2","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=2","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=0,v=3","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=1,v=3","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","0","yes","SYCL"
 "SYCL0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","1","yes","SYCL"
-"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","0","no","SYCL"
+"SYCL0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","0","yes","SYCL"
 "SYCL0","ARGSORT","type=f32,ne=[8,1,1,1],order=0","support","1","yes","SYCL"
 "SYCL0","ARGSORT","type=f32,ne=[16,10,10,10],order=0","support","1","yes","SYCL"
 "SYCL0","ARGSORT","type=f32,ne=[60,10,10,10],order=0","support","1","yes","SYCL"

examples/gguf/gguf.cpp

@@ -184,8 +184,13 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
             const char * name   = gguf_get_tensor_name  (ctx, i);
             const size_t size   = gguf_get_tensor_size  (ctx, i);
             const size_t offset = gguf_get_tensor_offset(ctx, i);
+            const auto   type   = gguf_get_tensor_type  (ctx, i);
 
-            printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);
+            const char * type_name  = ggml_type_name(type);
+            const size_t type_size  = ggml_type_size(type);
+            const size_t n_elements = size / type_size;
+
+            printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu, type = %s, n_elts = %zu\n", __func__, i, name, size, offset, type_name, n_elements);
         }
     }
 

ggml/CMakeLists.txt

@@ -168,7 +168,7 @@ option(GGML_RV_ZFH           "ggml: enable riscv zfh"        ON)
 option(GGML_RV_ZVFH          "ggml: enable riscv zvfh"       ON)
 option(GGML_RV_ZICBOP        "ggml: enable riscv zicbop"     ON)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
-option(GGML_VXE              "ggml: enable vxe"              ON)
+option(GGML_VXE              "ggml: enable vxe"              ${GGML_NATIVE})
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")

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

@@ -580,16 +580,19 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
             uint8_t *patmp = atmp;
             int vsums;
-            int tmp;
+            int tmp, t1, t2, t3, t4, t5, t6, t7;
             __asm__ __volatile__(
                 "vsetivli zero, 16, e8, m1\n\t"
                 "vmv.v.x v8, zero\n\t"
+                "lb zero, 15(%[sc])\n\t"
                 "vle8.v v1, (%[sc])\n\t"
+                "vle8.v v2, (%[bsums])\n\t"
+                "addi %[tmp], %[bsums], 16\n\t"
                 "vand.vi v0, v1, 0xF\n\t"
                 "vsrl.vi v1, v1, 4\n\t"
+                "vle8.v v3, (%[tmp])\n\t"
                 "vse8.v v0, (%[scale])\n\t"
                 "vsetivli zero, 16, e16, m2\n\t"
-                "vle16.v v2, (%[bsums])\n\t"
                 "vzext.vf2 v0, v1\n\t"
                 "vwmul.vv v4, v0, v2\n\t"
                 "vsetivli zero, 16, e32, m4\n\t"
@@ -608,46 +611,89 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
             for (int j = 0; j < QK_K/128; ++j) {
                 __asm__ __volatile__(
-                    "vsetvli zero, %[vl32], e8, m2\n\t"
+                    "lb zero, 31(%[q2])\n\t"
+                    "addi %[tmp], %[q2], 16\n\t"
+                    "addi %[t1], %[q8], 16\n\t"
+                    "vsetivli zero, 16, e8, m1\n\t"
                     "vle8.v v0, (%[q2])\n\t"
+                    "vle8.v v1, (%[tmp])\n\t"
                     "vsrl.vi v2, v0, 2\n\t"
+                    "vsrl.vi v3, v1, 2\n\t"
                     "vsrl.vi v4, v0, 4\n\t"
-                    "vsrl.vi v6, v0, 6\n\t"
-                    "vand.vi v0, v0, 0x3\n\t"
-                    "vand.vi v2, v2, 0x3\n\t"
-                    "vand.vi v4, v4, 0x3\n\t"
-                    "vsetvli zero, %[vl128], e8, m8\n\t"
+                    "addi %[tmp], %[q8], 32\n\t"
                     "vle8.v v8, (%[q8])\n\t"
-                    "vsetvli zero, %[vl64], e8, m4\n\t"
+                    "vle8.v v9, (%[t1])\n\t"
+                    "addi %[t1], %[t1], 32\n\t"
+                    "vsrl.vi v5, v1, 4\n\t"
+                    "vsrl.vi v6, v0, 6\n\t"
+                    "vsrl.vi v7, v1, 6\n\t"
+                    "vle8.v v10, (%[tmp])\n\t"
+                    "vle8.v v11, (%[t1])\n\t"
+                    "addi %[tmp], %[tmp], 32\n\t"
+                    "addi %[t1], %[t1], 32\n\t"
+                    "vand.vi v0, v0, 0x3\n\t"
+                    "vand.vi v1, v1, 0x3\n\t"
+                    "vand.vi v2, v2, 0x3\n\t"
+                    "vle8.v v12, (%[tmp])\n\t"
+                    "vle8.v v13, (%[t1])\n\t"
+                    "addi %[tmp], %[tmp], 32\n\t"
+                    "addi %[t1], %[t1], 32\n\t"
+                    "vand.vi v3, v3, 0x3\n\t"
+                    "vand.vi v4, v4, 0x3\n\t"
+                    "vand.vi v5, v5, 0x3\n\t"
+                    "vle8.v v14, (%[tmp])\n\t"
+                    "vle8.v v15, (%[t1])\n\t"
                     "vwmul.vv v16, v0, v8\n\t"
+                    "vwmul.vv v18, v1, v9\n\t"
+                    "vwmul.vv v20, v2, v10\n\t"
+                    "vwmul.vv v22, v3, v11\n\t"
                     "vwmul.vv v24, v4, v12\n\t"
-                    "vsetivli zero, 16, e16, m2\n\t"
+                    "vwmul.vv v26, v5, v13\n\t"
+                    "vwmul.vv v28, v6, v14\n\t"
+                    "vwmul.vv v30, v7, v15\n\t"
+                    "vsetivli zero, 8, e16, m1\n\t"
                     "vmv.v.x v0, zero\n\t"
-                    "vwredsum.vs v10, v16, v0\n\t"
+                    "lbu %[tmp], 0(%[scale])\n\t"
+                    "vwredsum.vs v8, v16, v0\n\t"
                     "vwredsum.vs v9, v18, v0\n\t"
-                    "vwredsum.vs v8, v20, v0\n\t"
-                    "vwredsum.vs v7, v22, v0\n\t"
-                    "vwredsum.vs v11, v24, v0\n\t"
-                    "vwredsum.vs v12, v26, v0\n\t"
-                    "vwredsum.vs v13, v28, v0\n\t"
-                    "vwredsum.vs v14, v30, v0\n\t"
+                    "lbu %[t1], 1(%[scale])\n\t"
+                    "vwredsum.vs v10, v20, v0\n\t"
+                    "vwredsum.vs v11, v22, v0\n\t"
+                    "lbu %[t2], 2(%[scale])\n\t"
+                    "vwredsum.vs v12, v24, v0\n\t"
+                    "vwredsum.vs v13, v26, v0\n\t"
+                    "lbu %[t3], 3(%[scale])\n\t"
+                    "vwredsum.vs v14, v28, v0\n\t"
+                    "vwredsum.vs v15, v30, v0\n\t"
+                    "lbu %[t4], 4(%[scale])\n\t"
+                    "vwredsum.vs v8, v17, v8\n\t"
+                    "vwredsum.vs v9, v19, v9\n\t"
+                    "lbu %[t5], 5(%[scale])\n\t"
+                    "vwredsum.vs v10, v21, v10\n\t"
+                    "vwredsum.vs v11, v23, v11\n\t"
+                    "lbu %[t6], 6(%[scale])\n\t"
+                    "vwredsum.vs v12, v25, v12\n\t"
+                    "vwredsum.vs v13, v27, v13\n\t"
+                    "lbu %[t7], 7(%[scale])\n\t"
+                    "vwredsum.vs v14, v29, v14\n\t"
+                    "vwredsum.vs v15, v31, v15\n\t"
                     "vsetivli zero, 4, e32, m1\n\t"
-                    "vslideup.vi v10, v9, 1\n\t"
-                    "vslideup.vi v8, v7, 1\n\t"
-                    "vslideup.vi v11, v12, 1\n\t"
-                    "vslideup.vi v13, v14, 1\n\t"
-                    "vslideup.vi v10, v8, 2\n\t"
-                    "vslideup.vi v11, v13, 2\n\t"
-                    "vsetivli zero, 8, e32, m2\n\t"
-                    "vle8.v v15, (%[scale])\n\t"
-                    "vzext.vf4 v12, v15\n\t"
-                    "vmul.vv v10, v10, v12\n\t"
-                    "vredsum.vs v0, v10, v0\n\t"
+                    "vmul.vx v0, v8, %[tmp]\n\t"
+                    "vmul.vx v1, v9, %[t1]\n\t"
+                    "vmacc.vx v0, %[t2], v10\n\t"
+                    "vmacc.vx v1, %[t3], v11\n\t"
+                    "vmacc.vx v0, %[t4], v12\n\t"
+                    "vmacc.vx v1, %[t5], v13\n\t"
+                    "vmacc.vx v0, %[t6], v14\n\t"
+                    "vmacc.vx v1, %[t7], v15\n\t"
                     "vmv.x.s %[tmp], v0\n\t"
-                    "add %[isum], %[isum], %[tmp]"
-                    : [tmp] "=&r" (tmp), [isum] "+&r" (isum)
+                    "vmv.x.s %[t1], v1\n\t"
+                    "add %[isum], %[isum], %[tmp]\n\t"
+                    "add %[isum], %[isum], %[t1]"
+                    : [tmp] "=&r" (tmp), [t1] "=&r" (t1), [t2] "=&r" (t2), [t3] "=&r" (t3)
+                    , [t4] "=&r" (t4), [t5] "=&r" (t5), [t6] "=&r" (t6), [t7] "=&r" (t7)
+                    , [isum] "+&r" (isum)
                     : [q2] "r" (q2), [scale] "r" (patmp), [q8] "r" (q8)
-                    , [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128)
                     : "memory"
                     , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
                     , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15"
@@ -929,7 +975,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const  int8_t * restrict q8 = y[i].qs;
 
             int8_t * scale = (int8_t *)utmp;
-            int tmp;
+            int tmp, t1, t2, t3, t4, t5, t6, t7;
             __asm__ __volatile__(
                 "vsetivli zero, 12, e8, m1\n\t"
                 "vle8.v v0, (%[s6b])\n\t"
@@ -967,19 +1013,23 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             int isum = 0;
             for (int j = 0; j < QK_K; j += 128) {
                 __asm__ __volatile__(
+                    "lb zero, 31(%[q3])\n\t"
                     "vsetvli zero, %[vl32], e8, m2, ta, mu\n\t"
                     "vle8.v v8, (%[q3])\n\t"
                     "vsrl.vi v10, v8, 2\n\t"
                     "vsrl.vi v12, v8, 4\n\t"
                     "vsrl.vi v14, v8, 6\n\t"
+                    "lb zero, 64(%[q8])\n\t"
                     "vand.vi v8, v8, 3\n\t"
                     "vand.vi v10, v10, 3\n\t"
                     "vand.vi v12, v12, 3\n\t"
                     "vle8.v v2, (%[qh])\n\t"
+                    "lb zero, 127(%[q8])\n\t"
                     "vand.vx v4, v2, %[m]\n\t"
                     "slli %[m], %[m], 1\n\t"
                     "vmseq.vx v0, v4, zero\n\t"
                     "vadd.vi v8, v8, -4, v0.t\n\t"
+                    "lb zero, 0(%[q8])\n\t"
                     "vand.vx v4, v2, %[m]\n\t"
                     "slli %[m], %[m], 1\n\t"
                     "vmseq.vx v0, v4, zero\n\t"
@@ -994,34 +1044,43 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                     "vadd.vi v14, v14, -4, v0.t\n\t"
                     "vsetvli zero, %[vl128], e8, m8\n\t"
                     "vle8.v v0, (%[q8])\n\t"
+                    "lb %[tmp], 0(%[scale])\n\t"
+                    "lb %[t1], 1(%[scale])\n\t"
+                    "lb %[t2], 2(%[scale])\n\t"
+                    "lb %[t3], 3(%[scale])\n\t"
                     "vsetvli zero, %[vl64], e8, m4\n\t"
                     "vwmul.vv v16, v0, v8\n\t"
                     "vwmul.vv v24, v4, v12\n\t"
                     "vsetivli zero, 16, e16, m2\n\t"
                     "vmv.v.x v0, zero\n\t"
-                    "vwredsum.vs v10, v16, v0\n\t"
+                    "vwredsum.vs v8, v16, v0\n\t"
+                    "lb %[t4], 4(%[scale])\n\t"
+                    "lb %[t5], 5(%[scale])\n\t"
                     "vwredsum.vs v9, v18, v0\n\t"
-                    "vwredsum.vs v8, v20, v0\n\t"
-                    "vwredsum.vs v7, v22, v0\n\t"
-                    "vwredsum.vs v11, v24, v0\n\t"
-                    "vwredsum.vs v12, v26, v0\n\t"
-                    "vwredsum.vs v13, v28, v0\n\t"
-                    "vwredsum.vs v14, v30, v0\n\t"
+                    "vwredsum.vs v10, v20, v0\n\t"
+                    "vwredsum.vs v11, v22, v0\n\t"
+                    "vwredsum.vs v12, v24, v0\n\t"
+                    "lb %[t6], 6(%[scale])\n\t"
+                    "lb %[t7], 7(%[scale])\n\t"
+                    "vwredsum.vs v13, v26, v0\n\t"
+                    "vwredsum.vs v14, v28, v0\n\t"
+                    "vwredsum.vs v15, v30, v0\n\t"
                     "vsetivli zero, 4, e32, m1\n\t"
-                    "vslideup.vi v10, v9, 1\n\t"
-                    "vslideup.vi v8, v7, 1\n\t"
-                    "vslideup.vi v11, v12, 1\n\t"
-                    "vslideup.vi v13, v14, 1\n\t"
-                    "vslideup.vi v10, v8, 2\n\t"
-                    "vslideup.vi v11, v13, 2\n\t"
-                    "vsetivli zero, 8, e32, m2\n\t"
-                    "vle8.v v15, (%[scale])\n\t"
-                    "vsext.vf4 v12, v15\n\t"
-                    "vmul.vv v10, v10, v12\n\t"
-                    "vredsum.vs v0, v10, v0\n\t"
+                    "vmul.vx v0, v8, %[tmp]\n\t"
+                    "vmul.vx v1, v9, %[t1]\n\t"
+                    "vmacc.vx v0, %[t2], v10\n\t"
+                    "vmacc.vx v1, %[t3], v11\n\t"
+                    "vmacc.vx v0, %[t4], v12\n\t"
+                    "vmacc.vx v1, %[t5], v13\n\t"
+                    "vmacc.vx v0, %[t6], v14\n\t"
+                    "vmacc.vx v1, %[t7], v15\n\t"
                     "vmv.x.s %[tmp], v0\n\t"
-                    "add %[isum], %[isum], %[tmp]"
-                    : [tmp] "=&r" (tmp), [m] "+&r" (m), [isum] "+&r" (isum)
+                    "vmv.x.s %[t1], v1\n\t"
+                    "add %[isum], %[isum], %[tmp]\n\t"
+                    "add %[isum], %[isum], %[t1]"
+                    : [tmp] "=&r" (tmp), [t1] "=&r" (t1), [t2] "=&r" (t2), [t3] "=&r" (t3)
+                    , [t4] "=&r" (t4), [t5] "=&r" (t5), [t6] "=&r" (t6), [t7] "=&r" (t7)
+                    , [m] "+&r" (m), [isum] "+&r" (isum)
                     : [vl128] "r" (128), [vl64] "r" (64), [vl32] "r" (32)
                     , [q3] "r" (q3), [qh] "r" (qh), [scale] "r" (scale), [q8] "r" (q8)
                     : "memory"

ggml/src/ggml-cpu/ops.cpp

@@ -7084,7 +7084,11 @@ static void ggml_compute_forward_conv_2d_dw_cwhn(
     const int64_t row_end = MIN(row_start + rows_per_thread, rows_total);
 
 #ifdef GGML_SIMD
-    const int64_t pkg_size = GGML_F32_EPR;
+    #if defined(__ARM_FEATURE_SVE)
+        const int64_t pkg_size = svcntw();
+    #else
+        const int64_t pkg_size = GGML_F32_EPR;
+    #endif
     const int64_t pkg_count = c / pkg_size;
     const int64_t c_pkg_end = pkg_count * pkg_size;
 #else

ggml/src/ggml-cuda/CMakeLists.txt

@@ -124,6 +124,7 @@ if (CUDAToolkit_FOUND)
 
     if (GGML_CUDA_DEBUG)
         list(APPEND CUDA_FLAGS -lineinfo)
+        add_compile_definitions(GGML_CUDA_DEBUG)
     endif()
 
     if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")

ggml/src/ggml-cuda/cpy.cu

@@ -7,6 +7,10 @@
 
 typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
 
+const int CUDA_CPY_TILE_DIM_2D = 32; // 2D tile dimension for transposed blocks
+const int CUDA_CPY_BLOCK_NM = 8;     // block size of 3rd dimension if available
+const int CUDA_CPY_BLOCK_ROWS = 8;   // block dimension for marching through rows
+
 template <cpy_kernel_t cpy_1>
 static __global__ void cpy_flt(const char * cx, char * cdst, const int ne,
                                const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
@@ -35,6 +39,55 @@ static __global__ void cpy_flt(const char * cx, char * cdst, const int ne,
     cpy_1(cx + x_offset, cdst + dst_offset);
 }
 
+template <typename T>
+static __global__ void cpy_flt_transpose(const char * cx, char * cdst, const int ne,
+                               const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                               const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                               const int nb12, const int nb13) {
+
+    const T* src = reinterpret_cast<const T*>(cx);
+    T* dst = reinterpret_cast<T*>(cdst);
+
+    const int64_t nmat = ne / (ne00 * ne01);
+    const int64_t n = ne00 * ne01;
+
+    const int x = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.x;
+    const int y = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
+    const int tx = blockIdx.y * CUDA_CPY_TILE_DIM_2D + threadIdx.x;  // transpose block offset
+    const int ty = blockIdx.x * CUDA_CPY_TILE_DIM_2D + threadIdx.y;
+
+    __shared__ float tile[CUDA_CPY_TILE_DIM_2D][CUDA_CPY_TILE_DIM_2D+1];
+
+#pragma unroll
+    for (int i = 0; i < CUDA_CPY_BLOCK_NM; ++i) {
+
+        const unsigned int imat = blockIdx.z * CUDA_CPY_BLOCK_NM + i;
+        if (imat >= nmat)
+            break;
+
+#pragma unroll
+        for (int j = 0; j < CUDA_CPY_TILE_DIM_2D; j += CUDA_CPY_BLOCK_ROWS) {
+            if(x < ne01 && y + j < ne00){
+                const int row = threadIdx.y+j;
+                const int col = threadIdx.x * sizeof(float)/sizeof(T);
+                T *tile2 = reinterpret_cast<T*>(tile[row]);
+                tile2[col] = src[imat*n + (y+j)*ne01 + x];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < CUDA_CPY_TILE_DIM_2D; j += CUDA_CPY_BLOCK_ROWS) {
+            if (ty + j < ne01 && tx < ne00) {
+                const int col = (threadIdx.y+j)*sizeof(float)/sizeof(T);
+                const T *tile2 = reinterpret_cast<const T*>(tile[threadIdx.x]);
+                dst[imat*n + (ty+j)*ne00 + tx] = tile2[col];
+            }
+        }
+    }
+}
+
 static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *)(cdsti);
 
@@ -136,15 +189,36 @@ cudaStream_t stream) {
         (cx, cdst, ne);
 }
 
-template<typename src_t, typename dst_t>
+template<typename src_t, typename dst_t, bool transposed = false>
 static void ggml_cpy_flt_cuda(
     const char * cx, char * cdst, const int ne,
     const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
     const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
 
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+    if (transposed) {
+        GGML_ASSERT(ne == ne00*ne01*ne02);  // ne[3] is 1 assumed
+        int ne00n, ne01n, ne02n;
+        if (nb00 <= nb02) { // most likely safe to handle nb00 = nb02 case here
+            ne00n = ne00;
+            ne01n = ne01;
+            ne02n = ne02;
+        } else if (nb00 > nb02) {
+            ne00n = ne00;
+            ne01n = ne01*ne02;
+            ne02n = 1;
+        }
+
+        dim3 dimGrid( (ne01n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D,
+                      (ne00n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D,
+                      (ne/(ne01n*ne00n) + CUDA_CPY_BLOCK_NM - 1) / CUDA_CPY_BLOCK_NM);
+        dim3 dimBlock(CUDA_CPY_TILE_DIM_2D, CUDA_CPY_BLOCK_ROWS, 1);
+        cpy_flt_transpose<dst_t><<<dimGrid, dimBlock, 0, stream>>>
+            (cx, cdst, ne, ne00n, ne01n, ne02n, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+    } else {
+        const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+        cpy_flt<cpy_1_flt<src_t, dst_t>><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+            (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+    }
 }
 
 static void ggml_cpy_f32_q8_0_cuda(
@@ -310,6 +384,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src1_ddc = (char *) src1->data;
 
     const bool contiguous_srcs = ggml_is_contiguous(src0) && ggml_is_contiguous(src1);
+    const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) && src0->ne[3] == 1;
 
     if (src0->type == src1->type && contiguous_srcs) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
@@ -322,7 +397,11 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
             CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
         }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_flt_cuda<float, float>           (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (can_be_transposed) {
+            ggml_cpy_flt_cuda<float, float, true> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
         if (contiguous_srcs) {
             ggml_cpy_flt_contiguous_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, main_stream);
@@ -361,7 +440,11 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
         ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_flt_cuda<half, half>               (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (can_be_transposed) {
+            ggml_cpy_flt_cuda<half, half, true> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<half, half>       (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
         if (contiguous_srcs) {
             ggml_cpy_flt_contiguous_cuda<half, nv_bfloat16>  (src0_ddc, src1_ddc, ne, main_stream);
@@ -375,7 +458,11 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
             ggml_cpy_flt_cuda<half, float>          (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
         }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16) {
-        ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        if (can_be_transposed) {
+            ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16, true> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        } else {
+            ggml_cpy_flt_cuda<nv_bfloat16, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
     } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F16) {
         if (contiguous_srcs) {
             ggml_cpy_flt_contiguous_cuda<nv_bfloat16, half>  (src0_ddc, src1_ddc, ne, main_stream);

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

@@ -27,7 +27,6 @@
 #include "ggml-cuda/mmq.cuh"
 #include "ggml-cuda/mmvf.cuh"
 #include "ggml-cuda/mmvq.cuh"
-#include "ggml-cuda/moe-expert-reduce.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
 #include "ggml-cuda/opt-step-sgd.cuh"
@@ -2113,7 +2112,7 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor) {
         src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
     const int cc      = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
+    use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
 
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
                        ggml_backend_buft_is_cuda_split(src1->buffer->buft);
@@ -2207,16 +2206,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
             const int cc            = ggml_cuda_info().devices[id].cc;
             const int warp_size     = ggml_cuda_info().devices[id].warp_size;
             use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
-            use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
+            use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
+            use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
             any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
         }
     } else {
         const int cc            = ggml_cuda_info().devices[ctx.device].cc;
         const int warp_size     = ggml_cuda_info().devices[ctx.device].warp_size;
         use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src1->ne[1], /*mul_mat_id=*/false);
-        use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src1->ne[1]);
+        use_mul_mat_f           = use_mul_mat_f             && ggml_cuda_should_use_mmf(src0->type, cc, warp_size, src0->ne, src0->nb, src1->ne[1], /*mul_mat_id=*/false);
+        use_mul_mat_vec_f       = use_mul_mat_vec_f         && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, src0->nb, src1->ne[1]);
         any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
     }
 
@@ -2287,7 +2286,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
             return;
         }
 
-        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src1->ne[2], /*mul_mat_id=*/true)) {
+        if (ggml_cuda_should_use_mmf(src0->type, cc, WARP_SIZE, src0->ne, src0->nb, src1->ne[2], /*mul_mat_id=*/true)) {
             ggml_cuda_mul_mat_f(ctx, src0, src1, ids, dst);
             return;
         }
@@ -3152,8 +3151,6 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
 
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];
-
-
 #ifdef GGML_CUDA_DEBUG
                 const int nodes_fused = i - prev_i - 1;
                 prev_i = i;
@@ -3199,31 +3196,6 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                         continue;
                     }
 
-                    if (node->op == GGML_OP_MUL) {
-                        int current_node = i + 1;
-                        int num_views    = 0;
-                        int num_adds     = 0;
-                        while (current_node < cgraph->n_nodes && cgraph->nodes[current_node]->op == GGML_OP_VIEW) {
-                            num_views++;
-                            current_node++;
-                        }
-
-                        while (current_node < cgraph->n_nodes && cgraph->nodes[current_node]->op == GGML_OP_ADD &&
-                                num_adds < num_views - 1) {
-                            num_adds++;
-                            current_node++;
-                        }
-
-                        if (num_adds == num_views - 1 && num_views > 0) {
-                            ggml_tensor * dst_node = cgraph->nodes[current_node - 1];
-                            if (ggml_cuda_should_use_moe_expert_reduce(cgraph, i, current_node)) {
-                                ggml_cuda_op_moe_expert_reduce(*cuda_ctx, node->src[0], node->src[1], dst_node);
-                                i += num_views + num_adds;
-                                continue;
-                            }
-                        }
-                    }
-
                     if (node->op == GGML_OP_ADD) {
                         int n_fuse = 0;
                         ggml_op ops[8];
@@ -3302,6 +3274,13 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                                 continue;
                             }
 
+                            // we don't support repeating adds
+                            if (bias_op == GGML_OP_ADD &&
+                                (!ggml_are_same_shape(gate_bias_n->src[0], gate_bias_n->src[1]) ||
+                                 !ggml_are_same_shape(up_bias_n->src[0], up_bias_n->src[1]))) {
+                                continue;
+                            }
+
                             const ggml_tensor * src0 = up_n->src[0];
                             const ggml_tensor * src1 = up_n->src[1];
                             const ggml_tensor * ids  = up_n->src[2];
@@ -3411,6 +3390,10 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                             continue;
                         }
 
+                        if (bias_op == GGML_OP_ADD && !ggml_are_same_shape(bias_node->src[0], bias_node->src[1])) {
+                            continue;
+                        }
+
                         ggml_cuda_mm_fusion_args_host fusion_data{};
                         fusion_data.x_bias = bias_tensor;
 

ggml/src/ggml-cuda/mmf.cu

@@ -119,15 +119,27 @@ void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * sr
     }
 }
 
-bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, const int src1_ncols, bool mul_mat_id) {
-
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne,
+        const size_t * src0_nb, const int src1_ncols, bool mul_mat_id) {
     if (ggml_is_quantized(type)) {
         return false;
     }
 
-    if (src0_ne[0] % (warp_size * (4/ggml_type_size(type))) != 0) {
+    const size_t ts = ggml_type_size(type);
+    if (src0_ne[0] % (warp_size * (4/ts)) != 0) {
         return false;
     }
+
+    if (src0_nb[0] != ts) {
+        return false;
+    }
+
+    // Pointers not aligned to the size of half2/nv_bfloat162/float2 would result in a crash:
+    for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
+        if (src0_nb[i] % (2*ts) != 0) {
+            return false;
+        }
+    }
     if (src0_ne[1] % MMF_ROWS_PER_BLOCK != 0) {
         return false;
     }

ggml/src/ggml-cuda/mmf.cuh

@@ -17,7 +17,7 @@ struct mmf_ids_data {
 
 void ggml_cuda_mul_mat_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
 
-bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const int src1_ncols, bool mul_mat_id);
+bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const int64_t * scr0_ne, const size_t * src0_nb, const int src1_ncols, bool mul_mat_id);
 
 template <typename T, int rows_per_block, int cols_per_block, int nwarps, bool has_ids>
 __launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)

ggml/src/ggml-cuda/mmq.cuh

@@ -3494,7 +3494,7 @@ static __global__ void mul_mat_q_stream_k_fixup(
     const int col_diff = col_high - col_low;
 
     for (int j = threadIdx.y*warp_size + threadIdx.x; j < mmq_x; j += nwarps*warp_size) {
-        ids_dst_shared[j] = ids_dst[col_low + j];
+        ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j];
     }
     __syncthreads();
 

ggml/src/ggml-cuda/mmvf.cu

@@ -716,10 +716,23 @@ void ggml_cuda_op_mul_mat_vec_f(
     GGML_UNUSED_VARS(ctx, src1, dst, src1_ddq_i, src1_ncols, src1_padded_row_size);
 }
 
-bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11) {
+bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, const size_t * src0_nb, int64_t ne11) {
     if (src0_ne[0] % 2 != 0) {
         return false;
     }
+
+    const size_t ts = ggml_type_size(type);
+    if (src0_nb[0] != ts) {
+        return false;
+    }
+
+    // Pointers not aligned to the size of half2/nv_bfloat162/float2 would result in a crash:
+    for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
+        if (src0_nb[i] % (2*ts) != 0) {
+            return false;
+        }
+    }
+
     switch (type) {
         case GGML_TYPE_F32:
             if (GGML_CUDA_CC_IS_NVIDIA(cc)) {

ggml/src/ggml-cuda/mmvf.cuh

@@ -9,4 +9,4 @@ void ggml_cuda_op_mul_mat_vec_f(
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
 
-bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11);
+bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0_ne, const size_t * src0_nb, int64_t ne11);

ggml/src/ggml-cuda/moe-expert-reduce.cu

@@ -1,168 +0,0 @@
-#include "moe-expert-reduce.cuh"
-
-// This kernel is a fusion of the expert weight reduce, common in MoE models
-
-template <int n_expert_used_template>
-__global__ void moe_expert_reduce_cuda(const float * __restrict__ experts,
-                                       const float * __restrict__ weights,
-                                       float * __restrict__ dst,
-                                       const int n_expert_used,
-                                       const int n_cols) {
-    const int row = blockIdx.x;
-    const int col = blockIdx.y * blockDim.x + threadIdx.x;
-    if (col >= n_cols) {
-        return;
-    }
-
-    experts += row * n_cols * n_expert_used;
-    weights += row * n_expert_used;
-    dst += row * n_cols;
-
-    float acc = 0.f;
-    if constexpr (n_expert_used_template == 0) {
-        for (int expert = 0; expert < n_expert_used; ++expert) {
-            ggml_cuda_mad(acc, experts[col], weights[expert]);
-            experts += n_cols;
-        }
-        dst[col] = acc;
-    } else {
-#pragma unroll
-        for (int i = 0; i < n_expert_used_template; ++i) {
-            ggml_cuda_mad(acc, experts[col], weights[i]);
-            experts += n_cols;
-        }
-        dst[col] = acc;
-    }
-}
-
-static void launch_moe_expert_reduce(ggml_backend_cuda_context & ctx,
-                                     const float *               experts,
-                                     const float *               weights,
-                                     float *                     dst,
-                                     const int                   n_expert_used,
-                                     const int                   n_cols,
-                                     const int                   n_rows) {
-    const int block_size = 32;
-
-    const int n_blocks_x = n_rows;
-    const int n_blocks_y = (n_cols + block_size - 1) / block_size;
-
-    dim3 block_dims(block_size);
-    dim3 grid_dims(n_blocks_x, n_blocks_y);
-
-    cudaStream_t stream = ctx.stream();
-    switch (n_expert_used) {
-        case 1:
-            moe_expert_reduce_cuda<1>
-                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
-            break;
-        case 2:
-            moe_expert_reduce_cuda<2>
-                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
-            break;
-        case 4:
-            moe_expert_reduce_cuda<4>
-                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
-            break;
-        case 6:
-            moe_expert_reduce_cuda<6>
-                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
-            break;
-        case 8:
-            moe_expert_reduce_cuda<8>
-                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
-            break;
-        case 16:
-            moe_expert_reduce_cuda<16>
-                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
-            break;
-        case 32:
-            moe_expert_reduce_cuda<32>
-                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
-            break;
-        case 64:
-            moe_expert_reduce_cuda<64>
-                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
-            break;
-        case 128:
-            moe_expert_reduce_cuda<128>
-                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
-            break;
-        default:
-            moe_expert_reduce_cuda<0>
-                <<<grid_dims, block_dims, 0, stream>>>(experts, weights, dst, n_expert_used, n_cols);
-            break;
-    }
-}
-
-bool ggml_cuda_should_use_moe_expert_reduce(const ggml_cgraph * cgraph, int start_index, int end_index) {
-    const ggml_tensor * mul = cgraph->nodes[start_index];
-
-    if (mul->op != GGML_OP_MUL || !ggml_is_contiguous(mul->src[0]) || !ggml_is_contiguous(mul->src[1])) {
-        return false;
-    }
-
-    int    current_node   = start_index + 1;
-    size_t current_offset = 0;
-
-    std::vector<const ggml_tensor *> view_nodes;
-    //check if all are views of the expert in increasing order
-    while (current_node < end_index && cgraph->nodes[current_node]->op == GGML_OP_VIEW) {
-        const ggml_tensor * node = cgraph->nodes[current_node];
-        if (node->view_src != mul) {
-            return false;
-        }
-        if (node->view_offs < current_offset) {
-            return false;
-        }
-        current_offset = node->view_offs;
-        current_node++;
-        view_nodes.push_back(node);
-    }
-
-    //check if all the adds are in increasing order
-    const ggml_tensor * prev_add_src = view_nodes.empty() ? nullptr : view_nodes[0];
-    int                 num_adds     = 0;
-    int                 num_views    = view_nodes.size();
-    while (current_node < end_index && cgraph->nodes[current_node]->op == GGML_OP_ADD) {
-        const ggml_tensor * add_node = cgraph->nodes[current_node];
-
-        bool is_first_op_ok  = num_views > num_adds ? add_node->src[0] == prev_add_src : false;
-        bool is_second_op_ok = num_views > num_adds ? add_node->src[1] == view_nodes[num_adds + 1] : false;
-
-        if (!is_first_op_ok || !is_second_op_ok) {
-            return false;
-        }
-        prev_add_src = add_node;
-
-        num_adds++;
-        current_node++;
-    }
-
-    if (num_views != num_adds + 1) {
-        return false;
-    }
-
-    return true;
-}
-
-void ggml_cuda_op_moe_expert_reduce(ggml_backend_cuda_context & ctx,
-                                    const ggml_tensor *         experts,
-                                    const ggml_tensor *         weights,
-                                    ggml_tensor *               dst) {
-    const int n_rows        = experts->ne[2];
-    const int n_expert_used = experts->ne[1];
-    const int n_cols        = experts->ne[0];
-
-    GGML_ASSERT(experts->type == GGML_TYPE_F32);
-    GGML_ASSERT(weights->type == GGML_TYPE_F32);
-    GGML_ASSERT(ggml_is_contiguous(experts));
-    GGML_ASSERT(ggml_is_contiguous(weights));
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-
-    const float * experts_d = (const float *) experts->data;
-    const float * weights_d = (const float *) weights->data;
-    float *       dst_d     = (float *) dst->data;
-
-    launch_moe_expert_reduce(ctx, experts_d, weights_d, dst_d, n_expert_used, n_cols, n_rows);
-}

ggml/src/ggml-cuda/moe-expert-reduce.cuh

@@ -1,11 +0,0 @@
-#include "common.cuh"
-#include "ggml.h"
-
-#include <initializer_list>
-
-void ggml_cuda_op_moe_expert_reduce(ggml_backend_cuda_context & ctx,
-                                    const ggml_tensor *         experts,
-                                    const ggml_tensor *         weights,
-                                    ggml_tensor *               dst);
-
-bool ggml_cuda_should_use_moe_expert_reduce(const ggml_cgraph * cgraph, int start_index, int end_index);

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -367,7 +367,13 @@ struct ggml_backend_hexagon_buffer_context {
     ggml_backend_hexagon_buffer_context(ggml_hexagon_session * sess, size_t size, bool repack) {
         size += 4 * 1024;  // extra page for padding
 
-        this->base = (uint8_t *) rpcmem_alloc2(RPCMEM_HEAP_ID_SYSTEM, RPCMEM_DEFAULT_FLAGS | RPCMEM_HEAP_NOREG, size);
+        if (rpcmem_alloc2) {
+            this->base = (uint8_t *) rpcmem_alloc2(RPCMEM_HEAP_ID_SYSTEM, RPCMEM_DEFAULT_FLAGS | RPCMEM_HEAP_NOREG, size);
+        } else {
+            GGML_LOG_INFO("ggml-hex: %s rpcmem_alloc2 not found, falling back to rpcmem_alloc\n", sess->name.c_str());
+            this->base = (uint8_t *) rpcmem_alloc(RPCMEM_HEAP_ID_SYSTEM, RPCMEM_DEFAULT_FLAGS | RPCMEM_HEAP_NOREG, size);
+        }
+
         if (!this->base) {
             GGML_LOG_ERROR("ggml-hex: %s failed to allocate buffer : size %zu\n", sess->name.c_str(), size);
             throw std::runtime_error("ggml-hex: rpcmem_alloc failed (see log for details)");
@@ -1679,12 +1685,13 @@ void ggml_hexagon_session::allocate(int dev_id) noexcept(false) {
     }
 
     // Get session URI
-    char htp_uri[256];
-    sprintf(htp_uri, "file:///libggml-htp-v%u.so?htp_iface_skel_handle_invoke&_modver=1.0", opt_arch);
 
     char session_uri[256];
     {
-        struct remote_rpc_get_uri u;
+        char htp_uri[256];
+        snprintf(htp_uri, sizeof(htp_uri), "file:///libggml-htp-v%u.so?htp_iface_skel_handle_invoke&_modver=1.0", opt_arch);
+
+        struct remote_rpc_get_uri u = {};
         u.session_id      = this->session_id;
         u.domain_name     = const_cast<char *>(CDSP_DOMAIN_NAME);
         u.domain_name_len = strlen(CDSP_DOMAIN_NAME);
@@ -1695,8 +1702,12 @@ void ggml_hexagon_session::allocate(int dev_id) noexcept(false) {
 
         int err = remote_session_control(FASTRPC_GET_URI, (void *) &u, sizeof(u));
         if (err != AEE_SUCCESS) {
-            GGML_LOG_ERROR("ggml-hex: failed to get URI for session %d : error 0x%x\n", dev_id, err);
-            throw std::runtime_error("ggml-hex: remote_session_control(get-uri) failed (see log for details)");
+            // fallback to single session uris
+            int htp_URI_domain_len = strlen(htp_uri) + MAX_DOMAIN_NAMELEN;
+
+            snprintf(session_uri, htp_URI_domain_len, "%s%s", htp_uri, my_domain->uri);
+
+            GGML_LOG_WARN("ggml-hex: failed to get URI for session %d : error 0x%x. Falling back to single session URI: %s\n", dev_id, err, session_uri);
         }
     }
 
@@ -3668,6 +3679,11 @@ ggml_hexagon_registry::ggml_hexagon_registry(ggml_backend_reg_t reg) {
         }
     }
 
+    if(opt_arch < 75) {
+        opt_ndev = 1;
+        GGML_LOG_WARN("ggml-hex: forcing ndev to 1 for SoCs archs lower than v75.\n");
+    }
+
     GGML_LOG_INFO("ggml-hex: Hexagon Arch version v%d\n", opt_arch);
 
     // Create devices / sessions

ggml/src/ggml-hexagon/htp-utils.h

@@ -64,6 +64,7 @@ extern "C" {
 #    pragma weak remote_handle64_control
 #    pragma weak fastrpc_mmap
 #    pragma weak fastrpc_munmap
+#    pragma weak rpcmem_alloc2
 #endif
 
 #if !defined(_WINDOWS)

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

@@ -35,7 +35,6 @@ struct ggml_metal {
     // additional, inference-time compiled pipelines
     ggml_metal_pipelines_t pipelines_ext;
 
-    bool use_bfloat;
     bool use_fusion;
     bool use_concurrency;
     bool use_graph_optimize;
@@ -121,11 +120,10 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
         }
     }
 
-    const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
+    //const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
 
     res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
-    res->use_bfloat      = props_dev->has_bfloat;
     res->use_fusion      = getenv("GGML_METAL_FUSION_DISABLE") == nil;
     res->use_concurrency = getenv("GGML_METAL_CONCURRENCY_DISABLE") == nil;
 
@@ -147,7 +145,6 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
 
     memset(res->fuse_cnt, 0, sizeof(res->fuse_cnt));
 
-    GGML_LOG_INFO("%s: use bfloat         = %s\n", __func__, res->use_bfloat         ? "true" : "false");
     GGML_LOG_INFO("%s: use fusion         = %s\n", __func__, res->use_fusion         ? "true" : "false");
     GGML_LOG_INFO("%s: use concurrency    = %s\n", __func__, res->use_concurrency    ? "true" : "false");
     GGML_LOG_INFO("%s: use graph optimize = %s\n", __func__, res->use_graph_optimize ? "true" : "false");

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

@@ -95,7 +95,9 @@ void ggml_metal_encoder_end_encoding(ggml_metal_encoder_t encoder);
 
 typedef struct ggml_metal_library * ggml_metal_library_t;
 
-ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev);
+ggml_metal_library_t ggml_metal_library_init            (ggml_metal_device_t dev);
+ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose);
+
 void ggml_metal_library_free(ggml_metal_library_t lib);
 
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline    (ggml_metal_library_t lib, const char * name);
@@ -193,6 +195,7 @@ struct ggml_metal_device_props {
     bool has_simdgroup_mm;
     bool has_unified_memory;
     bool has_bfloat;
+    bool has_tensor;
     bool use_residency_sets;
     bool use_shared_buffers;
 

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

@@ -21,8 +21,9 @@
 #define GGML_METAL_HAS_RESIDENCY_SETS 1
 #endif
 
-// overload of MTLGPUFamilyMetal3 (not available in some environments)
+// overload of MTLGPUFamilyMetalX (not available in some environments)
 static const NSInteger MTLGPUFamilyMetal3_GGML = 5001;
+static const NSInteger MTLGPUFamilyMetal4_GGML = 5002;
 
 // virtual address for GPU memory allocations
 static atomic_uintptr_t g_addr_device = 0x000000400ULL;
@@ -261,6 +262,10 @@ ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
                     [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
                 }
 
+                if (ggml_metal_device_get_props(dev)->has_tensor) {
+                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
+                }
+
 #if GGML_METAL_EMBED_LIBRARY
                 [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
 #endif
@@ -298,6 +303,72 @@ ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
     return res;
 }
 
+ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose) {
+    if (source == NULL) {
+        GGML_LOG_ERROR("%s: source is NULL\n", __func__);
+        return NULL;
+    }
+
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+    id<MTLLibrary> library = nil;
+    NSError * error = nil;
+
+    const int64_t t_start = ggml_time_us();
+
+    NSString * src = [[NSString alloc] initWithBytes:source
+                                              length:strlen(source)
+                                            encoding:NSUTF8StringEncoding];
+    if (!src) {
+        GGML_LOG_ERROR("%s: failed to create NSString from source\n", __func__);
+        return NULL;
+    }
+
+    @autoreleasepool {
+        NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+
+        MTLCompileOptions * options = [MTLCompileOptions new];
+        options.preprocessorMacros = prep;
+
+        library = [device newLibraryWithSource:src options:options error:&error];
+        if (error) {
+            if (verbose) {
+                GGML_LOG_ERROR("%s: error compiling source: %s\n", __func__, [[error description] UTF8String]);
+            } else {
+                GGML_LOG_ERROR("%s: error compiling source\n", __func__);
+            }
+            library = nil;
+        }
+
+        [options release];
+    }
+
+    [src release];
+
+    if (!library) {
+        if (verbose) {
+            GGML_LOG_ERROR("%s: failed to create Metal library from source\n", __func__);
+        }
+
+        return NULL;
+    }
+
+    if (verbose) {
+        GGML_LOG_INFO("%s: compiled in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
+    }
+
+    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
+    if (!res) {
+        GGML_LOG_ERROR("%s: calloc failed\n", __func__);
+        return NULL;
+    }
+
+    res->obj       = library;
+    res->device    = device;
+    res->pipelines = ggml_metal_pipelines_init();
+
+    return res;
+}
+
 void ggml_metal_library_free(ggml_metal_library_t lib) {
     if (!lib) {
         return;
@@ -345,9 +416,9 @@ ggml_metal_pipeline_t ggml_metal_library_compile_pipeline(ggml_metal_library_t l
         if (!mtl_function) {
             ggml_critical_section_end();
 
-            GGML_LOG_ERROR("%s: error: failed to compile pipeline: base = '%s', name = '%s'\n", __func__, base, name);
+            GGML_LOG_ERROR("%s: failed to compile pipeline: base = '%s', name = '%s'\n", __func__, base, name);
             if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+                GGML_LOG_ERROR("%s: %s\n", __func__, [[error description] UTF8String]);
             }
 
             return nil;
@@ -355,13 +426,21 @@ ggml_metal_pipeline_t ggml_metal_library_compile_pipeline(ggml_metal_library_t l
 
         res->obj = [lib->device newComputePipelineStateWithFunction:mtl_function error:&error];
 
-        ggml_metal_pipelines_add(lib->pipelines, name, res);
-
         [mtl_function release];
 
         GGML_LOG_DEBUG("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, name, (void *) res->obj,
                 (int) res->obj.maxTotalThreadsPerThreadgroup,
                 (int) res->obj.threadExecutionWidth);
+
+        if (res->obj.maxTotalThreadsPerThreadgroup == 0 || res->obj.threadExecutionWidth == 0) {
+            ggml_critical_section_end();
+
+            GGML_LOG_ERROR("%s: incompatible pipeline %s\n", __func__, name);
+
+            return nil;
+        }
+
+        ggml_metal_pipelines_add(lib->pipelines, name, res);
     }
 
     ggml_critical_section_end();
@@ -469,6 +548,126 @@ ggml_metal_device_t ggml_metal_device_init(void) {
 
             dev->props.has_bfloat  = [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
             dev->props.has_bfloat |= [dev->mtl_device supportsFamily:MTLGPUFamilyApple6];
+            if (getenv("GGML_METAL_BF16_DISABLE") != NULL) {
+                dev->props.has_bfloat = false;
+            }
+
+            dev->props.has_tensor = [dev->mtl_device supportsFamily:MTLGPUFamilyMetal4_GGML];
+            if (getenv("GGML_METAL_TENSOR_DISABLE") != NULL) {
+                dev->props.has_tensor = false;
+            }
+
+            // note: disable the tensor API by default for old chips because with the current implementation it is not useful
+            // - M2 Ultra:   ~5% slower
+            // - M4, M4 Max: no significant difference
+            //
+            // TODO: try to update the tensor API kernels to at least match the simdgroup performance
+            if (getenv("GGML_METAL_TENSOR_ENABLE") == NULL &&
+                ![[dev->mtl_device name] containsString:@"M5"] &&
+                ![[dev->mtl_device name] containsString:@"M6"]) {
+                GGML_LOG_WARN("%s: tensor API disabled for pre-M5 device\n", __func__);
+                dev->props.has_tensor = false;
+            }
+
+            // double-check that the tensor API compiles
+            if (dev->props.has_tensor) {
+                const char * src_tensor_f16 = "\n"
+                    "#include <metal_stdlib> \n"
+                    "#include <metal_tensor> \n"
+                    "#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h> \n"
+                    " \n"
+                    "using namespace metal; \n"
+                    "using namespace mpp::tensor_ops; \n"
+                    " \n"
+                    "kernel void dummy_kernel( \n"
+                    "    tensor<device  half, dextents<int32_t, 2>> A [[buffer(0)]], \n"
+                    "    tensor<device  half, dextents<int32_t, 2>> B [[buffer(1)]], \n"
+                    "    device float * C [[buffer(2)]], \n"
+                    "    uint2 tgid [[threadgroup_position_in_grid]]) \n"
+                    "{ \n"
+                    "    auto tA = A.slice(0, (int)tgid.y); \n"
+                    "    auto tB = B.slice((int)tgid.x, 0); \n"
+                    " \n"
+                    "    matmul2d< \n"
+                    "        matmul2d_descriptor(8, 8, dynamic_extent), \n"
+                    "        execution_simdgroups<4>> mm; \n"
+                    " \n"
+                    "    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>(); \n"
+                    " \n"
+                    "    auto sA = tA.slice(0, 0); \n"
+                    "    auto sB = tB.slice(0, 0); \n"
+                    "    mm.run(sB, sA, cT); \n"
+                    " \n"
+                    "    auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(4, 4)); \n"
+                    " \n"
+                    "    cT.store(tC); \n"
+                    "}";
+
+                GGML_LOG_INFO("%s: testing tensor API for f16 support\n", __func__);
+                ggml_metal_library_t lib = ggml_metal_library_init_from_source(dev, src_tensor_f16, false);
+                if (lib == NULL) {
+                    GGML_LOG_WARN("%s: - the tensor API is not supported in this environment - disabling\n", __func__);
+                    dev->props.has_tensor = false;
+                } else {
+                    ggml_metal_pipeline_t ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil);
+                    if (!ppl) {
+                        GGML_LOG_WARN("%s: - the tensor API is not supported in this environment - disabling\n", __func__);
+                        dev->props.has_tensor = false;
+                    }
+
+                    ggml_metal_library_free(lib);
+                }
+            }
+
+            // try to compile a dummy kernel to determine if the tensor API is supported for bfloat
+            if (dev->props.has_tensor && dev->props.has_bfloat) {
+                const char * src_tensor_bf16 = "\n"
+                    "#include <metal_stdlib> \n"
+                    "#include <metal_tensor> \n"
+                    "#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h> \n"
+                    " \n"
+                    "using namespace metal; \n"
+                    "using namespace mpp::tensor_ops; \n"
+                    " \n"
+                    "kernel void dummy_kernel( \n"
+                    "    tensor<device bfloat, dextents<int32_t, 2>> A [[buffer(0)]], \n"
+                    "    tensor<device bfloat, dextents<int32_t, 2>> B [[buffer(1)]], \n"
+                    "    device float * C [[buffer(2)]], \n"
+                    "    uint2 tgid [[threadgroup_position_in_grid]]) \n"
+                    "{ \n"
+                    "    auto tA = A.slice(0, (int)tgid.y); \n"
+                    "    auto tB = B.slice((int)tgid.x, 0); \n"
+                    " \n"
+                    "    matmul2d< \n"
+                    "        matmul2d_descriptor(8, 8, dynamic_extent), \n"
+                    "        execution_simdgroups<4>> mm; \n"
+                    " \n"
+                    "    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>(); \n"
+                    " \n"
+                    "    auto sA = tA.slice(0, 0); \n"
+                    "    auto sB = tB.slice(0, 0); \n"
+                    "    mm.run(sB, sA, cT); \n"
+                    " \n"
+                    "    auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(4, 4)); \n"
+                    " \n"
+                    "    cT.store(tC); \n"
+                    "}";
+
+                GGML_LOG_INFO("%s: testing tensor API for bfloat support\n", __func__);
+                ggml_metal_library_t lib = ggml_metal_library_init_from_source(dev, src_tensor_bf16, false);
+                if (lib == NULL) {
+                    GGML_LOG_WARN("%s: - the tensor API does not support bfloat - disabling bfloat support\n", __func__);
+                    dev->props.has_bfloat = false;
+                } else {
+                    ggml_metal_pipeline_t ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil);
+                    if (!ppl) {
+                        GGML_LOG_WARN("%s: - the tensor API does not support bfloat - disabling bfloat support\n", __func__);
+                        dev->props.has_bfloat = false;
+                    }
+
+                    ggml_metal_library_free(lib);
+                }
+            }
 
             dev->props.use_residency_sets = true;
 #if defined(GGML_METAL_HAS_RESIDENCY_SETS)
@@ -476,7 +675,6 @@ ggml_metal_device_t ggml_metal_device_init(void) {
 #endif
 
             dev->props.use_shared_buffers = dev->props.has_unified_memory;
-
             if (getenv("GGML_METAL_SHARED_BUFFERS_DISABLE") != NULL) {
                 dev->props.use_shared_buffers = false;
             }
@@ -529,6 +727,7 @@ ggml_metal_device_t ggml_metal_device_init(void) {
             GGML_LOG_INFO("%s: simdgroup matrix mul. = %s\n", __func__, dev->props.has_simdgroup_mm        ? "true" : "false");
             GGML_LOG_INFO("%s: has unified memory    = %s\n", __func__, dev->props.has_unified_memory      ? "true" : "false");
             GGML_LOG_INFO("%s: has bfloat            = %s\n", __func__, dev->props.has_bfloat              ? "true" : "false");
+            GGML_LOG_INFO("%s: has tensor            = %s\n", __func__, dev->props.has_tensor              ? "true" : "false");
             GGML_LOG_INFO("%s: use residency sets    = %s\n", __func__, dev->props.use_residency_sets      ? "true" : "false");
             GGML_LOG_INFO("%s: use shared buffers    = %s\n", __func__, dev->props.use_shared_buffers      ? "true" : "false");
 

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

@@ -9,6 +9,12 @@ __embed_ggml-common.h__
 
 #include <metal_stdlib>
 
+#ifdef GGML_METAL_HAS_TENSOR
+#include <metal_tensor>
+
+#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>
+#endif
+
 using namespace metal;
 
 #define MAX(x, y) ((x) > (y) ? (x) : (y))
@@ -1742,7 +1748,7 @@ kernel void kernel_op_sum_f32(
 
     float sumf = 0;
 
-    for (int64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
+    for (uint64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
         sumf += src0[i0];
     }
 
@@ -5467,6 +5473,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_dk576_dv512")]] kernel flash_at
 
 #undef FA_TYPES
 #undef FA_TYPES_BF
+#undef FA_TYPES_F32
 
 constant bool FC_flash_attn_ext_vec_has_mask  [[function_constant(FC_FLASH_ATTN_EXT_VEC + 0)]];
 constant bool FC_flash_attn_ext_vec_has_sinks [[function_constant(FC_FLASH_ATTN_EXT_VEC + 1)]];
@@ -6088,6 +6095,7 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_1_dk576_dv512")]] kernel flas
 template [[host_name("kernel_flash_attn_ext_vec_q8_0_dk576_dv512")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES,     block_q8_0, 8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 576, 512, 2>;
 
 #undef FA_TYPES
+#undef FA_TYPES_F32
 
 constant int32_t FC_flash_attn_ext_vec_reduce_DV  [[function_constant(FC_FLASH_ATTN_EXT_VEC_REDUCE + 0)]];
 constant int32_t FC_flash_attn_ext_vec_reduce_NWG [[function_constant(FC_FLASH_ATTN_EXT_VEC_REDUCE + 1)]];
@@ -8141,17 +8149,6 @@ kernel void kernel_set_rows_f(
 constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
 constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
 
-#define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
-#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
-#define BLOCK_SIZE_K 32
-#define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
-#define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
-#define THREAD_PER_BLOCK 128
-#define THREAD_PER_ROW 2 // 2 thread for each row in matrix A to load numbers
-#define THREAD_PER_COL 4 // 4 thread for each row in matrix B to load numbers
-#define SG_MAT_SIZE 64 // simdgroup matrix is of shape 8x8
-#define SG_MAT_ROW 8
-
 // each block_q contains 16*nl weights
 template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
 kernel void kernel_mul_mm(
@@ -8167,18 +8164,48 @@ kernel void kernel_mul_mm(
     threadgroup S0 * sa = (threadgroup S0 *)(shmem);
     threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
 
-    const int r0 = tgpig.y;
-    const int r1 = tgpig.x;
+    threadgroup float * sc = (threadgroup float *)(shmem);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
     const int im = tgpig.z;
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
 
     // if this block is of 64x32 shape or smaller
-    const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    const short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (args.ne1 - r1 < NR1) ? (args.ne1 - r1) : NR1;
 
     // a thread shouldn't load data outside of the matrix
-    const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
 
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int i12 = im%args.ne12;
+    const int i13 = im/args.ne12;
+
+    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 + lr1)
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
     S0_8x8 ma[4];
     S1_8x8 mb[2];
 
@@ -8187,36 +8214,36 @@ kernel void kernel_mul_mm(
     for (short i = 0; i < 8; i++){
         mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
     }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
 
-    short il = (tiitg % THREAD_PER_ROW);
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
 
-    const int i12 = im%args.ne12;
-    const int i13 = im/args.ne12;
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
 
-    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
-    const short    offset1 = il/nl;
-
-    device const block_q * x = (device const block_q *)(src0
-        + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
-
-    const short iy = (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL));
-
-    device const T1 * y = (device const T1 *)(src1
-        + args.nb13*i13
-        + args.nb12*i12
-        + args.nb11*(r1*BLOCK_SIZE_N + thread_col)
-        + args.nb10*iy);
-
-    for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
         // load data and store to threadgroup memory
         if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             // no need for dequantization
             for (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = loop_k + 16*il + i < args.ne00 ? ((device T0 *) x)[i] : 0;
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
             }
         } else {
             S0_4x4 temp_a;
@@ -8225,91 +8252,203 @@ kernel void kernel_mul_mm(
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             FOR_UNROLL (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
             }
         }
 
         if (FC_mul_mm_bc_inp) {
             for (short i = 0; i < 8; ++i) {
-                sb[32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL) + i] = loop_k + iy + i < args.ne00 ? (S1) ((device T1 *) y)[i] : 0;
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
             }
         } else {
-            *(threadgroup S1_2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (S1_2x4)(*((device T1_2x4 *) y));
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            const short dx = sx;
+            const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
         }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
         x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
-        y += BLOCK_SIZE_K;
+
+        y += NK;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
+#ifndef GGML_METAL_HAS_TENSOR
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup const S0 * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
-        threadgroup const S1 * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
 
-        #pragma unroll(4)
-        for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
             simdgroup_barrier(mem_flags::mem_none);
 
-            #pragma unroll(4)
-            for (short i = 0; i < 4; i++) {
-                simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
-            }
-
-            #pragma unroll(2)
-            for (short i = 0; i < 2; i++) {
-                simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
             }
 
             simdgroup_barrier(mem_flags::mem_none);
 
-            #pragma unroll(8)
-            for (short i = 0; i < 8; i++){
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
 
-            lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE;
-            lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE;
+            lsma += 8*64;
+            lsmb += 4*64;
         }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
     }
 
-    if (!FC_mul_mm_bc_out || ((r0 + 1) * BLOCK_SIZE_M <= args.ne0 && (r1 + 1) * BLOCK_SIZE_N <= args.ne1)) {
+    if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
         // if no bounds checks on the output are needed, we can directly write to device memory
+#ifdef GGML_METAL_HAS_TENSOR
         device float * C = (device float *) dst +
-            (BLOCK_SIZE_M * r0 + 32*(sgitg &  1)) + \
-            (BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+            r0 + \
+            r1 * args.ne0 + im*args.ne1*args.ne0;
+
+        auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(args.ne0, NR1));
+        cT.store(tC);
+#else
+        device float * C = (device float *) dst +
+            (r0 + 32*(sgitg &  1)) + \
+            (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
 
         for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0);
+            simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
         }
+#endif
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                     + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
+
+        threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+#ifdef GGML_METAL_HAS_TENSOR
+        auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+        cT.store(tC);
+#else
         for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
         }
+#endif
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         if (sgitg == 0) {
-            for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*args.ne0 + im*args.ne1*args.ne0;
+            for (int j = tiitg; j < nr1; j += NR1) {
+                device float  * D  = (device float  *) dst + r0 + (r1 + j)*args.ne0 + im*args.ne1*args.ne0;
                 device float4 * D4 = (device float4 *) D;
 
-                threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
+                threadgroup float  * C  = temp_str + (j*NR0);
                 threadgroup float4 * C4 = (threadgroup float4 *) C;
 
                 int i = 0;
-                for (; i < n_rows/4; i++) {
+                for (; i < nr0/4; i++) {
                     *(D4 + i) = *(C4 + i);
                 }
 
                 i *= 4;
-                for (; i < n_rows; i++) {
+                for (; i < nr0; i++) {
                     *(D + i) = *(C + i);
                 }
             }
@@ -8394,31 +8533,63 @@ kernel void kernel_mul_mm_id(
         ushort tiitg[[thread_index_in_threadgroup]],
         ushort tiisg[[thread_index_in_simdgroup]],
         ushort sgitg[[simdgroup_index_in_threadgroup]]) {
-
     threadgroup S0 * sa = (threadgroup S0 *)(shmem);
     threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
 
-    const int r0 = tgpig.y;
-    const int r1 = tgpig.x;
+    threadgroup float * sc = (threadgroup float *)(shmem);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
     const int im = tgpig.z; // expert
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
 
     device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
     device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
 
     const int32_t neh1 = tpe_u32[im];
 
-    if (r1*BLOCK_SIZE_N >= neh1) {
+    if (r1 >= neh1) {
         return;
     }
 
     // if this block is of 64x32 shape or smaller
-    const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    const short n_cols = (    neh1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (    neh1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;
 
     // a thread shouldn't load data outside of the matrix
-    const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
 
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int id = ids_i32[im*args.ne21 + r1 + lr1];
+
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
     S0_8x8 ma[4];
     S1_8x8 mb[2];
 
@@ -8427,39 +8598,36 @@ kernel void kernel_mul_mm_id(
     for (short i = 0; i < 8; i++){
         mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
     }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
 
-    short il = (tiitg % THREAD_PER_ROW);
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
 
-    const int id = ids_i32[im*args.ne21 + r1*BLOCK_SIZE_N + thread_col];
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
 
-    const short i11 = (id % args.ne20) % args.ne11;
-    const short i12 = (id / args.ne20);
-    const short i13 = 0;
-
-    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
-    const short    offset1 = il/nl;
-
-    device const block_q * x = (device const block_q *)(src0
-        + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
-
-    const short iy = (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL));
-
-    device const T1 * y = (device const T1 *)(src1
-        + args.nb13*i13
-        + args.nb12*i12
-        + args.nb11*i11
-        + args.nb10*iy);
-
-    for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
         // load data and store to threadgroup memory
         if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             // no need for dequantization
             for (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = loop_k + 16*il + i < args.ne00 ? ((device T0 *) x)[i] : 0;
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
             }
         } else {
             S0_4x4 temp_a;
@@ -8468,85 +8636,188 @@ kernel void kernel_mul_mm_id(
             threadgroup_barrier(mem_flags::mem_threadgroup);
 
             FOR_UNROLL (short i = 0; i < 16; i++) {
-                *(sa + SG_MAT_SIZE * ((tiitg/THREAD_PER_ROW/8) \
-                +                     (tiitg%THREAD_PER_ROW)*16 + (i/8)*8) \
-                +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
             }
         }
 
         if (FC_mul_mm_bc_inp) {
             for (short i = 0; i < 8; ++i) {
-                sb[32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL) + i] = loop_k + iy + i < args.ne00 ? (S1) ((device T1 *) y)[i] : 0;
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
             }
         } else {
-            *(threadgroup S1_2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = (S1_2x4)(*((device T1_2x4 *) y));
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            const short dx = sx;
+            const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
         }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
         x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
-        y += BLOCK_SIZE_K;
+
+        y += NK;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
+#ifndef GGML_METAL_HAS_TENSOR
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup const S0 * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
-        threadgroup const S1 * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
 
-        #pragma unroll(4)
-        for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
-            #pragma unroll(4)
-            for (short i = 0; i < 4; i++) {
-                simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
             }
 
             simdgroup_barrier(mem_flags::mem_none);
 
-            #pragma unroll(2)
-            for (short i = 0; i < 2; i++) {
-                simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
             }
 
-            #pragma unroll(8)
-            for (short i = 0; i < 8; i++){
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
 
-            lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE;
-            lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE;
+            lsma += 8*64;
+            lsmb += 4*64;
         }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
     }
 
+    // block is smaller than 64x32, we should avoid writing data outside of the matrix
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    threadgroup float * temp_str = ((threadgroup float *) shmem) \
-                                 + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
+#ifdef GGML_METAL_HAS_TENSOR
+    auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+    cT.store(tC);
+#else
+    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
 
-    #pragma unroll(8)
     for (short i = 0; i < 8; i++) {
-        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
     }
+#endif
 
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    for (short j = sgitg; j < n_cols; j += 4) {
-        const int id = ids_i32[im*args.ne21 + r1*BLOCK_SIZE_N + j];
+    for (short j = sgitg; j < nr1; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1 + j];
 
         const short ide = id % args.ne20;
         const short idt = id / args.ne20;
 
-        device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
         device float4 * D4 = (device float4 *) D;
 
-        threadgroup float  * C  = (threadgroup float  *) shmem + (j*BLOCK_SIZE_M);
+        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
         threadgroup float4 * C4 = (threadgroup float4 *) C;
 
         int i = tiisg;
-        for (; i < n_rows/4; i += 32) {
+        for (; i < nr0/4; i += 32) {
             *(D4 + i) = *(C4 + i);
         }
 
-        i = (4*(n_rows/4)) + tiisg;
-        for (; i < n_rows; i += 32) {
+        i = (4*(nr0/4)) + tiisg;
+        for (; i < nr0; i += 32) {
             *(D + i) = *(C + i);
         }
     }

ggml/src/ggml-sycl/concat.cpp

@@ -11,9 +11,13 @@
 //
 
 #include "concat.hpp"
-#include "common.hpp"
 
-static void concat_f32_dim0(const float *x, const float *y, float *dst,
+static inline size_t elem_size(ggml_type t) {
+    return ggml_type_size(t) / ggml_blck_size(t);
+}
+
+template <typename T>
+static void concat_T_dim0(const T *x, const T *y, T *dst,
                             const int ne0, const int ne00,
                             const sycl::nd_item<3> &item_ct1) {
   int nidx = item_ct1.get_local_id(2) +
@@ -36,7 +40,8 @@ static void concat_f32_dim0(const float *x, const float *y, float *dst,
   }
 }
 
-static void concat_f32_dim1(const float *x, const float *y, float *dst,
+template <typename T>
+static void concat_T_dim1(const T *x, const T *y, T *dst,
                             const int ne0, const int ne01,
                             const sycl::nd_item<3> &item_ct1) {
   int nidx = item_ct1.get_local_id(2) +
@@ -59,7 +64,8 @@ static void concat_f32_dim1(const float *x, const float *y, float *dst,
   }
 }
 
-static void concat_f32_dim2(const float *x, const float *y, float *dst,
+template <typename T>
+static void concat_T_dim2(const T *x, const T *y, T *dst,
                             const int ne0, const int ne02,
                             const sycl::nd_item<3> &item_ct1) {
   int nidx = item_ct1.get_local_id(2) +
@@ -82,45 +88,35 @@ static void concat_f32_dim2(const float *x, const float *y, float *dst,
   }
 }
 
-static void concat_f32_sycl(const float *x, const float *y, float *dst,
+template <typename T>
+static void concat_T_sycl(const T *x, const T *y, T *dst,
                             int ne00, int ne01, int ne02, int ne0, int ne1,
                             int ne2, int dim, queue_ptr stream) {
   int num_blocks = (ne0 + SYCL_CONCAT_BLOCK_SIZE - 1) / SYCL_CONCAT_BLOCK_SIZE;
   sycl::range<3> gridDim(ne2, ne1, num_blocks);
   switch (dim) {
   case 0:
-    stream->parallel_for(
-        sycl::nd_range<3>(gridDim *
-                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-          concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1);
-        });
-    break;
+      stream->parallel_for(sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+                        [=](sycl::nd_item<3> item_ct1) { concat_T_dim0<T>(x, y, dst, ne0, ne00, item_ct1); });
+      break;
   case 1:
-    stream->parallel_for(
-        sycl::nd_range<3>(gridDim *
-                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-          concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1);
-        });
-    break;
+      stream->parallel_for(sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+                        [=](sycl::nd_item<3> item_ct1) { concat_T_dim1<T>(x, y, dst, ne0, ne01, item_ct1); });
+      break;
   // dim >=2 will be dispatched to the default path
   default:
-    stream->parallel_for(
-        sycl::nd_range<3>(gridDim *
-                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-          concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1);
-        });
-    break;
+      stream->parallel_for(sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+                        [=](sycl::nd_item<3> item_ct1) { concat_T_dim2<T>(x, y, dst, ne0, ne02, item_ct1); });
+      break;
   }
 }
 
 // non-contiguous kernel (slow)
-static void concat_f32_sycl_non_cont(
+template<typename T>
+static void concat_T_sycl_non_cont(
     queue_ptr stream, const char *src0, const char *src1, char *dst,
     int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03, uint64_t nb00,
     uint64_t nb01, uint64_t nb02, uint64_t nb03, int64_t /*ne10*/,
@@ -137,24 +133,25 @@ static void concat_f32_sycl_non_cont(
       int64_t o[4] = { 0, 0, 0, 0 };
       o[dim]       = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
 
-      const float * x;
+      const T * x;
 
       for (int i0 = item_ct1.get_local_id(2); i0 < ne0; i0 += item_ct1.get_local_range(2)) {
           if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-              x = (const float *) (src0 + (i3) *nb03 + (i2) *nb02 + (i1) *nb01 + (i0) *nb00);
+              x = (const T *) (src0 + (i3) *nb03 + (i2) *nb02 + (i1) *nb01 + (i0) *nb00);
           } else {
-              x = (const float *) (src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 + (i1 - o[1]) * nb11 +
+              x = (const T *) (src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 + (i1 - o[1]) * nb11 +
                                    (i0 - o[0]) * nb10);
           }
 
-          float *y = (float *)(dst + i3 * nb3 + i2 * nb2 + i1 * nb1 + i0 * nb0);
+          T *y = (T *)(dst + i3 * nb3 + i2 * nb2 + i1 * nb1 + i0 * nb0);
 
           *y = *x;
       }
   });
 }
 
-void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+template <typename T>
+void concat_impl_sycl(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
     const ggml_tensor *  src0   = dst->src[0];
     const ggml_tensor *  src1   = dst->src[1];
@@ -163,15 +160,14 @@ void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
     const int32_t dim = ((int32_t *) dst->op_params)[0];
 
     if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
-        const float * src0_d = (const float *) src0->data;
-        const float * src1_d = (const float *) src1->data;
-
-        float * dst_d = (float *) dst->data;
-
+        const T * src0_d = (const T *) src0->data;
+        const T * src1_d = (const T *) src1->data;
+        T * dst_d = (T *) dst->data;
+        size_t type_size = elem_size(dst->type);
         if (dim != 3) {
             for (int i3 = 0; i3 < dst->ne[3]; i3++) {
-                concat_f32_sycl(src0_d + i3 * (src0->nb[3] / 4), src1_d + i3 * (src1->nb[3] / 4),
-                                dst_d + i3 * (dst->nb[3] / 4), src0->ne[0], src0->ne[1], src0->ne[2], dst->ne[0],
+                concat_T_sycl<T>(src0_d + i3 * (src0->nb[3] / type_size), src1_d + i3 * (src1->nb[3] / type_size),
+                                dst_d + i3 * (dst->nb[3] / type_size), src0->ne[0], src0->ne[1], src0->ne[2], dst->ne[0],
                                 dst->ne[1], dst->ne[2], dim, stream);
             }
         } else {
@@ -179,13 +175,28 @@ void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
             const size_t size1 = ggml_nbytes(src1);
 
             SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d, src0_d, size0).wait()));
-            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d + size0 / 4, src1_d, size1).wait()));
+            SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d + size0 / type_size, src1_d, size1).wait()));
         }
     } else {
-        concat_f32_sycl_non_cont(stream, (const char *) src0->data, (const char *) src1->data, (char *) dst->data,
+        concat_T_sycl_non_cont<T>(stream, (const char *) src0->data, (const char *) src1->data, (char *) dst->data,
                                  src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src0->nb[0], src0->nb[1],
                                  src0->nb[2], src0->nb[3], src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
                                  src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3], dst->ne[0], dst->ne[1], dst->ne[2],
                                  dst->ne[3], dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim);
     }
 }
+
+void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+
+    switch (dst->type) {
+    case GGML_TYPE_F32:
+        concat_impl_sycl<float>(ctx, dst);
+        break;
+    case GGML_TYPE_I32:
+        concat_impl_sycl<int32_t>(ctx, dst);
+        break;
+    default:
+    GGML_ASSERT(false && "ggml_sycl_op_concat: unsupported type");
+    break;
+    }
+}

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -4534,16 +4534,12 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 }
                 return false;
             }
-        case GGML_OP_CONCAT:
-            {
-                ggml_type src0_type = op->src[0]->type;
-                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
-            }
         case GGML_OP_REPEAT_BACK:
             {
                 ggml_type src0_type = op->src[0]->type;
                 return src0_type == GGML_TYPE_F32;
             }
+        case GGML_OP_CONCAT:
         case GGML_OP_DUP:
         case GGML_OP_ARGMAX:
         case GGML_OP_NONE:

ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.glsl

@@ -3,6 +3,9 @@
 
 #include "rte.glsl"
 #include "utils.glsl"
+#if RMS_NORM_ROPE_FUSION
+#include "rope_params.glsl"
+#endif
 
 layout (push_constant) uniform parameter
 {
@@ -12,11 +15,16 @@ layout (push_constant) uniform parameter
     uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
     uint misalign_offsets;
     float param1; float param2; int param3;
+#if RMS_NORM_ROPE_FUSION
+    rope_params rope;
+#endif
 } p;
 
+#if !RMS_NORM_ROPE_FUSION
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+#endif
 
 // true if src0/src1 are the same shape and the indices can be reused without additional modulus
 layout(constant_id = 0) const bool norepeat = false;

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

@@ -100,7 +100,6 @@ layout (push_constant) uniform parameter
 layout (constant_id = 0) const uint BLOCK_SIZE = 64;
 layout (constant_id = 1) const uint BM = 64;
 layout (constant_id = 2) const uint BN = 64;
-layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
 layout (constant_id = 4) const uint WM = 32;
 layout (constant_id = 5) const uint WN = 32;
 layout (constant_id = 6) const uint WMITER = 2;
@@ -109,6 +108,14 @@ layout (constant_id = 8) const uint TN = 2;
 layout (constant_id = 9) const uint TK = 1;  // Only needed for coopmat
 layout (constant_id = 10) const uint WARP = 32;
 
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+#define BK 32
+#define BK_STEP 4
+#else
+layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
+#define BK_STEP 2
+#endif
+
 #ifdef COOPMAT
 #define SHMEM_STRIDE (BK / 2 + 4)
 #else
@@ -244,8 +251,13 @@ void main() {
     }
 #else
     ACC_TYPE_VEC2 sums[WMITER * TM * WNITER * TN/2];
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+    FLOAT_TYPE_VEC4 cache_a[WMITER * TM];
+    FLOAT_TYPE_VEC4 cache_b;
+#else
     FLOAT_TYPE_VEC2 cache_a[WMITER * TM];
     FLOAT_TYPE_VEC2 cache_b;
+#endif
 
     [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN/2; i++) {
         sums[i] = ACC_TYPE_VEC2(0.0f, 0.0f);
@@ -283,24 +295,41 @@ void main() {
             }
         }
 #else
-        [[unroll]] for (uint i = 0; i < BK / 2; i++) {
+        [[unroll]] for (uint i = 0; i < BK / BK_STEP; i++) {
             // Load from shared into cache
             [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                 [[unroll]] for (uint j = 0; j < TM; j++) {
+                #if defined(DATA_A_F32) || defined(DATA_A_F16)
+                    cache_a[wsir * TM + j].xy = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + 2 * i    ];
+                    cache_a[wsir * TM + j].zw = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + 2 * i + 1];
+                #else
                     cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + i];
+                #endif
                 }
             }
 
             [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
                 [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+                #if defined(DATA_A_F32) || defined(DATA_A_F16)
+                    cache_b.xy = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + 2 * i    ];
+                    cache_b.zw = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + 2 * i + 1];
+                #else
                     cache_b = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + cc) * SHMEM_STRIDE + i];
+                #endif
 
                     [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                         [[unroll]] for (uint cr = 0; cr < TM / 2; cr++) {
                             // [WNITER][TN][WMITER][TM / 2] -> [wsic][cc][wsir][cr]
                             const uint sums_idx = (wsic * TN + cc) * WMITER * (TM / 2) + wsir * (TM / 2) + cr;
+                        #if defined(DATA_A_F32) || defined(DATA_A_F16)
+                            sums[sums_idx].x = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].y), ACC_TYPE(cache_b.y),
+                                               fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].z), ACC_TYPE(cache_b.z), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].w), ACC_TYPE(cache_b.w), sums[sums_idx].x))));
+                            sums[sums_idx].y = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].y), ACC_TYPE(cache_b.y),
+                                               fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].z), ACC_TYPE(cache_b.z), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].w), ACC_TYPE(cache_b.w), sums[sums_idx].y))));
+                        #else
                             sums[sums_idx].x = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr    ].y), ACC_TYPE(cache_b.y), sums[sums_idx].x));
                             sums[sums_idx].y = fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].x), ACC_TYPE(cache_b.x), fma(ACC_TYPE(cache_a[wsir * TM + 2 * cr + 1].y), ACC_TYPE(cache_b.y), sums[sums_idx].y));
+                        #endif
                         }
                     }
                 }

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

@@ -3,6 +3,32 @@
 #include "generic_binary_head.glsl"
 #include "types.glsl"
 
+#if RMS_NORM_ROPE_FUSION
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+
+// data is passed from rms_norm -> rope through shared memory.
+// rms_norm calls this data_d, rope calls this rope_data_a.
+// Binding 2 is not used
+shared FLOAT_TYPE rope_data_a[1024];
+#define data_d rope_data_a
+
+layout (binding = 3) readonly buffer R_Y {int rope_data_pos[];};
+layout (binding = 4) readonly buffer R_Z {float rope_data_ff[];};
+layout (binding = 5) writeonly buffer R_D {ROPE_D_TYPE rope_data_d[];};
+layout (binding = 6) readonly buffer R_I {uvec2 rope_data_i[];}; // indices for set_rows
+
+#include "rope_params.glsl"
+#include "rope_funcs.glsl"
+
+#define GGML_ROPE_TYPE_NORMAL 0
+#define GGML_ROPE_TYPE_NEOX   2
+#define GGML_ROPE_TYPE_MROPE  8
+#define GGML_ROPE_TYPE_VISION 24
+
+#endif
+
 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512
 
@@ -28,8 +54,12 @@ void rms_norm(uint num_iters) {
 
     uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset();
     uint32_t b_offset = src1_idx(0, row, channel, samp) + get_boffset();
+#if RMS_NORM_ROPE_FUSION
+    // Per-row offset in shared memory
+    uint32_t d_offset = 0;
+#else
     uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
-
+#endif
     FLOAT_TYPE sum = FLOAT_TYPE(0.0f); // partial sum for thread in warp
 
     [[unroll]] for (uint col = tid, idx = 0; idx < num_iters; col += BLOCK_SIZE, ++idx) {
@@ -79,6 +109,18 @@ void rms_norm(uint num_iters) {
             data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
         }
     }
+#if RMS_NORM_ROPE_FUSION
+    barrier();
+    rope_params rp = p.rope;
+    uint rope_row = (samp*nchannels + channel)*nrows + row;
+    for (uint t = 2*tid; t < ncols; t += 2*BLOCK_SIZE) {
+        if (rp.rope_mode == GGML_ROPE_TYPE_NEOX) {
+            rope_neox(t, rope_row, rp);
+        } else if (rp.rope_mode == GGML_ROPE_TYPE_NORMAL) {
+            rope_norm(t, rope_row, rp);
+        }
+    }
+#endif
 }
 
 void main() {

ggml/src/ggml-vulkan/vulkan-shaders/rope_funcs.glsl

@@ -0,0 +1,227 @@
+
+float rope_yarn_ramp(const float low, const float high, const uint i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+uint rope_a_coord(const uint i0, const uint i01, const uint i02, rope_params p) {
+#if RMS_NORM_ROPE_FUSION
+    // Per-row offset in shared memory
+    const uint ix = i0;
+#else
+    const uint ix = i02*p.nb02 + i01*p.nb01 + i0;
+#endif
+    return ix;
+}
+
+void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out float sin_theta, rope_params p) {
+    float mscale = p.attn_factor;
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = p.freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (p.ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(p.corr_dims[0], p.corr_dims[1], i0) * p.ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale);
+    }
+    // Backprogagation uses inverted rotation
+    if (p.is_back != 0) {
+        theta = -theta;
+    }
+    cos_theta = cos(theta) * mscale;
+    sin_theta = sin(theta) * mscale;
+}
+
+void rope_norm(const uint i0, const uint i1, rope_params p) {
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    // i1 is actually i2*nb2+i1, but the rows are contiguous
+    const uint i01 = i1 % ne1;
+    const uint i02 = i1 / ne1;
+
+    uint idst = i1*ne0 + i0;
+    const uint ix = rope_a_coord(i0, i01, i02, p);
+
+    // Fusion optimization: ROPE + VIEW + SET_ROWS..
+    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
+    if (p.set_rows_stride != 0) {
+        idst = i01*ne0 + i0;
+        idst += rope_data_i[i02].x * p.set_rows_stride;
+    }
+
+    if (i0 >= p.n_dims) {
+        rope_data_d[idst + 0] = ROPE_D_TYPE(rope_data_a[ix + 0]);
+        rope_data_d[idst + 1] = ROPE_D_TYPE(rope_data_a[ix + 1]);
+
+        return;
+    }
+
+    const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+
+    const float x0 = float(rope_data_a[ix + 0]);
+    const float x1 = float(rope_data_a[ix + 1]);
+
+    rope_data_d[idst + 0] = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + 1] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+}
+
+void rope_neox(const uint i0, const uint i1, rope_params p) {
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const uint i01 = i1 % ne1;
+    const uint i02 = i1 / ne1;
+
+    uint idst = i1*ne0 + i0/2;
+    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+
+    // Fusion optimization: ROPE + VIEW + SET_ROWS..
+    // The rope output is viewed as a 1D tensor and offset based on a row index in rope_data_i.
+    if (p.set_rows_stride != 0) {
+        idst = i01*ne0 + i0/2;
+        idst += rope_data_i[i02].x * p.set_rows_stride;
+    }
+
+    if (i0 >= p.n_dims) {
+        rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
+        rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);
+
+        return;
+    }
+
+    const float theta_base = rope_data_pos[i02] * pow(p.theta_scale, i0/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+
+    const float x0 = float(rope_data_a[ix + 0]);
+    const float x1 = float(rope_data_a[ix + p.n_dims/2]);
+
+    rope_data_d[idst + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+}
+
+
+void rope_multi(const uint i0, const uint i1, rope_params p) {
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
+    uint ne2 = p.ne02;
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const uint i01 = i1 % ne1;
+    const uint i02 = i1 / ne1;
+
+    const uint idst = i1*ne0 + i0/2;
+    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+
+    if (i0 >= p.n_dims) {
+        rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
+        rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);
+
+        return;
+    }
+
+    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 (p.is_imrope != 0) {
+        if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
+            theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+        } else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
+            theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+        } else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
+            theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+        } else {
+            theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+        }
+    } else {
+        if (sector < p.sections[0]) {
+            theta_base = rope_data_pos[i02]*pow(p.theta_scale, i0/2.0f);
+        }
+        else if (sector >= p.sections[0] && sector < sec_w) {
+            theta_base = rope_data_pos[i02 + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
+            theta_base = rope_data_pos[i02 + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w + p.sections[2]) {
+            theta_base = rope_data_pos[i02 + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
+        }
+    }
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+
+    const float x0 = float(rope_data_a[ix + 0]);
+    const float x1 = float(rope_data_a[ix + p.n_dims/2]);
+
+    rope_data_d[idst + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+}
+
+void rope_vision(const uint i0, const uint i1, rope_params p) {
+    uint ne0 = p.ncols;
+    uint ne1 = p.p_delta_rows;
+    uint ne2 = p.ne02;
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const uint i01 = i1 % ne1;
+    const uint i02 = i1 / ne1;
+
+    const uint idst = i1*ne0 + i0/2;
+    const uint ix = rope_a_coord(i0/2, i01, i02, p);
+
+    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 = rope_data_pos[i02]*pow(p.theta_scale, p0);
+    }
+    else if (sector >= p.sections[0] && sector < sec_w) {
+        const uint p0 = sector - p.sections[0];
+        theta_base = rope_data_pos[i02 + ne2]*pow(p.theta_scale, p0);
+    }
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+
+    const float x0 = float(rope_data_a[ix + 0]);
+    const float x1 = float(rope_data_a[ix + p.n_dims]);
+
+    rope_data_d[idst + 0]        = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + p.n_dims] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+}
+

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

@@ -3,56 +3,18 @@
 #extension GL_EXT_shader_16bit_storage : require
 
 #include "rte.glsl"
+#include "rope_params.glsl"
 
 layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
 
-layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer Y {int data_pos[];};
-layout (binding = 2) readonly buffer Z {float data_ff[];};
-layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
-layout (binding = 4) readonly buffer I {uvec2 data_i[];}; // indices for set_rows
+layout (binding = 0) readonly buffer X {A_TYPE rope_data_a[];};
+layout (binding = 1) readonly buffer Y {int rope_data_pos[];};
+layout (binding = 2) readonly buffer Z {float rope_data_ff[];};
+layout (binding = 3) writeonly buffer D {ROPE_D_TYPE rope_data_d[];};
+layout (binding = 4) readonly buffer I {uvec2 rope_data_i[];}; // indices for set_rows
+
 
 layout (push_constant) uniform parameter {
-    uint ncols;
-    uint n_dims;
-    float freq_scale;
-    uint p_delta_rows;
-    float freq_base;
-    float ext_factor;
-    float attn_factor;
-    float corr_dims[2];
-    float theta_scale;
-    uint has_ff;
-    uint ne02;
-    uint s1;
-    uint s2;
-    int sections[4];
-    uint is_imrope;
-    uint is_back;
-    uint set_rows_stride;
-} p;
+    rope_params pc;
+};
 
-float rope_yarn_ramp(const float low, const float high, const uint i0) {
-    const float y = (i0 / 2 - low) / max(0.001f, high - low);
-    return 1.0f - min(1.0f, max(0.0f, y));
-}
-
-void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out float sin_theta) {
-    float mscale = p.attn_factor;
-    // Get n-d rotational scaling corrected for extrapolation
-    float theta_interp = p.freq_scale * theta_extrap;
-    float theta = theta_interp;
-    if (p.ext_factor != 0.0f) {
-        float ramp_mix = rope_yarn_ramp(p.corr_dims[0], p.corr_dims[1], i0) * p.ext_factor;
-        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
-
-        // Get n-d magnitude scaling corrected for interpolation
-        mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale);
-    }
-    // Backprogagation uses inverted rotation
-    if (p.is_back != 0) {
-        theta = -theta;
-    }
-    cos_theta = cos(theta) * mscale;
-    sin_theta = sin(theta) * mscale;
-}

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

@@ -1,70 +1,11 @@
 #version 450
 
 #include "rope_head.glsl"
+#include "rope_funcs.glsl"
 
 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;
-
-    if (i0 >= p.n_dims) {
-        data_d[idst + i0/2 + 0] = data_a[ix + i0/2 + 0];
-        data_d[idst + i0/2 + 1] = data_a[ix + i0/2 + 1];
-
-        return;
-    }
-
-    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 (p.is_imrope != 0) {
-        if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
-            theta_base = data_pos[channel_x + ne2 * 1]*pow(p.theta_scale, i0/2.0f);
-        } else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
-            theta_base = data_pos[channel_x + ne2 * 2]*pow(p.theta_scale, i0/2.0f);
-        } else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
-            theta_base = data_pos[channel_x]*pow(p.theta_scale, i0/2.0f);
-        } else {
-            theta_base = data_pos[channel_x + ne2 * 3]*pow(p.theta_scale, i0/2.0f);
-        }
-    } else {
-        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);
+    // i1 is actually i2*nb2+i1, but the rows are contiguous
+    const uint i1 = gl_GlobalInvocationID.x;
+    rope_multi(i0, i1, pc);
 }

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

@@ -1,48 +1,11 @@
 #version 450
 
 #include "rope_head.glsl"
+#include "rope_funcs.glsl"
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-
-    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;
-
-    uint idst = row_dst*ne0 + i0/2;
-    const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0/2;
-
-    // Fusion optimization: ROPE + VIEW + SET_ROWS..
-    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
-    if (p.set_rows_stride != 0) {
-        idst = row_x*ne0 + i0/2;
-        idst += data_i[channel_x].x * p.set_rows_stride;
-    }
-
-    if (i0 >= p.n_dims) {
-        data_d[idst + i0/2 + 0] = D_TYPE(data_a[ix + i0/2 + 0]);
-        data_d[idst + i0/2 + 1] = D_TYPE(data_a[ix + i0/2 + 1]);
-
-        return;
-    }
-
-    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, 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);
+    // i1 is actually i2*nb2+i1, but the rows are contiguous
+    const uint i1 = gl_GlobalInvocationID.x;
+    rope_neox(i0, i1, pc);
 }

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

@@ -1,48 +1,11 @@
 #version 450
 
 #include "rope_head.glsl"
+#include "rope_funcs.glsl"
 
 void main() {
     const uint i0 = 2*gl_GlobalInvocationID.y;
-    uint ne0 = p.ncols;
-    uint ne1 = p.p_delta_rows;
-
-    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;
-
-    uint idst = row_dst*ne0 + i0;
-    const uint ix   = channel_x*p.s2 + row_x*p.s1 + i0;
-
-    // Fusion optimization: ROPE + VIEW + SET_ROWS..
-    // The rope output is viewed as a 1D tensor and offset based on a row index in data_i.
-    if (p.set_rows_stride != 0) {
-        idst = row_x*ne0 + i0;
-        idst += data_i[channel_x].x * p.set_rows_stride;
-    }
-
-    if (i0 >= p.n_dims) {
-        data_d[idst + 0] = D_TYPE(data_a[ix + 0]);
-        data_d[idst + 1] = D_TYPE(data_a[ix + 1]);
-
-        return;
-    }
-
-    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, i0, cos_theta, sin_theta);
-
-    const float x0 = float(data_a[ix + 0]);
-    const float x1 = float(data_a[ix + 1]);
-
-    data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta);
-    data_d[idst + 1] = D_TYPE(x0*sin_theta + x1*cos_theta);
+    // i1 is actually i2*nb2+i1, but the rows are contiguous
+    const uint i1 = gl_GlobalInvocationID.x;
+    rope_norm(i0, i1, pc);
 }

ggml/src/ggml-vulkan/vulkan-shaders/rope_params.glsl

@@ -0,0 +1,27 @@
+#if !defined(GGML_ROPE_PARAMS)
+#define GGML_ROPE_PARAMS
+
+#include "rte.glsl"
+
+struct rope_params {
+    uint rope_mode;
+    uint ncols;
+    uint n_dims;
+    float freq_scale;
+    uint p_delta_rows;
+    float freq_base;
+    float ext_factor;
+    float attn_factor;
+    float corr_dims[2];
+    float theta_scale;
+    uint has_ff;
+    uint ne02;
+    uint nb01;
+    uint nb02;
+    int sections[4];
+    uint is_imrope;
+    uint is_back;
+    uint set_rows_stride;
+};
+
+#endif // !defined(GGML_ROPE_PARAMS)

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

@@ -1,47 +1,11 @@
 #version 450
 
 #include "rope_head.glsl"
+#include "rope_funcs.glsl"
 
 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);
+    // i1 is actually i2*nb2+i1, but the rows are contiguous
+    const uint i1 = gl_GlobalInvocationID.x;
+    rope_vision(i0, i1, pc);
 }

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -695,6 +695,8 @@ void process_shaders() {
     string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("rms_norm_partials_f32", "rms_norm_partials.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("rms_norm_mul_rope_f32_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float"}, {"RMS_NORM_ROPE_FUSION", "1"}}));
+    string_to_spv("rms_norm_mul_rope_f32_f16_rte", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RMS_NORM_ROPE_FUSION", "1"}, {"RTE16", "1"}}));
     string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("l2_norm_f32", "l2_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
@@ -840,25 +842,25 @@ void process_shaders() {
     string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
     string_to_spv("soft_max_back_f32", "soft_max_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
 
-    string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-    string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
-    string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
+    string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
-    string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-    string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
-    string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
+    string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
-    string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-    string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
+    string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
-    string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
-    string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
-    string_to_spv("rope_vision_f16_rte", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
+    string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
+    string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
+    string_to_spv("rope_vision_f16_rte", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
 

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -15,6 +15,7 @@
 #include <condition_variable>
 #include <cstring>
 #include <iostream>
+#include <map>
 #include <mutex>
 #include <optional>
 #include <string>
@@ -73,6 +74,30 @@
 // For operations which process a row in parallel, this seems like a reasonable default
 #define WEBGPU_ROW_SPLIT_WG_SIZE 64
 
+// Matrix multiplication parameters
+
+// Register tiling parameters
+#define WEBGPU_MUL_MAT_TILE_M    8
+#define WEBGPU_MUL_MAT_TILE_N    8
+#define WEBGPU_MUL_MAT_WG_SIZE_M 8
+#define WEBGPU_MUL_MAT_WG_SIZE_N 8
+#define WEBGPU_MUL_MAT_TILE_K    32
+
+// Subgroup matrix parameters
+// The number of subgroups in the M dimension
+#define WEBGPU_MUL_MAT_SUBGROUP_M        2
+// The number of subgroups in the N dimension
+#define WEBGPU_MUL_MAT_SUBGROUP_N        2
+// The number of subgroup matrices each subgroup accumulates over
+#define WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M 4
+#define WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N 2
+
+// Matrix-vector multiplication parameters
+#define WEBGPU_MUL_MAT_VEC_WG_SIZE        256
+// Must be multiple of 4 to work with vectorized paths, and must divide mul_mat_vec wg size
+#define WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG 64
+#define WEBGPU_MUL_MAT_VEC_TILE_K         256
+
 /* End Constants */
 
 // This is a "fake" base pointer, since WebGPU buffers do not have pointers to their locations.
@@ -236,6 +261,10 @@ struct webgpu_context_struct {
     wgpu::Queue    queue;
     wgpu::Limits   limits;
 
+    bool                       supports_subgroup_matrix = false;
+    uint32_t                   subgroup_size;
+    wgpu::SubgroupMatrixConfig subgroup_matrix_config;
+
     // Separate this out from limits since on some Metal systems, the limit returned by
     // querying the limits is higher than the actual allowed maximum.
     uint32_t max_wg_size_x;
@@ -247,8 +276,13 @@ struct webgpu_context_struct {
     webgpu_buf_pool set_rows_error_buf_pool;
 
     webgpu_pipeline memset_pipeline;
+
+    std::map<int, std::map<int, std::map<int, webgpu_pipeline>>> mul_mat_pipelines;  // src0_type, src1_type, vectorized
+    std::map<int, std::map<int, std::map<int, webgpu_pipeline>>>
+        mul_mat_vec_pipelines;                                                       // src0_type, src1_type, vectorized
+
     webgpu_pipeline mul_mat_pipeline[30][2];
-    webgpu_pipeline set_rows_pipeline;
+    webgpu_pipeline set_rows_pipeline[1][2];  // dst->type, vectorized
     webgpu_pipeline get_rows_pipeline[30];
     webgpu_pipeline get_rows_f32_no_vec_pipeline;
     webgpu_pipeline cpy_pipeline[2][2];          // src type, dst type
@@ -309,16 +343,37 @@ struct ggml_backend_webgpu_context {
 struct ggml_backend_webgpu_buffer_context {
     webgpu_context webgpu_ctx;
     wgpu::Buffer   buffer;
+    std::string    label;
 
-    ggml_backend_webgpu_buffer_context(webgpu_context ctx, wgpu::Buffer buf) :
+    ggml_backend_webgpu_buffer_context(webgpu_context ctx, wgpu::Buffer buf, std::string lbl) :
         webgpu_ctx(std::move(ctx)),
-        buffer(std::move(buf)) {}
+        buffer(std::move(buf)),
+        label(std::move(lbl)) {}
 };
 
 /* End struct definitions */
 
 /* WebGPU object initializations */
 
+// Process a WGSL shader string, replacing tokens of the form {{KEY}} with
+// the corresponding values provided in `repls`.
+static std::string ggml_webgpu_process_shader_repls(const char *                               src,
+                                                    const std::map<std::string, std::string> & repls) {
+    if (!src) {
+        return std::string();
+    }
+    std::string s = src;
+    for (const auto & kv : repls) {
+        std::string token = "{{" + kv.first + "}}";
+        size_t      pos   = 0;
+        while ((pos = s.find(token, pos)) != std::string::npos) {
+            s.replace(pos, token.length(), kv.second);
+            pos += kv.second.length();
+        }
+    }
+    return s;
+}
+
 static void ggml_webgpu_create_pipeline(wgpu::Device &                           device,
                                         webgpu_pipeline &                        pipeline,
                                         const char *                             shader_code,
@@ -344,6 +399,30 @@ static void ggml_webgpu_create_pipeline(wgpu::Device &
     pipeline = { device.CreateComputePipeline(&pipeline_desc), label };
 }
 
+static webgpu_pipeline ggml_webgpu_create_pipeline2(wgpu::Device &                           device,
+                                                    const char *                             shader_code,
+                                                    const char *                             label,
+                                                    const std::vector<wgpu::ConstantEntry> & constants = {}) {
+    wgpu::ShaderSourceWGSL shader_source;
+    shader_source.code = shader_code;
+
+    wgpu::ShaderModuleDescriptor shader_desc;
+    shader_desc.nextInChain = &shader_source;
+
+    wgpu::ShaderModule shader_module = device.CreateShaderModule(&shader_desc);
+
+    wgpu::ComputePipelineDescriptor pipeline_desc;
+    pipeline_desc.label              = label;
+    pipeline_desc.compute.module     = shader_module;
+    pipeline_desc.compute.entryPoint = "main";   // Entry point in the WGSL code
+    pipeline_desc.layout             = nullptr;  // nullptr means auto layout
+    if (constants.size() > 0) {
+        pipeline_desc.compute.constants     = constants.data();
+        pipeline_desc.compute.constantCount = constants.size();
+    }
+    return { device.CreateComputePipeline(&pipeline_desc), label };
+}
+
 static void ggml_webgpu_create_buffer(wgpu::Device &    device,
                                       wgpu::Buffer &    buffer,
                                       size_t            size,
@@ -510,6 +589,7 @@ static webgpu_command ggml_backend_webgpu_build(webgpu_context &
                                                 std::vector<uint32_t>             params,
                                                 std::vector<wgpu::BindGroupEntry> bind_group_entries,
                                                 uint32_t                          wg_x,
+                                                uint32_t                          wg_y                = 1,
                                                 std::optional<webgpu_pool_bufs>   set_rows_error_bufs = std::nullopt) {
     webgpu_pool_bufs params_bufs = ctx->param_buf_pool.alloc_bufs();
 
@@ -555,7 +635,7 @@ static webgpu_command ggml_backend_webgpu_build(webgpu_context &
 #endif
     pass.SetPipeline(pipeline.pipeline);
     pass.SetBindGroup(0, bind_group);
-    pass.DispatchWorkgroups(wg_x, 1, 1);
+    pass.DispatchWorkgroups(wg_x, wg_y, 1);
     pass.End();
 
 #ifdef GGML_WEBGPU_GPU_PROFILE
@@ -764,10 +844,20 @@ static std::optional<webgpu_command> ggml_webgpu_set_rows(webgpu_context & ctx,
         { .binding = 3, .buffer = error_bufs.dev_buf, .offset = 0, .size = error_bufs.dev_buf.GetSize() }
     };
 
-    size_t   max_wg_size = ctx->max_wg_size_x;
-    uint32_t wg_x        = (src->ne[1] * src->ne[2] * src->ne[3] + max_wg_size - 1) / max_wg_size;
+    size_t max_wg_size = ctx->max_wg_size_x;
 
-    return ggml_backend_webgpu_build(ctx, ctx->set_rows_pipeline, params, entries, wg_x, error_bufs);
+    int             vectorized = src->ne[0] % 4 == 0;
+    webgpu_pipeline pipeline   = ctx->set_rows_pipeline[0][vectorized];
+    uint32_t        threads;
+    if (vectorized) {
+        threads = (src->ne[1] * src->ne[2] * src->ne[3]) * (src->ne[0] / 4);
+    } else {
+        threads = src->ne[0] * src->ne[1] * src->ne[2] * src->ne[3];
+    }
+
+    uint32_t wg_x = (threads + max_wg_size - 1) / max_wg_size;
+
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, 1, error_bufs);
 }
 
 static webgpu_command ggml_webgpu_get_rows(webgpu_context & ctx,
@@ -823,8 +913,8 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
         (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
-        (uint32_t) dst->ne[1],                                  // number of rows in result (M)
-        (uint32_t) dst->ne[0],                                  // number of columns in result (N)
+        (uint32_t) dst->ne[0],                                  // number of rows in result (M, transposed)
+        (uint32_t) dst->ne[1],                                  // number of columns in result (N)
         (uint32_t) src0->ne[0],                                 // number of columns in src0/src1 (K)
         (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),  // stride (elements/blocks) of src0 in dimension 1
         (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),  // stride (elements/blocks) of src1 in dimension 1
@@ -853,9 +943,67 @@ static webgpu_command ggml_webgpu_mul_mat(webgpu_context & ctx,
          .size    = ggml_webgpu_tensor_binding_size(ctx, dst)  },
     };
 
+    webgpu_pipeline pipeline = ctx->mul_mat_pipeline[src0->type][src1->type];
+
     uint32_t wg_x =
         (dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3] + WEBGPU_MUL_MAT_WG_SIZE - 1) / WEBGPU_MUL_MAT_WG_SIZE;
-    return ggml_backend_webgpu_build(ctx, ctx->mul_mat_pipeline[src0->type][src1->type], params, entries, wg_x);
+    uint32_t wg_y = 1;
+
+    bool use_fast = false;
+    switch (src1->type) {
+        case GGML_TYPE_F16:
+            use_fast = (src0->type == GGML_TYPE_F16);
+            break;
+        case GGML_TYPE_F32:
+            switch (src0->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                    use_fast = true;
+                    break;
+                default:
+                    break;
+            }
+            break;
+        default:
+            break;
+    }
+
+    if (use_fast) {
+        int vectorized = src0->ne[0] % 4 == 0 && dst->ne[0] % 4 == 0 && dst->ne[1] % 4 == 0;
+        if (dst->ne[1] == 1) {
+            // We don't support vectorized mul_mat_vec for quantized types
+            vectorized       = vectorized && (src0->type < 2);
+            pipeline         = ctx->mul_mat_vec_pipelines[src0->type][src1->type][vectorized];
+            uint32_t batches = dst->ne[2] * dst->ne[3];
+            uint32_t output_groups =
+                (dst->ne[0] + WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG - 1) / WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG;
+            uint32_t total_wg = output_groups * batches;
+            wg_x              = total_wg % ctx->limits.maxComputeWorkgroupsPerDimension;
+            wg_y              = (total_wg + ctx->limits.maxComputeWorkgroupsPerDimension - 1) /
+                   ctx->limits.maxComputeWorkgroupsPerDimension;
+        } else {
+            pipeline = ctx->mul_mat_pipelines[src0->type][src1->type][vectorized];
+            uint32_t wg_m;
+            uint32_t wg_n;
+            if (ctx->supports_subgroup_matrix) {
+                // The total number of subgroups/workgroups needed per matrix.
+                uint32_t wg_m_sg_tile =
+                    WEBGPU_MUL_MAT_SUBGROUP_M * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M * ctx->subgroup_matrix_config.M;
+                wg_m = (dst->ne[0] + wg_m_sg_tile - 1) / wg_m_sg_tile;
+                uint32_t wg_n_sg_tile =
+                    WEBGPU_MUL_MAT_SUBGROUP_N * WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N * ctx->subgroup_matrix_config.N;
+                wg_n = (dst->ne[1] + wg_n_sg_tile - 1) / wg_n_sg_tile;
+            } else {
+                uint32_t tile_m_s = WEBGPU_MUL_MAT_TILE_M * WEBGPU_MUL_MAT_WG_SIZE_M;
+                uint32_t tile_n_s = WEBGPU_MUL_MAT_TILE_N * WEBGPU_MUL_MAT_WG_SIZE_N;
+                wg_m              = (dst->ne[0] + tile_m_s - 1) / tile_m_s;
+                wg_n              = (dst->ne[1] + tile_n_s - 1) / tile_n_s;
+            }
+            wg_x = wg_m * wg_n * dst->ne[2] * dst->ne[3];
+        }
+    }
+    return ggml_backend_webgpu_build(ctx, pipeline, params, entries, wg_x, wg_y);
 }
 
 static webgpu_command ggml_webgpu_binary_op(webgpu_context &  ctx,
@@ -1336,11 +1484,11 @@ static void ggml_backend_webgpu_buffer_memset_tensor(ggml_backend_buffer_t buffe
 
     WEBGPU_CPU_PROFILE_TOTAL_START(memset_tensor);
 
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_memset_tensor(" << buffer << ", " << tensor << ", " << value << ", "
-                                                                 << offset << ", " << size << ")");
-
     ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
 
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_memset_tensor(" << buf_ctx->label << ", " << tensor << ", " << value
+                                                                 << ", " << offset << ", " << size << ")");
+
     size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
 
     // This is a trick to set all bytes of a u32 to the same 1 byte value.
@@ -1354,12 +1502,13 @@ static void ggml_backend_webgpu_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                                   const void *          data,
                                                   size_t                offset,
                                                   size_t                size) {
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", "
-                                                              << offset << ", " << size << ")");
     WEBGPU_CPU_PROFILE_TOTAL_START(set_tensor);
     ggml_backend_webgpu_buffer_context * buf_ctx    = (ggml_backend_webgpu_buffer_context *) buffer->context;
     webgpu_context                       webgpu_ctx = buf_ctx->webgpu_ctx;
 
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_set_tensor(" << buf_ctx->label << ", " << tensor << ", " << data
+                                                              << ", " << offset << ", " << size << ")");
+
     size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
 
     webgpu_ctx->queue.WriteBuffer(buf_ctx->buffer, total_offset, data, (size / 4) * 4);
@@ -1397,12 +1546,12 @@ static void ggml_backend_webgpu_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                                   void *                data,
                                                   size_t                offset,
                                                   size_t                size) {
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", "
-                                                              << offset << ", " << size << ")");
     WEBGPU_CPU_PROFILE_TOTAL_START(get_tensor);
-    ggml_backend_webgpu_buffer_context * buf_ctx    = (ggml_backend_webgpu_buffer_context *) buffer->context;
-    webgpu_context                       webgpu_ctx = buf_ctx->webgpu_ctx;
-    wgpu::Device                         device     = webgpu_ctx->device;
+    ggml_backend_webgpu_buffer_context * buf_ctx = (ggml_backend_webgpu_buffer_context *) buffer->context;
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_get_tensor(" << buf_ctx->label << ", " << tensor << ", " << data
+                                                              << ", " << offset << ", " << size << ")");
+    webgpu_context webgpu_ctx = buf_ctx->webgpu_ctx;
+    wgpu::Device   device     = webgpu_ctx->device;
 
     size_t total_offset = webgpu_tensor_offset(tensor) + tensor->view_offs + offset;
 
@@ -1473,16 +1622,20 @@ static const char * ggml_backend_webgpu_buffer_type_get_name(ggml_backend_buffer
 
 static ggml_backend_buffer_t ggml_backend_webgpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
                                                                           size_t                     size) {
-    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_type_alloc_buffer(" << size << ")");
+    static std::atomic<int> buffer_count;
+    int                     buffer_id = buffer_count++;
+    std::string             buf_name  = "tensor_buf" + std::to_string(buffer_id);
+    WEBGPU_LOG_DEBUG("ggml_backend_webgpu_buffer_type_alloc_buffer_" << buffer_id << ": " << size << " bytes");
     ggml_backend_webgpu_device_context * ctx = static_cast<ggml_backend_webgpu_device_context *>(buft->device->context);
 
     wgpu::Buffer buf;
     ggml_webgpu_create_buffer(ctx->webgpu_ctx->device, buf,
                               (size + WEBGPU_STORAGE_BUF_BINDING_MULT - 1) & ~(WEBGPU_STORAGE_BUF_BINDING_MULT - 1),
                               wgpu::BufferUsage::Storage | wgpu::BufferUsage::CopySrc | wgpu::BufferUsage::CopyDst,
-                              "allocated_buffer");
+                              buf_name.c_str());
 
-    ggml_backend_webgpu_buffer_context * buf_ctx = new ggml_backend_webgpu_buffer_context(ctx->webgpu_ctx, buf);
+    ggml_backend_webgpu_buffer_context * buf_ctx =
+        new ggml_backend_webgpu_buffer_context(ctx->webgpu_ctx, buf, buf_name);
 
     return ggml_backend_buffer_init(buft, ggml_backend_webgpu_buffer_interface, buf_ctx, size);
 }
@@ -1566,12 +1719,6 @@ static void ggml_webgpu_init_memset_pipeline(webgpu_context & webgpu_ctx) {
 }
 
 static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) {
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F32][GGML_TYPE_F32],
-                                wgsl_mul_mat_f32_f32, "mul_mat_f32_f32");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F16][GGML_TYPE_F16],
-                                wgsl_mul_mat_f16_f16, "mul_mat_f16_f16");
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_F16][GGML_TYPE_F32],
-                                wgsl_mul_mat_f16_f32, "mul_mat_f16_f32");
     ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_0][GGML_TYPE_F32],
                                 wgsl_mul_mat_q4_0_f32, "mul_mat_q4_0_f32");
     ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_Q4_1][GGML_TYPE_F32],
@@ -1610,11 +1757,143 @@ static void ggml_webgpu_init_mul_mat_pipeline(webgpu_context & webgpu_ctx) {
                                 wgsl_mul_mat_iq4_nl_f32, "mul_mat_iq4_nl_f32");
     ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->mul_mat_pipeline[GGML_TYPE_IQ4_XS][GGML_TYPE_F32],
                                 wgsl_mul_mat_iq4_xs_f32, "mul_mat_iq4_xs_f32");
+
+    if (webgpu_ctx->supports_subgroup_matrix) {
+        std::map<std::string, std::string> sg_matrix_repls;
+        sg_matrix_repls["WEBGPU_MAX_SUBGROUP_SIZE"] = std::to_string(webgpu_ctx->subgroup_size);
+        sg_matrix_repls["WEBGPU_TILE_K"]            = std::to_string(WEBGPU_MUL_MAT_TILE_K);
+        sg_matrix_repls["WEBGPU_SUBGROUP_M"]        = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_M);
+        sg_matrix_repls["WEBGPU_SUBGROUP_N"]        = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_N);
+        sg_matrix_repls["WEBGPU_SUBGROUP_MATRIX_M"] = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_MATRIX_M);
+        sg_matrix_repls["WEBGPU_SUBGROUP_MATRIX_N"] = std::to_string(WEBGPU_MUL_MAT_SUBGROUP_MATRIX_N);
+        sg_matrix_repls["WEBGPU_SG_MAT_M_SIZE"]     = std::to_string(webgpu_ctx->subgroup_matrix_config.M);
+        sg_matrix_repls["WEBGPU_SG_MAT_N_SIZE"]     = std::to_string(webgpu_ctx->subgroup_matrix_config.N);
+        sg_matrix_repls["WEBGPU_SG_MAT_K_SIZE"]     = std::to_string(webgpu_ctx->subgroup_matrix_config.K);
+
+        std::string proc_mul_mat_subgroup_matrix_f32_f32 =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f32_f32, sg_matrix_repls);
+        std::string proc_mul_mat_subgroup_matrix_f32_f32_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f32_f32_vec, sg_matrix_repls);
+        std::string proc_mul_mat_subgroup_matrix_f16_f32 =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f16_f32, sg_matrix_repls);
+        std::string proc_mul_mat_subgroup_matrix_f16_f32_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f16_f32_vec, sg_matrix_repls);
+        std::string proc_mul_mat_subgroup_matrix_f16_f16 =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f16_f16, sg_matrix_repls);
+        std::string proc_mul_mat_subgroup_matrix_f16_f16_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_f16_f16_vec, sg_matrix_repls);
+        std::string proc_mul_mat_subgroup_matrix_q4_0_f32 =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_q4_0_f32, sg_matrix_repls);
+        std::string proc_mul_mat_subgroup_matrix_q4_0_f32_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_subgroup_matrix_q4_0_f32_vec, sg_matrix_repls);
+
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline2(
+            webgpu_ctx->device, proc_mul_mat_subgroup_matrix_f32_f32.c_str(), "mul_mat_subgroup_matrix_f32_f32");
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][1] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_subgroup_matrix_f32_f32_vec.c_str(),
+                                         "mul_mat_subgroup_matrix_f32_f32_vec");
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline2(
+            webgpu_ctx->device, proc_mul_mat_subgroup_matrix_f16_f32.c_str(), "mul_mat_subgroup_matrix_f16_f32");
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][1] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_subgroup_matrix_f16_f32_vec.c_str(),
+                                         "mul_mat_subgroup_matrix_f16_f32_vec");
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][0] = ggml_webgpu_create_pipeline2(
+            webgpu_ctx->device, proc_mul_mat_subgroup_matrix_f16_f16.c_str(), "mul_mat_subgroup_matrix_f16_f16");
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][1] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_subgroup_matrix_f16_f16_vec.c_str(),
+                                         "mul_mat_subgroup_matrix_f16_f16_vec");
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline2(
+            webgpu_ctx->device, proc_mul_mat_subgroup_matrix_q4_0_f32.c_str(), "mul_mat_subgroup_matrix_q4_0_f32");
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][1] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_subgroup_matrix_q4_0_f32_vec.c_str(),
+                                         "mul_mat_subgroup_matrix_q4_0_f32_vec");
+    } else {
+        std::vector<wgpu::ConstantEntry> mul_mat_reg_tile_constants(3);
+        mul_mat_reg_tile_constants[0].key   = "TILE_K";
+        mul_mat_reg_tile_constants[0].value = WEBGPU_MUL_MAT_TILE_K;
+        mul_mat_reg_tile_constants[1].key   = "WORKGROUP_SIZE_M";
+        mul_mat_reg_tile_constants[1].value = WEBGPU_MUL_MAT_WG_SIZE_M;
+        mul_mat_reg_tile_constants[2].key   = "WORKGROUP_SIZE_N";
+        mul_mat_reg_tile_constants[2].value = WEBGPU_MUL_MAT_WG_SIZE_N;
+
+        std::map<std::string, std::string> reg_repls;
+        reg_repls["WEBGPU_TILE_M"] = std::to_string(WEBGPU_MUL_MAT_TILE_M);
+        reg_repls["WEBGPU_TILE_N"] = std::to_string(WEBGPU_MUL_MAT_TILE_N);
+
+        // Process each reg-tile shader with tile replacements.
+        // Keep the processed strings in-scope so .c_str() remains valid.
+        std::string proc_mul_mat_reg_tile_f32_f32 =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f32_f32, reg_repls);
+        std::string proc_mul_mat_reg_tile_f32_f32_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f32_f32_vec, reg_repls);
+        std::string proc_mul_mat_reg_tile_f16_f32 =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f16_f32, reg_repls);
+        std::string proc_mul_mat_reg_tile_f16_f32_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f16_f32_vec, reg_repls);
+        std::string proc_mul_mat_reg_tile_f16_f16 =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f16_f16, reg_repls);
+        std::string proc_mul_mat_reg_tile_f16_f16_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_f16_f16_vec, reg_repls);
+        std::string proc_mul_mat_reg_tile_q4_0_f32 =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_q4_0_f32, reg_repls);
+        std::string proc_mul_mat_reg_tile_q4_0_f32_vec =
+            ggml_webgpu_process_shader_repls(wgsl_mul_mat_reg_tile_q4_0_f32_vec, reg_repls);
+
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][0] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_reg_tile_f32_f32.c_str(),
+                                         "mul_mat_reg_tile_f32_f32", mul_mat_reg_tile_constants);
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][1] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_reg_tile_f32_f32_vec.c_str(),
+                                         "mul_mat_reg_tile_f32_f32_vec", mul_mat_reg_tile_constants);
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][0] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_reg_tile_f16_f32.c_str(),
+                                         "mul_mat_reg_tile_f16_f32", mul_mat_reg_tile_constants);
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][1] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_reg_tile_f16_f32_vec.c_str(),
+                                         "mul_mat_reg_tile_f16_f32_vec", mul_mat_reg_tile_constants);
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][0] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_reg_tile_f16_f16.c_str(),
+                                         "mul_mat_reg_tile_f16_f16", mul_mat_reg_tile_constants);
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][1] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_reg_tile_f16_f16_vec.c_str(),
+                                         "mul_mat_reg_tile_f16_f16_vec", mul_mat_reg_tile_constants);
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_reg_tile_q4_0_f32.c_str(),
+                                         "mul_mat_reg_tile_q4_0_f32", mul_mat_reg_tile_constants);
+        webgpu_ctx->mul_mat_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][1] =
+            ggml_webgpu_create_pipeline2(webgpu_ctx->device, proc_mul_mat_reg_tile_q4_0_f32_vec.c_str(),
+                                         "mul_mat_reg_tile_q4_0_f32_vec", mul_mat_reg_tile_constants);
+    }
+
+    std::vector<wgpu::ConstantEntry> mul_mat_vec_constants(3);
+    mul_mat_vec_constants[0].key   = "WORKGROUP_SIZE";
+    mul_mat_vec_constants[0].value = WEBGPU_MUL_MAT_VEC_WG_SIZE;
+    mul_mat_vec_constants[1].key   = "TILE_K";
+    mul_mat_vec_constants[1].value = WEBGPU_MUL_MAT_VEC_TILE_K;
+    mul_mat_vec_constants[2].key   = "OUTPUTS_PER_WG";
+    mul_mat_vec_constants[2].value = WEBGPU_MUL_MAT_VEC_OUTPUTS_PER_WG;
+
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline2(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f32_f32, "mul_mat_vec_f32_f32", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F32][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline2(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f32_f32_vec, "mul_mat_vec_f32_f32_vec", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline2(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f32, "mul_mat_vec_f16_f32", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F32][1] = ggml_webgpu_create_pipeline2(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f32_vec, "mul_mat_vec_f16_f32_vec", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][0] = ggml_webgpu_create_pipeline2(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f16, "mul_mat_vec_f16_f16", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_F16][GGML_TYPE_F16][1] = ggml_webgpu_create_pipeline2(
+        webgpu_ctx->device, wgsl_mul_mat_vec_f16_f16_vec, "mul_mat_vec_f16_f16_vec", mul_mat_vec_constants);
+    webgpu_ctx->mul_mat_vec_pipelines[GGML_TYPE_Q4_0][GGML_TYPE_F32][0] = ggml_webgpu_create_pipeline2(
+        webgpu_ctx->device, wgsl_mul_mat_vec_q4_0_f32, "mul_mat_vec_q4_0_f32", mul_mat_vec_constants);
 }
 
 static void ggml_webgpu_init_set_rows_pipeline(webgpu_context & webgpu_ctx) {
-    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->set_rows_pipeline, wgsl_set_rows, "set_rows",
-                                ggml_webgpu_wg_size_entry(webgpu_ctx->max_wg_size_x));
+    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->set_rows_pipeline[0][0], wgsl_set_rows_f16,
+                                "set_rows_f16", ggml_webgpu_wg_size_entry(webgpu_ctx->max_wg_size_x));
+    ggml_webgpu_create_pipeline(webgpu_ctx->device, webgpu_ctx->set_rows_pipeline[0][1], wgsl_set_rows_f16_vec,
+                                "set_rows_f16_vec", ggml_webgpu_wg_size_entry(webgpu_ctx->max_wg_size_x));
 }
 
 static void ggml_webgpu_init_get_rows_pipeline(webgpu_context & webgpu_ctx) {
@@ -1950,8 +2229,10 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
+            // TODO: support non-contiguous tensors, e.g. for MOE_EXPERT_REDUCE
+            // see https://github.com/ggml-org/llama.cpp/pull/16857
             supports_op = (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && (src0->type == op->type) &&
-                          (src1->type == op->type);
+                          (src1->type == op->type) && ggml_is_contiguous(src0) && ggml_is_contiguous(src1);
             break;
         case GGML_OP_CPY:
         case GGML_OP_CONT:
@@ -2103,7 +2384,13 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
 
     webgpu_context ctx = reg_ctx->webgpu_ctx;
 
-    wgpu::RequestAdapterOptions options = {};
+    // TODO: track need for these toggles: https://issues.chromium.org/issues/42251215
+    const char * const          adapterEnabledToggles[] = { "vulkan_enable_f16_on_nvidia", "use_vulkan_memory_model" };
+    wgpu::DawnTogglesDescriptor adapterTogglesDesc;
+    adapterTogglesDesc.enabledToggles     = adapterEnabledToggles;
+    adapterTogglesDesc.enabledToggleCount = 2;
+    wgpu::RequestAdapterOptions options   = {};
+    options.nextInChain                   = &adapterTogglesDesc;
     ctx->instance.WaitAny(ctx->instance.RequestAdapter(
                               &options, wgpu::CallbackMode::AllowSpontaneous,
                               [&ctx](wgpu::RequestAdapterStatus status, wgpu::Adapter adapter, const char * message) {
@@ -2119,16 +2406,63 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
     ctx->adapter.GetLimits(&ctx->limits);
     ctx->max_wg_size_x = 288;  // default value
 
-    wgpu::AdapterInfo info{};
+    wgpu::AdapterInfo                            info{};
+    wgpu::AdapterPropertiesSubgroupMatrixConfigs subgroup_matrix_configs{};
+    if (ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) {
+        info.nextInChain = &subgroup_matrix_configs;
+    }
     ctx->adapter.GetInfo(&info);
 
+    wgpu::SupportedFeatures features;
+    ctx->adapter.GetFeatures(&features);
+    // we require f16 support
+    GGML_ASSERT(ctx->adapter.HasFeature(wgpu::FeatureName::ShaderF16));
+
+    // Only support square f16 matrices of size 8 or 16 for now
+    bool valid_subgroup_matrix_config = false;
+    if (ctx->adapter.HasFeature(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix)) {
+        for (size_t i = 0; i < subgroup_matrix_configs.configCount; i++) {
+            const wgpu::SubgroupMatrixConfig config = subgroup_matrix_configs.configs[i];
+            if (config.M == config.N && config.N == config.K && (config.K == 8 || config.K == 16) &&
+                config.componentType == wgpu::SubgroupMatrixComponentType::F16 &&
+                config.resultComponentType == wgpu::SubgroupMatrixComponentType::F16) {
+                ctx->subgroup_matrix_config  = config;
+                valid_subgroup_matrix_config = true;
+                break;
+            }
+        }
+    }
+
+    // For subgroup matrix code to be the most efficient, we would like the subgroup size to be consistent and accurate.
+    // Unfortunately, that is not possible, so we use the maximum subgroup size reported by the adapter.
+    ctx->subgroup_size            = info.subgroupMaxSize;
+    ctx->supports_subgroup_matrix = valid_subgroup_matrix_config;
+
     // Initialize device
     std::vector<wgpu::FeatureName> required_features = { wgpu::FeatureName::ShaderF16,
                                                          wgpu::FeatureName::ImplicitDeviceSynchronization };
+    if (ctx->supports_subgroup_matrix) {
+        required_features.push_back(wgpu::FeatureName::Subgroups);
+        required_features.push_back(wgpu::FeatureName::ChromiumExperimentalSubgroupMatrix);
+    }
+
 #ifdef GGML_WEBGPU_GPU_PROFILE
     required_features.push_back(wgpu::FeatureName::TimestampQuery);
 #endif
 
+    // Enable Dawn-specific toggles to increase native performance
+    // TODO: Don't enable for WASM builds, they won't have an effect anyways
+    // TODO: Maybe WebGPU needs a "fast" mode where you can request compilers skip adding checks like these,
+    //       only for native performance?
+    const char * const deviceEnabledToggles[]  = { "skip_validation", "disable_robustness", "disable_workgroup_init",
+                                                   "disable_polyfills_on_integer_div_and_mod" };
+    const char * const deviceDisabledToggles[] = { "timestamp_quantization" };
+    wgpu::DawnTogglesDescriptor deviceTogglesDesc;
+    deviceTogglesDesc.enabledToggles      = deviceEnabledToggles;
+    deviceTogglesDesc.enabledToggleCount  = 4;
+    deviceTogglesDesc.disabledToggles     = deviceDisabledToggles;
+    deviceTogglesDesc.disabledToggleCount = 1;
+
     wgpu::DeviceDescriptor dev_desc;
     dev_desc.requiredLimits       = &ctx->limits;
     dev_desc.requiredFeatures     = required_features.data();
@@ -2146,6 +2480,7 @@ static ggml_backend_dev_t ggml_backend_webgpu_reg_get_device(ggml_backend_reg_t
             GGML_ABORT("ggml_webgpu: Device error! Reason: %d, Message: %s\n", static_cast<int>(reason),
                        std::string(message).c_str());
         });
+    dev_desc.nextInChain = &deviceTogglesDesc;
     ctx->instance.WaitAny(ctx->adapter.RequestDevice(
                               &dev_desc, wgpu::CallbackMode::AllowSpontaneous,
                               [ctx](wgpu::RequestDeviceStatus status, wgpu::Device device, wgpu::StringView message) {
@@ -2243,11 +2578,18 @@ ggml_backend_reg_t ggml_backend_webgpu_reg() {
     ctx.name         = GGML_WEBGPU_NAME;
     ctx.device_count = 1;
 
+    const char * const instanceEnabledToggles[] = { "allow_unsafe_apis" };
+
+    wgpu::DawnTogglesDescriptor instanceTogglesDesc;
+    instanceTogglesDesc.enabledToggles     = instanceEnabledToggles;
+    instanceTogglesDesc.enabledToggleCount = 1;
     wgpu::InstanceDescriptor               instance_descriptor{};
     std::vector<wgpu::InstanceFeatureName> instance_features = { wgpu::InstanceFeatureName::TimedWaitAny };
     instance_descriptor.requiredFeatures                     = instance_features.data();
     instance_descriptor.requiredFeatureCount                 = instance_features.size();
-    webgpu_ctx->instance                                     = wgpu::CreateInstance(&instance_descriptor);
+    instance_descriptor.nextInChain                          = &instanceTogglesDesc;
+
+    webgpu_ctx->instance = wgpu::CreateInstance(&instance_descriptor);
     GGML_ASSERT(webgpu_ctx->instance != nullptr);
 
     static ggml_backend_reg reg = {

ggml/src/ggml-webgpu/wgsl-shaders/embed_wgsl.py

@@ -72,9 +72,12 @@ def generate_variants(fname, input_dir, output_dir, outfile):
         except ValueError:
             decls_map = {}
 
-        with open(os.path.join(input_dir, "common_decls.tmpl"), "r", encoding="utf-8") as f:
-            common_decls = f.read()
-        decls_map.update(parse_decls(common_decls))
+        for fname in sorted(os.listdir(input_dir)):
+            if fname.endswith(".tmpl"):
+                tmpl_path = os.path.join(input_dir, fname)
+                with open(tmpl_path, "r", encoding="utf-8") as f_tmpl:
+                    decls = f_tmpl.read()
+                    decls_map.update(parse_decls(decls))
 
         shader_template = extract_block(text, "SHADER")
         for variant in variants:

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl

@@ -864,8 +864,8 @@ struct MulMatParams {
     broadcast3: u32
 };
 
-@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // N rows, K columns
-@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // M rows, K columns (transposed)
+@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // M rows, K columns
+@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // K rows, N columns (transposed)
 @group(0) @binding(2) var<storage, read_write> dst: array<f32>; // M rows, N columns
 
 @group(0) @binding(3) var<uniform> params: MulMatParams;
@@ -891,8 +891,8 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
 
     let dst2_rem = dst3_rem % dst2_stride;
 
-    let row = dst2_rem / params.n; // output row
-    let col = dst2_rem % params.n; // output column
+    let row = dst2_rem / params.m; // output row
+    let col = dst2_rem % params.m; // output column
 
     let src0_idx_base = params.offset_src0 + src03_idx * params.stride_03 + src02_idx * params.stride_02 + col * params.stride_01;
     let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12 + row * params.stride_11;
@@ -901,7 +901,7 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
     for (var i: u32 = 0u; i < params.k/{{BLOCK_SIZE}}; i = i + 1u) {
         sum += multiply_add(src0_idx_base, src1_idx_base, i);
     }
-    dst[params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.n + col] = sum;
+    dst[params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.m + col] = sum;
 }
 
 #end(SHADER)

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_decls.tmpl

@@ -0,0 +1,97 @@
+#decl(SHMEM_VEC)
+fn store_shmem(val: vec4<f16>, idx: u32) {
+    shmem[idx] = val.x;
+    shmem[idx + 1] = val.y;
+    shmem[idx + 2] = val.z;
+    shmem[idx + 3] = val.w;
+}
+#enddecl(SHMEM_VEC)
+
+#decl(SHMEM_SCALAR)
+fn store_shmem(val: f16, idx: u32) {
+    shmem[idx] = val;
+}
+#enddecl(SHMEM_SCALAR)
+
+#decl(INIT_SRC0_SHMEM_FLOAT)
+
+fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
+    for (var elem_idx = thread_id * {{VEC_SIZE}}; elem_idx < TILE_SRC0_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE * {{VEC_SIZE}}) {
+        let tile_m = elem_idx / TILE_K;
+        let tile_k = elem_idx % TILE_K;
+        let global_m = offset_m + tile_m;
+        let global_k = k_outer + tile_k;
+        let src0_idx = batch_offset + global_m * params.stride_01 + global_k;
+        let src0_val = select( // taking a slight performance hit to avoid oob
+            {{SRC0_TYPE}}(0.0),
+            src0[src0_idx/{{VEC_SIZE}}],
+            global_m < params.m && global_k < params.k);
+        store_shmem({{SHMEM_TYPE}}(src0_val), elem_idx);
+    }
+}
+
+#enddecl(INIT_SRC0_SHMEM_FLOAT)
+
+#decl(INIT_SRC1_SHMEM)
+
+fn init_shmem_src1(thread_id: u32, batch_offset: u32, offset_n: u32, k_outer: u32) {
+    for (var elem_idx = thread_id * {{VEC_SIZE}}; elem_idx < TILE_SRC1_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE * {{VEC_SIZE}}) {
+        let tile_n = elem_idx / TILE_K;
+        let tile_k = elem_idx % TILE_K;
+        let global_n = offset_n + tile_n;
+        let global_k = k_outer + tile_k;
+        let src1_idx = batch_offset + global_n * params.stride_11 + global_k;
+        let src1_val = select(
+            {{SRC1_TYPE}}(0.0),
+            src1[src1_idx/{{VEC_SIZE}}],
+            global_n < params.n && global_k < params.k);
+        store_shmem({{SHMEM_TYPE}}(src1_val), TILE_SRC0_SHMEM + elem_idx);
+    }
+}
+
+#enddecl(INIT_SRC1_SHMEM)
+
+#decl(INIT_SRC0_SHMEM_Q4_0)
+
+const BLOCK_SIZE = 32u;
+// the number of blocks per k-tile. Note that this currently only works if TILE_K is a multiple of BLOCK_SIZE, which may need to be rethought for larger quantized types.
+override BLOCKS_K = TILE_K/BLOCK_SIZE;
+const NQ = 16u;
+const F16_PER_BLOCK = 9u; // 1 scale + 8x4 packed weights
+const WEIGHTS_PER_F16 = 4u; // 4 weights per f16
+const F16_PER_THREAD = NQ / WEIGHTS_PER_F16;
+
+fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
+    for (var i = thread_id * NQ; i < TILE_SRC0_SHMEM; i += TOTAL_WORKGROUP_SIZE * NQ) {
+        let blck_idx = i / BLOCK_SIZE;
+        let block_offset = (i % BLOCK_SIZE) / WEIGHTS_PER_F16;
+        let shmem_idx = blck_idx * BLOCK_SIZE + block_offset * 2u;
+
+        let tile_m = blck_idx / BLOCKS_K;
+        let global_m = offset_m + tile_m;
+        let block_k = blck_idx % BLOCKS_K;
+        let global_k = k_outer / BLOCK_SIZE + block_k;
+
+        if (global_m < params.m && global_k < params.k / BLOCK_SIZE) {
+            let src0_idx = batch_offset + global_m * params.stride_01 + global_k;
+            let scale_idx = src0_idx * F16_PER_BLOCK;
+            let d = src0[scale_idx];
+
+            for (var j = 0u; j < F16_PER_THREAD; j += 2) {
+                let q_0 = src0[scale_idx + 1u + block_offset + j];
+                let q_1 = src0[scale_idx + 1u + block_offset + j + 1];
+
+                let q_packed = bitcast<u32>(vec2(q_0, q_1));
+                for (var k = 0u; k < 4u; k++) {
+                    let q_byte = get_byte(q_packed, k);
+                    let q_hi = (f16((q_byte >> 4) & 0xF) - 8.0) * d;
+                    let q_lo = (f16(q_byte & 0xF) - 8.0) * d;
+                    shmem[shmem_idx + j * 2 + k] = q_lo;
+                    shmem[shmem_idx + j * 2 + k + 16u] = q_hi;
+                }
+            }
+        }
+    }
+}
+
+#enddecl(INIT_SRC0_SHMEM_Q4_0)

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_reg_tile.tmpl.wgsl

@@ -0,0 +1,247 @@
+#define(VARIANTS)
+[
+  {
+    "SHADER_SUFFIX": "f32_f32_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "vec4<f32>",
+      "SRC1_TYPE" : "vec4<f32>",
+      "DST_TYPE" : "vec4<f32>",
+      "SHMEM_TYPE" : "vec4<f16>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "f32_f32",
+    "REPLS": {
+      "SRC0_TYPE" : "f32",
+      "SRC1_TYPE" : "f32",
+      "DST_TYPE" : "f32",
+      "SHMEM_TYPE" : "f16",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f32_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "vec4<f16>",
+      "SRC1_TYPE" : "vec4<f32>",
+      "DST_TYPE" : "vec4<f32>",
+      "SHMEM_TYPE" : "vec4<f16>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f32",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f32",
+      "DST_TYPE" : "f32",
+      "SHMEM_TYPE" : "f16",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f16_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "vec4<f16>",
+      "SRC1_TYPE" : "vec4<f16>",
+      "DST_TYPE" : "vec4<f32>",
+      "SHMEM_TYPE" : "vec4<f16>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f16",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f16",
+      "DST_TYPE" : "f32",
+      "SHMEM_TYPE" : "f16",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "q4_0_f32_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "vec4<f32>",
+      "DST_TYPE" : "vec4<f32>",
+      "SHMEM_TYPE" : "vec4<f16>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["BYTE_HELPERS", "VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_Q4_0", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "q4_0_f32",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f32",
+      "DST_TYPE" : "f32",
+      "SHMEM_TYPE" : "f16",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["BYTE_HELPERS", "SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_Q4_0", "INIT_SRC1_SHMEM"]
+  }
+]
+
+#end(VARIANTS)
+
+#define(DECLS)
+
+#decl(VEC)
+fn store_val(acc: array<array<f16, TILE_N>, TILE_M>, tn: u32, tm: u32) -> vec4<f32> {
+    return vec4<f32>(f32(acc[tm][tn]), f32(acc[tm + 1][tn]), f32(acc[tm + 2][tn]), f32(acc[tm + 3][tn]));
+}
+#enddecl(VEC)
+
+#decl(SCALAR)
+fn store_val(acc: array<array<f16, TILE_N>, TILE_M>, tn: u32, tm: u32) -> f32 {
+    return f32(acc[tm][tn]);
+}
+#enddecl(SCALAR)
+
+#end(DECLS)
+
+#define(SHADER)
+enable f16;
+
+struct MulMatParams {
+    offset_src0: u32,
+    offset_src1: u32,
+    offset_dst: u32,
+    m: u32,
+    n: u32,
+    k: u32,
+    stride_01: u32,
+    stride_11: u32,
+    stride_02: u32,
+    stride_12: u32,
+    stride_03: u32,
+    stride_13: u32,
+    bs02: u32,
+    bs03: u32,
+    broadcast2: u32,
+    broadcast3: u32
+};
+
+@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // M rows, K columns
+@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // K rows, N columns (transposed)
+@group(0) @binding(2) var<storage, read_write> dst: array<{{DST_TYPE}}>; // M rows, N columns (transposed)
+
+@group(0) @binding(3) var<uniform> params: MulMatParams;
+
+DECLS
+
+fn get_local_n(thread_id: u32) -> u32 {
+    return thread_id / WORKGROUP_SIZE_M;
+}
+fn get_local_m(thread_id: u32) -> u32 {
+    return thread_id % WORKGROUP_SIZE_M;
+}
+
+// TILE_M must be multiple of 4 for vec4 loads
+const TILE_M = {{WEBGPU_TILE_M}}u;
+const TILE_N = {{WEBGPU_TILE_N}}u;
+
+override WORKGROUP_SIZE_M: u32;
+override WORKGROUP_SIZE_N: u32;
+override TILE_K: u32;
+
+override TOTAL_WORKGROUP_SIZE = WORKGROUP_SIZE_M * WORKGROUP_SIZE_N;
+override TILE_SRC0_SHMEM = TILE_K * WORKGROUP_SIZE_M * TILE_M;
+override TILE_SRC1_SHMEM = TILE_K * WORKGROUP_SIZE_N * TILE_N;
+
+var<workgroup> shmem: array<f16, TILE_SRC0_SHMEM + TILE_SRC1_SHMEM>;
+
+@compute @workgroup_size(TOTAL_WORKGROUP_SIZE)
+fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
+        @builtin(local_invocation_id) local_id: vec3<u32>) {
+
+    let thread_id = local_id.x;
+    let local_m = get_local_m(thread_id);
+    let local_n = get_local_n(thread_id);
+
+    let wg_n_count = (params.n + WORKGROUP_SIZE_N * TILE_N - 1u) / (WORKGROUP_SIZE_N * TILE_N);
+    let wg_m_count = (params.m + WORKGROUP_SIZE_M * TILE_M - 1u) / (WORKGROUP_SIZE_M * TILE_M);
+    let wg_per_matrix = wg_m_count * wg_n_count;
+
+    let batch_idx = wg_id.x / wg_per_matrix;
+
+    let wg_in_batch = wg_id.x % wg_per_matrix;
+    let wg_m = wg_in_batch % wg_m_count;
+    let wg_n = wg_in_batch / wg_m_count;
+
+    let output_row_base = wg_m * WORKGROUP_SIZE_M * TILE_M + local_m * TILE_M;
+    let output_col_base = wg_n * WORKGROUP_SIZE_N * TILE_N + local_n * TILE_N;
+
+    let dst2_stride = params.m * params.n;
+    let dst3_stride = dst2_stride * params.bs02 * params.broadcast2;
+
+    let dst3_idx = batch_idx / (params.bs02 * params.broadcast2);
+    let src03_idx = dst3_idx / params.broadcast3;
+    let src13_idx = dst3_idx;
+    let dst2_idx = batch_idx % (params.bs02 * params.broadcast2);
+    let src02_idx = dst2_idx / params.broadcast2;
+    let src12_idx = dst2_idx;
+
+    let src0_batch_offset = params.offset_src0 + src03_idx * params.stride_03 + src02_idx * params.stride_02;
+    let src1_batch_offset = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12;
+
+    let offset_m = wg_m * WORKGROUP_SIZE_M * TILE_M;
+    let offset_n = wg_n * WORKGROUP_SIZE_N * TILE_N;
+
+    var acc: array<array<f16, TILE_N>, TILE_M>;
+
+    for (var k_outer = 0u; k_outer < params.k; k_outer += TILE_K) {
+
+        // see mul_mat_decls.tmpl
+        init_shmem_src0(thread_id, src0_batch_offset, offset_m, k_outer);
+        init_shmem_src1(thread_id, src1_batch_offset, offset_n, k_outer);
+
+        workgroupBarrier();
+
+        let k_end = min(TILE_K, params.k - k_outer);
+
+        for (var k_inner = 0u; k_inner < k_end; k_inner++) {
+            var src0_tile: array<f16, TILE_M>;
+            for (var tm = 0u; tm < TILE_M; tm++) {
+                let src0_m = local_m * TILE_M + tm;
+                let src0_idx = k_inner + src0_m * TILE_K;
+                src0_tile[tm] = shmem[src0_idx];
+            }
+            for (var tn = 0u; tn < TILE_N; tn++) {
+                let src1_n = local_n * TILE_N + tn;
+                let src1_idx = src1_n * TILE_K + k_inner;
+                let src1_val = shmem[TILE_SRC0_SHMEM + src1_idx];
+                for (var tm = 0u; tm < TILE_M; tm++) {
+                      acc[tm][tn] += src0_tile[tm] * src1_val;
+                }
+            }
+        }
+
+        workgroupBarrier();
+    }
+
+    let dst_batch_offset = params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride;
+
+    for (var tn = 0u; tn < TILE_N; tn++) {
+        let global_col = output_col_base + tn;
+        if (global_col < params.n) {
+            for (var tm = 0u; tm < TILE_M; tm += {{VEC_SIZE}}) {
+                let global_row = output_row_base + tm;
+                if (global_row < params.m) {
+                    let dst_idx = dst_batch_offset + global_col * params.m + global_row;
+                    dst[dst_idx/{{VEC_SIZE}}] = store_val(acc, tn, tm);
+                }
+            }
+        }
+    }
+}
+
+#end(SHADER)

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_subgroup_matrix.tmpl.wgsl

@@ -0,0 +1,302 @@
+#define(VARIANTS)
+[
+  {
+    "SHADER_SUFFIX": "f32_f32_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "vec4<f32>",
+      "SRC1_TYPE" : "vec4<f32>",
+      "DST_TYPE" : "vec4<f32>",
+      "SHMEM_TYPE" : "vec4<f16>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "f32_f32",
+    "REPLS": {
+      "SRC0_TYPE" : "f32",
+      "SRC1_TYPE" : "f32",
+      "DST_TYPE" : "f32",
+      "SHMEM_TYPE" : "f16",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f32_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "vec4<f16>",
+      "SRC1_TYPE" : "vec4<f32>",
+      "DST_TYPE" : "vec4<f32>",
+      "SHMEM_TYPE" : "vec4<f16>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f32",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f32",
+      "DST_TYPE" : "f32",
+      "SHMEM_TYPE" : "f16",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f16_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "vec4<f16>",
+      "SRC1_TYPE" : "vec4<f16>",
+      "DST_TYPE" : "vec4<f32>",
+      "SHMEM_TYPE" : "vec4<f16>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f16",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f16",
+      "DST_TYPE" : "f32",
+      "SHMEM_TYPE" : "f16",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "q4_0_f32_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "vec4<f32>",
+      "DST_TYPE" : "vec4<f32>",
+      "SHMEM_TYPE" : "vec4<f16>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["BYTE_HELPERS", "VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_Q4_0", "INIT_SRC1_SHMEM"]
+  },
+  {
+    "SHADER_SUFFIX": "q4_0_f32",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f32",
+      "DST_TYPE" : "f32",
+      "SHMEM_TYPE" : "f16",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["BYTE_HELPERS", "SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_Q4_0", "INIT_SRC1_SHMEM"]
+  }
+]
+
+#end(VARIANTS)
+
+#define(DECLS)
+
+#decl(VEC)
+fn store_dst(shmem_idx: u32, dst_idx: u32) {
+    dst[dst_idx] = vec4<f32>(
+        f32(shmem[shmem_idx]),
+        f32(shmem[shmem_idx + 1]),
+        f32(shmem[shmem_idx + 2]),
+        f32(shmem[shmem_idx + 3])
+    );
+}
+#enddecl(VEC)
+
+#decl(SCALAR)
+fn store_dst(shmem_idx: u32, dst_idx: u32) {
+    dst[dst_idx] = f32(shmem[shmem_idx]);
+}
+#enddecl(SCALAR)
+
+#end(DECLS)
+
+#define(SHADER)
+diagnostic(off, chromium.subgroup_matrix_uniformity);
+enable f16;
+enable subgroups;
+enable chromium_experimental_subgroup_matrix;
+
+struct MulMatParams {
+    offset_src0: u32,
+    offset_src1: u32,
+    offset_dst: u32,
+    m: u32,
+    n: u32,
+    k: u32,
+    stride_01: u32,
+    stride_11: u32,
+    stride_02: u32,
+    stride_12: u32,
+    stride_03: u32,
+    stride_13: u32,
+    bs02: u32,
+    bs03: u32,
+    broadcast2: u32,
+    broadcast3: u32
+};
+
+@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // M rows, K columns
+@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // K rows, N columns (transposed)
+@group(0) @binding(2) var<storage, read_write> dst: array<{{DST_TYPE}}>; // M rows, N columns (transposed)
+
+@group(0) @binding(3) var<uniform> params: MulMatParams;
+
+DECLS
+
+// Note: These are string interpolated at build time, cannot use override constants due to limitations in
+// current Dawn version type definitions/matrix load requirements for constant memory sizes.
+const SUBGROUP_M = {{WEBGPU_SUBGROUP_M}}u;
+const SUBGROUP_N = {{WEBGPU_SUBGROUP_N}}u;
+// For portability we assume the max subgroup size, meaning some subgroups will be masked out if the
+// runtime subgroup size is smaller.
+const MAX_SUBGROUP_SIZE = {{WEBGPU_MAX_SUBGROUP_SIZE}}u;
+
+const EXPECTED_SUBGROUPS = SUBGROUP_M * SUBGROUP_N;
+
+const SUBGROUP_MATRIX_M_SIZE = {{WEBGPU_SG_MAT_M_SIZE}}u;
+const SUBGROUP_MATRIX_N_SIZE = {{WEBGPU_SG_MAT_N_SIZE}}u;
+const SUBGROUP_MATRIX_K_SIZE = {{WEBGPU_SG_MAT_K_SIZE}}u;
+
+const SUBGROUP_MATRIX_M = {{WEBGPU_SUBGROUP_MATRIX_M}}u;
+const SUBGROUP_MATRIX_N = {{WEBGPU_SUBGROUP_MATRIX_N}}u;
+
+const TILE_K = {{WEBGPU_TILE_K}}u;
+
+const WG_M_SG_TILE_SIZE = SUBGROUP_M * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE;
+const WG_N_SG_TILE_SIZE = SUBGROUP_N * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE;
+
+const TOTAL_WORKGROUP_SIZE = SUBGROUP_M * SUBGROUP_N * MAX_SUBGROUP_SIZE;
+const TILE_SRC0_SHMEM = TILE_K * SUBGROUP_M * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE;
+const TILE_SRC1_SHMEM = TILE_K * SUBGROUP_N * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE;
+
+const SG_MAT_ACCUM_SHMEM = SUBGROUP_M * SUBGROUP_MATRIX_M * SUBGROUP_N * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_M_SIZE * SUBGROUP_MATRIX_N_SIZE;
+
+// We reuse shmem for accumulation matrices
+const SHMEM_SIZE = max(TILE_SRC0_SHMEM + TILE_SRC1_SHMEM, SG_MAT_ACCUM_SHMEM);
+
+var<workgroup> shmem: array<f16, SHMEM_SIZE>;
+
+@compute @workgroup_size(TOTAL_WORKGROUP_SIZE)
+fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
+        @builtin(local_invocation_id) local_id: vec3<u32>,
+        @builtin(subgroup_id) subgroup_id: u32) {
+
+    let thread_id = local_id.x;
+    let subgroup_m = subgroup_id % SUBGROUP_M;
+    let subgroup_n = subgroup_id / SUBGROUP_M;
+
+    let wg_m_count = (params.m + WG_M_SG_TILE_SIZE - 1) / WG_M_SG_TILE_SIZE;
+    let wg_n_count = (params.n + WG_N_SG_TILE_SIZE - 1) / WG_N_SG_TILE_SIZE;
+    let wg_per_matrix = wg_m_count * wg_n_count;
+
+    let batch_idx = wg_id.x / wg_per_matrix;
+
+    let wg_in_batch = wg_id.x % wg_per_matrix;
+    let wg_m = wg_in_batch % wg_m_count;
+    let wg_n = wg_in_batch / wg_m_count;
+
+    let dst2_stride = params.m * params.n;
+    let dst3_stride = dst2_stride * params.bs02 * params.broadcast2;
+
+    let dst3_idx = batch_idx / (params.bs02 * params.broadcast2);
+    let src03_idx = dst3_idx / params.broadcast3;
+    let src13_idx = dst3_idx;
+    let dst2_idx = batch_idx % (params.bs02 * params.broadcast2);
+    let src02_idx = dst2_idx / params.broadcast2;
+    let src12_idx = dst2_idx;
+
+    let src0_batch_offset = params.offset_src0 + src03_idx * params.stride_03 + src02_idx * params.stride_02;
+    let src1_batch_offset = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12;
+
+    let offset_m = wg_m * SUBGROUP_M * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE;
+    let offset_n = wg_n * SUBGROUP_N * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE;
+
+    var acc_sg_mat : array<array<subgroup_matrix_result<f16, SUBGROUP_MATRIX_N_SIZE, SUBGROUP_MATRIX_M_SIZE>, SUBGROUP_MATRIX_N>, SUBGROUP_MATRIX_M>;
+
+    for (var k_outer = 0u; k_outer < params.k; k_outer += TILE_K) {
+
+        // see mul_mat_decls.tmpl
+        init_shmem_src0(thread_id, src0_batch_offset, offset_m, k_outer);
+        init_shmem_src1(thread_id, src1_batch_offset, offset_n, k_outer);
+
+        workgroupBarrier();
+
+        if (subgroup_id < EXPECTED_SUBGROUPS) {
+
+            for (var k_inner = 0u; k_inner < TILE_K; k_inner += SUBGROUP_MATRIX_K_SIZE) {
+
+                let src0_shmem_idx_base = subgroup_m * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE * TILE_K + k_inner;
+                var src0_sg_mats: array<subgroup_matrix_left<f16, SUBGROUP_MATRIX_K_SIZE, SUBGROUP_MATRIX_M_SIZE>, SUBGROUP_MATRIX_M>;
+                for (var m = 0u; m < SUBGROUP_MATRIX_M; m++) {
+                    src0_sg_mats[m] = subgroupMatrixLoad<subgroup_matrix_left<f16, SUBGROUP_MATRIX_K_SIZE, SUBGROUP_MATRIX_M_SIZE>>(
+                        &shmem,
+                        src0_shmem_idx_base + m * SUBGROUP_MATRIX_M_SIZE * TILE_K,
+                        false,
+                        TILE_K
+                    );
+                }
+
+                let src1_shmem_idx_base = TILE_SRC0_SHMEM + subgroup_n * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE * TILE_K + k_inner;
+                for (var n = 0u; n < SUBGROUP_MATRIX_N; n++) {
+                    let src1_sg_mat = subgroupMatrixLoad<subgroup_matrix_right<f16, SUBGROUP_MATRIX_N_SIZE, SUBGROUP_MATRIX_K_SIZE>>(
+                        &shmem,
+                        src1_shmem_idx_base + n * SUBGROUP_MATRIX_N_SIZE * TILE_K,
+                        true,
+                        TILE_K
+                    );
+                    for (var m = 0u; m < SUBGROUP_MATRIX_M; m++) {
+                        acc_sg_mat[m][n] = subgroupMatrixMultiplyAccumulate(src0_sg_mats[m], src1_sg_mat, acc_sg_mat[m][n]);
+                    }
+                }
+            }
+        }
+
+        workgroupBarrier();
+    }
+
+    let dst_batch_offset = params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride;
+
+    // Stage the subgroup matrix tiles into shared memory
+    // This uses WG_M_SG_TILE_SIZE as the stride (number of columns in the workgroup tile).
+    let WG_TILE_STRIDE = WG_M_SG_TILE_SIZE;
+    let tile_row_base_local = subgroup_n * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE;
+    let tile_col_base_local = subgroup_m * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE;
+
+    if (subgroup_id < EXPECTED_SUBGROUPS) { // 2-5% performance hit :(
+        for (var n = 0u; n < SUBGROUP_MATRIX_N; n++) {
+            for (var m = 0u; m < SUBGROUP_MATRIX_M; m++) {
+                let local_row = tile_row_base_local + n * SUBGROUP_MATRIX_N_SIZE;
+                let local_col = tile_col_base_local + m * SUBGROUP_MATRIX_M_SIZE;
+                let out_base = local_row * WG_TILE_STRIDE + local_col;
+                subgroupMatrixStore(&shmem, out_base, acc_sg_mat[m][n], true, WG_TILE_STRIDE);
+            }
+        }
+    }
+
+    workgroupBarrier();
+
+    // Cooperative write: iterate over the entire workgroup tile
+    let tile_rows = WG_N_SG_TILE_SIZE;
+    let tile_cols = WG_M_SG_TILE_SIZE;
+    let total_tile_elems = tile_rows * tile_cols;
+    let tile_dst_row_base = wg_m * SUBGROUP_M * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE;
+    let tile_dst_col_base = wg_n * SUBGROUP_N * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE;
+
+    for (var idx = thread_id * {{VEC_SIZE}}; idx < total_tile_elems; idx += TOTAL_WORKGROUP_SIZE * {{VEC_SIZE}}) {
+        let local_row = idx % WG_TILE_STRIDE;
+        let local_col = idx / WG_TILE_STRIDE;
+
+        let global_row = tile_dst_row_base + local_row;
+        let global_col = tile_dst_col_base + local_col;
+
+        if (global_col < params.n && global_row < params.m) {
+            let dst_idx = dst_batch_offset + global_col * params.m + global_row;
+            store_dst(idx, dst_idx/{{VEC_SIZE}});
+        }
+    }
+}
+
+#end(SHADER)

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat_vec.tmpl.wgsl

@@ -0,0 +1,267 @@
+#define(VARIANTS)
+[
+  {
+    "SHADER_SUFFIX": "f32_f32_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "vec4<f32>",
+      "SRC1_TYPE" : "vec4<f32>",
+      "DST_TYPE": "vec4<f32>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["VEC", "MUL_ACC_FLOAT"]
+  },
+  {
+    "SHADER_SUFFIX": "f32_f32",
+    "REPLS": {
+      "SRC0_TYPE" : "f32",
+      "SRC1_TYPE" : "f32",
+      "DST_TYPE": "f32",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["SCALAR", "MUL_ACC_FLOAT"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f32_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "vec4<f16>",
+      "SRC1_TYPE" : "vec4<f32>",
+      "DST_TYPE": "vec4<f32>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["VEC", "MUL_ACC_FLOAT"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f32",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f32",
+      "DST_TYPE": "f32",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["SCALAR", "MUL_ACC_FLOAT"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f16_vec",
+    "REPLS": {
+      "SRC0_TYPE" : "vec4<f16>",
+      "SRC1_TYPE" : "vec4<f16>",
+      "DST_TYPE": "vec4<f32>",
+      "VEC_SIZE" : 4,
+    },
+    "DECLS": ["VEC", "MUL_ACC_FLOAT"]
+  },
+  {
+    "SHADER_SUFFIX": "f16_f16",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f16",
+      "DST_TYPE": "f32",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["SCALAR", "MUL_ACC_FLOAT"]
+  },
+  {
+    "SHADER_SUFFIX": "q4_0_f32",
+    "REPLS": {
+      "SRC0_TYPE" : "f16",
+      "SRC1_TYPE" : "f32",
+      "DST_TYPE": "f32",
+      "VEC_SIZE" : 1,
+    },
+    "DECLS": ["BYTE_HELPERS", "SCALAR", "MUL_ACC_Q4_0"]
+  }
+]
+
+#end(VARIANTS)
+
+#define(DECLS)
+
+#decl(VEC)
+fn inner_dot(src0_val: {{SRC0_TYPE}}, src1_val: {{SRC1_TYPE}}) -> f32 {
+    return f32(dot({{SRC1_TYPE}}(src0_val), src1_val));
+}
+
+fn store_val(group_base: u32) -> vec4<f32> {
+    return vec4<f32>(partial_sums[group_base],
+                     partial_sums[group_base + THREADS_PER_OUTPUT],
+                     partial_sums[group_base + THREADS_PER_OUTPUT * 2],
+                     partial_sums[group_base + THREADS_PER_OUTPUT * 3]);
+}
+#enddecl(VEC)
+
+#decl(SCALAR)
+fn inner_dot(src0_val: {{SRC0_TYPE}}, src1_val: {{SRC1_TYPE}}) -> f32 {
+    return f32(src0_val) * f32(src1_val);
+}
+
+fn store_val(group_base: u32) -> f32 {
+    return partial_sums[group_base];
+}
+#enddecl(SCALAR)
+
+#decl(MUL_ACC_FLOAT)
+
+fn mul_acc(tig:u32, tile_size: u32, idx_base: u32, k_outer: u32) -> f32 {
+    var local_sum = 0.0;
+    for (var i = tig * {{VEC_SIZE}}; i < tile_size; i += THREADS_PER_OUTPUT * {{VEC_SIZE}}) {
+        let a = src0[(idx_base + k_outer + i) / {{VEC_SIZE}}];
+        let b = shared_vector[i / {{VEC_SIZE}}];
+        local_sum += inner_dot(a, b);
+    }
+    return local_sum;
+}
+
+#enddecl(MUL_ACC_FLOAT)
+
+#decl(MUL_ACC_Q4_0)
+
+const BLOCK_SIZE = 32;
+const NQ = 16u; // number of weights per thread
+const F16_PER_BLOCK = 9u; // 1 scale + 8x4 packed weights
+const WEIGHTS_PER_F16 = 4u; // 4 weights per f16
+const F16_PER_THREAD = NQ / WEIGHTS_PER_F16;
+
+fn mul_acc(tig:u32, tile_size: u32, idx_base: u32, k_outer: u32) -> f32 {
+    var local_sum = 0.0;
+    for (var i = tig * NQ; i < tile_size; i += THREADS_PER_OUTPUT * NQ) {
+        let blck_idx = i / BLOCK_SIZE;
+        let block_offset = (i % BLOCK_SIZE) / WEIGHTS_PER_F16;
+        let scale_idx = (idx_base + k_outer / BLOCK_SIZE + blck_idx) * F16_PER_BLOCK;
+        // each f16 contains offsets [block_offset, block_offset + 1] and [block_offset + 16, block_offset + 17]
+        let shmem_idx = blck_idx * BLOCK_SIZE + block_offset * 2u;
+        let d = f32(src0[scale_idx]);
+        for (var j = 0u; j < F16_PER_THREAD; j += 2) {
+            let q_0 = src0[scale_idx + 1 + block_offset + j];
+            let q_1 = src0[scale_idx + 1 + block_offset + j + 1];
+            let q_packed = bitcast<u32>(vec2(q_0, q_1));
+            for (var k: u32 = 0; k < 4; k++) {
+                let q_byte = get_byte(q_packed, k);
+                let q_hi = (f32((q_byte >> 4) & 0xF) - 8.0) * d;
+                let q_lo = (f32(q_byte & 0xF) - 8.0) * d;
+                local_sum += q_lo * shared_vector[shmem_idx + j * 2 + k];
+                local_sum += q_hi * shared_vector[shmem_idx + j * 2 + k + 16];
+            }
+        }
+    }
+    return local_sum;
+}
+
+#enddecl(MUL_ACC_Q4_0)
+
+#end(DECLS)
+
+#define(SHADER)
+enable f16;
+
+DECLS
+
+struct MulMatParams {
+    offset_src0: u32,
+    offset_src1: u32,
+    offset_dst: u32,
+    m: u32,
+    n: u32,
+    k: u32,
+    stride_01: u32,
+    stride_11: u32,
+    stride_02: u32,
+    stride_12: u32,
+    stride_03: u32,
+    stride_13: u32,
+    bs02: u32,
+    bs03: u32,
+    broadcast2: u32,
+    broadcast3: u32
+};
+
+@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // Matrix (M x K)
+@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // Vector (K x 1, transposed)
+@group(0) @binding(2) var<storage, read_write> dst: array<{{DST_TYPE}}>;  // Result vector (transposed)
+
+@group(0) @binding(3) var<uniform> params: MulMatParams;
+
+override WORKGROUP_SIZE: u32;
+override TILE_K: u32;
+override OUTPUTS_PER_WG: u32;
+override THREADS_PER_OUTPUT = WORKGROUP_SIZE / OUTPUTS_PER_WG;
+
+// Shared memory for collaborative loading and reduction
+var<workgroup> shared_vector: array<{{SRC1_TYPE}}, TILE_K/{{VEC_SIZE}}>;  // Cache vector tile
+var<workgroup> partial_sums: array<f32, WORKGROUP_SIZE>;   // For reduction
+
+@compute @workgroup_size(WORKGROUP_SIZE)
+fn main(
+    @builtin(local_invocation_id) local_id: vec3<u32>,
+    @builtin(workgroup_id) wg_id: vec3<u32>,
+    @builtin(num_workgroups) num_wg: vec3<u32>) {
+    let thread_id = local_id.x;
+
+    // Handle batch dimensions
+    let total_batches = params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
+    let wg_linear = wg_id.y * num_wg.x + wg_id.x;
+    let output_groups = (params.m + OUTPUTS_PER_WG - 1u) / OUTPUTS_PER_WG;
+    let batch_idx = wg_linear / output_groups;
+    if (batch_idx >= total_batches) {
+        return;
+    }
+
+    // Which of the outputs does this thread belong to?
+    let thread_group = thread_id / THREADS_PER_OUTPUT;
+    let thread_in_group = thread_id % THREADS_PER_OUTPUT;
+
+    // Each workgroup computes OUTPUTS_PER_WG consecutive outputs
+    let output_row = (wg_linear % output_groups) * OUTPUTS_PER_WG + thread_group;
+
+    let dst2_stride = params.m * params.n;
+    let dst2_idx = batch_idx % (params.bs02 * params.broadcast2);
+    let dst3_stride = dst2_stride * params.bs02 * params.broadcast2;
+    let dst3_idx = batch_idx / (params.bs02 * params.broadcast2);
+    let src03_idx = dst3_idx / params.broadcast3;
+    let src13_idx = dst3_idx;
+    let src02_idx = dst2_idx / params.broadcast2;
+    let src12_idx = dst2_idx;
+
+    let src0_idx_base = params.offset_src0 + src03_idx * params.stride_03 + src02_idx * params.stride_02 + output_row * params.stride_01;
+    let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12;
+    let dst_idx = params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride + output_row;
+
+    var local_sum = 0.0;
+
+    // Each thread processes multiple K elements and accumulates
+    for (var k_tile = 0u; k_tile < params.k; k_tile += TILE_K) {
+        let tile_size = min(TILE_K, params.k - k_tile);
+
+        // Cooperatively load vector tile into shared memory (all threads)
+        for (var i = thread_id * {{VEC_SIZE}}; i < tile_size; i += WORKGROUP_SIZE * {{VEC_SIZE}}) {
+            shared_vector[i / {{VEC_SIZE}}] = src1[(src1_idx_base + k_tile + i) / {{VEC_SIZE}}];
+        }
+
+        workgroupBarrier();
+
+        if (output_row < params.m) {
+            local_sum += mul_acc(thread_in_group, tile_size, src0_idx_base, k_tile);
+        }
+
+        workgroupBarrier();
+    }
+
+    // Store partial sums and reduce within each partition
+    partial_sums[thread_id] = local_sum;
+    workgroupBarrier();
+    let group_base = thread_group * THREADS_PER_OUTPUT;
+    let thread_base = group_base + thread_in_group;
+    var offset = THREADS_PER_OUTPUT / 2;
+    while (offset > 0) {
+        if (thread_in_group < offset) {
+            partial_sums[thread_base] += partial_sums[thread_base + offset];
+        }
+        offset = offset / 2;
+        workgroupBarrier();
+    }
+
+    // Store back to global memory
+    if (output_row < params.m && thread_group % {{VEC_SIZE}} == 0 && thread_in_group == 0) {
+        dst[dst_idx / {{VEC_SIZE}}] = store_val(group_base);
+    }
+}
+#end(SHADER)

ggml/src/ggml-webgpu/wgsl-shaders/set_rows.tmpl.wgsl

@@ -1,13 +1,38 @@
+#define(VARIANTS)
+
+[
+  {
+    "SHADER_SUFFIX": "f16_vec",
+    "REPLS": {
+      "TYPE" : "vec4<f32>",
+      "DST_TYPE": "vec4<f16>",
+      "VEC_SIZE": 4
+    }
+  },
+  {
+    "SHADER_SUFFIX": "f16",
+    "REPLS": {
+      "TYPE" : "f32",
+      "DST_TYPE": "f16",
+      "VEC_SIZE": 1
+    }
+  }
+]
+
+#end(VARIANTS)
+
+#define(SHADER)
+
 enable f16;
 
 @group(0) @binding(0)
-var<storage, read_write> src: array<f32>;
+var<storage, read_write> src: array<{{TYPE}}>;
 
 @group(0) @binding(1)
 var<storage, read_write> idx: array<u32>;
 
 @group(0) @binding(2)
-var<storage, read_write> dst: array<f16>;
+var<storage, read_write> dst: array<{{DST_TYPE}}>;
 
 @group(0) @binding(3)
 var<storage, read_write> error: atomic<u32>;
@@ -47,10 +72,14 @@ var<uniform> params: Params;
 override wg_size: u32;
 @compute @workgroup_size(wg_size)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    if (gid.x >= params.n_rows * params.ne2 * params.ne3) {
+    if (gid.x >= (params.ne3 * params.ne2 * params.n_rows * params.ne0) / {{VEC_SIZE}}) {
         return;
     }
-    var i = gid.x;
+
+    // getting the row from gid
+    let elems_per_row = params.ne0 / {{VEC_SIZE}};
+    var i = gid.x / elems_per_row;
+
     let i_src3 = i / (params.ne2 * params.n_rows);
 
     i = i % (params.ne2 * params.n_rows);
@@ -75,7 +104,9 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let i_dst_row = params.offset_dst + idx_high_val * params.stride_dst1 + i_src2 * params.stride_dst2 + i_src3 * params.stride_dst3;
     let i_src_row = params.offset_src + i_src1 * params.stride_src1 + i_src2 * params.stride_src2 + i_src3 * params.stride_src3;
 
-    for (var i: u32 = 0; i < params.ne0; i++) {
-      dst[i_dst_row + i] = f16(src[i_src_row + i]);
-    }
+    let col_idx = (gid.x % elems_per_row);
+    dst[i_dst_row/{{VEC_SIZE}} + col_idx] = {{DST_TYPE}}(src[i_src_row/{{VEC_SIZE}} + col_idx]);
 }
+
+#end(SHADER)
+

gguf-py/gguf/constants.py

@@ -426,6 +426,7 @@ class MODEL_ARCH(IntEnum):
     APERTUS          = auto()
     COGVLM           = auto()
     MINIMAXM2        = auto()
+    PANGU_EMBED      = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -793,6 +794,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.APERTUS:          "apertus",
     MODEL_ARCH.MINIMAXM2:        "minimax-m2",
     MODEL_ARCH.COGVLM:           "cogvlm",
+    MODEL_ARCH.PANGU_EMBED:      "pangu-embedded",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -2958,6 +2960,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.VISEXP_UP,
         MODEL_TENSOR.VISEXP_DOWN,
     ],
+    MODEL_ARCH.PANGU_EMBED: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 
@@ -3013,6 +3029,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
     MODEL_ARCH.BAILINGMOE: [
         MODEL_TENSOR.ROPE_FREQS,
     ],
+    MODEL_ARCH.PANGU_EMBED: [
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+    ],
 }
 
 #

include/llama.h

@@ -463,6 +463,7 @@ extern "C" {
 
     // NOTE: After creating a llama_context, it is recommended to query the actual values using these functions
     //       In some cases the requested values via llama_context_params may differ from the actual values used by the context
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732
     LLAMA_API uint32_t llama_n_ctx      (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_ctx_seq  (const struct llama_context * ctx);
     LLAMA_API uint32_t llama_n_batch    (const struct llama_context * ctx);
@@ -485,6 +486,7 @@ extern "C" {
 
     LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_embd     (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_embd_inp (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
     LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);

scripts/sync-ggml.last

@@ -1 +1 @@
-e02fb860ccbba8967905bceff23b677e88105280
+7b6abb2b92fcef35cb01c6ce6ada9bd85306522d

src/CMakeLists.txt

@@ -99,6 +99,7 @@ add_library(llama
             models/openai-moe-iswa.cpp
             models/openelm.cpp
             models/orion.cpp
+            models/pangu-embedded.cpp
             models/phi2.cpp
             models/phi3.cpp
             models/plamo.cpp

src/llama-arch.cpp

@@ -107,6 +107,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_APERTUS,          "apertus"          },
     { LLM_ARCH_MINIMAX_M2,       "minimax-m2"       },
     { LLM_ARCH_COGVLM,           "cogvlm"           },
+    { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -2377,6 +2378,23 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
         },
     },
+    {
+        LLM_ARCH_PANGU_EMBED,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_COGVLM,
         {

src/llama-arch.h

@@ -111,6 +111,7 @@ enum llm_arch {
     LLM_ARCH_APERTUS,
     LLM_ARCH_MINIMAX_M2,
     LLM_ARCH_COGVLM,
+    LLM_ARCH_PANGU_EMBED,
     LLM_ARCH_UNKNOWN,
 };
 

src/llama-chat.cpp

@@ -73,6 +73,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "kimi-k2",           LLM_CHAT_TEMPLATE_KIMI_K2           },
     { "seed_oss",          LLM_CHAT_TEMPLATE_SEED_OSS          },
     { "grok-2",            LLM_CHAT_TEMPLATE_GROK_2            },
+    { "pangu-embedded",    LLM_CHAT_TEMPLATE_PANGU_EMBED       },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -213,6 +214,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_SEED_OSS;
     } else if (tmpl_contains("'Assistant: '  + message['content'] + '<|separator|>")) {
         return LLM_CHAT_TEMPLATE_GROK_2;
+    } else if (tmpl_contains(LU8("[unused9]系统：[unused10]"))) {
+        return LLM_CHAT_TEMPLATE_PANGU_EMBED;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -813,6 +816,35 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "Assistant:";
         }
+    }else if (tmpl == LLM_CHAT_TEMPLATE_PANGU_EMBED) {
+        // [unused9]系统：xxx[unused10]
+        // [unused9]用户：xxx[unused10]
+        // [unused9]助手：xxx[unused10]
+        // ...
+        for (size_t i = 0; i < chat.size(); ++i) {
+            const auto & msg = chat[i];
+            const std::string & role = msg->role;
+            const std::string & content = msg->content;
+
+            if (i == 0 && role != "system") {
+                ss << "[unused9]系统：[unused10]";
+            }
+
+            if (role == "system") {
+                ss << "[unused9]系统：" << content << "[unused10]";
+            } else if (role == "user") {
+                ss << "[unused9]用户：" << content << "[unused10]";
+            } else if (role == "assistant") {
+                ss << "[unused9]助手：" << content << "[unused10]";
+            } else if (role == "tool") {
+                ss << "[unused9]工具：" << content << "[unused10]";
+            } else if (role == "function") {
+                ss << "[unused9]方法：" << content << "[unused10]";
+            }
+        }
+        if (add_ass) {
+            ss << "[unused9]助手：";
+        }
     } else {
         // template not supported
         return -1;

src/llama-chat.h

@@ -53,6 +53,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_KIMI_K2,
     LLM_CHAT_TEMPLATE_SEED_OSS,
     LLM_CHAT_TEMPLATE_GROK_2,
+    LLM_CHAT_TEMPLATE_PANGU_EMBED,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

src/llama-context.cpp

@@ -21,6 +21,8 @@ llama_context::llama_context(
               llama_context_params params) :
     model(model),
     balloc(std::make_unique<llama_batch_allocr>(model.hparams.n_pos_per_embd())) {
+    // TODO warning when creating llama_context with awkward ctx size that is not a power of 2,
+    //     may need to be backend-dependent
     LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
 
     t_start_us = model.t_start_us;
@@ -112,10 +114,14 @@ llama_context::llama_context(
         }
     }
 
+    // ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732
+    cparams.n_ctx = GGML_PAD(cparams.n_ctx, 256);
+
     if (cparams.kv_unified) {
         cparams.n_ctx_seq = cparams.n_ctx;
     } else {
         cparams.n_ctx_seq = cparams.n_ctx / cparams.n_seq_max;
+        cparams.n_ctx_seq = GGML_PAD(cparams.n_ctx_seq, 256);
 
         if (cparams.n_ctx_seq == 0) {
             throw std::runtime_error("n_ctx_seq == 0");
@@ -821,7 +827,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
 
     const auto & hparams = model.hparams;
 
-    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd_inp();
     const int64_t n_vocab = model.vocab.n_tokens();
 
     // note: during encode, we always pass the full sequence starting from pos = 0
@@ -990,7 +996,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const auto & hparams = model.hparams;
 
     const int64_t n_vocab = vocab.n_tokens();
-    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_embd  = hparams.n_embd_inp();
 
     // when computing embeddings, all tokens are output
     const bool output_all = cparams.embeddings;
@@ -2148,7 +2154,7 @@ void llama_context::opt_epoch_iter(
             batch.logits  [pos_batch]    = true;
         }
 
-        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
+        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd_inp(), cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
             LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
             return;
         }

src/llama-graph.cpp

@@ -1142,7 +1142,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
 // input embeddings with optional lora
 ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
-    const int64_t n_embd = hparams.n_embd;
+    const int64_t n_embd = hparams.n_embd_inp();
 
     auto inp = std::make_unique<llm_graph_input_embd>();
 
@@ -1279,7 +1279,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
     //    return cur;
     //}
 
-    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd;
+    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp();
     const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
     cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);

src/llama-hparams.cpp

@@ -60,6 +60,16 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
     return n_head/n_head_kv;
 }
 
+uint32_t llama_hparams::n_embd_inp() const {
+    uint32_t n_embd_inp = n_embd;
+
+    if (n_deepstack_layers > 0) {
+        n_embd_inp += n_embd * n_deepstack_layers;
+    }
+
+    return n_embd_inp;
+}
+
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 

src/llama-hparams.h

@@ -227,6 +227,9 @@ struct llama_hparams {
 
     uint32_t n_gqa(uint32_t il = 0) const;
 
+    // dimension of main + auxiliary input embeddings
+    uint32_t n_embd_inp() const;
+
     // dimension of key embeddings across all k-v heads
     uint32_t n_embd_k_gqa(uint32_t il = 0) const;
 

src/llama-kv-cache-iswa.cpp

@@ -45,7 +45,9 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
 
     const uint32_t size_base = kv_size;
 
-    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch, n_pad));
+    // note: the SWA cache is always padded to 256 for performance
+    //       https://github.com/ggml-org/llama.cpp/issues/17037
+    uint32_t size_swa = GGML_PAD(std::min(size_base, hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256);
 
     // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
     if (swa_full) {

src/llama-model.cpp

@@ -276,8 +276,8 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
             } break;
         case GGML_OP_IM2COL:
             {
-                const int n_embd = hparams.n_embd;
-                ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd, w->ne[1], 1, 1);
+                const int n_embd_inp = hparams.n_embd_inp();
+                ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd_inp, w->ne[1], 1, 1);
                 op_tensor = ggml_im2col(ctx, w, b, 1, 0, 0, 0, 1, 0, false, GGML_TYPE_F16);
             } break;
         case GGML_OP_SCALE:
@@ -1039,9 +1039,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     case 64: type = LLM_TYPE_32B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
-                // since vision model stacks deepstack features along feature dim
-                // we also create a fake "n_embd" for text model to be the main embd + deepstack embds
-                hparams.n_embd *= hparams.n_deepstack_layers + 1;
             } break;
         case LLM_ARCH_QWEN3MOE:
             {
@@ -1065,9 +1062,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     case 94: type = LLM_TYPE_235B_A22B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
-                // since vision model stacks deepstack features along feature dim
-                // we also create a fake "n_embd" for text model to be the main embd + deepstack embds
-                hparams.n_embd *= hparams.n_deepstack_layers + 1;
             } break;
         case LLM_ARCH_PHI2:
             {
@@ -2177,6 +2171,15 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_PANGU_EMBED:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 26: type = LLM_TYPE_1B; break; // openPangu-Embedded-1B-V1.1
+                    case 34: type = LLM_TYPE_7B; break; // openPangu-Embedded-7B-V1.1
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -3332,10 +3335,6 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             case LLM_ARCH_QWEN3:
             case LLM_ARCH_QWEN3VL:
                 {
-                    // for model loading, the weights only have the main embd
-                    // so we need to divide by the number of deepstack layers + 1
-                    // n_embd is const int so we declare a new variable
-                    int64_t n_embd = hparams.n_embd / (hparams.n_deepstack_layers + 1);
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
                     // output
@@ -3371,10 +3370,6 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             case LLM_ARCH_QWEN3MOE:
             case LLM_ARCH_QWEN3VLMOE:
                 {
-                    // for model loading, the weights only have the main embd
-                    // so we need to divide by the number of deepstack layers + 1
-                    // n_embd is const int so we declare a new variable
-                    int64_t n_embd = hparams.n_embd / (hparams.n_deepstack_layers + 1);
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
                     // output
@@ -6263,6 +6258,50 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.visexp_ffn_up   = create_tensor(tn(LLM_TENSOR_VISEXP_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
                     }
                 } break;
+            case LLM_ARCH_PANGU_EMBED:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+                    // if output is NULL, init from the input tok embed
+                    if (output == NULL) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        // weight tensors
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        // bias tensors
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd_head_k * n_head}, 0);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, 0);
+                        layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
+                            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                        } else {
+                            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                        }
+
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -6482,6 +6521,7 @@ void llama_model::print_info() const {
     if (!hparams.vocab_only) {
         LLAMA_LOG_INFO("%s: n_ctx_train      = %u\n",     __func__, hparams.n_ctx_train);
         LLAMA_LOG_INFO("%s: n_embd           = %u\n",     __func__, hparams.n_embd);
+        LLAMA_LOG_INFO("%s: n_embd_inp       = %u\n",     __func__, hparams.n_embd_inp());
         LLAMA_LOG_INFO("%s: n_layer          = %u\n",     __func__, hparams.n_layer);
         LLAMA_LOG_INFO("%s: n_head           = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head(il);    }, hparams.n_layer).c_str());
         LLAMA_LOG_INFO("%s: n_head_kv        = %s\n",     __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
@@ -7260,6 +7300,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_cogvlm>(*this, params);
             } break;
+        case LLM_ARCH_PANGU_EMBED:
+            {
+                llm = std::make_unique<llm_build_pangu_embedded>(*this, params);
+            }break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -7323,6 +7367,10 @@ int32_t llama_model_n_embd(const llama_model * model) {
     return model->hparams.n_embd;
 }
 
+int32_t llama_model_n_embd_inp(const llama_model * model) {
+    return model->hparams.n_embd_inp();
+}
+
 int32_t llama_model_n_layer(const llama_model * model) {
     return model->hparams.n_layer;
 }
@@ -7479,6 +7527,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_APERTUS:
         case LLM_ARCH_MINIMAX_M2:
         case LLM_ARCH_COGVLM:
+        case LLM_ARCH_PANGU_EMBED:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:

src/models/models.h

@@ -361,6 +361,10 @@ struct llm_build_orion : public llm_graph_context {
     llm_build_orion(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_pangu_embedded : public llm_graph_context {
+    llm_build_pangu_embedded(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_phi2 : public llm_graph_context {
     llm_build_phi2(const llama_model & model, const llm_graph_params & params);
 };

src/models/pangu-embedded.cpp

@@ -0,0 +1,121 @@
+#include "models.h"
+
+
+llm_build_pangu_embedded::llm_build_pangu_embedded(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self attention
+        {
+            // compute Q and K and RoPE them
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+            cb(Qcur, "Qcur", il);
+
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+            Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+            cb(Kcur, "Kcur", il);
+
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+            Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+            cb(Vcur, "Vcur", il);
+
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+                model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                NULL,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    if (model.output_b != nullptr) {
+        cur = ggml_add(ctx0, cur, model.output_b);
+    }
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

src/models/qwen3vl-moe.cpp

@@ -1,9 +1,8 @@
 #include "models.h"
 
 llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_full = hparams.n_embd; // main embd + deepstack embds
     const size_t n_deepstack_layers = hparams.n_deepstack_layers;
-    const int64_t n_embd = n_embd_full / (n_deepstack_layers + 1);
+    const int64_t n_embd = hparams.n_embd;
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);

src/models/qwen3vl.cpp

@@ -1,13 +1,10 @@
 #include "models.h"
 
 llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-
-    const int64_t n_embd_full = hparams.n_embd; // main embd + deepstack embds
     const size_t n_deepstack_layers = hparams.n_deepstack_layers;
-    const int64_t n_embd = n_embd_full / (n_deepstack_layers + 1);
+    const int64_t n_embd = hparams.n_embd;
     const int64_t n_embd_head = hparams.n_embd_head_v;
 
-
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
     GGML_ASSERT(n_embd_head == hparams.n_rot);
 

tests/test-backend-ops.cpp

@@ -2294,6 +2294,79 @@ struct test_rope_set_rows : public test_case {
     }
 };
 
+// GGML_OP_RMS_NORM + GGML_OP_MUL + GGML_OP_ROPE (+ GGML_OP_VIEW + GGML_OP_SET_ROWS)
+struct test_rms_norm_mul_rope : public test_case {
+    const std::array<int64_t, 4> ne;
+    const float eps;
+    const bool multi_add; // test a sequence of adds feeding into rms_norm
+    const bool set_rows;
+    int mode;
+
+    std::string op_desc(ggml_tensor * t) override {
+        GGML_UNUSED(t);
+        return "RMS_NORM_MUL_ROPE";
+    }
+
+    bool run_whole_graph() override { return true; }
+
+    std::string vars() override {
+        return VARS_TO_STR5(ne, eps, multi_add, set_rows, mode);
+    }
+
+    test_rms_norm_mul_rope(std::array<int64_t, 4> ne, float eps = 1e-6f, bool multi_add = false,
+                           bool set_rows = false, int mode = GGML_ROPE_TYPE_NORMAL)
+        : ne(ne), eps(eps), multi_add(multi_add), set_rows(set_rows), mode(mode) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, ne[0], ne[1], ne[2], 1);
+        ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, ne[0], ne[1], ne[2], 1);
+        ggml_tensor * c = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, ne[0], ne[1], ne[2], 1);
+
+        if (multi_add) {
+            a = ggml_add(ctx, ggml_add(ctx, a, b), c);
+        }
+
+        a = ggml_mul(ctx, ggml_rms_norm(ctx, a, eps), b);
+
+        ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne[2]);
+
+        ggml_tensor * rope = ggml_rope(ctx, a, pos, ne[0], mode);
+
+        ggml_tensor * out;
+
+        if (set_rows) {
+            ggml_tensor * view = ggml_view_2d(ctx, rope, ne[0] * ne[1], ne[2], rope->nb[2], 0);
+
+            ggml_tensor * dst = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, ne[0] * ne[1], ne[2] * ne[3], 1, 1);
+            ggml_set_name(dst, "dst");
+
+            ggml_tensor * row_idxs = ggml_new_tensor_3d(ctx, GGML_TYPE_I64, ne[2], 1, 1);
+            ggml_set_name(row_idxs, "row_idxs");
+
+            out = ggml_set_rows(ctx, dst, view, row_idxs);
+            ggml_set_name(out, "out");
+        } else {
+            out = rope;
+        }
+
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (t->type == GGML_TYPE_I64 || t->type == GGML_TYPE_I32) {
+                if (ggml_is_view_op(t->op)) {
+                    continue;
+                }
+
+                init_set_rows_row_ids(t, ne[2]);
+            } else {
+                init_tensor_uniform(t);
+            }
+        }
+    }
+};
+
 // GGML_OP_ARGMAX
 struct test_argmax : public test_case {
     const ggml_type type;
@@ -2576,9 +2649,10 @@ struct test_cpy : public test_case {
     const std::array<int64_t, 4> permute_dst;
     bool _src_use_permute;
     bool _dst_use_permute;
+    bool _src_transpose;
 
     std::string vars() override {
-        return VARS_TO_STR5(type_src, type_dst, ne, permute_src, permute_dst);
+        return VARS_TO_STR6(type_src, type_dst, ne, permute_src, permute_dst, _src_transpose);
     }
 
     double max_nmse_err() override {
@@ -2616,10 +2690,12 @@ struct test_cpy : public test_case {
     test_cpy(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {10, 10, 10, 1},
             std::array<int64_t, 4> permute_src = {0, 0, 0, 0},
-            std::array<int64_t, 4> permute_dst = {0, 0, 0, 0})
+            std::array<int64_t, 4> permute_dst = {0, 0, 0, 0},
+            bool transpose_src = false)
         : type_src(type_src), type_dst(type_dst), ne(ne), permute_src(permute_src), permute_dst(permute_dst),
           _src_use_permute(permute_src[0] + permute_src[1] + permute_src[2] + permute_src[3] > 0),
-          _dst_use_permute(permute_dst[0] + permute_dst[1] + permute_dst[2] + permute_dst[3] > 0) {}
+          _dst_use_permute(permute_dst[0] + permute_dst[1] + permute_dst[2] + permute_dst[3] > 0),
+          _src_transpose(transpose_src){}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
@@ -2631,6 +2707,11 @@ struct test_cpy : public test_case {
             ggml_set_name(src, "src_permuted");
         }
 
+        if (_src_transpose) {
+            src = ggml_transpose(ctx, src);
+            ggml_set_name(src, "src_transposed");
+        }
+
         ggml_tensor * dst = ggml_new_tensor(ctx, type_dst, 4, src->ne);
         ggml_set_name(dst, "dst");
 
@@ -3377,11 +3458,11 @@ struct test_mul_mat : public test_case {
     const std::array<int64_t, 2> bs;  // dims 3 and 4
     const std::array<int64_t, 2> nr;  // repeat in dims 3 and 4
     const std::array<int64_t, 4> per; // permutation of dimensions
-    const bool v; // whether a and b are non-contiguous views
+    const int64_t k_v; // size of k in memory, resulting in a non-contiguous view for k_v > k, no view for k_v == 0
     const uint32_t o; // number of outputs
 
     std::string vars() override {
-        return VARS_TO_STR10(type_a, type_b, m, n, k, bs, nr, per, v, o);
+        return VARS_TO_STR10(type_a, type_b, m, n, k, bs, nr, per, k_v, o);
     }
 
     double max_nmse_err() override {
@@ -3402,8 +3483,8 @@ struct test_mul_mat : public test_case {
             std::array<int64_t, 2> bs = {10, 10},
             std::array<int64_t, 2> nr = {2, 2},
             std::array<int64_t, 4> per = {0, 1, 2, 3},
-            bool v = false, uint32_t o = 1)
-        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), v(v), o(o) {}
+            int64_t k_v = 0, uint32_t o = 1)
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), k_v(k_v), o(o) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         // C^T = A * B^T: (k, m) * (k, n) => (m, n)
@@ -3413,7 +3494,7 @@ struct test_mul_mat : public test_case {
         const int npermuted = (per[0] != 0) + (per[1] != 1) + (per[2] != 2) + (per[3] != 3);
         if (npermuted > 0) {
             GGML_ASSERT(npermuted == 2);
-            GGML_ASSERT(!v); // not handled
+            GGML_ASSERT(k_v == 0); // not handled
             GGML_ASSERT(!ggml_is_quantized(type_a) || per[0] == 0);
             GGML_ASSERT(!ggml_is_quantized(type_b) || per[0] == 0);
 
@@ -3437,29 +3518,21 @@ struct test_mul_mat : public test_case {
             ggml_set_name(a, "a_permuted");
             ggml_set_name(b, "b_permuted");
         } else {
-            if (v) {
-                a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0],       bs[1]);
-                b = ggml_new_tensor_4d(ctx, type_b, k*2, n, bs[0]*nr[0], bs[1]*nr[1]);
+            const int64_t k_physical = k_v == 0 ? k : k_v;
+            a = ggml_new_tensor_4d(ctx, type_a, k_physical, m, bs[0],       bs[1]);
+            b = ggml_new_tensor_4d(ctx, type_b, k_physical, n, bs[0]*nr[0], bs[1]*nr[1]);
 
-                if (!ggml_is_quantized(type_a)) {
-                    if (bs[1] == 1 && nr[1] == 1) {
-                        ggml_set_param(a);
-                    }
-                    ggml_set_param(b);
+            if (!ggml_is_quantized(type_a)) {
+                if (bs[1] == 1 && nr[1] == 1) {
+                    ggml_set_param(a);
                 }
+                ggml_set_param(b);
+            }
 
+            if (k_v != 0) {
+                GGML_ASSERT(k_v > k);
                 a = ggml_view_4d(ctx, a, k, m, bs[0],       bs[1],       a->nb[1], a->nb[2], a->nb[3], 0);
                 b = ggml_view_4d(ctx, b, k, n, bs[0]*nr[0], bs[1]*nr[1], b->nb[1], b->nb[2], b->nb[3], 0);
-            } else {
-                a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
-                b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
-
-                if (!ggml_is_quantized(type_a)) {
-                    if (bs[1] == 1 && nr[1] == 1) {
-                        ggml_set_param(a);
-                    }
-                    ggml_set_param(b);
-                }
             }
             ggml_set_name(a, "a");
             ggml_set_name(b, "b");
@@ -4809,60 +4882,6 @@ struct test_topk_moe: public test_case {
     }
 };
 
-struct test_moe_expert_reduce : public test_case {
-    const int64_t n_embd;
-    const int64_t n_tokens;
-    const int64_t n_expert_used;
-
-    test_moe_expert_reduce(int64_t n_embd = 64, int64_t n_tokens = 5, int64_t n_expert_used = 4)
-        : n_embd(n_embd), n_tokens(n_tokens), n_expert_used(n_expert_used) {
-        GGML_ASSERT(n_expert_used > 1);
-    }
-
-    std::string vars() override {
-        return VARS_TO_STR3(n_embd, n_tokens, n_expert_used);
-    }
-
-    std::string op_desc(ggml_tensor * t) override {
-        GGML_UNUSED(t);
-        return "MOE_EXPERT_REDUCE";
-    }
-
-    bool run_whole_graph() override { return true; }
-
-    ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * experts = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, n_expert_used, n_tokens);
-        ggml_set_name(experts, "experts");
-
-        ggml_tensor * weights = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 1, n_expert_used, n_tokens);
-        ggml_set_name(weights, "weights");
-
-        ggml_tensor * weighted = ggml_mul(ctx, experts, weights);
-        ggml_set_name(weighted, "weighted_experts");
-
-        std::vector<ggml_tensor *> expert_views(n_expert_used);
-        for (int64_t i = 0; i < n_expert_used; ++i) {
-            expert_views[i] = ggml_view_2d(ctx, weighted, n_embd, n_tokens, weighted->nb[2], i * weighted->nb[1]);
-
-            std::string name = "expert_view_" + std::to_string(i);
-            ggml_set_name(expert_views[i], name.c_str());
-            ggml_build_forward_expand(gf, expert_views[i]);
-        }
-
-        ggml_tensor * moe_out = expert_views[0];
-        for (int64_t i = 1; i < n_expert_used; ++i) {
-            moe_out = ggml_add(ctx, moe_out, expert_views[i]);
-
-            std::string name = "expert_add_" + std::to_string(i - 1);
-            ggml_set_name(moe_out, name.c_str());
-        }
-
-        ggml_set_name(moe_out, "moe_out");
-
-        return moe_out;
-    }
-};
-
 struct test_mul_mat_vec_fusion : public test_case {
     const ggml_type type;
     const ggml_glu_op glu_op;
@@ -4911,8 +4930,10 @@ struct test_mul_mat_vec_fusion : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         if (!use_id) {
-            std::array<int64_t, 4> ne = {k, m, 1, 1};
-            std::array<int64_t, 4> ne0 = {k, n, 1, 1};
+            const int              channels = 4;
+            const int              samples  = 2;
+            std::array<int64_t, 4> ne       = { k, m, channels, samples };
+            std::array<int64_t, 4> ne0      = { k, n, channels, samples };
 
             ggml_tensor * cur  = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne.data());
             ggml_tensor * gate = with_gate ? ggml_new_tensor(ctx, type, 4, ne0.data()) : nullptr;
@@ -4920,14 +4941,14 @@ struct test_mul_mat_vec_fusion : public test_case {
 
             ggml_tensor * ffn_up = ggml_mul_mat(ctx, up, cur);
             if (with_bias) {
-                std::array<int64_t, 4> bias_ne = {ffn_up->ne[0], 1, 1, 1};
+                std::array<int64_t, 4> bias_ne = { ffn_up->ne[0], 1, channels, samples };
                 ggml_tensor * up_bias = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, bias_ne.data());
                 ffn_up = ggml_add(ctx, ffn_up, up_bias);
             }
 
             ggml_tensor * ffn_gate = with_gate ? ggml_mul_mat(ctx, gate, cur) : nullptr;
             if (with_bias && with_gate) {
-                std::array<int64_t, 4> bias_ne = {ffn_gate->ne[0], 1, 1, 1};
+                std::array<int64_t, 4> bias_ne   = { ffn_gate->ne[0], 1, channels, samples };
                 ggml_tensor * gate_bias = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, bias_ne.data());
                 ffn_gate = ggml_add(ctx, ffn_gate, gate_bias);
             }
@@ -6641,6 +6662,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_I32, {256, 2, 3, 4}, {1, 0, 2, 3}));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_I32, GGML_TYPE_F32, {256, 2, 3, 4}));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_I32, GGML_TYPE_F32, {256, 2, 3, 4}, {1, 0, 2, 3}));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {256, 4, 3, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {256, 4, 3, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {256, 4, 3, 3}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_BF16, GGML_TYPE_BF16, {256, 4, 3, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {256, 4, 1, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {256, 4, 1, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_BF16, GGML_TYPE_BF16, {256, 4, 1, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {256, 1, 4, 1}, {1, 2, 0, 3}, {0, 0, 0, 0}));
 
     test_cases.emplace_back(new test_cont());
     test_cases.emplace_back(new test_cont(GGML_TYPE_F32, {2, 1, 1 ,1}));
@@ -6743,6 +6772,22 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         }
     }
 
+    for (auto multi_add : {false, true}) {
+        for (auto set_rows : {false, true}) {
+            for (auto rope : {GGML_ROPE_TYPE_NORMAL, GGML_ROPE_TYPE_NEOX}) {
+                test_cases.emplace_back(new test_rms_norm_mul_rope({768, 1, 1, 1}, 1e-6f, multi_add, set_rows, rope));
+                test_cases.emplace_back(new test_rms_norm_mul_rope({768, 3, 1, 1}, 1e-6f, multi_add, set_rows, rope));
+                test_cases.emplace_back(new test_rms_norm_mul_rope({768, 3, 5, 1}, 1e-6f, multi_add, set_rows, rope));
+                test_cases.emplace_back(new test_rms_norm_mul_rope({128, 32, 2, 1}, 1e-6f, multi_add, set_rows, rope));
+                test_cases.emplace_back(new test_rms_norm_mul_rope({128, 4, 2, 1}, 1e-6f, multi_add, set_rows, rope));
+                test_cases.emplace_back(new test_rms_norm_mul_rope({128, 32, 50, 1}, 1e-6f, multi_add, set_rows, rope));
+                test_cases.emplace_back(new test_rms_norm_mul_rope({128, 4, 50, 1}, 1e-6f, multi_add, set_rows, rope));
+                test_cases.emplace_back(new test_rms_norm_mul_rope({8192, 2, 2, 1}, 1e-6f, multi_add, set_rows, rope));
+                test_cases.emplace_back(new test_rms_norm_mul_rope({8192, 2, 2, 1}, 1e-6f, multi_add, set_rows, rope));
+            }
+        }
+    }
+
     test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, {64, 5, 4, 3}, 1e-12f));
 
     for (int64_t d_conv : {3, 4}) {
@@ -6886,7 +6931,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45,  64, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 193, {1,  1}, {4, 1}, {0, 2, 1, 3}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67,  {1,  1}, {4, 1}, {0, 2, 1, 3}));
-    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 32, 32, { 1,  1}, {1, 1}, {0, 1, 2, 3}, true, 3));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 32, 32, { 1,  1}, {1, 1}, {0, 1, 2, 3}, 64, 3));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 77, 77, {12,1}, {1,1}));
 
 #if 0
@@ -6912,7 +6957,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                     for (uint32_t k = 0; k < 2; ++k) {
                         for (ggml_type type: {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_F32}) {
                             test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 1056 + m, 1, 128 + k,  {bs,  bs2}, {nr, 1}, {0, 2, 1, 3}));
-                            test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 128 + m,  1, 1056 + k, {bs,  bs2}, {nr, 1}, {0, 1, 2, 3}, true));
+                            test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 128 + m,  1, 1056 + k, {bs,  bs2}, {nr, 1}, {0, 1, 2, 3}, 2*1056 + k));
                         }
                     }
                 }
@@ -7316,10 +7361,6 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_topk_moe({ 8, 22, 1, 1 }, 4, /*with_norm*/ false, /*delayed_softmax*/ true));
     test_cases.emplace_back(new test_topk_moe({ 32, 22, 1, 1 }, 8, /*with_norm*/ false, /*delayed_softmax*/ true));
 
-    test_cases.emplace_back(new test_moe_expert_reduce(1024, 5, 4));
-    test_cases.emplace_back(new test_moe_expert_reduce(80, 3, 6));
-    test_cases.emplace_back(new test_moe_expert_reduce(80, 3, 7));
-
 #if 0
     // these tests are disabled to save execution time, sbut they can be handy for debugging
     test_cases.emplace_back(new test_llama(2, true));
@@ -7385,6 +7426,18 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32,  GGML_TYPE_Q4_0, {8192, 512, 2, 1}));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_Q4_0, GGML_TYPE_F32,  {8192, 512, 2, 1}));
 
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {768*1024, 256, 1, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {768*1024, 256, 1, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {768, 1024, 256, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_BF16, GGML_TYPE_BF16, {768, 1024, 256, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}));
+
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {768*1024, 256, 1, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {768, 1024, 256, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {768*1024, 256, 1, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {768, 1024, 256, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+    test_cases.emplace_back(new test_cpy(GGML_TYPE_BF16, GGML_TYPE_BF16, {768, 1024, 256, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
+
+
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {4096, 4096, 5, 1}, false, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {12888, 256, 5, 1}, false, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 4096, 5, 1}, false, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
@@ -7405,7 +7458,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_pad_reflect_1d(GGML_TYPE_F32, {3000, 384, 4, 1}));
 
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 16416, 1, 128, {8,  1}, {4, 1}, {0, 2, 1, 3}));
-    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 1, 16416, {8,  1}, {4, 1}, {0, 1, 2, 3}, true));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 1, 16416, {8,  1}, {4, 1}, {0, 1, 2, 3}, 2*16416));
 
     for (int bs : {1, 2, 3, 4, 5, 8, 512}) {
         for (ggml_type type_a : all_types) {

tools/llama-bench/llama-bench.cpp

@@ -1919,6 +1919,12 @@ struct sql_printer : public printer {
     }
 };
 
+struct ctx_state {
+    int depth = 0; // in tokens
+
+    std::vector<uint8_t> buf; // the llama_context state buffer
+};
+
 static bool test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
@@ -2051,6 +2057,10 @@ int main(int argc, char ** argv) {
     llama_model *               lmodel    = nullptr;
     const cmd_params_instance * prev_inst = nullptr;
 
+    // store the llama_context state at the previous depth that we performed a test
+    // ref: https://github.com/ggml-org/llama.cpp/pull/16944#issuecomment-3478151721
+    ctx_state cstate;
+
     int  params_idx   = 0;
     auto params_count = params_instances.size();
     for (const auto & inst : params_instances) {
@@ -2134,14 +2144,37 @@ int main(int argc, char ** argv) {
             llama_memory_clear(llama_get_memory(ctx), false);
 
             if (t.n_depth > 0) {
-                if (params.progress) {
-                    fprintf(stderr, "llama-bench: benchmark %d/%zu: depth run %d/%d\n", params_idx, params_count,
-                            i + 1, params.reps);
+                bool is_cached = t.n_depth == cstate.depth;
+
+                if (is_cached) {
+                    // if previously we have computed at this depth, just restore the state
+                    const size_t ret = llama_state_seq_set_data(ctx, cstate.buf.data(), cstate.buf.size(), 0);
+                    if (ret == 0) {
+                        // if the old state is incompatible with the current context - reprocess from scratch
+                        is_cached = false;
+                    }
                 }
-                bool res = test_prompt(ctx, t.n_depth, t.n_batch, t.n_threads);
-                if (!res) {
-                    fprintf(stderr, "%s: error: failed to run depth\n", __func__);
-                    exit(1);
+
+                if (!is_cached) {
+                    if (params.progress) {
+                        fprintf(stderr, "llama-bench: benchmark %d/%zu: depth run %d/%d\n", params_idx, params_count,
+                                i + 1, params.reps);
+                    }
+                    bool res = test_prompt(ctx, t.n_depth, t.n_batch, t.n_threads);
+                    if (!res) {
+                        fprintf(stderr, "%s: error: failed to run depth\n", __func__);
+                        exit(1);
+                    }
+
+                    // store the context state for reuse in later runs
+                    cstate.depth = t.n_depth;
+                    cstate.buf.resize(llama_state_seq_get_size(ctx, 0));
+                    llama_state_seq_get_data(ctx, cstate.buf.data(), cstate.buf.size(), 0);
+                } else {
+                    if (params.progress) {
+                        fprintf(stderr, "llama-bench: benchmark %d/%zu: depth run %d/%d (cached)\n", params_idx, params_count,
+                                i + 1, params.reps);
+                    }
                 }
             }
 

tools/mtmd/clip.cpp

@@ -1083,16 +1083,24 @@ struct clip_graph {
     }
 
     ggml_cgraph * build_minicpmv() {
-        const int batch_size = 1;
-
         GGML_ASSERT(model.class_embedding == nullptr);
-        const int n_pos = n_patches;
+        const int n_pos       = n_patches;
+        const int n_embd_proj = clip_n_mmproj_embd(ctx);
 
         // position embeddings for the projector (not for ViT)
-        int n_output_dim = clip_n_mmproj_embd(ctx);
-        ggml_tensor * pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_output_dim, n_pos, batch_size);
-        ggml_set_name(pos_embed, "pos_embed");
-        ggml_set_input(pos_embed);
+        // see: https://huggingface.co/openbmb/MiniCPM-o-2_6/blob/main/resampler.py#L70
+        // base frequency omega
+        ggml_tensor * omega = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_embd_proj / 4);
+        ggml_set_name(omega, "omega");
+        ggml_set_input(omega);
+
+        // 2D input positions (using float for sinusoidal embeddings)
+        ggml_tensor * pos_h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_pos);
+        ggml_set_name(pos_h, "pos_h");
+        ggml_set_input(pos_h);
+        ggml_tensor * pos_w = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_pos);
+        ggml_set_name(pos_w, "pos_w");
+        ggml_set_input(pos_w);
 
         // for selecting learned pos embd, used by ViT
         struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_pos);
@@ -1103,7 +1111,7 @@ struct clip_graph {
 
         ggml_tensor * inp = build_inp();
         ggml_tensor * embeddings = build_vit(
-                                inp, n_patches,
+                                inp, n_pos,
                                 NORM_TYPE_NORMAL,
                                 hparams.ffn_op,
                                 learned_pos_embd,
@@ -1115,17 +1123,39 @@ struct clip_graph {
         ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);
 
         // norm
-        q = build_norm(q, model.mm_model_ln_q_w, model.mm_model_ln_q_b, NORM_TYPE_NORMAL, eps, -1);
+        q = build_norm(q, model.mm_model_ln_q_w,  model.mm_model_ln_q_b,  NORM_TYPE_NORMAL, eps, -1);
         v = build_norm(v, model.mm_model_ln_kv_w, model.mm_model_ln_kv_b, NORM_TYPE_NORMAL, eps, -1);
 
+        // calculate sinusoidal pos embd
+        ggml_tensor * pos_embed = nullptr;
+        {
+            // outer product
+            ggml_tensor * omega_b = ggml_repeat_4d(ctx0, omega, omega->ne[0], n_pos, 1, 1); // n_pos rows
+            ggml_tensor * theta_x = ggml_mul(ctx0, omega_b, pos_w);
+            ggml_tensor * theta_y = ggml_mul(ctx0, omega_b, pos_h);
+            // sin and cos
+            ggml_tensor * pos_embd_x = ggml_concat(
+                ctx0,
+                ggml_sin(ctx0, theta_x),
+                ggml_cos(ctx0, theta_x),
+                0 // concat on first dim
+            );
+            ggml_tensor * pos_embd_y = ggml_concat(
+                ctx0,
+                ggml_sin(ctx0, theta_y),
+                ggml_cos(ctx0, theta_y),
+                0 // concat on first dim
+            );
+            pos_embed = ggml_concat(ctx0, pos_embd_x, pos_embd_y, 0);
+        }
+
         // k = v + pos_embed
         ggml_tensor * k = ggml_add(ctx0, v, pos_embed);
 
         // attention
         {
-            int n_embd = clip_n_mmproj_embd(ctx);
             const int d_head = 128;
-            int n_head = n_embd/d_head;
+            int n_head = n_embd_proj/d_head;
             // Use actual config value if available, otherwise fall back to hardcoded values
             int num_query = ctx->model.hparams.minicpmv_query_num;
             ggml_tensor * Q = ggml_add(ctx0,
@@ -2761,6 +2791,7 @@ struct clip_model_loader {
                     {
                         // ref: https://huggingface.co/mistral-community/pixtral-12b/blob/main/preprocessor_config.json
                         // TODO: verify the image_min_tokens
+                        hparams.n_merge = 1; // the original pixtral does not use patch merging
                         hparams.rope_theta = 10000.0f;
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
                         hparams.set_limit_image_tokens(8, 1024);
@@ -2790,14 +2821,8 @@ struct clip_model_loader {
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
                         get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern, model.proj_type == PROJECTOR_TYPE_QWEN25VL); // only 2.5 requires it
                         // ref: https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/blob/main/preprocessor_config.json
-                        // the actual max limit is 12845056/14/14/2/2/4 = 4096 tokens
-                        // but we set a lower value to avoid OOM
-                        // TODO: make it configurable by user
-                        // TODO (2): bbox coordinates become inaccurate with small number of tokens,
-                        //           therefore we need to increase the min_tokens
-                        //           see: https://github.com/ggml-org/llama.cpp/issues/16842#issuecomment-3475144858
-                        hparams.set_limit_image_tokens(8, 2048);
-                        hparams.set_warmup_n_tokens(256); // avoid OOM on warmup
+                        hparams.set_limit_image_tokens(8, 4096);
+                        hparams.set_warmup_n_tokens(46*46); // avoid OOM on warmup
                         const int warn_min_pixels = 1024 * hparams.n_merge * hparams.n_merge * hparams.patch_size * hparams.patch_size;
                         if (hparams.image_min_pixels < warn_min_pixels) {
                             LOG_WRN("%s: Qwen-VL models require at minimum 1024 image tokens to function correctly on grounding tasks\n", __func__);
@@ -4569,92 +4594,6 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
     return n_patches;
 }
 
-static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
-    assert(embed_dim % 2 == 0);
-    int H = pos.size();
-    int W = pos[0].size();
-
-    std::vector<float> omega(embed_dim / 2);
-    for (int i = 0; i < embed_dim / 2; ++i) {
-        omega[i] = 1.0 / pow(10000.0, static_cast<float>(i) / (embed_dim / 2));
-    }
-
-    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
-    for (int h = 0; h < H; ++h) {
-        for (int w = 0; w < W; ++w) {
-            for (int d = 0; d < embed_dim / 2; ++d) {
-                float out_value = pos[h][w] * omega[d];
-                emb[h][w][d] = sin(out_value);
-                emb[h][w][d + embed_dim / 2] = cos(out_value);
-            }
-        }
-    }
-
-    return emb;
-}
-
-static std::vector<std::vector<std::vector<float>>> get_2d_sincos_pos_embed_from_grid(int embed_dim, const std::vector<std::vector<std::vector<float>>> & grid) {
-    assert(embed_dim % 2 == 0);
-    std::vector<std::vector<std::vector<float>>> emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[0]); // (H, W, D/2)
-    std::vector<std::vector<std::vector<float>>> emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[1]); // (H, W, D/2)
-
-    int H = emb_h.size();
-    int W = emb_h[0].size();
-    std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
-
-    for (int h = 0; h < H; ++h) {
-        for (int w = 0; w < W; ++w) {
-            for (int d = 0; d < embed_dim / 2; ++d) {
-                emb[h][w][d] = emb_h[h][w][d];
-                emb[h][w][d + embed_dim / 2] = emb_w[h][w][d];
-            }
-        }
-    }
-    return emb;
-}
-
-static std::vector<std::vector<float>> get_2d_sincos_pos_embed(int embed_dim, const std::pair<int, int> image_size) {
-    int grid_h_size = image_size.first;
-    int grid_w_size = image_size.second;
-
-    std::vector<float> grid_h(grid_h_size);
-    std::vector<float> grid_w(grid_w_size);
-
-    for (int i = 0; i < grid_h_size; ++i) {
-        grid_h[i] = static_cast<float>(i);
-    }
-    for (int i = 0; i < grid_w_size; ++i) {
-        grid_w[i] = static_cast<float>(i);
-    }
-
-    std::vector<std::vector<float>> grid(grid_h_size, std::vector<float>(grid_w_size));
-    for (int h = 0; h < grid_h_size; ++h) {
-        for (int w = 0; w < grid_w_size; ++w) {
-            grid[h][w] = grid_w[w];
-        }
-    }
-    std::vector<std::vector<std::vector<float>>> grid_2d = {grid, grid};
-    for (int h = 0; h < grid_h_size; ++h) {
-        for (int w = 0; w < grid_w_size; ++w) {
-            grid_2d[0][h][w] = grid_h[h];
-            grid_2d[1][h][w] = grid_w[w];
-        }
-    }
-
-    std::vector<std::vector<std::vector<float>>> pos_embed_3d = get_2d_sincos_pos_embed_from_grid(embed_dim, grid_2d);
-
-    int H = image_size.first;
-    int W = image_size.second;
-    std::vector<std::vector<float>> pos_embed_2d(H * W, std::vector<float>(embed_dim));
-    for (int h = 0; h < H; ++h) {
-        for (int w = 0; w < W; ++w) {
-            pos_embed_2d[w * H + h] = pos_embed_3d[h][w];
-        }
-    }
-
-    return pos_embed_2d;
-}
-
 bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
     clip_image_f32_batch imgs;
     clip_image_f32_ptr img_copy(clip_image_f32_init());
@@ -4793,27 +4732,33 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                 }
                 set_input_i32("positions", positions);
 
-                // inspired from resampler of Qwen-VL:
-                //    -> https://huggingface.co/Qwen/Qwen-VL/tree/main
-                //    -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
-                int embed_dim = clip_n_mmproj_embd(ctx);
-
-                // TODO @ngxson : this is very inefficient, can we do this using ggml_sin and ggml_cos?
-                auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
-
-                std::vector<float> pos_embed(embed_dim * pos_w * pos_h);
-                for(int i = 0; i < pos_w * pos_h; ++i){
-                    for(int j = 0; j < embed_dim; ++j){
-                        pos_embed[i * embed_dim + j] = pos_embed_t[i][j];
-                    }
+                // inputs for resampler projector
+                // set the 2D positions (using float for sinusoidal embedding)
+                int n_patches_per_col = image_size_width / patch_size;
+                std::vector<float> pos_data(n_pos);
+                // dimension H
+                for (int i = 0; i < n_pos; i++) {
+                    pos_data[i] = static_cast<float>(i / n_patches_per_col);
                 }
-
-                set_input_f32("pos_embed", pos_embed);
+                set_input_f32("pos_h", pos_data);
+                // dimension W
+                for (int i = 0; i < n_pos; i++) {
+                    pos_data[i] = static_cast<float>(i % n_patches_per_col);
+                }
+                set_input_f32("pos_w", pos_data);
+                // base frequency omega
+                const float base_freq   = 10000.0f;
+                const int   n_embd_proj = clip_n_mmproj_embd(ctx);
+                std::vector<float> omega(n_embd_proj / 4);
+                for (int i = 0; i < n_embd_proj / 4; ++i) {
+                    omega[i] = 1.0f / std::pow(base_freq, static_cast<float>(i) / (n_embd_proj / 4));
+                }
+                set_input_f32("omega", omega);
             } break;
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN3VL:
             {
-                const int merge_ratio = 2;
+                const int merge_ratio = hparams.n_merge;
                 const int pw = image_size_width  / patch_size;
                 const int ph = image_size_height / patch_size;
                 std::vector<int> positions(n_pos * 4);

tools/mtmd/mtmd.cpp

@@ -101,16 +101,17 @@ static clip_flash_attn_type mtmd_get_clip_flash_attn_type(enum llama_flash_attn_
 }
 
 mtmd_context_params mtmd_context_params_default() {
-    mtmd_context_params params;
-    params.use_gpu = true;
-    params.print_timings = true;
-    params.n_threads = 4;
-    params.verbosity = GGML_LOG_LEVEL_INFO;
-    params.image_marker = MTMD_DEFAULT_IMAGE_MARKER;
-    params.media_marker = mtmd_default_marker();
-    params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO;
-    params.image_min_tokens = -1;
-    params.image_max_tokens = -1;
+    mtmd_context_params params {
+        /* use_gpu           */ true,
+        /* print_timings     */ true,
+        /* n_threads         */ 4,
+        /* verbosity         */ GGML_LOG_LEVEL_INFO,
+        /* image_marker      */ MTMD_DEFAULT_IMAGE_MARKER,
+        /* media_marker      */ mtmd_default_marker(),
+        /* flash_attn_type   */ LLAMA_FLASH_ATTN_TYPE_AUTO,
+        /* image_min_tokens  */ -1,
+        /* image_max_tokens  */ -1,
+    };
     return params;
 }
 
@@ -162,7 +163,7 @@ struct mtmd_context {
         print_timings(ctx_params.print_timings),
         n_threads    (ctx_params.n_threads),
         media_marker (ctx_params.media_marker),
-        n_embd_text  (llama_model_n_embd(text_model))
+        n_embd_text  (llama_model_n_embd_inp(text_model))
     {
         if (std::string(ctx_params.image_marker) != MTMD_DEFAULT_IMAGE_MARKER) {
             throw std::runtime_error("custom image_marker is not supported anymore, use media_marker instead");
@@ -172,13 +173,13 @@ struct mtmd_context {
             throw std::runtime_error("media_marker must not be empty");
         }
 
-        clip_context_params ctx_clip_params;
-        ctx_clip_params.use_gpu          = ctx_params.use_gpu;
-        ctx_clip_params.verbosity        = ctx_params.verbosity;
-        ctx_clip_params.flash_attn_type  = mtmd_get_clip_flash_attn_type(ctx_params.flash_attn_type);
-        // custom image token limits
-        ctx_clip_params.image_min_tokens = ctx_params.image_min_tokens;
-        ctx_clip_params.image_max_tokens = ctx_params.image_max_tokens;
+        clip_context_params ctx_clip_params {
+            /* use_gpu           */ ctx_params.use_gpu,
+            /* verbosity         */ ctx_params.verbosity,
+            /* flash_attn_type   */ CLIP_FLASH_ATTN_TYPE_AUTO,
+            /* image_min_tokens  */ ctx_params.image_min_tokens,
+            /* image_max_tokens  */ ctx_params.image_max_tokens,
+        };
 
         auto res = clip_init(mmproj_fname, ctx_clip_params);
         ctx_v = res.ctx_v;

tools/server/README.md

@@ -277,7 +277,7 @@ For more details, please refer to [multimodal documentation](../../docs/multimod
 
 ## Web UI
 
-The project includes a web-based user interface that enables interaction with the model through the `/chat/completions` endpoint.
+The project includes a web-based user interface that enables interaction with the model through the `/v1/chat/completions` endpoint.
 
 The web UI is developed using:
 - `react` framework for frontend development
@@ -512,7 +512,7 @@ These words will not be included in the completion, so make sure to add them to
 
 `timings_per_token`: Include prompt processing and text generation speed information in each response.  Default: `false`
 
-`return_progress`: Include prompt processing progress in `stream` mode. The progress will be contained inside `prompt_progress` with 3 values: `total`, `cache` and `processed`. The overall progress is `processed/total`, while the actual timed progress is `(processed-cache)/(total-cache)`. Default: `false`
+`return_progress`: Include prompt processing progress in `stream` mode. The progress will be contained inside `prompt_progress` with 4 values: `total`, `cache`, `processed`, and `time_ms`. The overall progress is `processed/total`, while the actual timed progress is `(processed-cache)/(total-cache)`. The `time_ms` field contains the elapsed time in milliseconds since prompt processing started. Default: `false`
 
 `post_sampling_probs`: Returns the probabilities of top `n_probs` tokens after applying sampling chain.
 

tools/server/server.cpp

@@ -2400,7 +2400,7 @@ struct server_context {
 
         add_bos_token = llama_vocab_get_add_bos(vocab);
 
-        if (!params_base.speculative.model.path.empty() || !params_base.speculative.model.hf_repo.empty()) {
+        if (params_base.has_speculative()) {
             SRV_INF("loading draft model '%s'\n", params_base.speculative.model.path.c_str());
 
             auto params_dft = params_base;
@@ -2476,7 +2476,7 @@ struct server_context {
                 SRV_WRN("%s\n", "cache_reuse is not supported by multimodal, it will be disabled");
             }
 
-            if (!params_base.speculative.model.path.empty()) {
+            if (params_base.has_speculative()) {
                 SRV_ERR("%s\n", "err: speculative decode is not supported by multimodal");
                 return false;
             }
@@ -2520,6 +2520,7 @@ struct server_context {
             if (model_dft) {
                 slot.batch_spec = llama_batch_init(params_base.speculative.n_max + 1, 0, 1);
 
+                // TODO: rework speculative decoding [TAG_SERVER_SPEC_REWORK]
                 slot.ctx_dft = llama_init_from_model(model_dft, cparams_dft);
                 if (slot.ctx_dft == nullptr) {
                     SRV_ERR("%s", "failed to create draft context\n");
@@ -2822,9 +2823,12 @@ struct server_context {
                 send_error(task, "Failed to parse grammar", ERROR_TYPE_INVALID_REQUEST);
                 return false;
             }
+
+            SLT_INF(slot, "sampler chain: %s\n", common_sampler_print(slot.smpl).c_str());
         }
 
         // initialize draft batch
+        // TODO: rework speculative decoding [TAG_SERVER_SPEC_REWORK]
         if (slot.ctx_dft) {
             llama_batch_free(slot.batch_spec);
 
@@ -3074,7 +3078,7 @@ struct server_context {
             res->progress.total     = slot.task->n_tokens();
             res->progress.cache     = slot.n_prompt_tokens_cache;
             res->progress.processed = slot.prompt.tokens.size();
-            res->progress.time_ms   = (ggml_time_us() - slot.t_start_process_prompt / 1000);
+            res->progress.time_ms   = (ggml_time_us() - slot.t_start_process_prompt) / 1000;
         } else {
             res->content = tkn.text_to_send;
             res->tokens  = { tkn.tok };
@@ -3830,7 +3834,9 @@ struct server_context {
                             // the largest pos_min required for a checkpoint to be useful
                             const auto pos_min_thold = std::max(0, n_past - n_swa);
 
-                            if (n_past > 0 && n_past < slot.prompt.n_tokens()) {
+                            // note: disallow with mtmd contexts for now
+                            //       https://github.com/ggml-org/llama.cpp/issues/17043
+                            if (!mctx && n_past > 0 && n_past < slot.prompt.n_tokens()) {
                                 const auto pos_min = llama_memory_seq_pos_min(llama_get_memory(ctx), slot.id);
                                 if (pos_min == -1) {
                                     SLT_ERR(slot, "n_past = %d, slot.prompt.tokens.size() = %d, seq_id = %d, pos_min = %d\n", n_past, (int) slot.prompt.tokens.size(), slot.id, pos_min);
@@ -4291,6 +4297,8 @@ struct server_context {
             }
 
             // do speculative decoding
+            // TODO: rework to have a single draft llama_context shared across all slots [TAG_SERVER_SPEC_REWORK]
+            //       perform the speculative drafting for all sequences at the same time in a single batch
             for (auto & slot : slots) {
                 if (!slot.is_processing() || !slot.can_speculate()) {
                     continue;
@@ -4445,8 +4453,10 @@ int main(int argc, char ** argv) {
 
     // TODO: should we have a separate n_parallel parameter for the server?
     //       https://github.com/ggml-org/llama.cpp/pull/16736#discussion_r2483763177
-    if (params.n_parallel == 1 && params.kv_unified == false) {
-        LOG_WRN("%s: setting n_parallel = 4 and kv_unified = true\n", __func__);
+    // TODO: this is a common configuration that is suitable for most local use cases
+    //       however, overriding the parameters is a bit confusing - figure out something more intuitive
+    if (params.n_parallel == 1 && params.kv_unified == false && !params.has_speculative()) {
+        LOG_WRN("%s: setting n_parallel = 4 and kv_unified = true (add -kvu to disable this)\n", __func__);
 
         params.n_parallel = 4;
         params.kv_unified = true;

tools/server/tests/unit/test_speculative.py

@@ -77,10 +77,10 @@ def test_different_draft_min_draft_max():
 
 def test_slot_ctx_not_exceeded():
     global server
-    server.n_ctx = 64
+    server.n_ctx = 256
     server.start()
     res = server.make_request("POST", "/completion", data={
-        "prompt": "Hello " * 56,
+        "prompt": "Hello " * 248,
         "temperature": 0.0,
         "top_k": 1,
         "speculative.p_min": 0.0,
@@ -91,19 +91,19 @@ def test_slot_ctx_not_exceeded():
 
 def test_with_ctx_shift():
     global server
-    server.n_ctx = 64
+    server.n_ctx = 256
     server.enable_ctx_shift = True
     server.start()
     res = server.make_request("POST", "/completion", data={
-        "prompt": "Hello " * 56,
+        "prompt": "Hello " * 248,
         "temperature": 0.0,
         "top_k": 1,
-        "n_predict": 64,
+        "n_predict": 256,
         "speculative.p_min": 0.0,
     })
     assert res.status_code == 200
     assert len(res.body["content"]) > 0
-    assert res.body["tokens_predicted"] == 64
+    assert res.body["tokens_predicted"] == 256
     assert res.body["truncated"] == True