vulkan: fuse rms_norm + mul + rope (+ view + set_rows) (#16977)

This change combines the rms_norm+mul and rope+view+set_rows fusions to
allow fusing the whole sequence together. This comes up in Qwen3, Bailing,
and some other models.

.devops/s390x.Dockerfile

@@ -24,8 +24,9 @@ RUN --mount=type=cache,target=/root/.ccache \
         -DCMAKE_C_COMPILER_LAUNCHER=ccache \
         -DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
         -DLLAMA_BUILD_TESTS=OFF \
-        -DGGML_BACKEND_DL=OFF \
         -DGGML_NATIVE=OFF \
+        -DGGML_BACKEND_DL=ON \
+        -DGGML_CPU_ALL_VARIANTS=ON \
         -DGGML_BLAS=ON \
         -DGGML_BLAS_VENDOR=OpenBLAS && \
     cmake --build build --config Release -j $(nproc) && \
@@ -103,6 +104,7 @@ FROM base AS light
 WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
+COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
 COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin
 
 ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]
@@ -116,6 +118,7 @@ ENV LLAMA_ARG_HOST=0.0.0.0
 WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
+COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
 COPY --from=collector /llama.cpp/bin/llama-server /llama.cpp/bin
 
 EXPOSE 8080

.github/labeler.yml

@@ -76,6 +76,10 @@ ggml:
     - changed-files:
         - any-glob-to-any-file:
             - ggml/**
+model:
+    - changed-files:
+        - any-glob-to-any-file:
+            - src/models/**
 nix:
     - changed-files:
         - any-glob-to-any-file:

.github/workflows/build-linux-cross.yml

@@ -4,49 +4,49 @@ on:
   workflow_call:
 
 jobs:
-  ubuntu-24-riscv64-cpu-cross:
-    runs-on: ubuntu-24.04
+  # ubuntu-24-riscv64-cpu-cross:
+  #   runs-on: ubuntu-24.04
 
-    steps:
-      - uses: actions/checkout@v4
-      - name: Setup Riscv
-        run: |
-          sudo dpkg --add-architecture riscv64
+  #   steps:
+  #     - uses: actions/checkout@v4
+  #     - name: Setup Riscv
+  #       run: |
+  #         sudo dpkg --add-architecture riscv64
 
-          # Add arch-specific repositories for non-amd64 architectures
-          cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
-          deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
-          EOF
+  #         # Add arch-specific repositories for non-amd64 architectures
+  #         cat << EOF | sudo tee /etc/apt/sources.list.d/riscv64-ports.list
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+  #         deb [arch=riscv64] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+  #         EOF
 
-          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+  #         sudo apt-get update || true    ;# Prevent failure due to missing URLs.
 
-          sudo apt-get install -y --no-install-recommends \
-                  build-essential \
-                  gcc-14-riscv64-linux-gnu \
-                  g++-14-riscv64-linux-gnu
+  #         sudo apt-get install -y --no-install-recommends \
+  #                 build-essential \
+  #                 gcc-14-riscv64-linux-gnu \
+  #                 g++-14-riscv64-linux-gnu
 
-      - name: Build
-        run: |
-          cmake -B build -DLLAMA_CURL=OFF \
-                         -DCMAKE_BUILD_TYPE=Release \
-                         -DGGML_OPENMP=OFF \
-                         -DLLAMA_BUILD_EXAMPLES=ON \
-                         -DLLAMA_BUILD_TOOLS=ON \
-                         -DLLAMA_BUILD_TESTS=OFF \
-                         -DCMAKE_SYSTEM_NAME=Linux \
-                         -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
-                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
-                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
-                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
-                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+  #     - name: Build
+  #       run: |
+  #         cmake -B build -DLLAMA_CURL=OFF \
+  #                        -DCMAKE_BUILD_TYPE=Release \
+  #                        -DGGML_OPENMP=OFF \
+  #                        -DLLAMA_BUILD_EXAMPLES=ON \
+  #                        -DLLAMA_BUILD_TOOLS=ON \
+  #                        -DLLAMA_BUILD_TESTS=OFF \
+  #                        -DCMAKE_SYSTEM_NAME=Linux \
+  #                        -DCMAKE_SYSTEM_PROCESSOR=riscv64 \
+  #                        -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+  #                        -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+  #                        -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+  #                        -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+  #                        -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
-          cmake --build build --config Release -j $(nproc)
+  #         cmake --build build --config Release -j $(nproc)
 
   # ubuntu-24-riscv64-vulkan-cross:
   #   runs-on: ubuntu-24.04

.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

.github/workflows/docker.yml

@@ -40,7 +40,7 @@ jobs:
           # https://github.com/ggml-org/llama.cpp/issues/11888
           #- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
           - { tag: "cpu",    dockerfile: ".devops/cpu.Dockerfile",    platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-          - { tag: "cuda",   dockerfile: ".devops/cuda.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
+          - { tag: "cuda",   dockerfile: ".devops/cuda.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "musa",   dockerfile: ".devops/musa.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "intel",  dockerfile: ".devops/intel.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
           - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }

.github/workflows/release.yml

@@ -134,8 +134,8 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
-          - build: 's390x-z15' # z15 because our CI runners are on z15
-            os: ubuntu-22.04-s390x
+          - build: 's390x'
+            os: ubuntu-24.04-s390x
           # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
           # - build: 'arm64'
           #   os: ubuntu-22.04-arm

CODEOWNERS

@@ -65,7 +65,7 @@
 /ggml/src/ggml-impl.h                   @ggerganov @slaren
 /ggml/src/ggml-metal/                   @ggerganov
 /ggml/src/ggml-opencl/                  @lhez @max-krasnyansky
-/ggml/src/ggml-hexagon/                 @max-krasnyansky
+/ggml/src/ggml-hexagon/                 @max-krasnyansky @lhez
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
 /ggml/src/ggml-rpc/                     @rgerganov
@@ -89,6 +89,7 @@
 /src/llama-model-loader.*               @slaren
 /src/llama-model.*                      @CISC
 /src/llama-vocab.*                      @CISC
+/src/models/                            @CISC
 /tests/                                 @ggerganov
 /tests/test-backend-ops.cpp             @slaren
 /tests/test-thread-safety.cpp           @slaren

README.md

@@ -17,14 +17,13 @@ LLM inference in C/C++
 
 ## Hot topics
 
-- **[guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)**
-- **[[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)**
+- **[guide : using the new WebUI of llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/16938)**
+- [guide : running gpt-oss with llama.cpp](https://github.com/ggml-org/llama.cpp/discussions/15396)
+- [[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)
 - Support for the `gpt-oss` model with native MXFP4 format has been added | [PR](https://github.com/ggml-org/llama.cpp/pull/15091) | [Collaboration with NVIDIA](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss) | [Comment](https://github.com/ggml-org/llama.cpp/discussions/15095)
-- Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
 - Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim
-- Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123
 - Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669
 - Hugging Face GGUF editor: [discussion](https://github.com/ggml-org/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 

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/chat.cpp

@@ -313,7 +313,6 @@ json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msg
         }
         if (!msg.reasoning_content.empty()) {
             jmsg["reasoning_content"] = msg.reasoning_content;
-            jmsg["thinking"] = msg.reasoning_content; // gpt-oss
         }
         if (!msg.tool_name.empty()) {
             jmsg["name"] = msg.tool_name;
@@ -1810,7 +1809,23 @@ static void common_chat_parse_deepseek_v3_1(common_chat_msg_parser & builder) {
 
 static common_chat_params common_chat_params_init_gpt_oss(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    auto prompt = apply(tmpl, inputs);
+
+    // Copy reasoning to the "thinking" field as expected by the gpt-oss template
+    auto adjusted_messages = json::array();
+    for (const auto & msg : inputs.messages) {
+        auto has_reasoning_content = msg.contains("reasoning_content") && msg.at("reasoning_content").is_string();
+        auto has_tool_calls = msg.contains("tool_calls") && msg.at("tool_calls").is_array();
+
+        if (has_reasoning_content && has_tool_calls) {
+            auto adjusted_message = msg;
+            adjusted_message["thinking"] = msg.at("reasoning_content");
+            adjusted_messages.push_back(adjusted_message);
+        } else {
+            adjusted_messages.push_back(msg);
+        }
+    }
+
+    auto prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
 
     // Check if we need to replace the return token with end token during
     // inference and without generation prompt. For more details see:

common/common.h

@@ -406,6 +406,8 @@ struct common_params {
     bool mmproj_use_gpu = true;     // use GPU for multimodal model
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
+    int image_min_tokens = -1;
+    int image_max_tokens = -1;
 
     // finetune
     struct lr_opt lr;
@@ -505,6 +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

@@ -1054,6 +1054,9 @@ class TextModel(ModelBase):
         if chkhsh == "53e325976a6e142379c19b09afcae354f2f496f147afa8f9e189a33fe4e3024e":
             # ref: https://huggingface.co/ibm-granite/granite-docling-258M
             res = "granite-docling"
+        if chkhsh == "f4f37b6c8eb9ea29b3eac6bb8c8487c5ab7885f8d8022e67edc1c68ce8403e95":
+            # ref: https://huggingface.co/MiniMaxAI/MiniMax-M2
+            res = "minimax-m2"
 
         if res is None:
             logger.warning("\n")
@@ -1528,7 +1531,7 @@ class MmprojModel(ModelBase):
             self.gguf_writer.add_vision_embedding_length(self.find_vparam(["hidden_size"]))
             self.gguf_writer.add_vision_feed_forward_length(self.find_vparam(["intermediate_size"]))
             self.gguf_writer.add_vision_block_count(self.find_vparam(self.n_block_keys))
-            self.gguf_writer.add_vision_head_count(self.find_vparam(["num_attention_heads"]))
+            self.gguf_writer.add_vision_head_count(self.find_vparam(["num_attention_heads", "num_heads"]))
 
             # preprocessor config
             image_mean = _MISTRAL_COMMON_DATASET_MEAN if self.is_mistral_format else self.preprocessor_config["image_mean"]
@@ -3852,7 +3855,43 @@ class Qwen2MoeModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         # process the experts separately
         name = name.replace("language_model.", "") # InternVL
-        if name.startswith("mlp") or name.startswith("vision_model") or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector"):
+
+        # handle aggregated expert tensors
+        # GGUF stores dimensions reversed from PyTorch, so:
+        # PyTorch (A,B,C) -> GGUF writes [C,B,A] -> GGML reads ne={C,B,A}
+        # Input shapes from HF: (n_expert, n_ff_exp, n_embd) or (n_expert, n_embd, n_ff_exp)
+        # Expected GGML ne: {n_embd, n_ff_exp, n_expert} for gate/up, {n_ff_exp, n_embd, n_expert} for down
+        if name.endswith("mlp.experts.down_proj") or name.endswith("mlp.experts.down_proj.weight"):
+            mapped = f"{name}.weight" if not name.endswith(".weight") else name
+            # Input: (n_expert=128, n_ff_exp=768, n_embd=2048)
+            # Want GGML ne: {n_ff_exp, n_embd, n_expert} = {768, 2048, 128}
+            # Need PyTorch: (128, 2048, 768) [reversed of GGML]
+            # So: permute(0, 2, 1): (128, 768, 2048) -> (128, 2048, 768)
+            permuted = data_torch.permute(0, 2, 1).contiguous()
+            return [(self.map_tensor_name(mapped), permuted)]
+
+        if name.endswith("mlp.experts.gate_up_proj") or name.endswith("mlp.experts.gate_up_proj.weight"):
+            if data_torch.ndim < 3 or data_torch.shape[-1] % 2 != 0:
+                raise ValueError(f"Unexpected gate_up_proj shape for {name}: {tuple(data_torch.shape)}")
+            split_dim = data_torch.shape[-1] // 2
+            gate = data_torch[..., :split_dim].contiguous()
+            up = data_torch[..., split_dim:].contiguous()
+            # Input gate/up: (n_expert=128, n_embd=2048, n_ff_exp=768)
+            # Want GGML ne: {n_embd, n_ff_exp, n_expert} = {2048, 768, 128}
+            # Need PyTorch: (128, 768, 2048) [reversed of GGML]
+            # So: permute(0, 2, 1): (128, 2048, 768) -> (128, 768, 2048)
+            base_name = name.removesuffix(".weight")
+            base = base_name.rsplit('.', 1)[0]
+            mapped_gate = f"{base}.gate_proj.weight"
+            mapped_up = f"{base}.up_proj.weight"
+            perm_gate = gate.permute(0, 2, 1).contiguous()
+            perm_up = up.permute(0, 2, 1).contiguous()
+            return [
+                (self.map_tensor_name(mapped_gate), perm_gate),
+                (self.map_tensor_name(mapped_up), perm_up),
+            ]
+
+        if name.startswith("mlp") or name.startswith("vision_model") or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector") or name.startswith("model.visual"):
             # skip visual tensors
             return []
         if name.find("experts") != -1:
@@ -4004,6 +4043,187 @@ class Qwen3MoeModel(Qwen2MoeModel):
         super().set_vocab()
 
 
+@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration")
+class Qwen3VLVisionModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        # Compute image_size if not present
+        if "image_size" not in self.hparams_vision:
+            # For Qwen3VL/Qwen3VLMoe, compute from num_position_embeddings
+            num_pos = self.hparams_vision.get("num_position_embeddings", 2304)
+            patch_size = self.hparams_vision.get("patch_size", 16)
+            # num_position_embeddings = (image_size / patch_size) ** 2
+            # So image_size = sqrt(num_position_embeddings) * patch_size
+            image_size = int(num_pos**0.5 * patch_size)
+            self.hparams_vision["image_size"] = image_size
+
+        # Rename config values for compatibility
+        self.hparams_vision["num_attention_heads"] = self.hparams_vision.get("num_heads")
+        self.hparams_vision["num_hidden_layers"] = self.hparams_vision.get("depth")
+
+        self.is_deepstack_layers = [False] * int(self.hparams_vision["num_hidden_layers"] or 0)
+        for idx in self.hparams_vision.get("deepstack_visual_indexes", []):
+            self.is_deepstack_layers[idx] = True
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.QWEN3VL)
+        self.gguf_writer.add_vision_use_gelu(True)
+
+        if self.hparams_vision is not None:
+            merge_size = self.hparams_vision.get("spatial_merge_size")
+            if merge_size is not None:
+                self.gguf_writer.add_vision_spatial_merge_size(int(merge_size))
+
+        # Use text config's rms_norm_eps for vision attention layernorm eps
+        rms_norm_eps = self.global_config.get("text_config", {}).get("rms_norm_eps", 1e-6)
+        self.gguf_writer.add_vision_attention_layernorm_eps(rms_norm_eps)
+
+        if self.is_deepstack_layers:
+            self.gguf_writer.add_vision_is_deepstack_layers(self.is_deepstack_layers)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        assert self.hparams_vision is not None
+        # Skip text model tensors - they go in the text model file
+        if name.startswith("model.language_model.") or name.startswith("lm_head."):
+            return []
+
+        if name.startswith("model.visual."):
+            name = name.replace("model.visual.", "visual.", 1)
+
+        if name.startswith("visual.deepstack_merger_list."):
+            prefix, rest = name.split(".", maxsplit=3)[2:]
+            # prefix is the layer index, convert to absolute clip layer index!
+            idx = self.hparams_vision.get("deepstack_visual_indexes", [])[int(prefix)]
+            target = rest
+
+            tensor_type: gguf.MODEL_TENSOR
+            if target.startswith("norm."):
+                tensor_type = gguf.MODEL_TENSOR.V_DS_NORM
+                suffix = target.split(".", 1)[1]
+            elif target.startswith("linear_fc1."):
+                tensor_type = gguf.MODEL_TENSOR.V_DS_FC1
+                suffix = target.split(".", 1)[1]
+            elif target.startswith("linear_fc2."):
+                tensor_type = gguf.MODEL_TENSOR.V_DS_FC2
+                suffix = target.split(".", 1)[1]
+            else:
+                raise ValueError(f"Unexpected deepstack tensor: {name}")
+
+            new_name = self.format_tensor_name(tensor_type, idx, suffix=f".{suffix}")
+            return [(new_name, data_torch)]
+
+        if name.startswith("visual.merger."):
+            suffix = name.split(".", 2)[2]
+            if suffix.startswith("linear_fc"):
+                fc_idx_str, tail = suffix.split(".", 1)
+                fc_num = int(fc_idx_str.replace("linear_fc", ""))
+                # Qwen3VL has linear_fc1 and linear_fc2
+                # Map to indices 0 and 2 (matching Qwen2VL which uses indices 0 and 2)
+                if fc_num == 1:
+                    fc_idx = 0
+                elif fc_num == 2:
+                    fc_idx = 2
+                else:
+                    raise ValueError(f"unexpected fc index {fc_num} in {name}")
+                new_name = self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, fc_idx, suffix=f".{tail}")
+            elif suffix.startswith("norm."):
+                new_name = self.format_tensor_name(gguf.MODEL_TENSOR.V_POST_NORM, suffix=f".{suffix.split('.', 1)[1]}")
+            else:
+                raise ValueError(f"Unexpected merger tensor: {name}")
+            return [(new_name, data_torch)]
+
+        if name == "visual.patch_embed.proj.weight":
+            # split Conv3D into Conv2Ds along temporal dimension
+            c1, c2, kt, _, _ = data_torch.shape
+            del c1, c2
+            if kt != 2:
+                raise ValueError("Current implementation only supports temporal_patch_size of 2")
+            return [
+                (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight", data_torch[:, :, 0, ...]),
+                (gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".weight.1", data_torch[:, :, 1, ...]),
+            ]
+
+        if name == "visual.patch_embed.proj.bias":
+            # Include the bias - it's used by the C++ code
+            return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_ENC_EMBD_PATCH] + ".bias", data_torch)]
+
+        if name.startswith("visual."):
+            return [(self.map_tensor_name(name), data_torch)]
+
+        # Fall back to parent class for other tensors
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Qwen3VLForConditionalGeneration")
+class Qwen3VLTextModel(Qwen3Model):
+    model_arch = gguf.MODEL_ARCH.QWEN3VL
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # Handle MRoPE (Multi-axis Rotary Position Embedding) for Qwen3-VL
+        text_config = self.hparams.get("text_config", {})
+        # rope_scaling is deprecated in V5, use rope_parameters instead
+        rope_scaling = text_config.get("rope_scaling") or text_config.get("rope_parameters") or {}
+
+        if rope_scaling.get("mrope_section"):
+            # mrope_section contains [time, height, width] dimensions
+            mrope_section = rope_scaling["mrope_section"]
+            # Pad to 4 dimensions [time, height, width, extra]
+            while len(mrope_section) < 4:
+                mrope_section.append(0)
+            self.gguf_writer.add_rope_dimension_sections(mrope_section[:4])
+
+            logger.info(f"MRoPE sections: {mrope_section[:4]}")
+
+        vision_config = self.hparams.get("vision_config", {})
+        deepstack_layer_num = len(vision_config.get("deepstack_visual_indexes", []))
+        self.gguf_writer.add_num_deepstack_layers(deepstack_layer_num)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision tensors - they go in the mmproj file
+        if name.startswith("model.visual."):
+            return []
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Qwen3VLMoeForConditionalGeneration")
+class Qwen3VLMoeTextModel(Qwen3MoeModel):
+    model_arch = gguf.MODEL_ARCH.QWEN3VLMOE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # Handle MRoPE (Multi-axis Rotary Position Embedding) for Qwen3-VL
+        text_config = self.hparams.get("text_config", {})
+        # rope_scaling is deprecated in V5, use rope_parameters instead
+        rope_scaling = text_config.get("rope_scaling") or text_config.get("rope_parameters") or {}
+
+        if rope_scaling.get("mrope_section"):
+            # mrope_section contains [time, height, width] dimensions
+            mrope_section = rope_scaling["mrope_section"]
+            # Pad to 4 dimensions [time, height, width, extra]
+            while len(mrope_section) < 4:
+                mrope_section.append(0)
+            self.gguf_writer.add_rope_dimension_sections(mrope_section[:4])
+
+            logger.info(f"MRoPE sections: {mrope_section[:4]}")
+
+        vision_config = self.hparams.get("vision_config", {})
+        deepstack_layer_num = len(vision_config.get("deepstack_visual_indexes", []))
+        self.gguf_writer.add_num_deepstack_layers(deepstack_layer_num)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision tensors - they go in the mmproj file
+        if name.startswith("model.visual."):
+            return []
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("GPT2LMHeadModel")
 class GPT2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GPT2
@@ -6909,6 +7129,100 @@ class DeepseekV2Model(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("MiniMaxM2ForCausalLM")
+class MiniMaxM2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.MINIMAXM2
+    _experts_cache: dict[int, dict[str, Tensor]] = {}
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["num_experts"] = self.hparams["num_local_experts"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if self.hparams["scoring_func"] == "sigmoid":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        elif self.hparams["scoring_func"] == "softmax":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
+        else:
+            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
+
+        self.gguf_writer.add_expert_feed_forward_length(self.find_hparam(["intermediate_size"]))
+        self.gguf_writer.add_rope_dimension_count(self.find_hparam(["rotary_dim"]))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        # merge expert weights
+        if 'experts' in name:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            expert_cache = self._experts_cache.setdefault(bid, {})
+            expert_cache[name] = data_torch
+            expert_weights = ["w1", "w2", "w3"]
+
+            # not enough expert weights to merge
+            if len(expert_cache) < n_experts * len(expert_weights):
+                return []
+
+            tensors: list[tuple[str, Tensor]] = []
+            for w_name in expert_weights:
+                datas: list[Tensor] = []
+
+                for xid in range(n_experts):
+                    ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{w_name}.weight"
+                    datas.append(expert_cache[ename])
+                    del expert_cache[ename]
+
+                data_torch = torch.stack(datas, dim=0)
+                merged_name = f"model.layers.{bid}.block_sparse_moe.experts.{w_name}.weight"
+                new_name = self.map_tensor_name(merged_name)
+                tensors.append((new_name, data_torch))
+
+            del self._experts_cache[bid]
+            return tensors
+
+        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
@@ -9493,6 +9807,144 @@ class KimiVLModel(MmprojModel):
 
         return [] # skip other tensors
 
+
+@ModelBase.register("CogVLMForCausalLM")
+class CogVLMVisionModel(MmprojModel):
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-6))
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.COGVLM)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        if not name.startswith("model.vision."):
+            return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@ModelBase.register("CogVLMForCausalLM")
+class CogVLMModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.COGVLM
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # block vision tensors
+        if name.startswith("model.vision."):
+            return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@ModelBase.register("JanusForConditionalGeneration")
+class JanusProModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.LLAMA  # reuse Llama arch
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision, aligner, and generation tensors
+        skip_prefixes = (
+            'model.vision_model.',
+            'model.aligner.',
+            'model.vqmodel.',
+            'model.generation_embeddings.',
+            'model.generation_aligner.',
+            'model.generation_head.',
+        )
+        if name.startswith(skip_prefixes):
+            return []
+
+        if name.startswith('model.language_model.'):
+            name = name.replace('model.language_model.', 'model.')
+        elif name.startswith('language_model.'):
+            name = name.replace('language_model.', '')
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("JanusForConditionalGeneration")
+class JanusProVisionModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        if "intermediate_size" not in self.hparams_vision:
+            mlp_ratio = self.hparams_vision.get("mlp_ratio")
+            hidden_size = self.hparams_vision.get("hidden_size")
+            if mlp_ratio is not None and hidden_size is not None:
+                self.hparams_vision["intermediate_size"] = int(round(hidden_size * mlp_ratio))
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.JANUS_PRO)
+
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))
+
+        hidden_act = str(self.hparams_vision.get("hidden_act", "")).lower()
+        if hidden_act == "gelu":
+            self.gguf_writer.add_vision_use_gelu(True)
+        elif hidden_act == "silu":
+            self.gguf_writer.add_vision_use_silu(True)
+
+    def _map_aligner_tensor(self, data_torch: Tensor, name: str) -> Iterable[tuple[str, Tensor]]:
+        """Map aligner tensors to projector format"""
+        suffix = ".bias" if name.endswith(".bias") else ".weight"
+
+        if name.startswith("model.aligner."):
+            local_name = name[len("model.aligner."):]
+        elif name.startswith("aligner."):
+            local_name = name[len("aligner."):]
+        else:
+            raise ValueError(f"Unsupported Janus aligner prefix: {name}")
+
+        if local_name.startswith("fc1."):
+            mm_index = 0
+        elif local_name.startswith("hidden_layers."):
+            parts = local_name.split(".", 2)
+            if len(parts) < 3:
+                raise ValueError(f"Unexpected Janus aligner tensor name: {name}")
+            mm_index = int(parts[1]) + 1
+        else:
+            raise ValueError(f"Unsupported Janus aligner tensor: {name}")
+
+        tensor_name = self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, mm_index, suffix=suffix)
+        return [(tensor_name, data_torch)]
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # Skip language model tensors as they will be handled by `JanusProModel`
+        if name.startswith(('model.language_model.', 'language_model.')):
+            return []
+
+        # Skip generation-related components
+        skip_generation_prefixes = (
+            'model.vqmodel.',
+            'vqmodel.',
+            'model.generation_embeddings.',
+            'generation_embeddings.',
+            'model.generation_aligner.',
+            'generation_aligner.',
+            'model.generation_head.',
+            'generation_head.',
+        )
+        if name.startswith(skip_generation_prefixes):
+            return []
+
+        # Handle aligner tensors
+        if name.startswith(('model.aligner.', 'aligner.')):
+            return list(self._map_aligner_tensor(data_torch, name))
+
+        # Handle vision tensors
+        if name.startswith(('model.vision_model.', 'vision_model.')):
+            return [(self.map_tensor_name(name), data_torch)]
+
+        return []
+
+
 ###### CONVERSION LOGIC ######
 
 

convert_hf_to_gguf_update.py

@@ -141,6 +141,7 @@ models = [
     {"name": "mellum",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/JetBrains/Mellum-4b-base", },
     {"name": "bailingmoe2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-mini-base-2.0", },
     {"name": "granite-docling",  "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ibm-granite/granite-docling-258M", },
+    {"name": "minimax-m2",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/MiniMaxAI/MiniMax-M2", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -435,7 +436,7 @@ for model in models:
             tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
         else:
             tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-    except OSError as e:
+    except (OSError, TypeError) as e:
         logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
         continue  # Skip this model and continue with the next one in the loop
 

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/docker.md

@@ -7,9 +7,9 @@
 ## Images
 We have three Docker images available for this project:
 
-1. `ghcr.io/ggml-org/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
-2. `ghcr.io/ggml-org/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggml-org/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
+1. `ghcr.io/ggml-org/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
+2. `ghcr.io/ggml-org/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
+3. `ghcr.io/ggml-org/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`, `linux/s390x`)
 
 Additionally, there the following images, similar to the above:
 

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 | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
@@ -79,12 +79,12 @@ Legend:
 |                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
-|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
 |                             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

@@ -5637,25 +5637,25 @@
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000000,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000000","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000000,inplace=0","support","1","yes","SYCL"
-"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000000","support","0","no","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000000","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000001","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000001,inplace=0","support","1","yes","SYCL"
-"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000001","support","0","no","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000001","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000100","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000100,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000100","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.000100,inplace=0","support","1","yes","SYCL"
-"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000100","support","0","no","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.000100","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.100000","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.100000,inplace=0","support","1","yes","SYCL"
 "SYCL0","NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.100000","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=1,eps=0.100000,inplace=0","support","1","yes","SYCL"
-"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.100000","support","0","no","SYCL"
+"SYCL0","RMS_NORM_BACK","type=f32,ne=[64,5,4,3],eps=0.100000","support","1","yes","SYCL"
 "SYCL0","L2_NORM","type=f32,ne=[64,5,4,3]","support","1","yes","SYCL"
 "SYCL0","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=1","support","1","yes","SYCL"
 "SYCL0","RMS_NORM_MUL_ADD","type=f32,ne=[64,5,4,3],eps=0.000000,broadcast=0,multi_add=0","support","1","yes","SYCL"
@@ -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);
         }
     }
 

examples/model-conversion/scripts/causal/run-org-model.py

@@ -138,6 +138,9 @@ if model_path is None:
         "Model path must be specified either via --model-path argument or MODEL_PATH environment variable"
     )
 
+
+print("Loading model and tokenizer using AutoTokenizer:", model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
 config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
 
 print("Model type:       ", config.model_type)
@@ -147,10 +150,6 @@ print("Number of layers: ", config.num_hidden_layers)
 print("BOS token id:     ", config.bos_token_id)
 print("EOS token id:     ", config.eos_token_id)
 
-print("Loading model and tokenizer using AutoTokenizer:", model_path)
-tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
-config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
-
 if unreleased_model_name:
     model_name_lower = unreleased_model_name.lower()
     unreleased_module_path = (
@@ -171,7 +170,7 @@ if unreleased_model_name:
         exit(1)
 else:
     model = AutoModelForCausalLM.from_pretrained(
-        model_path, device_map="auto", offload_folder="offload", trust_remote_code=True
+        model_path, device_map="auto", offload_folder="offload", trust_remote_code=True, config=config
     )
 
 for name, module in model.named_modules():

ggml/include/ggml.h

@@ -242,6 +242,7 @@
 #define GGML_ROPE_TYPE_NEOX   2
 #define GGML_ROPE_TYPE_MROPE  8
 #define GGML_ROPE_TYPE_VISION 24
+#define GGML_ROPE_TYPE_IMROPE 40 // binary: 101000
 
 #define GGML_MROPE_SECTIONS   4
 
@@ -2107,6 +2108,7 @@ extern "C" {
     enum ggml_scale_mode {
         GGML_SCALE_MODE_NEAREST  = 0,
         GGML_SCALE_MODE_BILINEAR = 1,
+        GGML_SCALE_MODE_BICUBIC  = 2,
 
         GGML_SCALE_MODE_COUNT
     };

ggml/src/CMakeLists.txt

@@ -308,6 +308,10 @@ function(ggml_add_cpu_backend_variant tag_name)
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
+        foreach (feat VXE2 NNPA)
+            set(GGML_INTERNAL_${feat} OFF)
+        endforeach()
+
         foreach (feat ${ARGN})
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
@@ -377,9 +381,8 @@ if (GGML_CPU_ALL_VARIANTS)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-            ggml_add_cpu_backend_variant(s390x_z15  Z15 VXE)
-            # ggml_add_cpu_backend_variant(s390x_z16  Z16 VXE)
-            # ggml_add_cpu_backend_variant(s390x_z17  Z17 VXE)
+            ggml_add_cpu_backend_variant(z15    Z15 VXE2)
+            ggml_add_cpu_backend_variant(z16    Z16 VXE2 NNPA)
         else()
             message(FATAL_ERROR "Unsupported s390x target OS: ${CMAKE_SYSTEM_NAME}")
         endif()

ggml/src/ggml-cpu/CMakeLists.txt

@@ -504,11 +504,18 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             endforeach()
         endif()
 
-        if (GGML_VXE OR GGML_INTERNAL_VXE)
-            message(STATUS "VX/VXE/VXE2 enabled")
+        if (GGML_VXE OR GGML_INTERNAL_VXE2)
+            message(STATUS "VXE2 enabled")
             list(APPEND ARCH_FLAGS -mvx -mzvector)
-            list(APPEND ARCH_DEFINITIONS GGML_VXE)
+            list(APPEND ARCH_DEFINITIONS GGML_USE_VXE2)
         endif()
+
+        if (GGML_INTERNAL_NNPA)
+            message(STATUS "NNPA enabled")
+            list(APPEND ARCH_DEFINITIONS GGML_USE_NNPA)
+        endif()
+
+        ggml_add_cpu_backend_features(${GGML_CPU_NAME} s390 ${ARCH_DEFINITIONS})
     elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
         message(STATUS "Wasm detected")
         list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)

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

@@ -700,7 +700,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     for (; ib + 1 < nb; ib += 2) {
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
+        const float ft0 = GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d);
+        const __m128 d_0_1 = (__m128)(v4f32){ft0, ft0, ft0, ft0};
 
         const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
 
@@ -714,11 +715,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         bx_1 = __lsx_vsub_b(bx_1, off);
         const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
 
-        //_mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
-        //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
-
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
+        const float ft1 = GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d);
+        const __m128 d_2_3 = (__m128)(v4f32){ft1, ft1, ft1, ft1};
 
         const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
 

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/arch/s390/cpu-feats.cpp

@@ -0,0 +1,50 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__s390x__)
+#include <sys/auxv.h>
+
+// find hwcap bits in asm/elf.h
+#ifndef HWCAP_VXRS_EXT2
+#define HWCAP_VXRS_EXT2 (1 << 15)
+#endif
+
+#ifndef HWCAP_NNPA
+#define HWCAP_NNPA (1 << 20)
+#endif
+
+struct s390x_features {
+    bool has_vxe2 = false;
+    bool has_nnpa = false;
+
+    s390x_features() {
+        uint32_t hwcap = getauxval(AT_HWCAP);
+        // NOTE: use hwcap2 with DFLT for z17 and later
+        // uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+        has_vxe2 = !!(hwcap & HWCAP_VXRS_EXT2);
+        has_nnpa = !!(hwcap & HWCAP_NNPA);
+    }
+};
+
+static int ggml_backend_cpu_s390x_score() {
+    int score = 1;
+    s390x_features sf;
+
+// IBM z15 / LinuxONE 3
+#ifdef GGML_USE_VXE2
+    if (!sf.has_vxe2) { return 0; }
+    score += 1 << 1;
+#endif
+
+// IBM z16 / LinuxONE 4 and z17 / LinuxONE 5
+#ifdef GGML_USE_NNPA
+    if (!sf.has_nnpa) { return 0; }
+    score += 1 << 2;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_s390x_score)
+
+#endif  // __s390x__

ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -500,13 +500,15 @@ inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
 
 #endif
 
-#if defined(__loongarch_asx)
+#if defined(__loongarch_sx)
 /* float type data load instructions */
 static __m128 __lsx_vreplfr2vr_s(const float val) {
     v4f32 res = {val, val, val, val};
     return (__m128)res;
 }
+#endif
 
+#if defined(__loongarch_asx)
 static __m256 __lasx_xvreplfr2vr_s(const float val) {
     v8f32 res = {val, val, val, val, val, val, val, val};
     return (__m256)res;

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

@@ -1613,13 +1613,8 @@ static void ggml_compute_forward_mul_mat_id(
             chunk_size = 64;
         }
 
-#if defined(__aarch64__)
-        // disable for ARM
-        const bool disable_chunking = true;
-#else
         // disable for NUMA
         const bool disable_chunking = ggml_is_numa();
-#endif // defined(__aarch64__)
 
         int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size;
         int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;

ggml/src/ggml-cpu/ops.cpp

@@ -5474,7 +5474,7 @@ static void ggml_rope_cache_init(
 }
 
 static void ggml_mrope_cache_init(
-     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool indep_sects,
+     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool is_imrope, bool indep_sects,
      float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
      float * cache, float sin_sign, float theta_scale) {
     // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
@@ -5509,14 +5509,26 @@ static void ggml_mrope_cache_init(
         }
 
         float theta = theta_t;
-        if (sector >= sections[0] && sector < sec_w) {
-            theta = theta_h;
-        }
-        else if (sector >= sec_w && sector < sec_w + sections[2]) {
-            theta = theta_w;
-        }
-        else if (sector >= sec_w + sections[2]) {
-            theta = theta_e;
+        if (is_imrope) { // qwen3vl apply interleaved mrope
+            if (sector % 3 == 1 && sector < 3 * sections[1]) {
+                theta = theta_h;
+            } else if (sector % 3 == 2 && sector < 3 * sections[2]) {
+                theta = theta_w;
+            } else if (sector % 3 == 0 && sector < 3 * sections[0]) {
+                theta = theta_t;
+            } else {
+                theta = theta_e;
+            }
+        } else {
+            if (sector >= sections[0] && sector < sec_w) {
+                theta = theta_h;
+            }
+            else if (sector >= sec_w && sector < sec_w + sections[2]) {
+                theta = theta_w;
+            }
+            else if (sector >= sec_w + sections[2]) {
+                theta = theta_e;
+            }
         }
 
         rope_yarn(
@@ -5589,6 +5601,7 @@ static void ggml_compute_forward_rope_f32(
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, multimodal rotary position embedding
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -5627,7 +5640,7 @@ static void ggml_compute_forward_rope_f32(
                 const int64_t p_w = pos[i2 + ne2 * 2];
                 const int64_t p_e = pos[i2 + ne2 * 3];
                 ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
                     freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
             }
 
@@ -5775,6 +5788,7 @@ static void ggml_compute_forward_rope_f16(
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -5813,7 +5827,7 @@ static void ggml_compute_forward_rope_f16(
                 const int64_t p_w = pos[i2 + ne2 * 2];
                 const int64_t p_e = pos[i2 + ne2 * 3];
                 ggml_mrope_cache_init(
-                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    p_t, p_h, p_w, p_e, sections, is_imrope, is_vision,
                     freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
             }
 
@@ -7070,7 +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
@@ -7493,10 +7511,17 @@ static void ggml_compute_forward_upscale_f32(
     float sf1 = (float)ne1/src0->ne[1];
     float sf2 = (float)ne2/src0->ne[2];
     float sf3 = (float)ne3/src0->ne[3];
+    float pixel_offset = 0.5f;
 
     const int32_t mode_flags = ggml_get_op_params_i32(dst, 0);
     const ggml_scale_mode mode = (ggml_scale_mode) (mode_flags & 0xFF);
 
+    if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+        pixel_offset = 0.0f;
+        sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
+        sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
+    }
+
     if (mode == GGML_SCALE_MODE_NEAREST) {
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
@@ -7516,13 +7541,6 @@ static void ggml_compute_forward_upscale_f32(
             }
         }
     } else if (mode == GGML_SCALE_MODE_BILINEAR) {
-        float pixel_offset = 0.5f;
-        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
-            pixel_offset = 0.0f;
-            sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
-            sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
-        }
-
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
             for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
@@ -7557,6 +7575,51 @@ static void ggml_compute_forward_upscale_f32(
 
                         const float val = a*(1 - dx)*(1 - dy) + b*dx*(1 - dy) + c*(1 - dx)*dy + d*dx*dy;
 
+                        float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+                        *y_dst = val;
+                    }
+                }
+            }
+        }
+    } else if (mode == GGML_SCALE_MODE_BICUBIC) {
+        // https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
+        const float a = -0.75f; // use alpha = -0.75 (same as PyTorch)
+        auto weight1 = [a](float x) { return ((a + 2) * x - (a + 3)) * x * x + 1; };
+        auto weight2 = [a](float x) { return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a; };
+        auto bicubic = [=](float p0, float p1, float p2, float p3, float x) {
+            const float w0 = weight2(x + 1);
+            const float w1 = weight1(x + 0);
+            const float w2 = weight1(1 - x);
+            const float w3 = weight2(2 - x);
+            return p0*w0 + p1*w1 + p2*w2 + p3*w3;
+        };
+
+        for (int64_t i3 = 0; i3 < ne3; i3++) {
+            const int64_t i03 = i3 / sf3;
+            for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+                const int64_t i02 = i2 / sf2;
+                for (int64_t i1 = 0; i1 < ne1; i1++) {
+                    const float y = ((float)i1 + pixel_offset) / sf1 - pixel_offset;
+                    const int64_t y0 = (int64_t)floorf(y);
+                    const float dy = y - (float)y0;
+
+                    for (int64_t i0 = 0; i0 < ne0; i0++) {
+                        const float x = ((float)i0 + pixel_offset) / sf0 - pixel_offset;
+                        const int64_t x0 = (int64_t)floorf(x);
+                        const float dx = x - (float)x0;
+
+                        auto p = [=](int64_t x_off, int64_t y_off) -> float {
+                            int64_t i00 = std::max(int64_t(0), std::min(x0 + x_off, ne00 - 1));
+                            int64_t i01 = std::max(int64_t(0), std::min(y0 + y_off, ne01 - 1));
+                            return *(const float *)((const char *)src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                        };
+
+                        const float val = bicubic(
+                            bicubic(p(-1,-1), p(0,-1), p(1,-1), p(2,-1), dx),
+                            bicubic(p(-1, 0), p(0, 0), p(1, 0), p(2, 0), dx),
+                            bicubic(p(-1, 1), p(0, 1), p(1, 1), p(2, 1), dx),
+                            bicubic(p(-1, 2), p(0, 2), p(1, 2), p(2, 2), dx), dy);
+
                         float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
                         *y_dst = val;
                     }
@@ -7909,10 +7972,10 @@ void ggml_compute_forward_argsort(
 
 // ggml_compute_forward_flash_attn_ext
 
-static void ggml_compute_forward_flash_attn_ext_f16(
+static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
         const ggml_compute_params * params,
-        ggml_tensor * dst) {
-
+        ggml_tensor * dst,
+        int ir0, int ir1) {
     const ggml_tensor * q     = dst->src[0];
     const ggml_tensor * k     = dst->src[1];
     const ggml_tensor * v     = dst->src[2];
@@ -7928,9 +7991,6 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
     GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
-    const int ith = params->ith;
-    const int nth = params->nth;
-
     const int64_t DK = nek0;
     const int64_t DV = nev0;
     const int64_t N  = neq1;
@@ -7964,16 +8024,6 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 
     // parallelize by q rows using ggml_vec_dot_f32
 
-    // total rows in q
-    const int nr = neq1*neq2*neq3;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
     float scale         = 1.0f;
     float max_bias      = 0.0f;
     float logit_softcap = 0.0f;
@@ -8000,6 +8050,8 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     GGML_ASSERT((                            q_to_vec_dot) && "fattn: unsupported K-type");
     GGML_ASSERT((v->type == GGML_TYPE_F32 || v_to_float  ) && "fattn: unsupported V-type");
 
+    int ith = params->ith;
+
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
         // q indices
@@ -8147,6 +8199,91 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     }
 }
 
+static void ggml_compute_forward_flash_attn_ext_f16(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * q     = dst->src[0];
+    const ggml_tensor * k     = dst->src[1];
+    const ggml_tensor * v     = dst->src[2];
+
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
+
+    const int64_t DK = nek0;
+    const int64_t DV = nev0;
+    const int64_t N  = neq1;
+
+    GGML_ASSERT(ne0 == DV);
+    GGML_ASSERT(ne2 == N);
+
+    // input tensor rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
+
+    GGML_ASSERT(neq0 == DK);
+    GGML_ASSERT(nek0 == DK);
+    GGML_ASSERT(nev0 == DV);
+
+    GGML_ASSERT(neq1 == N);
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // parallelize by q rows using ggml_vec_dot_f32
+
+    // total rows in q
+    const int64_t nr = neq1*neq2*neq3;
+
+    // rows per thread
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    // disable for NUMA
+    const bool disable_chunking = ggml_is_numa();
+
+    // 4x chunks per thread
+    int nth_scaled = nth * 4;
+    int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+    int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+    if (nth == 1 || nchunk < nth || disable_chunking) {
+        nchunk = nth;
+    }
+
+    if (ith == 0) {
+        // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+        ggml_threadpool_chunk_set(params->threadpool, nth);
+    }
+
+    ggml_barrier(params->threadpool);
+
+    // The number of elements in each chunk
+    const int64_t dr = (nr + nchunk - 1) / nchunk;
+
+    // The first chunk comes from our thread_id, the rest will get auto-assigned.
+    int current_chunk = ith;
+
+    while (current_chunk < nchunk) {
+        const int64_t ir0 = dr * current_chunk;
+        const int64_t ir1 = MIN(ir0 + dr, nr);
+
+        ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+
+        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+    }
+}
+
 void ggml_compute_forward_flash_attn_ext(
         const ggml_compute_params * params,
         ggml_tensor * dst) {

ggml/src/ggml-cpu/repack.cpp

@@ -1600,6 +1600,32 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         return false;
     }
 
+    void forward_mul_mat_one_chunk(ggml_compute_params * params, ggml_tensor * op, int64_t src0_start, int64_t src0_end) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const void * src1_wdata      = params->wdata;
+        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
+
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (ne11 > 3) {
+            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                    (float *) ((char *) dst->data) + src0_start, ne01,
+                    (const char *) src0->data + src0_start * nb01,
+                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
+        }
+        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
+            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
+                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
+                    (const char *) src0->data + src0_start * nb01,
+                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
+                    src0_end - src0_start);
+        }
+    }
+
     void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
         const ggml_tensor * src0 = op->src[0];
         const ggml_tensor * src1 = op->src[1];
@@ -1643,31 +1669,62 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10);
         }
 
+        // disable for NUMA
+        const bool disable_chunking = ggml_is_numa();
+
+        // 4x chunks per thread
+        int64_t nr = ggml_nrows(op->src[0]);
+        int nth_scaled = nth * 4;
+        int64_t chunk_size = (nr + nth_scaled - 1) / nth_scaled;
+        int64_t nchunk     = (nr + chunk_size - 1) / chunk_size;
+
+        // Ensure minimum chunk size to avoid alignment issues with high thread counts
+        // Minimum chunk size should be at least NB_COLS to prevent overlapping chunks after alignment
+        const int64_t min_chunk_size = NB_COLS;
+        if (nchunk > 0 && (nr / nchunk) < min_chunk_size && nr >= min_chunk_size) {
+            nchunk = (nr + min_chunk_size - 1) / min_chunk_size;
+        }
+
+        if (nth == 1 || nchunk < nth || disable_chunking) {
+            nchunk = nth;
+        }
+
+        // Ensure nchunk doesn't exceed the number of rows divided by minimum chunk size
+        // This prevents creating too many tiny chunks that could overlap after alignment
+        const int64_t max_nchunk = (nr + min_chunk_size - 1) / min_chunk_size;
+        if (nchunk > max_nchunk) {
+            nchunk = max_nchunk;
+        }
+
+        if (ith == 0) {
+            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+            ggml_threadpool_chunk_set(params->threadpool, nth);
+        }
+
         ggml_barrier(params->threadpool);
 
-        const void * src1_wdata      = params->wdata;
-        const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10);
-        int64_t      src0_start      = (ith * ne01) / nth;
-        int64_t      src0_end        = ((ith + 1) * ne01) / nth;
-        src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
-        src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
-        if (src0_start >= src0_end) {
-            return;
-        }
+        // The first chunk comes from our thread_id, the rest will get auto-assigned.
+        int current_chunk = ith;
 
-        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
-        if (ne11 > 3) {
-            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
-        }
-        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
-            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00,
-                    (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
-                    (const char *) src0->data + src0_start * nb01,
-                    (const char *) src1_wdata + (src1_col_stride * iter), 1,
-                    src0_end - src0_start);
+        while (current_chunk < nchunk) {
+            int64_t src0_start = (current_chunk * ne01) / nchunk;
+            int64_t src0_end   = ((current_chunk + 1) * ne01) / nchunk;
+
+            // Align boundaries to NB_COLS - round up to ensure all data is included
+            // The chunk size limiting above ensures chunks are large enough to prevent overlaps
+            src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
+            src0_end   = (src0_end   % NB_COLS) ? src0_end   + NB_COLS - (src0_end   % NB_COLS) : src0_end;
+            if (src0_end > ne01) {
+                src0_end = ne01;
+            }
+
+            if (src0_start >= src0_end) {
+                break;
+            }
+
+            forward_mul_mat_one_chunk(params, dst, src0_start, src0_end);
+
+            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
         }
     }
 
@@ -1772,8 +1829,12 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             int64_t src0_cur_start = (ith * ne01) / nth;
             int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
 
+            // Align boundaries to NB_COLS - round up to ensure all data is included
             src0_cur_start = (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start;
             src0_cur_end   = (src0_cur_end   % NB_COLS) ? src0_cur_end   + NB_COLS - (src0_cur_end   % NB_COLS) : src0_cur_end;
+            if (src0_cur_end > ne01) {
+                src0_cur_end = ne01;
+            }
 
             if (src0_cur_start >= src0_cur_end) {
                 return;

ggml/src/ggml-cpu/simd-mappings.h

@@ -956,7 +956,7 @@ do {                                                              \
 
 #define GGML_F32Cx8          __m256
 #define GGML_F32Cx8_ZERO    (__m256)__lasx_xvldi(0)
-#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x))
+#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplfr2vr_s((x))
 
 static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) {
     __m256i a;
@@ -999,34 +999,34 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 
 #define GGML_F32x4         __m128
 #define GGML_F32x4_ZERO    (__m128)__lsx_vldi(0)
-#define GGML_F32x4_SET1(x) (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32x4_SET1(x) (__m128)__lsx_vreplfr2vr_s((x))
 #define GGML_F32x4_LOAD(x) (__m128)__lsx_vld((x), 0)
 #define GGML_F32x4_STORE(x, y)   __lsx_vst(y, x, 0)
 #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
 #define GGML_F32x4_ADD     __lsx_vfadd_s
 #define GGML_F32x4_MUL     __lsx_vfmul_s
-#define GGML_F32x4_REDUCE(res, x)                                                     \
-{                                                                                     \
-    int offset = GGML_F32_ARR >> 1;                                                   \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    offset >>= 1;                                                                     \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    offset >>= 1;                                                                     \
-    for (int i = 0; i < offset; ++i) {                                                \
-        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
-    }                                                                                 \
-    __m128i tmp     = __lsx_vsrli_d((__m128i) x[0], 32);                              \
-    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]);                    \
-    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
-    const __m128 t0 = (__m128)__lsx_vshuf4i_w(tmp, 0x88);                                     \
-    tmp             = __lsx_vsrli_d((__m128i) t0, 32);                                \
-    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, t0);                      \
-    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
-    res             = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \
+
+#define GGML_F32x4_REDUCE(res, x)                               \
+{                                                               \
+    int offset = GGML_F32_ARR >> 1;                             \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    offset >>= 1;                                               \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    offset >>= 1;                                               \
+    for (int i = 0; i < offset; ++i) {                          \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                \
+    }                                                           \
+    __m128i t0 = __lsx_vpickev_w((__m128i)x[0], (__m128i)x[0]); \
+    __m128i t1 = __lsx_vpickod_w((__m128i)x[0], (__m128i)x[0]); \
+    __m128 t2 = __lsx_vfadd_s((__m128)t0, (__m128)t1);          \
+    __m128i t3 = __lsx_vpickev_w((__m128i)t2, (__m128i)t2);     \
+    __m128i t4 = __lsx_vpickod_w((__m128i)t2, (__m128i)t2);     \
+    __m128 t5 = __lsx_vfadd_s((__m128)t3, (__m128)t4);          \
+    res = (ggml_float) ((v4f32)t5)[0];                          \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -1068,7 +1068,7 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 
 #define GGML_F32Cx4             __m128
 #define GGML_F32Cx4_ZERO        (__m128)__lsx_vldi(0)
-#define GGML_F32Cx4_SET1(x)     (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32Cx4_SET1(x)     (__m128)__lsx_vreplfr2vr_s((x))
 #define GGML_F32Cx4_LOAD(x)     (__m128)__lsx_f16x4_load(x)
 #define GGML_F32Cx4_STORE(x, y) __lsx_f16x4_store(x, y)
 #define GGML_F32Cx4_FMA         GGML_F32x4_FMA

ggml/src/ggml-cuda/argsort.cu

@@ -87,7 +87,7 @@ template<ggml_sort_order order>
 static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
     // bitonic sort
     int col = threadIdx.x;
-    int row = blockIdx.y;
+    int row = blockIdx.x;
 
     if (col >= ncols_pad) {
         return;
@@ -151,7 +151,7 @@ static void argsort_f32_i32_cuda_bitonic(const float *   x,
     const int ncols_pad = next_power_of_2(ncols);
 
     const dim3 block_dims(ncols_pad, 1, 1);
-    const dim3 block_nums(1, nrows, 1);
+    const dim3 block_nums(nrows, 1, 1);
     const size_t shared_mem = ncols_pad * sizeof(int);
 
     // FIXME: this limit could be raised by ~2-4x on Ampere or newer

ggml/src/ggml-cuda/common.cuh

@@ -224,6 +224,11 @@ static const char * cu_get_error_str(CUresult err) {
 #define AMD_MFMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
+// The Volta instructions are in principle available on Turing or newer but they are effectively unusable:
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#define VOLTA_MMA_AVAILABLE
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define TURING_MMA_AVAILABLE
 #endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
@@ -278,7 +283,10 @@ static bool amd_mfma_available(const int cc) {
 #endif //!defined(GGML_HIP_NO_MMQ_MFMA)
 }
 
-// Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
+static bool volta_mma_available(const int cc) {
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_VOLTA;
+}
+
 static bool turing_mma_available(const int cc) {
     return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING;
 }
@@ -625,8 +633,11 @@ static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 // and a shift:
 //
 // n/d = (mulhi(n, mp) + n) >> L;
-static const uint3 init_fastdiv_values(uint32_t d) {
-    GGML_ASSERT(d != 0);
+static const uint3 init_fastdiv_values(uint64_t d_64) {
+    GGML_ASSERT(d_64 != 0);
+    GGML_ASSERT(d_64 <= std::numeric_limits<uint32_t>::max());
+
+    uint32_t d = (uint32_t)d_64;
 
     // compute L = ceil(log2(d));
     uint32_t L = 0;

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/fattn-tile.cu

@@ -14,6 +14,10 @@ void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_cuda_flash_attn_ext_tile_case< 64,  64>(ctx, dst);
         } break;
+        case  72: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 72,  72>(ctx, dst);
+        } break;
         case  80: {
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_cuda_flash_attn_ext_tile_case< 80,  80>(ctx, dst);

ggml/src/ggml-cuda/fattn-tile.cuh

@@ -6,7 +6,7 @@
 // nbatch_K == number of K columns to load in parallel for KQ calculation
 
 // TODO optimize kernel parameters for FP16 NVIDIA (P100)
-// TODO optimize kernel parameters for head sizes 40, 80, 96, 112
+// TODO optimize kernel parameters for head sizes 40, 72, 80, 96, 112
 
 // The ROCm compiler cannot handle templating in __launch_bounds__.
 // As a workaround, define a macro to package the kernel parameters as uint32_t:
@@ -32,6 +32,12 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  64,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  64,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  64,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  64,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  64,  40)
@@ -80,6 +86,12 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 16, 128, 3,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
@@ -130,6 +142,13 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
@@ -185,6 +204,13 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 32, 128, 4,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 64,  64, 64, 128, 5,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  2,  64, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  4, 128, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72,  8, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 16, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 32, 256, 2,  32,  72)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE( 72,  72, 64, 256, 2,  32,  72)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  2,  64, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  4, 128, 2,  32,  40)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE( 80,  80,  8, 256, 2,  32,  40)
@@ -723,7 +749,7 @@ static __global__ void flash_attn_tile(
 
     if (
 #ifdef GGML_USE_WMMA_FATTN
-            (ncols2 != 1 && DV != 40 && DV != 512) ||
+            (ncols2 != 1 && DV != 40 && DV != 72 && DV != 512) ||
 #endif // GGML_USE_WMMA_FATTN
             (use_logit_softcap && !(DV == 128 || DV == 256))
     ) {
@@ -1198,6 +1224,7 @@ void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor
 
 extern DECL_FATTN_TILE_CASE( 40,  40);
 extern DECL_FATTN_TILE_CASE( 64,  64);
+extern DECL_FATTN_TILE_CASE( 72,  72);
 extern DECL_FATTN_TILE_CASE( 80,  80);
 extern DECL_FATTN_TILE_CASE( 96,  96);
 extern DECL_FATTN_TILE_CASE(112, 112);

ggml/src/ggml-cuda/fattn.cu

@@ -223,6 +223,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     switch (K->ne[0]) {
         case  40:
         case  64:
+        case  72:
         case  80:
         case  96:
         case 128:
@@ -275,7 +276,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
     // If Turing tensor cores available, use them:
-    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40) {
+    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72) {
         if (can_use_vector_kernel) {
             if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE && Q->ne[1] == 1 && Q->ne[3] == 1 && !(gqa_ratio > 4 && K->ne[1] >= 8192)) {
@@ -301,7 +302,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     }
 
     // Use the WMMA kernel if possible:
-    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 576) {
+    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 576) {
         if (can_use_vector_kernel && Q->ne[1] <= 2) {
             return BEST_FATTN_KERNEL_VEC;
         }

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

@@ -27,6 +27,7 @@
 #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"
@@ -50,6 +51,7 @@
 #include "ggml-cuda/upscale.cuh"
 #include "ggml-cuda/wkv.cuh"
 #include "ggml-cuda/gla.cuh"
+#include "ggml-cuda/set.cuh"
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
 #include "ggml.h"
@@ -2111,7 +2113,15 @@ 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);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
 
     //we only support fusion for ncols_dst = 1
     if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
@@ -2152,6 +2162,15 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
         return false;
     }
 
+
+    const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
+                       ggml_backend_buft_is_cuda_split(src1->buffer->buft);
+
+    //TODO: add support for fusion for split buffers
+    if (split) {
+        return false;
+    }
+
     return use_mul_mat_vec_q;
 }
 
@@ -2188,16 +2207,16 @@ 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);
     }
 
@@ -2268,7 +2287,7 @@ 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;
         }
@@ -2416,6 +2435,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SET_ROWS:
             ggml_cuda_op_set_rows(ctx, dst);
             break;
+        case GGML_OP_SET:
+            ggml_cuda_op_set(ctx, dst);
+            break;
         case GGML_OP_DUP:
             ggml_cuda_dup(ctx, dst);
             break;
@@ -2494,6 +2516,18 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_XIELU:
                     ggml_cuda_op_xielu(ctx, dst);
                     break;
+                case GGML_UNARY_OP_FLOOR:
+                    ggml_cuda_op_floor(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_CEIL:
+                    ggml_cuda_op_ceil(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_ROUND:
+                    ggml_cuda_op_round(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_TRUNC:
+                    ggml_cuda_op_trunc(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -2974,7 +3008,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
         ggml_cuda_topk_moe_ops(/*with_norm=*/false, /*delayed_softmax=*/true);
 
     if (ops.size() == topk_moe_ops_with_norm.size() &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 8 })) {
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx];
         ggml_tensor * weights = cgraph->nodes[node_idx + 9];
 
@@ -2993,7 +3027,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
     }
 
     if (ops.size() == topk_moe_ops_delayed_softmax.size() &&
-        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2, node_idx + 5 })) {
+        ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) {
         ggml_tensor * softmax = cgraph->nodes[node_idx + 4];
         ggml_tensor * weights = cgraph->nodes[node_idx + 5];
 
@@ -3114,9 +3148,20 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
         // With the use of CUDA graphs, the execution will be performed by the graph launch.
         if (!use_cuda_graph || cuda_graph_update_required) {
 
+            [[maybe_unused]] int prev_i = 0;
+
             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;
+                if (nodes_fused > 0) {
+                    GGML_LOG_INFO("nodes_fused: %d\n", nodes_fused);
+                }
+#endif
+
                 if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
                     continue;
                 }
@@ -3154,6 +3199,31 @@ 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];
@@ -3728,6 +3798,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_EXP:
                 case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_FLOOR:
+                case GGML_UNARY_OP_CEIL:
+                case GGML_UNARY_OP_ROUND:
+                case GGML_UNARY_OP_TRUNC:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -3842,6 +3916,13 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                        op->src[0]->type == GGML_TYPE_F32 &&
                        (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
             } break;
+        case GGML_OP_SET:
+            {
+                const ggml_type t = op->type;
+                return (t == GGML_TYPE_F32 || t == GGML_TYPE_I32) &&
+                    t == op->src[0]->type &&
+                    t == op->src[1]->type;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;

ggml/src/ggml-cuda/mma.cuh

@@ -18,6 +18,10 @@
 
 #include "common.cuh"
 
+// On Volta each warp is doing 4 8x8 mma operations in parallel.
+// The basic memory layout for a 32x8 output tile is to stack 4 input tiles in I direction and to mirror the B tile.
+// However, the i indices in this file are by default permuted to simplify the index calculations.
+// #define GGML_CUDA_MMA_NO_VOLTA_PERM
 
 #if CUDART_VERSION >= 11080
 
@@ -73,6 +77,15 @@ namespace ggml_cuda_mma {
         static constexpr int ne = I * J / 64;
         T x[ne] = {0};
 
+        static constexpr __device__ bool supported() {
+            if (I == 64 && J ==  2) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 32 && J ==  4) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J == 32) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
                 return threadIdx.x % 16;
@@ -85,7 +98,8 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 32 && J == 32) {
                 return 4 * (threadIdx.x / 32) + 8 * (l / 4) + (l % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
@@ -101,22 +115,67 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 32 && J == 32) {
                 return threadIdx.x % 32;
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        static constexpr int ne = I * J / 32;
+        T x[ne] = {0};
+
+        static constexpr __device__ bool supported() {
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 32 && J == 8) {
+#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
+                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (l & 2) | (threadIdx.x % 2);
+#else
+                return (l & 2) | (threadIdx.x & ~2);
+#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 32 && J == 8) {
+                return (threadIdx.x & 2) | (l & (4 + 1));
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 #else
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 8 && (J == 4 || J == 8)) {
+            if constexpr (I == 8 && J == 4) {
+                return threadIdx.x / 4;
+            } else if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 8 + threadIdx.x / 4;
+                return ((l / 2) * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 16) {
-                return ((l / 2) % 2) * 8 + threadIdx.x / 4;
+                return (((l / 2) % 2) * 8) | (threadIdx.x / 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return tile<16, 8, T>::get_i(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
@@ -124,13 +183,16 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 8 && J == 8) {
-                return 4 * l + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 8) {
-                return 2 * (threadIdx.x % 4) + l % 2;
+                return ((threadIdx.x % 4) * 2) | (l % 2);
             } else if constexpr (I == 16 && J == 16) {
-                return 8 * (l / 4) + 2 * (threadIdx.x % 4) + l % 2;
+                return ((l / 4) * 8) | ((threadIdx.x % 4) * 2) | (l % 2);
+            } else if constexpr (I == 32 && J == 8) {
+                return tile<16, 8, T>::get_j(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 #endif // defined(GGML_USE_HIP)
@@ -140,32 +202,83 @@ namespace ggml_cuda_mma {
     struct tile<I_, J_, half2> {
         static constexpr int I  = I_;
         static constexpr int J  = J_;
+
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        static constexpr int ne = I == 8 && J == 8 ? I * J / (WARP_SIZE/4) : I * J / WARP_SIZE;
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  8) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 8 && J == 8) {
+                return ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+            } else if constexpr (I == 32 && J == 8) {
+#ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
+                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+#else
+                return threadIdx.x;
+#endif // GGML_CUDA_MMA_NO_VOLTA_PERM
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr ((I == 8 || I == 32) && J == 8) {
+                return l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+#else
         static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  8) return true;
+            if (I == 16 && J == 16) return true;
+            if (I == 32 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return l * 8 + threadIdx.x / 4;
+                return (l * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return (l % 2) * 8 + threadIdx.x / 4;
+                return ((l % 2) * 8) | (threadIdx.x / 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return ((l / 4) * 16) | ((l % 2) * 8) | (threadIdx.x / 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return l * 4 + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 4 + threadIdx.x % 4;
+                return ((l / 2) * 4) | (threadIdx.x % 4);
+            } else if constexpr (I == 32 && J == 8) {
+                return ((l & 2) * 2) | (threadIdx.x % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
     };
 
     template <int I_, int J_>
@@ -175,27 +288,36 @@ namespace ggml_cuda_mma {
         static constexpr int ne = I * J / WARP_SIZE;
         nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
 
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  8) return true;
+            if (I == 16 && J ==  4) return true;
+            if (I == 16 && J ==  8) return true;
+            return false;
+        }
+
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return l * 8 + threadIdx.x / 4;
+                return (l * 8) | (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return (l % 2) * 8 + threadIdx.x / 4;
+                return ((l % 2) * 8) | (threadIdx.x / 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return l * 4 + threadIdx.x % 4;
+                return (l * 4) | (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return (l / 2) * 4 + threadIdx.x % 4;
+                return ((l / 2) * 4) | (threadIdx.x % 4);
             } else {
-                static_assert(I == -1 && J == -1, "template specialization not implemented");
+                NO_DEVICE_CODE;
+                return -1;
             }
         }
     };
@@ -263,8 +385,12 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
 #else
-        load_generic(xs0, stride);
-        GGML_UNUSED(t);
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
+#else
+        load_generic(t, xs0, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -277,11 +403,35 @@ namespace ggml_cuda_mma {
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
             : "l"(xs));
+#else
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #else
         load_generic(t, xs0, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
+    template <typename T>
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<32, 8, T> & t, const T * __restrict__ xs0, const int stride) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if 1
+        // TODO: more generic handling
+        static_assert(sizeof(T) == 4, "bad type size");
+        ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 4, xs0 + t.get_i(4)*stride + 4);
+#else
+        load_generic(t, xs0, stride);
+#endif // 1
+#else
+        tile<16, 8, T> * t16 = (tile<16, 8, T> *) &t;
+        load_ldmatrix(t16[0], xs0 +  0*stride, stride);
+        load_ldmatrix(t16[1], xs0 + 16*stride, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+    }
+
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix_trans(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
@@ -546,4 +696,43 @@ namespace ggml_cuda_mma {
         NO_DEVICE_CODE;
 #endif // AMD_MFMA_AVAILABLE
     }
+
+    template <typename T1, typename T2, int J, int K>
+    static __device__ __forceinline__ void mma(
+            tile<32, J, T1> & D, const tile<32, K, T2> & A, const tile<J, K, T2> & B) {
+        tile<16, J, T1> * D16 = (tile<16, J, T1> *) &D;
+        tile<16, K, T2> * A16 = (tile<16, K, T2> *) &A;
+        mma(D16[0], A16[0], B);
+        mma(D16[1], A16[1], B);
+    }
+
+    static __device__ __forceinline__ void mma(
+            tile<32, 8, float> & D, const tile<32, 8, half2> & A, const tile<8, 8, half2> & B) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0]), "r"(Bxi[1]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]), "r"(Bxi[3]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[4]), "r"(Axi[5]), "r"(Bxi[4]), "r"(Bxi[5]));
+        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
+            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
+            : "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7]));
+#else
+        tile<16, 8, float> * D16 = (tile<16, 8, float> *) &D;
+        tile<16, 8, half2> * A16 = (tile<16, 8, half2> *) &A;
+        mma(D16[0], A16[0], B);
+        mma(D16[1], A16[1], B);
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+    }
 }

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;
     }
@@ -148,7 +160,7 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
         case GGML_TYPE_F32:
             return ampere_mma_available(cc);
         case GGML_TYPE_F16:
-            return turing_mma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc);
         case GGML_TYPE_BF16:
             return ampere_mma_available(cc);
         default:

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)
@@ -28,9 +28,19 @@ static __global__ void mul_mat_f(
         const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
+    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
+    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
+
+    if (!I_16_supported && !I_32_supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work but 16 is ~1% faster.
+
+    typedef tile<I_preferred, 8, T>     tile_A;
+    typedef tile<8,           8, T>     tile_B;
+    typedef tile<I_preferred, 8, float> tile_C;
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -232,7 +242,6 @@ static __global__ void mul_mat_f(
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
-
 //This kernel is for larger batch sizes of mul_mat_id
 template <typename T, int rows_per_block, int cols_per_block, int nwarps>
 __launch_bounds__(ggml_cuda_get_physical_warp_size()*nwarps, 1)
@@ -245,9 +254,19 @@ static __global__ void mul_mat_f_ids(
         const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
         const uint3 sis1_fd, const uint3 nch_fd) {
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    typedef tile<16, 8, T>     tile_A;
-    typedef tile< 8, 8, T>     tile_B;
-    typedef tile<16, 8, float> tile_C;
+    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
+    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
+
+    if (!I_16_supported && !I_32_supported) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work butr 16 is ~1% faster.
+
+    typedef tile<I_preferred, 8, T>     tile_A;
+    typedef tile<8,           8, T>     tile_B;
+    typedef tile<I_preferred, 8, float> tile_C;
 
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
@@ -533,7 +552,8 @@ void mul_mat_f_cuda(
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         cudaStream_t stream, const mmf_ids_data * ids_data) {
-    typedef tile<16, 8, T>     tile_A;
+    typedef tile<16, 8, T>     tile_A_16;
+    typedef tile<32, 8, T>     tile_A_32;
     typedef tile< 8, 8, T>     tile_B;
 
     GGML_ASSERT(ncols_x      % 2 == 0);
@@ -544,7 +564,8 @@ void mul_mat_f_cuda(
     const int64_t channel_ratio = nchannels_dst / nchannels_x;
     const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
 
-    const int device = ggml_cuda_get_device();
+    const int device    = ggml_cuda_get_device();
+    const int cc        = ggml_cuda_info().devices[device].cc;
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
 
     int64_t nwarps_best     = 1;
@@ -559,7 +580,7 @@ void mul_mat_f_cuda(
     }
 
     constexpr int rows_per_block = MMF_ROWS_PER_BLOCK;
-    const int nbytes_shared_iter = nwarps_best * tile_A::I * (warp_size + 4) * 4;
+    const int nbytes_shared_iter = nwarps_best * (volta_mma_available(cc) ? tile_A_32::I : tile_A_16::I) * (warp_size + 4) * 4;
     const int nbytes_shared_combine = GGML_PAD(cols_per_block, tile_B::I) * (nwarps_best*rows_per_block + 4) * 4;
     const int nbytes_shared = std::max(nbytes_shared_iter, nbytes_shared_combine);
     const int nbytes_slotmap = ids ? GGML_PAD(cols_per_block, 16) * sizeof(int) : 0;

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/mmvq.cu

@@ -190,12 +190,30 @@ static __global__ void mul_mat_vec_q(
 
     const uint32_t channel_bias = ids ? channel_x : channel_dst;
 
+    float x_biases[ncols_dst]    = { 0.0f };
+    float gate_biases[ncols_dst] = { 0.0f };
     if constexpr (has_fusion) {
         if (use_bias) {
             x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
+            // 1. Hide latency by prefetching bias and gate here
+            // 2. load only on threads that won't die after partial sum calculation
+            if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
+                (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    x_biases[j] = x_bias[j * stride_col_dst + threadIdx.x];
+                }
+            }
         }
         if (use_gate_bias) {
             gate_bias = gate_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
+            if (threadIdx.x < rows_per_cuda_block && threadIdx.y == 0 &&
+                (rows_per_cuda_block == 1 || uint32_t(row0 + threadIdx.x) < stride_col_dst)) {
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    gate_biases[j] = gate_bias[j * stride_col_dst + threadIdx.x];
+                }
+            }
         }
     }
 
@@ -283,12 +301,12 @@ static __global__ void mul_mat_vec_q(
             float result = tmp[j][threadIdx.x];
             if constexpr (has_fusion) {
                 if (use_bias) {
-                    result += x_bias[j*stride_col_dst + threadIdx.x];
+                    result += x_biases[j];
                 }
                 if (use_gate) {
                     float gate_value = tmp_gate[j][threadIdx.x];
                     if (use_gate_bias) {
-                        gate_value += gate_bias[j*stride_col_dst + threadIdx.x];
+                        gate_value += gate_biases[j];
                     }
                     switch (active_glu) {
                         case GGML_GLU_OP_SWIGLU:

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

@@ -0,0 +1,168 @@
+#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

@@ -0,0 +1,11 @@
+#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-cuda/rope.cu

@@ -125,7 +125,7 @@ template<bool forward, bool has_ff, typename T>
 static __global__ void rope_multi(
         const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2,
         const int n_dims, const int32_t * pos, const float freq_scale, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors, const mrope_sections sections) {
+        const rope_corr_dims corr_dims, const float theta_scale, const float * freq_factors, const mrope_sections sections, const bool is_imrope) {
     const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
     if (i0 >= ne0) {
@@ -152,17 +152,29 @@ static __global__ void rope_multi(
     const int sector = (i0 / 2) % sect_dims;
 
     float theta_base = 0.0;
-    if (sector < sections.v[0]) {
-        theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sections.v[0] && sector < sec_w) {
-        theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-        theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sec_w + sections.v[2]) {
-        theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+    if (is_imrope) {
+        if (sector % 3 == 1 && sector < 3 * sections.v[1]) { // h
+            theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
+        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) { // w
+            theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
+        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) { // t
+            theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+        } else {
+            theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+        }
+    } else {
+        if (sector < sections.v[0]) {
+            theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sections.v[0] && sector < sec_w) {
+            theta_base = pos[channel_x + ne2 * 1]*powf(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 2]*powf(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w + sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+        }
     }
 
     const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
@@ -276,7 +288,7 @@ template<bool forward, typename T>
 static void rope_multi_cuda(
         const T * x, T * dst, const int ne0, const int ne1, const int ne2, const int s1, const int s2, const int n_dims, const int nr,
         const int32_t * pos, const float freq_scale, const float freq_base, const float ext_factor, const float attn_factor,
-        const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, cudaStream_t stream) {
+        const rope_corr_dims corr_dims, const float * freq_factors, const mrope_sections sections, const bool is_imrope, cudaStream_t stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
     const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
@@ -287,11 +299,11 @@ static void rope_multi_cuda(
     if (freq_factors == nullptr) {
         rope_multi<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
             x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections);
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
     } else {
         rope_multi<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
             x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections);
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
     }
 }
 
@@ -369,6 +381,7 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -406,11 +419,11 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst)
         if (src0->type == GGML_TYPE_F32) {
             rope_multi_cuda<forward>(
                 (const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope, stream);
         } else if (src0->type == GGML_TYPE_F16) {
             rope_multi_cuda<forward>(
                 (const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale,
-                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+                freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, is_imrope, stream);
         } else {
             GGML_ABORT("fatal error");
         }

ggml/src/ggml-cuda/set-rows.cu

@@ -4,30 +4,53 @@
 typedef void (*set_rows_kernel_t)(const char * src, char * dst);
 
 // Generic quantized set_rows kernel template
-template<typename idx_t, typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
-static __global__ void k_set_rows_quant(
-        const float * __restrict__ src0, const idx_t * __restrict__ src1, block_type * __restrict__ dst,
-        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
-        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
-        const int64_t s01, const int64_t s02, const int64_t s03,
-        const int64_t s10, const int64_t s11, const int64_t s12,
-        const int64_t s1, const int64_t s2, const int64_t s3) {
-
+template <typename idx_t, typename block_type, int qk, void (*quantize_func)(const float *, block_type *)>
+static __global__ void k_set_rows_quant(const float * __restrict__ src0,
+                                        const idx_t * __restrict__ src1,
+                                        block_type * __restrict__ dst,
+                                        const int64_t ne_total,
+                                        const int64_t ne10,
+                                        const int64_t ne11,
+                                        const int64_t ne12,
+                                        const int64_t ne13,
+                                        const int64_t s01,
+                                        const int64_t s02,
+                                        const int64_t s03,
+                                        const int64_t s10,
+                                        const int64_t s11,
+                                        const int64_t s12,
+                                        const int64_t s1,
+                                        const int64_t s2,
+                                        const int64_t s3,
+                                        const uint3   ne00,
+                                        const uint3   ne01,
+                                        const uint3   ne02,
+                                        const uint3   ne11_fd,
+                                        const uint3   ne12_fd) {
     const int64_t i = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
-    const int64_t ne_total = (ne00 * ne01 * ne02 * ne03) / qk;
 
     if (i >= ne_total) {
         return;
     }
 
     const int64_t i_base = i * qk;
-    const int64_t i03 = i_base / (ne00 * ne01 * ne02);
-    const int64_t i02 = (i_base - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
-    const int64_t i01 = (i_base - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
-    const int64_t i00 = i_base - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+    uint32_t      tmp    = (uint32_t) i_base;
+    uint2         div_mod;
 
-    const int64_t i12 = i03 % ne12;
-    const int64_t i11 = i02 % ne11;
+    div_mod           = fast_div_modulo(tmp, ne00);
+    const int64_t i00 = div_mod.y;
+    tmp               = div_mod.x;
+
+    div_mod           = fast_div_modulo(tmp, ne01);
+    const int64_t i01 = div_mod.y;
+    tmp               = div_mod.x;
+
+    div_mod           = fast_div_modulo(tmp, ne02);
+    const int64_t i02 = div_mod.y;
+    const int64_t i03 = div_mod.x;
+
+    const int64_t i12 = fastmodulo((uint32_t) i03, ne12_fd);
+    const int64_t i11 = fastmodulo((uint32_t) i02, ne11_fd);
     const int64_t i10 = i01;
 
     const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
@@ -41,6 +64,8 @@ static __global__ void k_set_rows_quant(
     quantize_func(src_block, dst_block);
 
     GGML_UNUSED(ne10);
+    GGML_UNUSED(ne11);
+    GGML_UNUSED(ne12);
     GGML_UNUSED(ne13);
 }
 
@@ -71,40 +96,65 @@ static void set_rows_cuda_quant(
     const int64_t s2  = nb2;
     const int64_t s3  = nb3;
 
-    if (ne_total > 0) {
+    if (ne_total > 0 && ne00 > 0 && ne01 > 0 && ne02 > 0 && ne11 > 0 && ne12 > 0) {
+        const uint3 ne00_fd = init_fastdiv_values((uint32_t) ne00);
+        const uint3 ne01_fd = init_fastdiv_values((uint32_t) ne01);
+        const uint3 ne02_fd = init_fastdiv_values((uint32_t) ne02);
+        const uint3 ne11_fd = init_fastdiv_values((uint32_t) ne11);
+        const uint3 ne12_fd = init_fastdiv_values((uint32_t) ne12);
+
         k_set_rows_quant<idx_t, block_type, qk, quantize_func><<<grid_size, block_size, 0, stream>>>(
-            src0_d, src1_d, dst_d,
-            ne00, ne01, ne02, ne03,
-            ne10, ne11, ne12, ne13,
-            s01, s02, s03,
-            s10, s11, s12,
-            s1, s2, s3);
+            src0_d, src1_d, dst_d, ne_total, ne10, ne11, ne12, ne13, s01, s02, s03, s10, s11, s12, s1, s2, s3, ne00_fd,
+            ne01_fd, ne02_fd, ne11_fd, ne12_fd);
     }
 }
 
-template<typename src_t, typename idx_t, typename dst_t>
-static __global__ void k_set_rows(
-        const src_t * __restrict__ src0, const idx_t * __restrict__ src1, dst_t * __restrict__ dst,
-        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
-        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
-        const int64_t s01, const int64_t s02, const int64_t s03,
-        const int64_t s10, const int64_t s11, const int64_t s12,
-        const int64_t s1, const int64_t s2, const int64_t s3) {
-
+template <typename src_t, typename idx_t, typename dst_t>
+static __global__ void k_set_rows(const src_t * __restrict__ src0,
+                                  const idx_t * __restrict__ src1,
+                                  dst_t * __restrict__ dst,
+                                  const int64_t ne_total,
+                                  const int64_t ne10,
+                                  const int64_t ne11,
+                                  const int64_t ne12,
+                                  const int64_t ne13,
+                                  const int64_t s01,
+                                  const int64_t s02,
+                                  const int64_t s03,
+                                  const int64_t s10,
+                                  const int64_t s11,
+                                  const int64_t s12,
+                                  const int64_t s1,
+                                  const int64_t s2,
+                                  const int64_t s3,
+                                  const uint3   ne00,
+                                  const uint3   ne01,
+                                  const uint3   ne02,
+                                  const uint3   ne11_fd,
+                                  const uint3   ne12_fd) {
     const int64_t i = int64_t(blockDim.x) * blockIdx.x + threadIdx.x;
-    const int64_t ne_total = ne00 * ne01 * ne02 * ne03;
 
     if (i >= ne_total) {
         return;
     }
 
-    const int64_t i03 = i / (ne00 * ne01 * ne02);
-    const int64_t i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
-    const int64_t i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01) / ne00;
-    const int64_t i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne00 * ne01 - i01 * ne00;
+    uint32_t tmp = (uint32_t) i;
+    uint2    div_mod;
 
-    const int64_t i12 = i03 % ne12;
-    const int64_t i11 = i02 % ne11;
+    div_mod           = fast_div_modulo(tmp, ne00);
+    const int64_t i00 = div_mod.y;
+    tmp               = div_mod.x;
+
+    div_mod           = fast_div_modulo(tmp, ne01);
+    const int64_t i01 = div_mod.y;
+    tmp               = div_mod.x;
+
+    div_mod           = fast_div_modulo(tmp, ne02);
+    const int64_t i02 = div_mod.y;
+    const int64_t i03 = div_mod.x;
+
+    const int64_t i12 = fastmodulo((uint32_t) i03, ne12_fd);
+    const int64_t i11 = fastmodulo((uint32_t) i02, ne11_fd);
     const int64_t i10 = i01;
 
     const int64_t dst_row = *(src1 + i10*s10 + i11*s11 + i12*s12);
@@ -115,6 +165,8 @@ static __global__ void k_set_rows(
     dst_row_ptr[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
 
     GGML_UNUSED(ne10);
+    GGML_UNUSED(ne11);
+    GGML_UNUSED(ne12);
     GGML_UNUSED(ne13);
 }
 
@@ -144,14 +196,16 @@ static void set_rows_cuda(
     const int64_t s2  = nb2/sizeof(dst_t);
     const int64_t s3  = nb3/sizeof(dst_t);
 
-    if (ne_total > 0) {
-        k_set_rows<<<grid_size, block_size, 0, stream>>>(
-            src0_d, src1_d, dst_d,
-            ne00, ne01, ne02, ne03,
-            ne10, ne11, ne12, ne13,
-            s01, s02, s03,
-            s10, s11, s12,
-            s1, s2, s3);
+    if (ne_total > 0 && ne00 > 0 && ne01 > 0 && ne02 > 0 && ne11 > 0 && ne12 > 0) {
+        const uint3 ne00_fd = init_fastdiv_values((uint32_t) ne00);
+        const uint3 ne01_fd = init_fastdiv_values((uint32_t) ne01);
+        const uint3 ne02_fd = init_fastdiv_values((uint32_t) ne02);
+        const uint3 ne11_fd = init_fastdiv_values((uint32_t) ne11);
+        const uint3 ne12_fd = init_fastdiv_values((uint32_t) ne12);
+
+        k_set_rows<<<grid_size, block_size, 0, stream>>>(src0_d, src1_d, dst_d, ne_total, ne10, ne11, ne12, ne13, s01,
+                                                         s02, s03, s10, s11, s12, s1, s2, s3, ne00_fd, ne01_fd, ne02_fd,
+                                                         ne11_fd, ne12_fd);
     }
 }
 

ggml/src/ggml-cuda/set.cu

@@ -0,0 +1,39 @@
+#include "set.cuh"
+#include "cpy.cuh"
+
+void ggml_cuda_op_set(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32));
+    GGML_ASSERT(src1->type == src0->type);
+    GGML_ASSERT(dst ->type == src0->type);
+
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+
+    const size_t nb1    = ((int32_t *) dst->op_params)[0];
+    const size_t nb2    = ((int32_t *) dst->op_params)[1];
+    const size_t nb3    = ((int32_t *) dst->op_params)[2];
+    const size_t offset = ((int32_t *) dst->op_params)[3];
+    const bool   inplace= (bool)     ((int32_t *) dst->op_params)[4];
+
+    if (!inplace) {
+        ggml_cuda_cpy(ctx, src0, dst);
+    }
+
+    ggml_tensor dst_view = *dst;
+    dst_view.data  = (void *)((char *)dst->data + offset);
+    dst_view.ne[0] = src1->ne[0];
+    dst_view.ne[1] = src1->ne[1];
+    dst_view.ne[2] = src1->ne[2];
+    dst_view.ne[3] = src1->ne[3];
+
+    dst_view.nb[0] = ggml_element_size(dst);
+    dst_view.nb[1] = nb1;
+    dst_view.nb[2] = nb2;
+    dst_view.nb[3] = nb3;
+
+    ggml_cuda_cpy(ctx, src1, &dst_view);
+}

ggml/src/ggml-cuda/set.cuh

@@ -0,0 +1,7 @@
+#pragma once
+
+#include "common.cuh"
+
+#define CUDA_SET_BLOCK_SIZE 256
+
+void ggml_cuda_op_set(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq72-dv72.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(72, 72);

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

@@ -3,7 +3,7 @@
 from glob import glob
 import os
 
-HEAD_SIZES_KQ = [40, 64, 80, 96, 112, 128, 256, 576]
+HEAD_SIZES_KQ = [40, 64, 72, 80, 96, 112, 128, 256, 576]
 
 TYPES_KV = ["GGML_TYPE_F16", "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0"]
 
@@ -81,6 +81,8 @@ for ncols in [8, 16, 32, 64]:
             for head_size_kq in HEAD_SIZES_KQ:
                 if head_size_kq == 40:
                     continue
+                if head_size_kq == 72:
+                    continue
                 if head_size_kq != 576 and ncols2 == 16:
                     continue
                 if head_size_kq == 576 and ncols2 != 16:

ggml/src/ggml-cuda/unary.cu

@@ -85,6 +85,22 @@ static __device__ __forceinline__ float op_elu(float x) {
     return (x > 0.f) ? x : expm1f(x);
 }
 
+static __device__ __forceinline__ float op_floor(float x) {
+    return floorf(x);
+}
+
+static __device__ __forceinline__ float op_ceil(float x) {
+    return ceilf(x);
+}
+
+static __device__ __forceinline__ float op_round(float x) {
+    return round(x);
+}
+
+static __device__ __forceinline__ float op_trunc(float x) {
+    return trunc(x);
+}
+
 template <float (*op)(float), typename T>
 static __global__ void unary_op_kernel(const T * x, T * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
@@ -201,6 +217,22 @@ void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_unary<op_elu>(ctx, dst);
 }
+
+void ggml_cuda_op_floor(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_floor>(ctx, dst);
+}
+
+void ggml_cuda_op_ceil(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_ceil>(ctx, dst);
+}
+
+void ggml_cuda_op_round(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_round>(ctx, dst);
+}
+
+void ggml_cuda_op_trunc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary<op_trunc>(ctx, dst);
+}
 /* gated ops */
 
 template <float (*op)(float), typename T>

ggml/src/ggml-cuda/unary.cuh

@@ -63,6 +63,14 @@ void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_elu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_floor(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_ceil(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_round(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_trunc(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -217,6 +217,9 @@ struct ggml_hexagon_session {
     void allocate(int dev_id) noexcept(false);
     void release() noexcept(true);
 
+    void enqueue(struct htp_general_req &req, struct dspqueue_buffer *bufs, uint32_t n_bufs, bool sync = false);
+    void flush();
+
     ggml_backend_buffer_type buffer_type;
     ggml_backend_buffer_type repack_buffer_type;
 
@@ -237,15 +240,37 @@ struct ggml_hexagon_session {
     uint32_t         prof_pkts;
 };
 
-// Packet callback
-static void htp_packet_callback(dspqueue_t queue, AEEResult error, void * context) {
-    auto sess = static_cast<ggml_hexagon_session *>(context);
+void ggml_hexagon_session::enqueue(struct htp_general_req &req, struct dspqueue_buffer *bufs, uint32_t n_bufs, bool sync) {
+    // Bump pending flag (cleared in the session::flush once we get the responce)
+    this->op_pending++;  // atomic inc
+
+    int err = dspqueue_write(this->queue,
+                             0,                       // flags - the framework will autoset this
+                             n_bufs,                  // number of buffers
+                             bufs,                    // buffer references
+                             sizeof(req),
+                             (const uint8_t *) &req,  // Message
+                             1000000                  // Timeout
+    );
+
+    if (err != 0) {
+        GGML_ABORT("ggml-hex: %s dspqueue_write failed: 0x%08x\n", this->name.c_str(), (unsigned) err);
+    }
+
+    if (sync) {
+        flush();
+    }
+}
+
+// Flush HTP response queue i.e wait for all outstanding requests to complete
+void ggml_hexagon_session::flush() {
+    dspqueue_t q = this->queue;
 
     // Repeatedly read packets from the queue until it's empty. We don't
     // necessarily get a separate callback for each packet, and new packets
     // may arrive while we're processing the previous one.
 
-    while (1) {
+    while (this->op_pending) {
         struct htp_general_rsp rsp;
         uint32_t               rsp_size;
         uint32_t               flags;
@@ -253,22 +278,23 @@ static void htp_packet_callback(dspqueue_t queue, AEEResult error, void * contex
         struct dspqueue_buffer bufs[HTP_MAX_PACKET_BUFFERS];
         uint32_t               n_bufs;
 
-        // Read packet from queue
-        int err = dspqueue_read_noblock(queue, &flags,
-                                        HTP_MAX_PACKET_BUFFERS,  // Maximum number of buffer references
-                                        &n_bufs,                 // Number of buffer references
-                                        bufs,                    // Buffer references
-                                        sizeof(rsp),             // Max message length
-                                        &rsp_size,               // Message length
-                                        (uint8_t *) &rsp);
+        // Read response packet from queue
+        int err = dspqueue_read(q, &flags,
+                                   HTP_MAX_PACKET_BUFFERS,  // Maximum number of buffer references
+                                   &n_bufs,                 // Number of buffer references
+                                   bufs,                    // Buffer references
+                                   sizeof(rsp),             // Max message length
+                                   &rsp_size,               // Message length
+                                   (uint8_t *) &rsp,
+                                   1000000);                // Timeout
 
-        if (err == AEE_EWOULDBLOCK) {
-            // Consumed all packets available for now
-            return;
+        if (err == AEE_EEXPIRED) {
+            // TODO: might need to bail out if the HTP is stuck on something
+            continue;
         }
 
         if (err != 0) {
-            GGML_ABORT("ggml-hex: dspqueue_read_noblock failed: 0x%08x\n", (unsigned) err);
+            GGML_ABORT("ggml-hex: dspqueue_read failed: 0x%08x\n", (unsigned) err);
         }
 
         // Basic sanity checks
@@ -281,21 +307,15 @@ static void htp_packet_callback(dspqueue_t queue, AEEResult error, void * contex
             // TODO: handle errors
         }
 
-        // FIXME: update profiling implementation
-        sess->prof_usecs  = rsp.prof_usecs;
-        sess->prof_cycles = rsp.prof_cycles;
-        sess->prof_pkts   = rsp.prof_pkts;
+        // TODO: update profiling implementation, currently only works for opt_opsync mode
+        this->prof_usecs  = rsp.prof_usecs;
+        this->prof_cycles = rsp.prof_cycles;
+        this->prof_pkts   = rsp.prof_pkts;
 
-        sess->op_pending--;  // atomic dec
+        this->op_pending--;  // atomic dec
     }
 }
 
-// Error callback - simply terminates with an error. Used where we don't
-// expect errors.
-[[noreturn]] static void htp_error_callback(dspqueue_t queue, AEEResult error, void * context) {
-    GGML_ABORT("ggml-hex: dspcall general error 0x%x: for queue %p\n", error, (void *) queue);
-}
-
 // ** backend buffers
 
 struct ggml_backend_hexagon_buffer_type_context {
@@ -347,7 +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)");
@@ -656,6 +682,15 @@ static void repack_q4_0_q4x4x2(ggml_tensor * t, const void * data, size_t size)
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q4_0x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to read more data than is available in the source buffer 'data'
+    // or write more than the tensor can hold.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -667,7 +702,8 @@ static void repack_q4_0_q4x4x2(ggml_tensor * t, const void * data, size_t size)
 
     init_row_q4x4x2((block_q4_0 *) buf_pd, t->ne[0]);  // init padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) data + (i * row_size);
         uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
 
@@ -676,6 +712,25 @@ static void repack_q4_0_q4x4x2(ggml_tensor * t, const void * data, size_t size)
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
+
+        // re-init the row because we are potentially copying a partial row
+        init_row_q4x4x2((block_q4_0 *) buf_pd, t->ne[0]);
+
+        // Copy only the remaining bytes from the source.
+        memcpy(buf_pd, src, n_rem_bytes);
+
+        // Repack the entire buffer
+        repack_row_q4x4x2((uint8_t *) buf_rp, (const block_q4_0 *) buf_pd, t->ne[0]);
+
+        // Write only the corresponding remaining bytes to the destination tensor.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -688,6 +743,14 @@ static void repack_q4x4x2_q4_0(void * data, const ggml_tensor * t, size_t size)
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q4_0x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to copy more data than the tensor actually contains.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -699,7 +762,8 @@ static void repack_q4x4x2_q4_0(void * data, const ggml_tensor * t, size_t size)
 
     memset(buf_pd, 0, row_size_pd);  // clear-out padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
         uint8_t *       dst = (uint8_t *) data + (i * row_size);
 
@@ -708,6 +772,20 @@ static void repack_q4x4x2_q4_0(void * data, const ggml_tensor * t, size_t size)
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) data + (i * row_size);
+
+        // We still need to read and unpack the entire source row because quantization is block-based.
+        memcpy(buf_pd, src, row_size);
+        unpack_row_q4x4x2((block_q4_0 *) buf_rp, (const uint8_t *) buf_pd, t->ne[0]);
+
+        // But we only copy the remaining number of bytes to the destination.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -930,6 +1008,15 @@ static void repack_q8_0_q8x4x2(ggml_tensor * t, const void * data, size_t size)
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q8_0x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to read more data than is available in the source buffer 'data'
+    // or write more than the tensor can hold.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -941,7 +1028,8 @@ static void repack_q8_0_q8x4x2(ggml_tensor * t, const void * data, size_t size)
 
     init_row_q8x4x2((block_q8_0 *) buf_pd, t->ne[0]);  // init padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) data + (i * row_size);
         uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
 
@@ -950,6 +1038,25 @@ static void repack_q8_0_q8x4x2(ggml_tensor * t, const void * data, size_t size)
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
+
+        // re-init the row because we are potentially copying a partial row
+        init_row_q8x4x2((block_q8_0 *) buf_pd, t->ne[0]);
+
+        // Copy only the remaining bytes from the source.
+        memcpy(buf_pd, src, n_rem_bytes);
+
+        // Repack the entire buffer
+        repack_row_q8x4x2((uint8_t *) buf_rp, (const block_q8_0 *) buf_pd, t->ne[0]);
+
+        // Write only the corresponding remaining bytes to the destination tensor.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -962,6 +1069,14 @@ static void repack_q8x4x2_q8_0(void * data, const ggml_tensor * t, size_t size)
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q8_0x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to copy more data than the tensor actually contains.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -973,7 +1088,8 @@ static void repack_q8x4x2_q8_0(void * data, const ggml_tensor * t, size_t size)
 
     memset(buf_pd, 0, row_size_pd);  // clear-out padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
         uint8_t *       dst = (uint8_t *) data + (i * row_size);
 
@@ -982,6 +1098,20 @@ static void repack_q8x4x2_q8_0(void * data, const ggml_tensor * t, size_t size)
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) data + (i * row_size);
+
+        // We still need to read and unpack the entire source row because quantization is block-based.
+        memcpy(buf_pd, src, row_size);
+        unpack_row_q8x4x2((block_q8_0 *) buf_rp, (const uint8_t *) buf_pd, t->ne[0]);
+
+        // But we only copy the remaining number of bytes to the destination.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -1229,6 +1359,15 @@ static void repack_mxfp4_mxfp4x4x2(ggml_tensor * t, const void * data, size_t si
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_MXFP4x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to read more data than is available in the source buffer 'data'
+    // or write more than the tensor can hold.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -1240,7 +1379,8 @@ static void repack_mxfp4_mxfp4x4x2(ggml_tensor * t, const void * data, size_t si
 
     init_row_mxfp4x4x2((block_mxfp4 *) buf_pd, t->ne[0]);  // init padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) data + (i * row_size);
         uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
 
@@ -1249,6 +1389,25 @@ static void repack_mxfp4_mxfp4x4x2(ggml_tensor * t, const void * data, size_t si
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) t->data + (i * row_size);
+
+        // re-init the row because we are potentially copying a partial row
+        init_row_mxfp4x4x2((block_mxfp4 *) buf_pd, t->ne[0]);
+
+        // Copy only the remaining bytes from the source.
+        memcpy(buf_pd, src, n_rem_bytes);
+
+        // Repack the entire buffer (partial data + zero padding).
+        repack_row_mxfp4x4x2((uint8_t *) buf_rp, (const block_mxfp4 *) buf_pd, t->ne[0]);
+
+        // Write only the corresponding remaining bytes to the destination tensor.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -1261,6 +1420,14 @@ static void repack_mxfp4x4x2_mxfp4(void * data, const ggml_tensor * t, size_t si
     size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_MXFP4x4x2));  // extra elements for the pad
     size_t row_size_rp = row_size * 2;  // extra space for tmp pad (if any)
 
+    // Ensure we don't try to copy more data than the tensor actually contains.
+    const size_t total_tensor_size = (size_t)nrows * row_size;
+    const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;
+
+    // Calculate how many full rows and how many remaining bytes we need to process.
+    const int64_t n_full_rows = n_bytes_to_copy / row_size;
+    const size_t  n_rem_bytes = n_bytes_to_copy % row_size;
+
     void * buf_pd = ggml_aligned_malloc(row_size_pd);
     GGML_ASSERT(buf_pd != NULL);
 
@@ -1272,7 +1439,8 @@ static void repack_mxfp4x4x2_mxfp4(void * data, const ggml_tensor * t, size_t si
 
     memset(buf_pd, 0, row_size_pd);  // clear-out padded buffer to make sure the tail is all zeros
 
-    for (int64_t i = 0; i < nrows; i++) {
+    // 1. Process all the full rows
+    for (int64_t i = 0; i < n_full_rows; i++) {
         const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
         uint8_t *       dst = (uint8_t *) data + (i * row_size);
 
@@ -1281,6 +1449,20 @@ static void repack_mxfp4x4x2_mxfp4(void * data, const ggml_tensor * t, size_t si
         memcpy(dst, buf_rp, row_size);
     }
 
+    // 2. Process the final, potentially partial, row
+    if (n_rem_bytes > 0) {
+        const int64_t i = n_full_rows;
+        const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
+        uint8_t *       dst = (uint8_t *) data + (i * row_size);
+
+        // We still need to read and unpack the entire source row because the format is block-based.
+        memcpy(buf_pd, src, row_size);
+        unpack_row_mxfp4x4x2((block_mxfp4 *) buf_rp, (const uint8_t *) buf_pd, t->ne[0]);
+
+        // But we only copy the remaining number of bytes to the destination to respect the size limit.
+        memcpy(dst, buf_rp, n_rem_bytes);
+    }
+
     ggml_aligned_free(buf_pd, row_size_pd);
     ggml_aligned_free(buf_rp, row_size_rp);
 }
@@ -1299,19 +1481,19 @@ static void ggml_backend_hexagon_buffer_set_tensor(ggml_backend_buffer_t buffer,
     switch (tensor->type) {
         case GGML_TYPE_Q4_0:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_q4_0_q4x4x2(tensor, data, size);
             break;
 
         case GGML_TYPE_Q8_0:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_q8_0_q8x4x2(tensor, data, size);
             break;
 
         case GGML_TYPE_MXFP4:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_mxfp4_mxfp4x4x2(tensor, data, size);
             break;
 
@@ -1335,19 +1517,19 @@ static void ggml_backend_hexagon_buffer_get_tensor(ggml_backend_buffer_t buffer,
     switch (tensor->type) {
         case GGML_TYPE_Q4_0:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_q4x4x2_q4_0(data, tensor, size);
             break;
 
         case GGML_TYPE_Q8_0:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_q8x4x2_q8_0(data, tensor, size);
             break;
 
         case GGML_TYPE_MXFP4:
             GGML_ASSERT(offset == 0);
-            GGML_ASSERT(size == ggml_nbytes(tensor));
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
             repack_mxfp4x4x2_mxfp4(data, tensor, size);
             break;
 
@@ -1503,12 +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);
@@ -1519,8 +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);
         }
     }
 
@@ -1564,7 +1751,8 @@ void ggml_hexagon_session::allocate(int dev_id) noexcept(false) {
                           0,              // Flags
                           128 * 1024,     // Request  queue size (in bytes)
                           64 * 1024,      // Response queue size (in bytes)
-                          htp_packet_callback, htp_error_callback,
+                          nullptr,        // Read packet callback (we handle reads explicitly)
+                          nullptr,        // Error callback (we handle errors during reads)
                           (void *) this,  // Callback context
                           &queue);
     if (err != 0) {
@@ -2205,7 +2393,7 @@ static void ggml_hexagon_mul_mat(const struct ggml_tensor * op, uint32_t flags)
     bufs[0].ptr    = src0->data;
     bufs[0].offset = (uint8_t *) src0->data - src0_buf->base;
     bufs[0].size   = ggml_nbytes(src0);
-    bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_REF;
+    bufs[0].flags  = 0;
 
     // Second buffer Input Activations. This is a buffer that the CPU
     // writes and the DSP reads, so we'll need to flush CPU caches and
@@ -2215,8 +2403,7 @@ static void ggml_hexagon_mul_mat(const struct ggml_tensor * op, uint32_t flags)
     bufs[1].ptr    = src1->data;
     bufs[1].offset = (uint8_t *) src1->data - src1_buf->base;
     bufs[1].size   = ggml_nbytes(src1);
-    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                     DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                      DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
 
     // Third buffer Output Activations. We'll handle DSP
@@ -2227,7 +2414,7 @@ static void ggml_hexagon_mul_mat(const struct ggml_tensor * op, uint32_t flags)
     bufs[2].ptr    = dst->data;
     bufs[2].offset = (uint8_t *) dst->data - dst_buf->base;
     bufs[2].size   = ggml_nbytes(dst);
-    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_REF | DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
+    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
 
     // Primary DSP session from the src0 (normally weight) tensor
     auto sess = src0_buf->sess;
@@ -2255,27 +2442,7 @@ static void ggml_hexagon_mul_mat(const struct ggml_tensor * op, uint32_t flags)
     }
 
     if ((opt_opmask & HTP_OPMASK_QUEUE)) {
-        // Bump pending flag (cleared in the callback once we get the responce)
-        sess->op_pending++;  // atomic inc
-
-        int err = dspqueue_write(sess->queue,
-                                 0,                       // flags - the framework will autoset this
-                                 3,                       // number of buffers
-                                 bufs,                    // buffer references
-                                 sizeof(req),
-                                 (const uint8_t *) &req,  // Message
-                                 1000000                  // Timeout
-        );
-
-        if (err != 0) {
-            GGML_ABORT("ggml-hex: %s dspqueue_write failed: 0x%08x\n", sess->name.c_str(), (unsigned) err);
-        }
-    }
-
-    if (opt_opsync) {
-        while (sess->op_pending) {
-            ;
-        }
+        sess->enqueue(req, bufs, 3, opt_opsync);
     }
 
     t2 = ggml_time_us();
@@ -2331,7 +2498,7 @@ static void ggml_hexagon_mul_mat_id(const struct ggml_tensor * op, uint32_t flag
     bufs[0].ptr    = src0->data;
     bufs[0].offset = (uint8_t *) src0->data - src0_buf->base;
     bufs[0].size   = ggml_nbytes(src0);
-    bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_REF;
+    bufs[0].flags  = 0;
 
     // Second buffer Input Activations. This is a buffer that the CPU
     // writes and the DSP reads, so we'll need to flush CPU caches and
@@ -2341,8 +2508,7 @@ static void ggml_hexagon_mul_mat_id(const struct ggml_tensor * op, uint32_t flag
     bufs[1].ptr    = src1->data;
     bufs[1].offset = (uint8_t *) src1->data - src1_buf->base;
     bufs[1].size   = ggml_nbytes(src1);
-    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                     DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                      DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
 
     // Third buffer expert IDs. This is a buffer that the CPU
@@ -2353,8 +2519,7 @@ static void ggml_hexagon_mul_mat_id(const struct ggml_tensor * op, uint32_t flag
     bufs[2].ptr    = src2->data;
     bufs[2].offset = (uint8_t *) src2->data - src2_buf->base;
     bufs[2].size   = ggml_nbytes(src2);
-    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                     DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                      DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
 
     // Forth buffer Output Activations. We'll handle DSP
@@ -2365,7 +2530,7 @@ static void ggml_hexagon_mul_mat_id(const struct ggml_tensor * op, uint32_t flag
     bufs[3].ptr    = dst->data;
     bufs[3].offset = (uint8_t *) dst->data - dst_buf->base;
     bufs[3].size   = ggml_nbytes(dst);
-    bufs[3].flags  = (DSPQUEUE_BUFFER_FLAG_REF | DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
+    bufs[3].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
 
     // Primary DSP session from the src0 (normally weight) tensor
     auto sess = src0_buf->sess;
@@ -2394,27 +2559,7 @@ static void ggml_hexagon_mul_mat_id(const struct ggml_tensor * op, uint32_t flag
     }
 
     if ((opt_opmask & HTP_OPMASK_QUEUE)) {
-        // Bump pending flag (cleared in the callback once we get the responce)
-        sess->op_pending++;  // atomic inc
-
-        int err = dspqueue_write(sess->queue,
-                                 0,                       // flags - the framework will autoset this
-                                 4,                       // number of buffers
-                                 bufs,                    // buffer references
-                                 sizeof(req),
-                                 (const uint8_t *) &req,  // Message
-                                 1000000                  // Timeout
-        );
-
-        if (err != 0) {
-            GGML_ABORT("ggml-hex: %s dspqueue_write failed: 0x%08x\n", sess->name.c_str(), (unsigned) err);
-        }
-    }
-
-    if (opt_opsync) {
-        while (sess->op_pending) {
-            ;
-        }
+        sess->enqueue(req, bufs, 4, opt_opsync);
     }
 
     t2 = ggml_time_us();
@@ -2487,8 +2632,7 @@ static void ggml_hexagon_binary(const struct ggml_tensor * op, uint32_t flags) {
     bufs[0].ptr    = src0->data;
     bufs[0].offset = (uint8_t *) src0->data - src0_buf->base;
     bufs[0].size   = ggml_nbytes(src0);
-    bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                     DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                      DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP;
 
     // Second buffer = Second Operand of Binary op
@@ -2500,8 +2644,7 @@ static void ggml_hexagon_binary(const struct ggml_tensor * op, uint32_t flags) {
     bufs[1].ptr    = src1->data;
     bufs[1].offset = (uint8_t *) src1->data - src1_buf->base;
     bufs[1].size   = ggml_nbytes(src1);
-    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                     DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                      DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
 
     // Third buffer = Output Activations. We'll handle DSP
@@ -2512,7 +2655,7 @@ static void ggml_hexagon_binary(const struct ggml_tensor * op, uint32_t flags) {
     bufs[2].ptr    = dst->data;
     bufs[2].offset = (uint8_t *) dst->data - dst_buf->base;
     bufs[2].size   = ggml_nbytes(dst);
-    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_REF | DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
+    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
 
     // Primary DSP session from the src0 tensor
     ggml_hexagon_session * sess = src0_buf->sess;
@@ -2540,26 +2683,7 @@ static void ggml_hexagon_binary(const struct ggml_tensor * op, uint32_t flags) {
     }
 
     if ((opt_opmask & HTP_OPMASK_QUEUE)) {
-        // Bump pending flag (cleared in the callback once we get the responce)
-        sess->op_pending++;  // atomic inc
-
-        int err = dspqueue_write(sess->queue,
-                                 0,                       // flags - the framework will autoset this
-                                 3,                       // number of buffers
-                                 bufs,                    // buffer references
-                                 sizeof(req),
-                                 (const uint8_t *) &req,  // Message
-                                 1000000);                // Timeout
-
-        if (0 != err) {
-            GGML_ABORT("ggml-hex: %s dspqueue_write failed: 0x%08x\n", sess->name.c_str(), (unsigned) err);
-        }
-    }
-
-    if (opt_opsync) {
-        while (sess->op_pending) {
-            ;
-        }
+        sess->enqueue(req, bufs, 3, opt_opsync);
     }
 
     t2 = ggml_time_us();
@@ -2624,8 +2748,7 @@ static void ggml_hexagon_add_id(const struct ggml_tensor * op, uint32_t flags) {
     bufs[0].ptr    = src0->data;
     bufs[0].offset = (uint8_t *) src0->data - src0_buf->base;
     bufs[0].size   = ggml_nbytes(src0);
-    bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                     DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                      DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP;
 
     // Second buffer = experts bias
@@ -2633,8 +2756,7 @@ static void ggml_hexagon_add_id(const struct ggml_tensor * op, uint32_t flags) {
     bufs[1].ptr    = src1->data;
     bufs[1].offset = (uint8_t *) src1->data - src1_buf->base;
     bufs[1].size   = ggml_nbytes(src1);
-    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                     DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                      DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
 
     // Third buffer = activated experts
@@ -2642,8 +2764,7 @@ static void ggml_hexagon_add_id(const struct ggml_tensor * op, uint32_t flags) {
     bufs[2].ptr    = src2->data;
     bufs[2].offset = (uint8_t *) src2->data - src2_buf->base;
     bufs[2].size   = ggml_nbytes(src2);
-    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                     DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                      DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
 
     // Forth buffer = output activations
@@ -2651,7 +2772,7 @@ static void ggml_hexagon_add_id(const struct ggml_tensor * op, uint32_t flags) {
     bufs[3].ptr    = dst->data;
     bufs[3].offset = (uint8_t *) dst->data - dst_buf->base;
     bufs[3].size   = ggml_nbytes(dst);
-    bufs[3].flags  = (DSPQUEUE_BUFFER_FLAG_REF | DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
+    bufs[3].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
 
     // Primary DSP session from the src0 tensor
     ggml_hexagon_session * sess = src0_buf->sess;
@@ -2681,26 +2802,7 @@ static void ggml_hexagon_add_id(const struct ggml_tensor * op, uint32_t flags) {
     }
 
     if ((opt_opmask & HTP_OPMASK_QUEUE)) {
-        // Bump pending flag (cleared in the callback once we get the responce)
-        sess->op_pending++;  // atomic inc
-
-        int err = dspqueue_write(sess->queue,
-                                 0,                       // flags - the framework will autoset this
-                                 4,                       // number of buffers
-                                 bufs,                    // buffer references
-                                 sizeof(req),
-                                 (const uint8_t *) &req,  // Message
-                                 1000000);                // Timeout
-
-        if (0 != err) {
-            GGML_ABORT("ggml-hex: %s dspqueue_write failed: 0x%08x\n", sess->name.c_str(), (unsigned) err);
-        }
-    }
-
-    if (opt_opsync) {
-        while (sess->op_pending) {
-            ;
-        }
+        sess->enqueue(req, bufs, 4, opt_opsync);
     }
 
     t2 = ggml_time_us();
@@ -2798,8 +2900,7 @@ static void ggml_hexagon_unary(const struct ggml_tensor * op, uint32_t flags) {
     bufs[n_bufs].ptr    = src0->data;
     bufs[n_bufs].offset = (uint8_t *) src0->data - src0_buf->base;
     bufs[n_bufs].size   = ggml_nbytes(src0);
-    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                          DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                           DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP;
     ++n_bufs;
 
@@ -2814,8 +2915,7 @@ static void ggml_hexagon_unary(const struct ggml_tensor * op, uint32_t flags) {
         bufs[n_bufs].ptr    = src1->data;
         bufs[n_bufs].offset = (uint8_t *) src1->data - src1_buf->base;
         bufs[n_bufs].size   = ggml_nbytes(src1);
-        bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                              DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+        bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                               DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
         ++n_bufs;
     }
@@ -2830,7 +2930,7 @@ static void ggml_hexagon_unary(const struct ggml_tensor * op, uint32_t flags) {
     bufs[n_bufs].ptr    = dst->data;
     bufs[n_bufs].offset = (uint8_t *) dst->data - dst_buf->base;
     bufs[n_bufs].size   = ggml_nbytes(dst);
-    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_REF | DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
+    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
     ++n_bufs;
 
     // Primary DSP session from the src0 tensor
@@ -2863,26 +2963,7 @@ static void ggml_hexagon_unary(const struct ggml_tensor * op, uint32_t flags) {
     }
 
     if ((opt_opmask & HTP_OPMASK_QUEUE)) {
-        // Bump pending flag (cleared in the callback once we get the responce)
-        sess->op_pending++;  // atomic inc
-
-        int err = dspqueue_write(sess->queue,
-                                 0,                       // flags - the framework will autoset this
-                                 n_bufs,                  // number of buffers
-                                 bufs,                    // buffer references
-                                 sizeof(req),
-                                 (const uint8_t *) &req,  // Message
-                                 1000000);                // Timeout
-
-        if (0 != err) {
-            GGML_ABORT("ggml-hex: %s dspqueue_write failed: 0x%08x\n", sess->name.c_str(), (unsigned) err);
-        }
-    }
-
-    if (opt_opsync) {
-        while (sess->op_pending) {
-            ;
-        }
+        sess->enqueue(req, bufs, n_bufs, opt_opsync);
     }
 
     t2 = ggml_time_us();
@@ -2956,8 +3037,7 @@ static void ggml_hexagon_rope(const struct ggml_tensor * op, uint32_t flags) {
     bufs[n_bufs].ptr    = src0->data;
     bufs[n_bufs].offset = (uint8_t *) src0->data - src0_buf->base;
     bufs[n_bufs].size   = ggml_nbytes(src0);
-    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                          DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                           DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP;
     ++n_bufs;
 
@@ -2971,8 +3051,7 @@ static void ggml_hexagon_rope(const struct ggml_tensor * op, uint32_t flags) {
     bufs[n_bufs].ptr    = src1->data;
     bufs[n_bufs].offset = (uint8_t *) src1->data - src1_buf->base;
     bufs[n_bufs].size   = ggml_nbytes(src1);
-    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                          DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                           DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
     ++n_bufs;
 
@@ -2987,8 +3066,7 @@ static void ggml_hexagon_rope(const struct ggml_tensor * op, uint32_t flags) {
         bufs[n_bufs].ptr    = src2->data;
         bufs[n_bufs].offset = (uint8_t *) src2->data - src2_buf->base;
         bufs[n_bufs].size   = ggml_nbytes(src2);
-        bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_REF |                   // Take a reference
-                              DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush CPU
+        bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush CPU
                               DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate DSP
         ++n_bufs;
     }
@@ -3003,7 +3081,7 @@ static void ggml_hexagon_rope(const struct ggml_tensor * op, uint32_t flags) {
     bufs[n_bufs].ptr    = dst->data;
     bufs[n_bufs].offset = (uint8_t *) dst->data - dst_buf->base;
     bufs[n_bufs].size   = ggml_nbytes(dst);
-    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_REF | DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
+    bufs[n_bufs].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER);
     ++n_bufs;
 
     // Primary DSP session from the src0 tensor
@@ -3036,26 +3114,7 @@ static void ggml_hexagon_rope(const struct ggml_tensor * op, uint32_t flags) {
     }
 
     if ((opt_opmask & HTP_OPMASK_QUEUE)) {
-        // Bump pending flag (cleared in the callback once we get the responce)
-        sess->op_pending++;  // atomic inc
-
-        int err = dspqueue_write(sess->queue,
-                                 0,                       // flags - the framework will autoset this
-                                 n_bufs,                  // number of buffers
-                                 bufs,                    // buffer references
-                                 sizeof(req),
-                                 (const uint8_t *) &req,  // Message
-                                 1000000);                // Timeout
-
-        if (0 != err) {
-            GGML_ABORT("ggml-hex: %s dspqueue_write failed: 0x%08x\n", sess->name.c_str(), (unsigned) err);
-        }
-    }
-
-    if (opt_opsync) {
-        while (sess->op_pending) {
-            ;
-        }
+        sess->enqueue(req, bufs, n_bufs, opt_opsync);
     }
 
     t2 = ggml_time_us();
@@ -3200,9 +3259,7 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
     }
 
     // Wait until all pending ops complete
-    while (sess->op_pending) {
-        ;
-    }
+    sess->flush();
 
     return GGML_STATUS_SUCCESS;
 }
@@ -3213,9 +3270,7 @@ static void ggml_backend_hexagon_synchronize(ggml_backend_t backend) {
     HEX_VERBOSE("ggml-hex: %s synchronize\n", sess->name.c_str());
 
     // Wait until all pending ops complete
-    while (sess->op_pending) {
-        ;
-    }
+    sess->flush();
 }
 
 struct node_info {
@@ -3624,6 +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-hexagon/htp/main.c

@@ -395,28 +395,14 @@ static void proc_matmul_req(struct htp_context *     ctx,
                             struct htp_general_req * req,
                             struct dspqueue_buffer * bufs,
                             size_t                   n_bufs) {
-    // Prep response buffer structs (needed for error responses, etc)
-    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
-    memset(rsp_bufs, 0, sizeof(rsp_bufs));
-    rsp_bufs[0].fd     = bufs[0].fd;
-    rsp_bufs[0].ptr    = bufs[0].ptr;
-    rsp_bufs[0].size   = bufs[0].size;
-    rsp_bufs[0].offset = bufs[0].offset;
-    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-    rsp_bufs[1].fd     = bufs[1].fd;
-    rsp_bufs[1].ptr    = bufs[1].ptr;
-    rsp_bufs[1].size   = bufs[1].size;
-    rsp_bufs[1].offset = bufs[1].offset;
-    rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+    struct dspqueue_buffer rsp_bufs[1];
 
     // We had written to the output buffer, we'd also need to flush it
-    rsp_bufs[2].fd     = bufs[2].fd;
-    rsp_bufs[2].ptr    = bufs[2].ptr;
-    rsp_bufs[2].size   = bufs[2].size;
-    rsp_bufs[2].offset = bufs[2].offset;
-    rsp_bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
-                         DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+    rsp_bufs[0].fd     = bufs[2].fd;
+    rsp_bufs[0].ptr    = bufs[2].ptr;
+    rsp_bufs[0].size   = bufs[2].size;
+    rsp_bufs[0].offset = bufs[2].offset;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
                          DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
 
     // Setup Op context
@@ -444,41 +430,21 @@ static void proc_matmul_req(struct htp_context *     ctx,
     }
 
     profile_stop(&prof);
-    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 3, &prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
 static void proc_matmul_id_req(struct htp_context *     ctx,
                                struct htp_general_req * req,
                                struct dspqueue_buffer * bufs,
                                size_t                   n_bufs) {
-    // Prep response buffer structs (needed for error responses, etc)
-    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
-    memset(rsp_bufs, 0, sizeof(rsp_bufs));
-    rsp_bufs[0].fd     = bufs[0].fd;
-    rsp_bufs[0].ptr    = bufs[0].ptr;
-    rsp_bufs[0].size   = bufs[0].size;
-    rsp_bufs[0].offset = bufs[0].offset;
-    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-    rsp_bufs[1].fd     = bufs[1].fd;
-    rsp_bufs[1].ptr    = bufs[1].ptr;
-    rsp_bufs[1].size   = bufs[1].size;
-    rsp_bufs[1].offset = bufs[1].offset;
-    rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-    rsp_bufs[2].fd     = bufs[2].fd;
-    rsp_bufs[2].ptr    = bufs[2].ptr;
-    rsp_bufs[2].size   = bufs[2].size;
-    rsp_bufs[2].offset = bufs[2].offset;
-    rsp_bufs[2].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+    struct dspqueue_buffer rsp_bufs[1];
 
     // We had written to the output buffer, we'd also need to flush it
-    rsp_bufs[3].fd     = bufs[3].fd;
-    rsp_bufs[3].ptr    = bufs[3].ptr;
-    rsp_bufs[3].size   = bufs[3].size;
-    rsp_bufs[3].offset = bufs[3].offset;
-    rsp_bufs[3].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
-                         DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+    rsp_bufs[0].fd     = bufs[3].fd;
+    rsp_bufs[0].ptr    = bufs[3].ptr;
+    rsp_bufs[0].size   = bufs[3].size;
+    rsp_bufs[0].offset = bufs[3].offset;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
                          DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
 
     // Setup Op context
@@ -508,32 +474,18 @@ static void proc_matmul_id_req(struct htp_context *     ctx,
     }
 
     profile_stop(&prof);
-    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 4, &prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
 static void proc_binary_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
-    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
-    memset(rsp_bufs, 0, sizeof(rsp_bufs));
-
-    rsp_bufs[0].fd     = bufs[0].fd;
-    rsp_bufs[0].ptr    = bufs[0].ptr;
-    rsp_bufs[0].offset = bufs[0].offset;
-    rsp_bufs[0].size   = bufs[0].size;
-    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-    rsp_bufs[1].fd     = bufs[1].fd;
-    rsp_bufs[1].ptr    = bufs[1].ptr;
-    rsp_bufs[1].offset = bufs[1].offset;
-    rsp_bufs[1].size   = bufs[1].size;
-    rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+    struct dspqueue_buffer rsp_bufs[1];
 
     // We had written to the output buffer, we'd also need to flush it
-    rsp_bufs[2].fd     = bufs[2].fd;
-    rsp_bufs[2].ptr    = bufs[2].ptr;
-    rsp_bufs[2].offset = bufs[2].offset;
-    rsp_bufs[2].size   = bufs[2].size;
-    rsp_bufs[2].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
-                         DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+    rsp_bufs[0].fd     = bufs[2].fd;
+    rsp_bufs[0].ptr    = bufs[2].ptr;
+    rsp_bufs[0].offset = bufs[2].offset;
+    rsp_bufs[0].size   = bufs[2].size;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
                          DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
 
     // Setup Op context
@@ -561,38 +513,18 @@ static void proc_binary_req(struct htp_context * ctx, struct htp_general_req * r
     }
 
     profile_stop(&prof);
-    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 3, &prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
 static void proc_add_id_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
-    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
-    memset(rsp_bufs, 0, sizeof(rsp_bufs));
-
-    rsp_bufs[0].fd     = bufs[0].fd;
-    rsp_bufs[0].ptr    = bufs[0].ptr;
-    rsp_bufs[0].offset = bufs[0].offset;
-    rsp_bufs[0].size   = bufs[0].size;
-    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-    rsp_bufs[1].fd     = bufs[1].fd;
-    rsp_bufs[1].ptr    = bufs[1].ptr;
-    rsp_bufs[1].offset = bufs[1].offset;
-    rsp_bufs[1].size   = bufs[1].size;
-    rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-    rsp_bufs[2].fd     = bufs[2].fd;
-    rsp_bufs[2].ptr    = bufs[2].ptr;
-    rsp_bufs[2].offset = bufs[2].offset;
-    rsp_bufs[2].size   = bufs[2].size;
-    rsp_bufs[2].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
+    struct dspqueue_buffer rsp_bufs[1];
 
     // We had written to the output buffer, we'd also need to flush it
-    rsp_bufs[3].fd     = bufs[3].fd;
-    rsp_bufs[3].ptr    = bufs[3].ptr;
-    rsp_bufs[3].offset = bufs[3].offset;
-    rsp_bufs[3].size   = bufs[3].size;
-    rsp_bufs[3].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
-                         DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+    rsp_bufs[0].fd     = bufs[3].fd;
+    rsp_bufs[0].ptr    = bufs[3].ptr;
+    rsp_bufs[0].offset = bufs[3].offset;
+    rsp_bufs[0].size   = bufs[3].size;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
                          DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
 
     // Setup Op context
@@ -622,26 +554,18 @@ static void proc_add_id_req(struct htp_context * ctx, struct htp_general_req * r
     }
 
     profile_stop(&prof);
-    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 4, &prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
 static void proc_unary_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
     struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
-    memset(rsp_bufs, 0, sizeof(rsp_bufs));
-
-    rsp_bufs[0].fd     = bufs[0].fd;
-    rsp_bufs[0].ptr    = bufs[0].ptr;
-    rsp_bufs[0].offset = bufs[0].offset;
-    rsp_bufs[0].size   = bufs[0].size;
-    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
 
     // We had written to the output buffer, we'd also need to flush it
-    rsp_bufs[1].fd     = bufs[1].fd;
-    rsp_bufs[1].ptr    = bufs[1].ptr;
-    rsp_bufs[1].offset = bufs[1].offset;
-    rsp_bufs[1].size   = bufs[1].size;
-    rsp_bufs[1].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
-                         DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
+    rsp_bufs[0].fd     = bufs[1].fd;
+    rsp_bufs[0].ptr    = bufs[1].ptr;
+    rsp_bufs[0].offset = bufs[1].offset;
+    rsp_bufs[0].size   = bufs[1].size;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
                          DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
 
     // Setup Op context
@@ -669,7 +593,7 @@ static void proc_unary_req(struct htp_context * ctx, struct htp_general_req * re
     }
 
     profile_stop(&prof);
-    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 2, &prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
 static void proc_activations_req(struct htp_context *     ctx,
@@ -677,33 +601,16 @@ static void proc_activations_req(struct htp_context *     ctx,
                                  struct dspqueue_buffer * bufs,
                                  uint32_t                 n_bufs) {
     struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
-    memset(rsp_bufs, 0, sizeof(rsp_bufs));
 
-    rsp_bufs[0].fd     = bufs[0].fd;
-    rsp_bufs[0].ptr    = bufs[0].ptr;
-    rsp_bufs[0].offset = bufs[0].offset;
-    rsp_bufs[0].size   = bufs[0].size;
-    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-    int write_idx = 1;
-    if (3 == n_bufs) {
-        rsp_bufs[1].fd     = bufs[1].fd;
-        rsp_bufs[1].ptr    = bufs[1].ptr;
-        rsp_bufs[1].offset = bufs[1].offset;
-        rsp_bufs[1].size   = bufs[1].size;
-        rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-        write_idx = 2;
-    }
+    int write_idx = (n_bufs == 3) ? 2 : 1;
 
     // We had written to the output buffer, we'd also need to flush it
-    rsp_bufs[write_idx].fd     = bufs[write_idx].fd;
-    rsp_bufs[write_idx].ptr    = bufs[write_idx].ptr;
-    rsp_bufs[write_idx].offset = bufs[write_idx].offset;
-    rsp_bufs[write_idx].size   = bufs[write_idx].size;
-    rsp_bufs[write_idx].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
-                                 DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
-                                 DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+    rsp_bufs[0].fd     = bufs[write_idx].fd;
+    rsp_bufs[0].ptr    = bufs[write_idx].ptr;
+    rsp_bufs[0].offset = bufs[write_idx].offset;
+    rsp_bufs[0].size   = bufs[write_idx].size;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
+                          DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT); // Invalidate CPU
 
     // Setup Op context
     struct htp_ops_context octx = { 0 };
@@ -742,7 +649,7 @@ static void proc_activations_req(struct htp_context *     ctx,
     }
 
     profile_stop(&prof);
-    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, n_bufs, &prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
 static void proc_rope_req(struct htp_context *     ctx,
@@ -750,39 +657,16 @@ static void proc_rope_req(struct htp_context *     ctx,
                           struct dspqueue_buffer * bufs,
                           uint32_t                 n_bufs) {
     struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
-    memset(rsp_bufs, 0, sizeof(rsp_bufs));
 
-    rsp_bufs[0].fd     = bufs[0].fd;
-    rsp_bufs[0].ptr    = bufs[0].ptr;
-    rsp_bufs[0].offset = bufs[0].offset;
-    rsp_bufs[0].size   = bufs[0].size;
-    rsp_bufs[0].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-    rsp_bufs[1].fd     = bufs[1].fd;
-    rsp_bufs[1].ptr    = bufs[1].ptr;
-    rsp_bufs[1].offset = bufs[1].offset;
-    rsp_bufs[1].size   = bufs[1].size;
-    rsp_bufs[1].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-    int write_idx = 2;
-    if (4 == n_bufs) {
-        rsp_bufs[write_idx].fd     = bufs[write_idx].fd;
-        rsp_bufs[write_idx].ptr    = bufs[write_idx].ptr;
-        rsp_bufs[write_idx].offset = bufs[write_idx].offset;
-        rsp_bufs[write_idx].size   = bufs[write_idx].size;
-        rsp_bufs[write_idx].flags  = DSPQUEUE_BUFFER_FLAG_DEREF;  // Release reference
-
-        write_idx++;
-    }
+    int write_idx = (n_bufs == 4) ? 3 : 2;
 
     // We had written to the output buffer, we'd also need to flush it
-    rsp_bufs[write_idx].fd     = bufs[write_idx].fd;
-    rsp_bufs[write_idx].ptr    = bufs[write_idx].ptr;
-    rsp_bufs[write_idx].offset = bufs[write_idx].offset;
-    rsp_bufs[write_idx].size   = bufs[write_idx].size;
-    rsp_bufs[write_idx].flags  = (DSPQUEUE_BUFFER_FLAG_DEREF |                 // Release reference
-                                 DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |          // Flush NSP
-                                 DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT);  // Invalidate CPU
+    rsp_bufs[0].fd     = bufs[write_idx].fd;
+    rsp_bufs[0].ptr    = bufs[write_idx].ptr;
+    rsp_bufs[0].offset = bufs[write_idx].offset;
+    rsp_bufs[0].size   = bufs[write_idx].size;
+    rsp_bufs[0].flags  = (DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER |         // Flush HTP
+                          DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT); // Invalidate CPU
 
     // Setup Op context
     struct htp_ops_context octx = { 0 };
@@ -819,7 +703,7 @@ static void proc_rope_req(struct htp_context *     ctx,
     }
 
     profile_stop(&prof);
-    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, n_bufs, &prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }
 
 static void htp_packet_callback(dspqueue_t queue, int error, void * context) {

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

@@ -43,46 +43,46 @@ static inline int32_t htp_is_one_chunk(void * addr, uint32_t n, uint32_t chunk_s
 }
 
 static inline void htp_dump_int8_line(char * pref, const int8_t * x, int n) {
-    char str[1024], *p = str;
-    p += sprintf(p, "%s: ", pref);
-    for (int i = 0; i < 16; i++) {
-        p += sprintf(p, "%d, ", x[i]);
+    char str[1024], *p = str, *p_end = str + sizeof(str);
+    p += snprintf(p, p_end - p, "%s: ", pref);
+    for (int i = 0; i < n && p < p_end; i++) {
+        p += snprintf(p, p_end - p, "%d, ", x[i]);
     }
     FARF(HIGH, "%s\n", str);
 }
 
 static inline void htp_dump_uint8_line(char * pref, const uint8_t * x, uint32_t n) {
-    char str[1024], *p = str;
-    p += sprintf(p, "%s: ", pref);
-    for (int i = 0; i < n; i++) {
-        p += sprintf(p, "%d, ", x[i]);
+    char str[1024], *p = str, *p_end = str + sizeof(str);
+    p += snprintf(p, p_end - p, "%s: ", pref);
+    for (int i = 0; i < n && p < p_end; i++) {
+        p += snprintf(p, p_end - p, "%d, ", x[i]);
     }
     FARF(HIGH, "%s\n", str);
 }
 
 static inline void htp_dump_int32_line(char * pref, const int32_t * x, uint32_t n) {
-    char str[1024], *p = str;
-    p += sprintf(p, "%s: ", pref);
+    char str[1024], *p = str, *p_end = str + sizeof(str);
+    p += snprintf(p, p_end - p, "%s: ", pref);
     for (int i = 0; i < n; i++) {
-        p += sprintf(p, "%d, ", (int) x[i]);
+        p += snprintf(p, p_end - p, "%d, ", (int) x[i]);
     }
     FARF(HIGH, "%s\n", str);
 }
 
 static inline void htp_dump_fp16_line(char * pref, const __fp16 * x, uint32_t n) {
-    char str[1024], *p = str;
-    p += sprintf(p, "%s: ", pref);
+    char str[1024], *p = str, *p_end = str + sizeof(str);
+    p += snprintf(p, p_end - p, "%s: ", pref);
     for (int i = 0; i < n; i++) {
-        p += sprintf(p, "%.6f, ", (float) x[i]);
+        p += snprintf(p, p_end - p, "%.6f, ", (float) x[i]);
     }
     FARF(HIGH, "%s\n", str);
 }
 
 static inline void htp_dump_fp32_line(char * pref, const float * x, uint32_t n) {
-    char str[1024], *p = str;
-    p += sprintf(p, "%s: ", pref);
+    char str[1024], *p = str, *p_end = str + sizeof(str);
+    p += snprintf(p, p_end - p, "%s: ", pref);
     for (int i = 0; i < n; i++) {
-        p += sprintf(p, "%.6f, ", x[i]);
+        p += snprintf(p, p_end - p, "%.6f, ", x[i]);
     }
     FARF(HIGH, "%s\n", str);
 }

ggml/src/ggml-impl.h

@@ -682,6 +682,7 @@ static inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph * cgraph,
 #endif
 
 #ifdef __cplusplus
+#include <array>
 #include <initializer_list>
 #include <vector>
 
@@ -697,6 +698,21 @@ inline bool ggml_can_fuse_subgraph(const struct ggml_cgraph *          cgraph,
     return ggml_can_fuse_subgraph(cgraph, start_idx, ops.size(), ops.begin(), outputs.begin(), outputs.size());
 }
 
+// Return true if the edges in the graph match expectations.
+inline bool ggml_check_edges(const struct ggml_cgraph *                cgraph,
+                             int                                       start_idx,
+                             std::initializer_list<std::array<int, 3>> edges) {
+    for (const auto & edge : edges) {
+        int dst_node = edge[0];
+        int src_idx  = edge[1];
+        int src_node = edge[2];
+        if (cgraph->nodes[start_idx + dst_node]->src[src_idx] != cgraph->nodes[start_idx + src_node]) {
+            return false;
+        }
+    }
+    return true;
+}
+
 // expose GGUF internals for test code
 GGML_API size_t gguf_type_size(enum gguf_type type);
 GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);

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.cpp

@@ -677,7 +677,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id_map0(ggml_metal_
     char name[256];
 
     snprintf(base, 256, "kernel_mul_mm_id_map0_ne20_%d", ne20);
-    snprintf(name, 256, "%s", base);
+    snprintf(name, 256, "%s_ne02=%d", base, ne02);
 
     ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
     if (res) {
@@ -1332,11 +1332,12 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rope(ggml_metal_library_t
 
     const bool is_neox   = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope  = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_neox) {
         snprintf(base, 256, "kernel_rope_neox_%s", ggml_type_name(op->src[0]->type));
-    } else if (is_mrope && !is_vision) {
+    } else if ((is_mrope || is_imrope) && !is_vision) {
         GGML_ASSERT(op->src[1]->ne[0]*4 >= op->src[0]->ne[2]); // need at least 4 pos per token
         snprintf(base, 256, "kernel_rope_multi_%s", ggml_type_name(op->src[0]->type));
     } else if (is_vision) {
@@ -1346,14 +1347,20 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rope(ggml_metal_library_t
         snprintf(base, 256, "kernel_rope_norm_%s", ggml_type_name(op->src[0]->type));
     }
 
-    snprintf(name, 256, "%s", base);
+    snprintf(name, 256, "%s_imrope=%d", base, is_imrope ? 1 : 0);
 
     ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
     if (res) {
         return res;
     }
 
-    res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+    ggml_metal_cv_set_bool(cv, is_imrope, FC_ROPE + 0);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);
 
     return res;
 }

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");
 
@@ -707,6 +906,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             if (op->src[0]->ne[0] != 32 &&
                 op->src[0]->ne[0] != 40 &&
                 op->src[0]->ne[0] != 64 &&
+                op->src[0]->ne[0] != 72 &&
                 op->src[0]->ne[0] != 80 &&
                 op->src[0]->ne[0] != 96 &&
                 op->src[0]->ne[0] != 112 &&

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

@@ -76,6 +76,7 @@
 #define FC_FLASH_ATTN_EXT_VEC_REDUCE   500
 #define FC_MUL_MV                      600
 #define FC_MUL_MM                      700
+#define FC_ROPE                        800
 
 // op-specific constants
 #define OP_FLASH_ATTN_EXT_NQPTG 8

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];
     }
 
@@ -3709,6 +3715,8 @@ template [[host_name("kernel_mul_mv_bf16_f32_short")]]  kernel mul_mv_t_t_short_
 template [[host_name("kernel_mul_mv_bf16_bf16_short")]] kernel mul_mv_t_t_short_t kernel_mul_mv_t_t_short<bfloat, bfloat>;
 #endif
 
+constant bool FC_rope_is_imrope [[function_constant(FC_ROPE + 0)]];
+
 static float rope_yarn_ramp(const float low, const float high, const int i0) {
     const float y = (i0 / 2 - low) / max(0.001f, high - low);
     return 1.0f - min(1.0f, max(0.0f, y));
@@ -3889,14 +3897,26 @@ kernel void kernel_rope_multi(
             const int sector    = ic % sect_dims;
 
             float theta_base;
-            if (sector < args.sect_0) {
-                theta_base = (float) pos[i2];
-            } else if (sector < sec_w01) {
-                theta_base = (float) pos[i2 + args.ne02];
-            } else if (sector < sec_w012) {
-                theta_base = (float) pos[i2 + args.ne02 * 2];
+            if (FC_rope_is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * args.sect_1) { // h
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * args.sect_2) { // w
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                    theta_base = (float) pos[i2 + args.ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
             } else {
-                theta_base = (float) pos[i2 + args.ne02 * 3];
+                if (sector < args.sect_0) {
+                    theta_base = (float) pos[i2];
+                } else if (sector < sec_w01) {
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector < sec_w012) {
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else {
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
             }
             // end of mrope
 
@@ -5348,6 +5368,7 @@ typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, hal
 template [[host_name("kernel_flash_attn_ext_f32_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  32,  32>;
 template [[host_name("kernel_flash_attn_ext_f32_dk40_dv40"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  40,  40>;
 template [[host_name("kernel_flash_attn_ext_f32_dk64_dv64"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  64,  64>;
+template [[host_name("kernel_flash_attn_ext_f32_dk72_dv72"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  72,  72>;
 template [[host_name("kernel_flash_attn_ext_f32_dk80_dv80"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  80,  80>;
 template [[host_name("kernel_flash_attn_ext_f32_dk96_dv96"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  96,  96>;
 template [[host_name("kernel_flash_attn_ext_f32_dk112_dv112")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  112, 112>;
@@ -5360,6 +5381,7 @@ template [[host_name("kernel_flash_attn_ext_f32_dk576_dv512")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  32,  32>;
 template [[host_name("kernel_flash_attn_ext_f16_dk40_dv40"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  40,  40>;
 template [[host_name("kernel_flash_attn_ext_f16_dk64_dv64"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  64,  64>;
+template [[host_name("kernel_flash_attn_ext_f16_dk72_dv72"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  72,  72>;
 template [[host_name("kernel_flash_attn_ext_f16_dk80_dv80"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  80,  80>;
 template [[host_name("kernel_flash_attn_ext_f16_dk96_dv96"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  96,  96>;
 template [[host_name("kernel_flash_attn_ext_f16_dk112_dv112")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  112, 112>;
@@ -5373,6 +5395,7 @@ template [[host_name("kernel_flash_attn_ext_f16_dk576_dv512")]]  kernel flash_at
 template [[host_name("kernel_flash_attn_ext_bf16_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_bf16_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
@@ -5386,6 +5409,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q4_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 112, 112>;
@@ -5398,6 +5422,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q4_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q4_1_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 112, 112>;
@@ -5410,6 +5435,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q5_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 112, 112>;
@@ -5422,6 +5448,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q5_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q5_1_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 112, 112>;
@@ -5434,6 +5461,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_dk576_dv512")]] kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 40,  40>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_q8_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80,  80>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk96_dv96"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 96,  96>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk112_dv112")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 112, 112>;
@@ -5445,6 +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)]];
@@ -6066,6 +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)]];
@@ -8119,17 +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(
@@ -8145,18 +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];
 
@@ -8165,36 +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;
@@ -8203,91 +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);
                 }
             }
@@ -8372,31 +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];
 
@@ -8405,39 +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;
@@ -8446,85 +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-opencl/ggml-opencl.cpp

@@ -8399,6 +8399,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     const bool is_neox = mode & 2;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+    const int  is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
 
     if (is_mrope) {
         GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
@@ -8489,9 +8490,14 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     CL_CHECK(clSetKernelArg(kernel, 30, sizeof(float),    &attn_factor));
     CL_CHECK(clSetKernelArg(kernel, 31, sizeof(float),    &beta_fast));
     CL_CHECK(clSetKernelArg(kernel, 32, sizeof(float),    &beta_slow));
+    // both mrope and vision kernels have sections
     if (is_mrope || is_vision) {
         CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int32_t)*4, &sections));
     }
+    // only mrope has is_imrope
+    if (is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 34, sizeof(int), &is_imrope));
+    }
 
     size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
     size_t local_work_size[] = {(size_t)nth, 1, 1};

ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl

@@ -79,8 +79,8 @@ kernel void kernel_mul_mm_f16_f32_l4_lm(
 
     for (int block = 0; block < ne00; block += BK) {
         for (int l = 0; l < BM; l += loadstride_a) {
-            if (loadc_a + l < ne01) {
-            const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+            if (ir*BM + loadc_a + l < ne01) {
+                const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
                 buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0;
                 buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1;
                 buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = src0[idx].s2;
@@ -94,7 +94,7 @@ kernel void kernel_mul_mm_f16_f32_l4_lm(
         }
 
         for (int l = 0; l < BN; l += loadstride_b) {
-            if (loadc_b + l < ne11) {
+            if (ic*BN + loadc_b + l < ne11) {
                 const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
                 buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
                 buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;

ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl

@@ -79,7 +79,7 @@ kernel void kernel_mul_mm_f32_f32_l4_lm(
 
     for (int block = 0; block < ne00; block += BK) {
         for (int l = 0; l < BM; l += loadstride_a) {
-            if (loadc_a + l < ne01) {
+            if (ir*BM + loadc_a + l < ne01) {
                 const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
                 buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0;
                 buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1;
@@ -94,7 +94,7 @@ kernel void kernel_mul_mm_f32_f32_l4_lm(
         }
 
         for (int l = 0; l < BN; l += loadstride_b) {
-            if (loadc_b + l < ne11) {
+            if (ic*BN + loadc_b + l < ne11) {
                 const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
                 buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
                 buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;

ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl

@@ -78,7 +78,7 @@ kernel void kernel_mul_mm_q8_0_f32_l4_lm(
 
     for (int block = 0; block < ne00; block += BK) {
         for (int l = 0; l < BM; l += loadstride_a) {
-            if (loadc_a + l < ne01) {
+            if (ir*BM + loadc_a + l < ne01) {
                 int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
                 int ib  = idx / 8;
                 int iqs = idx % 8;
@@ -101,7 +101,7 @@ kernel void kernel_mul_mm_q8_0_f32_l4_lm(
         }
 
         for (int l = 0; l < BN; l += loadstride_b) {
-            if (loadc_b + l < ne11) {
+            if (ic*BN + loadc_b + l < ne11) {
                 int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
                 buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
                 buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;

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

@@ -392,7 +392,8 @@ kernel void kernel_rope_multi_f32(
         float attn_factor,
         float beta_fast,
         float beta_slow,
-        int4 sections
+        int4 sections,
+        int  is_imrope
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -419,17 +420,29 @@ kernel void kernel_rope_multi_f32(
             const int sector = (i0 / 2) % sect_dims;
             float theta_base = 0.0f;
 
-            if (sector < sections.s0) {
-                theta_base = pos[i2];
-            }
-            else if (sector >= sections.s0 && sector < sec_w) {
-                theta_base = pos[i2 + ne2 * 1];
-            }
-            else if (sector >= sec_w && sector < sec_w + sections.s2) {
-                theta_base = pos[i2 + ne2 * 2];
-            }
-            else if (sector >= sec_w + sections.s2) {
-                theta_base = pos[i2 + ne2 * 3];
+            if (is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * sections.s1) { // h
+                    theta_base = (float) pos[i2 + ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * sections.s2) { // w
+                    theta_base = (float) pos[i2 + ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * sections.s0) { // t
+                    theta_base = (float) pos[i2 + ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + ne02 * 3];
+                }
+            } else {
+                if (sector < sections.s0) {
+                    theta_base = pos[i2];
+                }
+                else if (sector >= sections.s0 && sector < sec_w) {
+                    theta_base = pos[i2 + ne2 * 1];
+                }
+                else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 2];
+                }
+                else if (sector >= sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 3];
+                }
             }
 
             const float theta = theta_base * pow(freq_base, inv_ndims*i0);
@@ -490,7 +503,8 @@ kernel void kernel_rope_multi_f16(
         float attn_factor,
         float beta_fast,
         float beta_slow,
-        int4 sections
+        int4 sections,
+        int  is_imrope
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -517,17 +531,29 @@ kernel void kernel_rope_multi_f16(
             const int sector = (i0 / 2) % sect_dims;
             float theta_base = 0.0f;
 
-            if (sector < sections.s0) {
-                theta_base = pos[i2];
-            }
-            else if (sector >= sections.s0 && sector < sec_w) {
-                theta_base = pos[i2 + ne2 * 1];
-            }
-            else if (sector >= sec_w && sector < sec_w + sections.s2) {
-                theta_base = pos[i2 + ne2 * 2];
-            }
-            else if (sector >= sec_w + sections.s2) {
-                theta_base = pos[i2 + ne2 * 3];
+            if (is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * sections.s1) { // h
+                    theta_base = (float) pos[i2 + ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * sections.s2) { // w
+                    theta_base = (float) pos[i2 + ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * sections.s0) { // t
+                    theta_base = (float) pos[i2 + ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + ne02 * 3];
+                }
+            } else {
+                if (sector < sections.s0) {
+                    theta_base = pos[i2];
+                }
+                else if (sector >= sections.s0 && sector < sec_w) {
+                    theta_base = pos[i2 + ne2 * 1];
+                }
+                else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 2];
+                }
+                else if (sector >= sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 3];
+                }
             }
 
             const float theta = theta_base * pow(freq_base, inv_ndims*i0);

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

@@ -42,6 +42,7 @@
 #include "ggml-sycl/backend.hpp"
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/element_wise.hpp"
+#include "ggml-sycl/norm.hpp"
 #include "ggml-sycl/presets.hpp"
 #include "ggml-sycl/gemm.hpp"
 #include "ggml-sycl/set_rows.hpp"
@@ -2637,6 +2638,11 @@ static void ggml_sycl_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * ds
     ggml_sycl_op_rms_norm(ctx, dst);
 }
 
+static void ggml_sycl_rms_norm_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    ggml_sycl_op_rms_norm_back(ctx, dst);
+}
+
 static void ggml_sycl_l2_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     ggml_sycl_op_l2_norm(ctx, dst);
@@ -3827,6 +3833,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_LEAKY_RELU:
             ggml_sycl_leaky_relu(ctx, dst);
             break;
+        case GGML_OP_RMS_NORM_BACK:
+            ggml_sycl_rms_norm_back(ctx, dst);
+            break;
         case GGML_OP_RMS_NORM:
             ggml_sycl_rms_norm(ctx, dst);
             break;
@@ -4525,16 +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:
@@ -4571,6 +4576,8 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_RMS_NORM:
             return ((op->src[0]->ne[0] % WARP_SIZE) == 0);
+        case GGML_OP_RMS_NORM_BACK:
+            return ((op->src[0]->ne[0] % WARP_SIZE) == 0);
         case GGML_OP_SCALE:
             return true;
         case GGML_OP_CONT:

ggml/src/ggml-sycl/norm.cpp

@@ -480,6 +480,162 @@ void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     rms_norm_f32_sycl(src0_dd, dst_dd, ne00, ne01, ne02, ne03, s01, s02, s03, eps, main_stream, ctx.device);
 }
 
+void ggml_sycl_op_rms_norm_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); // dz
+    GGML_ASSERT(dst->src[1]->type == GGML_TYPE_F32); // x
+    GGML_ASSERT(dst->type         == GGML_TYPE_F32);
+
+    float eps = 1e-5f;
+    std::memcpy(&eps, dst->op_params, sizeof(float));
+    if (!(eps > 0.0f) || !std::isfinite(eps)) eps = 1e-5f;
+
+    const float * g_base  = static_cast<const float *>(dst->src[0]->data); // dz
+    const float * x_base  = static_cast<const float *>(dst->src[1]->data); // x
+          float * dx_base = static_cast<      float *>(dst->data);
+
+    const int64_t D  = dst->ne[0];
+    const int64_t n1 = dst->ne[1], n2 = dst->ne[2], n3 = dst->ne[3]; (void) n3;
+    const int64_t N  = ggml_nrows(dst);
+    if (D == 0 || N == 0) return;
+
+    const ggml_tensor *G = dst->src[0];
+    const ggml_tensor *X = dst->src[1];
+    const int ts = (int) ggml_type_size(X->type);
+    GGML_ASSERT((size_t) X->nb[0]   == (size_t) ts);
+    GGML_ASSERT((size_t) G->nb[0]   == (size_t) ts);
+    GGML_ASSERT((size_t) dst->nb[0] == (size_t) ts);
+
+    const int64_t xs1 = X->nb[1] / ts, xs2 = X->nb[2] / ts, xs3 = X->nb[3] / ts;
+    const int64_t gs1 = G->nb[1] / ts, gs2 = G->nb[2] / ts, gs3 = G->nb[3] / ts;
+    const int64_t ds1 = dst->nb[1] / ts, ds2 = dst->nb[2] / ts, ds3 = dst->nb[3] / ts;
+
+    dpct::queue_ptr q = ctx.stream();
+
+    // work-group size: multiple of WARP_SIZE, capped by device and 256, and not larger than D
+    const int device_max_wg = ggml_sycl_info().max_work_group_sizes[ctx.device];
+    auto roundup = [](int v, int m) { return ((v + m - 1) / m) * m; };
+    int wg_cap = 256;
+    if (device_max_wg > 0) wg_cap = std::min(wg_cap, device_max_wg);
+    int WG = std::max(WARP_SIZE, std::min(roundup((int)std::min<int64_t>(D, wg_cap), WARP_SIZE), wg_cap));
+
+    // FP32 path: per-thread compensated accumulation + hierarchical reduction
+    q->submit([&](sycl::handler &cgh) {
+        const int nwarps_loc = std::max(1, WG / WARP_SIZE);
+        // store one partial value per warp (xx and xg) for cross-warp reduction
+        auto l_xx   = sycl::local_accessor<sycl::float2, 1>(sycl::range<1>(nwarps_loc), cgh);
+        auto l_xg   = sycl::local_accessor<sycl::float2, 1>(sycl::range<1>(nwarps_loc), cgh);
+
+        cgh.parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, N) * sycl::range<3>(1, 1, WG),
+                              sycl::range<3>(1, 1, WG)),
+            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                const int row = item_ct1.get_group(2);
+                const int tid = item_ct1.get_local_id(2);
+
+                const int64_t i1 = row % n1;
+                const int64_t i2 = (row / n1) % n2;
+                const int64_t i3 = row / (n1 * n2);
+
+                const float *__restrict x_row = x_base + i3 * xs3 + i2 * xs2 + i1 * xs1;
+                const float *__restrict g_row = g_base + i3 * gs3 + i2 * gs2 + i1 * gs1;
+                float *__restrict d_row       = dx_base + i3 * ds3 + i2 * ds2 + i1 * ds1;
+
+                // per-thread accumulation (compensated by default)
+                float sum_xx = 0.f, sum_xg = 0.f;
+#ifndef GGML_SYCL_RMS_BACK_FAST
+                float c_xx = 0.f, c_xg = 0.f;
+#endif
+                for (int64_t col = tid; col < D; col += WG) {
+                    const float xv = x_row[col];
+                    const float gv = g_row[col];
+#ifdef GGML_SYCL_RMS_BACK_FAST
+                    sum_xx += xv * xv;
+                    sum_xg += xv * gv;
+#else
+                    float y1 = xv * xv - c_xx;
+                    float t1 = sum_xx + y1;
+                    c_xx = (t1 - sum_xx) - y1;
+                    sum_xx = t1;
+
+                    float y2 = xv * gv - c_xg;
+                    float t2 = sum_xg + y2;
+                    c_xg = (t2 - sum_xg) - y2;
+                    sum_xg = t2;
+#endif
+                }
+
+                // warp-level reduction
+                sycl::float2 xx = sycl::float2(sum_xx,
+#ifndef GGML_SYCL_RMS_BACK_FAST
+                    c_xx
+#else
+                    0.f
+#endif
+                );
+                sycl::float2 xg = sycl::float2(sum_xg,
+#ifndef GGML_SYCL_RMS_BACK_FAST
+                    c_xg
+#else
+                    0.f
+#endif
+                );
+                xx = warp_reduce_sum(xx, item_ct1);
+                xg = warp_reduce_sum(xg, item_ct1);
+
+                // cross-warp reduction using local memory (single barrier)
+                const auto sub_group = item_ct1.get_sub_group();
+                const auto sg_id     = sub_group.get_group_linear_id();
+                const auto wi_in_sg  = sub_group.get_local_linear_id();
+                const int nthreads   = item_ct1.get_local_range(2);
+                const int nwarps     = nthreads / WARP_SIZE;
+
+                sycl::float2 xx_total = xx;
+                sycl::float2 xg_total = xg;
+                if (nwarps > 1) {
+                    if (wi_in_sg == 0) {
+                        l_xx[sg_id] = xx;
+                        l_xg[sg_id] = xg;
+                    }
+                    item_ct1.barrier(sycl::access::fence_space::local_space);
+
+                    if (sg_id == 0) {
+                        const unsigned wi_u = wi_in_sg;
+                        sycl::float2 xx_first = (wi_u < static_cast<unsigned>(nwarps)) ? l_xx[wi_u] : sycl::float2(0.f, 0.f);
+                        sycl::float2 xg_first = (wi_u < static_cast<unsigned>(nwarps)) ? l_xg[wi_u] : sycl::float2(0.f, 0.f);
+                        xx_total = warp_reduce_sum(xx_first, item_ct1);
+                        xg_total = warp_reduce_sum(xg_first, item_ct1);
+                    } else {
+                        // other subgroups keep their local totals; they'll be ignored
+                        xx_total = xx;
+                        xg_total = xg;
+                    }
+                    // ensure all threads see the first-subgroup result via broadcast below
+                }
+
+                // compute inv_r and coeff once per row and broadcast to the whole work-group
+                float inv_r = 0.f;
+                float coeff = 0.f;
+                if (tid == 0) {
+                    const float sum_xx_f  = xx_total.x() + xx_total.y();
+                    const float sum_xdz_f = xg_total.x() + xg_total.y();
+                    const float mean_eps  = sum_xx_f / (float) D + eps;
+                    const float sum_eps   = sum_xx_f + eps * (float) D;
+                    inv_r = sycl::rsqrt(mean_eps);
+                    coeff = -sum_xdz_f / sum_eps;
+                }
+                inv_r = sycl::group_broadcast(item_ct1.get_group(), inv_r);
+                coeff = sycl::group_broadcast(item_ct1.get_group(), coeff);
+
+                for (int64_t col = tid; col < D; col += WG) {
+                    d_row[col] = (g_row[col] + coeff * x_row[col]) * inv_r;
+                }
+            });
+    });
+
+}
+
 void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);

ggml/src/ggml-sycl/norm.hpp

@@ -19,6 +19,8 @@ void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
 void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
+void ggml_sycl_op_rms_norm_back(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
+
 void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);
 
 void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst);

ggml/src/ggml-sycl/repeat_back.cpp

@@ -2,26 +2,43 @@
 
 #include "common.hpp"
 
-void ggml_sycl_op_repeat_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+#define GGML_ASSERT_TENSOR_FITS_INT(t) \
+    GGML_ASSERT((t)->ne[0] < INT_MAX && (t)->ne[1] < INT_MAX && (t)->ne[2] < INT_MAX && (t)->ne[3] < INT_MAX)
 
+void ggml_sycl_op_repeat_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     const float * src0_dd = (const float *) dst->src[0]->data;
     float *       dst_dd  = (float *) dst->data;
 
-    const int64_t ne0 = dst->ne[0], ne1 = dst->ne[1], ne2 = dst->ne[2], ne3 = dst->ne[3];
-    const int64_t ne00 = dst->src[0]->ne[0], ne01 = dst->src[0]->ne[1], ne02 = dst->src[0]->ne[2],
-                  ne03 = dst->src[0]->ne[3];
+    GGML_ASSERT_TENSOR_FITS_INT(dst);
+    GGML_ASSERT_TENSOR_FITS_INT(dst->src[0]);
 
-    const int nr0 = (int) (ne00 / ne0);
-    const int nr1 = (int) (ne01 / ne1);
-    const int nr2 = (int) (ne02 / ne2);
-    const int nr3 = (int) (ne03 / ne3);
+    const int ne0 = dst->ne[0], ne1 = dst->ne[1], ne2 = dst->ne[2], ne3 = dst->ne[3];
+    const int ne00 = dst->src[0]->ne[0], ne01 = dst->src[0]->ne[1], ne02 = dst->src[0]->ne[2],
+              ne03 = dst->src[0]->ne[3];
 
-    const size_t total      = ne0 * ne1 * ne2 * ne3;
-    const int    BLOCK_SIZE = 256;
-    const int    num_blocks = (total + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    const int nr0 = ne00 / ne0;
+    const int nr1 = ne01 / ne1;
+    const int nr2 = ne02 / ne2;
+    const int nr3 = ne03 / ne3;
+
+    const int nb0 = dst->src[0]->nb[0];
+    const int nb1 = dst->src[0]->nb[1];
+    const int nb2 = dst->src[0]->nb[2];
+    const int nb3 = dst->src[0]->nb[3];
+
+    const char * base = (const char *) src0_dd;
+
+    const size_t  total      = (size_t) ne0 * ne1 * ne2 * ne3;
+    constexpr int BLOCK_SIZE = 256;
+    const int     num_blocks = (total + BLOCK_SIZE - 1) / BLOCK_SIZE;
+
+    const float inv_ne0      = 1.0f / ne0;
+    const float inv_ne_01    = 1.0f / (ne0 * ne1);
+    const float inv_ne_012   = 1.0f / (ne0 * ne1 * ne2);
+    const int   repeat_count = nr0 * nr1 * nr2 * nr3;
 
     queue_ptr stream = ctx.stream();
 
@@ -33,24 +50,27 @@ void ggml_sycl_op_repeat_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst
                 return;
             }
 
-            const int i0 = i % ne0;
-            const int i1 = (i / ne0) % ne1;
-            const int i2 = (i / (ne0 * ne1)) % ne2;
-            const int i3 = i / (ne0 * ne1 * ne2);
+            const int i3 = (int) (i * inv_ne_012);
+            const int i2 = (int) (i * inv_ne_01) - i3 * ne2;
+            const int i1 = (int) (i * inv_ne0) - (int) (i * inv_ne_01) * ne1;
+            const int i0 = i - (int) (i * inv_ne0) * ne0;
 
+            int   j0 = 0, j1 = 0, j2 = 0, j3 = 0;
             float acc = 0.0f;
 
-            for (int j3 = 0; j3 < nr3; ++j3) {
-                for (int j2 = 0; j2 < nr2; ++j2) {
-                    for (int j1 = 0; j1 < nr1; ++j1) {
-                        for (int j0 = 0; j0 < nr0; ++j0) {
-                            acc += src0_dd[(i0 + j0 * ne0) + (i1 + j1 * ne1) * ne00 + (i2 + j2 * ne2) * ne00 * ne01 +
-                                           (i3 + j3 * ne3) * ne00 * ne01 * ne02];
-                        }
-                    }
-                }
-            }
+            for (int j = 0; j < repeat_count; ++j) {
+                const float * ptr = (const float *) (base + (i0 + j0 * ne0) * nb0 + (i1 + j1 * ne1) * nb1 +
+                    (i2 + j2 * ne2) * nb2 + (i3 + j3 * ne3) * nb3);
+                acc += *ptr;
 
+                int carry = (++j0 >= nr0);
+                j0 -= carry * nr0;
+                carry = (carry && (++j1 >= nr1));
+                j1 -= carry * nr1;
+                carry = (carry && (++j2 >= nr2));
+                j2 -= carry * nr2;
+                j3 += carry;
+            }
             dst_dd[i] = acc;
         });
 }

ggml/src/ggml-sycl/rope.cpp

@@ -119,7 +119,7 @@ static void rope_multi(const T * x, T * dst, const int ne0, const int ne1, const
                         const size_t s2, const int n_dims, const int32_t * pos, const float freq_scale,
                         const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
                         const float theta_scale, const float * freq_factors, const mrope_sections sections,
-                        const sycl::nd_item<3> & item_ct1) {
+                        const bool is_imrope, const sycl::nd_item<3> & item_ct1) {
     // get index pos
     const int i0 = 2 * (item_ct1.get_group(1) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1));
     if (i0 >= ne0) {
@@ -143,17 +143,29 @@ static void rope_multi(const T * x, T * dst, const int ne0, const int ne1, const
 
 
     float theta_base = 0.0;
-    if (sector < sections.v[0]) {
-        theta_base = pos[channel_x]*sycl::pow(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sections.v[0] && sector < sec_w) {
-        theta_base = pos[channel_x + ne2 * 1]*sycl::pow(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-        theta_base = pos[channel_x + ne2 * 2]*sycl::pow(theta_scale, i0/2.0f);
-    }
-    else if (sector >= sec_w + sections.v[2]) {
-        theta_base = pos[channel_x + ne2 * 3]*sycl::pow(theta_scale, i0/2.0f);
+    if (is_imrope) {
+        if (sector % 3 == 1 && sector < 3 * sections.v[1]) {
+            theta_base = pos[channel_x + ne2 * 1]*sycl::pow(theta_scale, i0/2.0f);
+        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 2]*sycl::pow(theta_scale, i0/2.0f);
+        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) {
+            theta_base = pos[channel_x]*sycl::pow(theta_scale, i0/2.0f);
+        } else {
+            theta_base = pos[channel_x + ne2 * 3]*sycl::pow(theta_scale, i0/2.0f);
+        }
+    } else {
+        if (sector < sections.v[0]) {
+            theta_base = pos[channel_x]*sycl::pow(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sections.v[0] && sector < sec_w) {
+            theta_base = pos[channel_x + ne2 * 1]*sycl::pow(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 2]*sycl::pow(theta_scale, i0/2.0f);
+        }
+        else if (sector >= sec_w + sections.v[2]) {
+            theta_base = pos[channel_x + ne2 * 3]*sycl::pow(theta_scale, i0/2.0f);
+        }
     }
 
     const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
@@ -281,7 +293,7 @@ static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1,
                              const size_t s2, const int n_dims, const int nr, const int32_t * pos,
                              const float freq_scale, const float freq_base, const float ext_factor,
                              const float attn_factor, const rope_corr_dims corr_dims, const float * freq_factors,
-                             const mrope_sections sections, queue_ptr stream) {
+                             const mrope_sections sections, const bool is_imrope, queue_ptr stream) {
     GGML_ASSERT(ne0 % 2 == 0);
     const sycl::range<3>    block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
     const int               n_blocks_y = ceil_div(ne0, (2 * SYCL_ROPE_BLOCK_SIZE));
@@ -297,12 +309,12 @@ static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1,
     if (freq_factors == nullptr) {
         stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_multi<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
-                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
+                                  corr_dims, theta_scale, freq_factors, sections, is_imrope, item_ct1);
         });
     } else {
         stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_multi<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
-                                 corr_dims, theta_scale, freq_factors, sections, item_ct1);
+                                 corr_dims, theta_scale, freq_factors, sections, is_imrope, item_ct1);
         });
     }
 }
@@ -381,6 +393,7 @@ inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
     const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
     const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
 
     if (is_mrope) {
@@ -422,11 +435,11 @@ inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
         if (dst->src[0]->type == GGML_TYPE_F16) {
             rope_multi_sycl((const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, ne01, ne02, s01,
                 s02, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                freq_factors, sections, main_stream);
+                freq_factors, sections, is_imrope, main_stream);
         } else if (dst->src[0]->type == GGML_TYPE_F32) {
             rope_multi_sycl((const float *) dst->src[0]->data, (float *) dst->data, ne00, ne01, ne02, s01, s02, n_dims,
                              nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections,
-                             main_stream);
+                             is_imrope, main_stream);
         } else {
             GGML_ABORT("Fatal error: Tensor type unsupported!");
         }

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

@@ -14,6 +14,7 @@ layout (binding = 1)          buffer D {int data_d[];};
 
 layout (push_constant) uniform parameter {
     uint ncols;
+    uint nrows;
     uint order;
 } p;
 
@@ -26,10 +27,9 @@ void swap(uint idx0, uint idx1) {
     dst_row[idx1] = tmp;
 }
 
-void argsort(bool needs_bounds_check) {
+void argsort(bool needs_bounds_check, const uint row) {
     // bitonic sort
     const int col = int(gl_LocalInvocationID.x);
-    const uint row = gl_WorkGroupID.y;
 
     const uint row_offset = row * p.ncols;
 
@@ -72,8 +72,16 @@ void argsort(bool needs_bounds_check) {
 
 void main() {
     if (p.ncols == BLOCK_SIZE) {
-        argsort(false);
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            argsort(false, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
     } else {
-        argsort(true);
+        uint row = gl_WorkGroupID.y;
+        while (row < p.nrows) {
+            argsort(true, row);
+            row += gl_WorkGroupSize.y * gl_NumWorkGroups.y;
+        }
     }
 }

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

@@ -437,7 +437,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
 #if defined(DATA_A_MXFP4)
 vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2(kvalues_mxfp4[vui & 0xF], kvalues_mxfp4[vui >> 4]);
+    return vec2(kvalues_mxfp4[vui & 0xF], kvalues_mxfp4[vui >> 4]) * 0.5;
 }
 vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
     vec2 v0 = dequantize(ib, iqs, a_offset);
@@ -488,9 +488,9 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
 
     const uvec2 qs = uvec2(data_a[a_offset + ib].qs[qsi], data_a[a_offset + ib].qs[qsi + 1]);
     const uint scales = data_a[a_offset + ib].scales[scalesi];
-    const vec2 d = vec2(data_a[a_offset + ib].d);
+    const vec2 dm = vec2(data_a[a_offset + ib].dm);
 
-    return d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
+    return dm.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - dm.y * float(scales >> 4);
 }
 vec2 get_dm(uint ib, uint a_offset) {
     return vec2(1, 0);
@@ -529,7 +529,7 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     const uint is = 2 * n + b;                 // 0..7
     const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
 
-    const vec2 loadd = vec2(data_a[a_offset + ib].d);
+    const vec2 loadd = vec2(data_a[a_offset + ib].dm);
 
     const uint scidx0 = (is < 4) ? is : (is + 4);
     const uint scidx1 = (is < 4) ? is : (is - 4);
@@ -567,7 +567,7 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
 
     const uint8_t hm = uint8_t(1 << (iqs / 16));
 
-    const vec2 loadd = vec2(data_a[a_offset + ib].d);
+    const vec2 loadd = vec2(data_a[a_offset + ib].dm);
 
     const uint scidx0 = (is < 4) ? is : (is + 4);
     const uint scidx1 = (is < 4) ? is : (is - 4);

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

@@ -120,7 +120,7 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ2
 float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
     decodeBufQ2_K_packed16 bl16 = decodeBufQ2_K_packed16(bl);
-    const f16vec2 d = bl.block.d;
+    const f16vec2 dm = bl.block.dm;
     const uint idx = coordInBlock[1];
 
     const uint scalesi = (idx & 0xF0) >> 4;             // 0..15
@@ -131,7 +131,7 @@ float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2
     qs = unpack8(qs)[idx & 1];
 
     const uint scales = bl.block.scales[scalesi];
-    float16_t ret = d.x * float16_t(scales & 0xF) * float16_t(qs) - d.y * float16_t(scales >> 4);
+    float16_t ret = dm.x * float16_t(scales & 0xF) * float16_t(qs) - dm.y * float16_t(scales >> 4);
     return ret;
 }
 
@@ -680,7 +680,7 @@ float16_t dequantFuncMXFP4(const in decodeBufMXFP4 bl, const in uint blockCoords
     uint32_t qs = bl.block.qs[iqs];
     qs >>= shift;
     qs &= 0xF;
-    float16_t ret = float16_t(kvalues_mxfp4[qs] * d);
+    float16_t ret = float16_t(kvalues_mxfp4[qs] * d * 0.5);
     return ret;
 }
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/dequant_mxfp4.comp

@@ -26,7 +26,7 @@ void main() {
     const float d = e8m0_to_fp32(data_a[ib].e);
 
     [[unroll]] for (uint l = 0; l < 8; ++l) {
-        data_b[b_idx + l +  0] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]);
-        data_b[b_idx + l + 16] = D_TYPE(d * kvalues_mxfp4[data_a[ib].qs[q_idx + l] >>  4]);
+        data_b[b_idx + l +  0] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] & 0xF]));
+        data_b[b_idx + l + 16] = D_TYPE(d * 0.5 * float(kvalues_mxfp4[data_a[ib].qs[q_idx + l] >>  4]));
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp

@@ -24,8 +24,8 @@ void main() {
         const uint ql_idx = 32 * ip + il;
         const uint8_t qs = data_a[i].qs[32 * ip + il];
 
-        FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
-        FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+        FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].dm.x);
+        FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].dm.y);
         data_b[y_idx +  0] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+0] & 0xF) * ((qs >> 0) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+0] >> 4));
         data_b[y_idx + 32] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+2] & 0xF) * ((qs >> 2) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+2] >> 4));
         data_b[y_idx + 64] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+4] & 0xF) * ((qs >> 4) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+4] >> 4));

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp

@@ -20,8 +20,8 @@ void main() {
         const uint is = 2 * il;
         const uint n = 4;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].dm.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].dm.y);
 
         const uint y_idx = ib * QUANT_K + 64 * il + n * ir;
         const uint qs_idx = 32*il + n * ir;

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp

@@ -19,8 +19,8 @@ void main() {
         const uint ir = tid % 16;
         const uint is = 2 * il;
 
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].dm.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].dm.y);
 
         const uint y_idx = ib * QUANT_K + 64 * il + 2 * ir;
         const uint qs_idx = 32*il + 2 * ir;

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_mat_vec_base.glsl

@@ -28,8 +28,11 @@ layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
 #endif
 
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+layout (binding = 3) readonly buffer Bias {D_TYPE data_bias[];};
+
 #ifdef MUL_MAT_ID
-layout (binding = 3) readonly buffer IDS {int data_ids[];};
+layout (binding = 4) readonly buffer IDS {int data_ids[];};
 #endif
 
 #include "dequant_funcs.glsl"
@@ -45,6 +48,8 @@ layout (push_constant) uniform parameter
     uint batch_stride_b;
     uint batch_stride_d;
 
+    uint enable_bias;
+
 #ifdef MUL_MAT_ID
     uint nei0;
     uint ne11;
@@ -56,6 +61,10 @@ layout (push_constant) uniform parameter
 #endif
 } p;
 
+#ifdef MUL_MAT_ID
+uint expert_id;
+#endif
+
 void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #ifdef MUL_MAT_ID
     const uint expert_idx = gl_GlobalInvocationID.y;
@@ -75,7 +84,7 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
         batch_idx_a = i03 * p.ne02 + i02;
     }
 #else
-    const uint expert_id = data_ids[expert_idx];
+    expert_id = data_ids[expert_idx];
 #endif
 
     a_offset =
@@ -113,6 +122,13 @@ void reduce_result(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t
     if (tid == 0) {
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                if (p.enable_bias != 0) {
+#ifdef MUL_MAT_ID
+                    temp[j][n] += FLOAT_TYPE(data_bias[expert_id*p.stride_d + first_row + n]);
+#else
+                    temp[j][n] += FLOAT_TYPE(data_bias[j*p.batch_stride_d + d_offset + first_row + n]);
+#endif
+                }
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
             }
         }
@@ -148,6 +164,13 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
                 [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
                     temp[j][n] += tmpsh[j][n][s];
                 }
+                if (p.enable_bias != 0) {
+#ifdef MUL_MAT_ID
+                    temp[j][n] += FLOAT_TYPE(data_bias[expert_id*p.stride_d + first_row + n]);
+#else
+                    temp[j][n] += FLOAT_TYPE(data_bias[j*p.batch_stride_d + d_offset + first_row + n]);
+#endif
+                }
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(temp[j][n]);
             }
         }
@@ -173,6 +196,13 @@ void reduce_result(FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offs
     if (tid == 0) {
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                if (p.enable_bias != 0) {
+#ifdef MUL_MAT_ID
+                    tmpsh[j][n][0] += FLOAT_TYPE(data_bias[expert_id*p.stride_d + first_row + n]);
+#else
+                    tmpsh[j][n][0] += FLOAT_TYPE(data_bias[j*p.batch_stride_d + d_offset + first_row + n]);
+#endif
+                }
                 data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][0]);
             }
         }

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

@@ -15,6 +15,8 @@ layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
 layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
 layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
 
+layout (binding = 3) readonly buffer Bias {D_TYPE data_bias[];};
+
 layout (push_constant) uniform parameter
 {
     uint ncols_x;
@@ -29,6 +31,7 @@ layout (push_constant) uniform parameter
     uint nb03;
     uint nb13;
     uint nb23;
+    uint enable_bias;
 } p;
 
 shared FLOAT_TYPE tmp[BLOCK_SIZE];
@@ -117,6 +120,9 @@ void main() {
     }
 
     if (tid == 0) {
+        if (p.enable_bias != 0) {
+            tmp[0] += FLOAT_TYPE(data_bias[idst]);
+        }
         dst[idst] = tmp[0];
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp

@@ -17,6 +17,8 @@ layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
 layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];};
 layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
 
+layout (binding = 3) readonly buffer Bias {D_TYPE data_bias[];};
+
 layout(constant_id = 0) const int BLOCK_SIZE = 32;
 // gqa_ratio is in the range [1,8]
 layout(constant_id = 1) const uint gqa_ratio = 1;
@@ -29,6 +31,7 @@ layout (push_constant) uniform parameter
     uint nchannels_y;
     uint b_offset;
     uint d_offset;
+    uint enable_bias;
 } p;
 
 #if !USE_SUBGROUP_ADD
@@ -148,6 +151,9 @@ void main() {
         [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) {
             // dst is not transposed and not permuted
             const uint idst = (channel + c)*nrows_dst + row_dst;
+            if (p.enable_bias != 0) {
+                temp[c] += FLOAT_TYPE(data_bias[idst]);
+            }
             dst[idst] = temp[c];
         }
     }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -41,9 +41,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
         const vec4 qs_u32_4 = vec4(unpack8((qs_u32 >> 4) & 0x03030303));
         const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303));
 
-        vec2 d = vec2(data_a[ib0 + i].d);
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+        const FLOAT_TYPE_VEC2 dm = vec2(data_a[ib0 + i].dm);
 
         [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
             vec2 b0 =   vec2(data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 +  0]);
@@ -75,7 +73,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
                        fma(FLOAT_TYPE(b96[l]),  sccache2[csel][ix][6 + 8*v_im],
                        fma(FLOAT_TYPE(b112[l]), sccache2[csel][ix][7 + 8*v_im], sum2))))))));
             }
-            temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
+            temp[j][n] = fma(dm.x, sum1, fma(-dm.y, sum2, temp[j][n]));
         }
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -14,9 +14,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
-        vec2 d = vec2(data_a[ib0 + i].d);
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+        const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
 
         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
         const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
@@ -81,7 +79,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
                 fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
                 fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
                 fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
-            temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
+            temp[j][n] = fma(dm.x, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dm.y, smin, temp[j][n]));
         }
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -14,9 +14,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
-        vec2 d = vec2(data_a[ib0 + i].d);
-        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+        const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
 
         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
         const uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
@@ -113,7 +111,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
               fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
               fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
                   (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
-            temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
+            temp[j][n] = fma(dm.x, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dm.y, smin, temp[j][n]));
         }
     }
 }

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

@@ -120,81 +120,11 @@ shared FLOAT_TYPE_VEC2 buf_b[BN * SHMEM_STRIDE];
 
 #define NUM_WARPS (BLOCK_SIZE / WARP)
 
-#ifdef MUL_MAT_ID
-shared u16vec2 row_ids[BN];
-uint _ne1;
-
-#ifdef MUL_MAT_ID_USE_SUBGROUPS
-shared uvec4 ballots_sh[NUM_WARPS];
-
-void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
-    _ne1 = 0;
-    uint num_elements = p.nei1 * p.nei0;
-    uint nei0shift = findLSB(p.nei0);
-
-    uint ids[16];
-    uint iter = 0;
-
-    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
-        // prefetch up to 16 elements
-        if (iter == 0) {
-            [[unroll]] for (uint k = 0; k < 16; ++k) {
-                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
-                bool in_range = i < num_elements;
-                uint ii1;
-                if (nei0_is_pow2) {
-                    ii1 = i >> nei0shift;
-                } else {
-                    ii1 = i / p.nei0;
-                }
-                uint ii0 = i - ii1 * p.nei0;
-                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
-            }
-        }
-        uint i = j + gl_LocalInvocationIndex;
-        bool in_range = i < num_elements;
-        uint ii1;
-        if (nei0_is_pow2) {
-            ii1 = i >> nei0shift;
-        } else {
-            ii1 = i / p.nei0;
-        }
-        uint ii0 = i - ii1 * p.nei0;
-        uint id = ids[iter++];
-        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
-
-        ballots_sh[gl_SubgroupID] = ballot;
-        barrier();
-
-        uint subgroup_base = 0;
-        uint total = 0;
-        for (uint k = 0; k < gl_NumSubgroups; ++k) {
-            if (k == gl_SubgroupID) {
-                subgroup_base = total;
-            }
-            total += subgroupBallotBitCount(ballots_sh[k]);
-        }
-        barrier();
-
-        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
-        if (in_range && id == expert_idx && _ne1 + idx >= ic * BN && _ne1 + idx < (ic + 1) * BN) {
-            row_ids[_ne1 + idx - ic * BN] = u16vec2(ii0, ii1);
-        }
-        _ne1 += total;
-        iter &= 15;
-        if (_ne1 >= (ic + 1) * BN) {
-            break;
-        }
-    }
-    barrier();
-}
-#endif // MUL_MAT_ID_USE_SUBGROUPS
-#endif // MUL_MAT_ID
-
 #ifdef COOPMAT
 shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
 #endif
 
+#include "mul_mm_id_funcs.glsl"
 #include "mul_mm_funcs.glsl"
 
 void main() {

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

@@ -134,15 +134,15 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint ib = idx / 128;                         // 2 values per idx
             const uint iqs = idx % 128;                        // 0..127
 
-            const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
+            const uint qsi = (iqs / 64) * 16 + (iqs % 16);     // 0..15
             const uint scalesi = iqs / 8;                      // 0..15
             const uint qsshift = ((iqs % 64) / 16) * 2;        // 0,2,4,6
 
-            const uvec2 qs = uvec2(data_a[ib].qs[qsi], data_a[ib].qs[qsi + 1]);
+            const uvec2 qs = uvec2(unpack8(data_a_packed16[ib].qs[qsi]));
             const uint scales = data_a[ib].scales[scalesi];
-            const vec2 d = vec2(data_a[ib].d);
+            const vec2 dm = vec2(data_a[ib].dm);
 
-            const vec2 v = d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
+            const vec2 v = dm.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - dm.y * float(scales >> 4);
 
             buf_a[buf_idx] = FLOAT_TYPE_VEC2(v.xy);
 #elif defined(DATA_A_Q3_K)
@@ -179,7 +179,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint is = 2 * n + b;                 // 0..7
             const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
 
-            const vec2 loadd = vec2(data_a[ib].d);
+            const vec2 loadd = vec2(data_a[ib].dm);
 
             const uint scidx0 = (is < 4) ? is : (is + 4);
             const uint scidx1 = (is < 4) ? is : (is - 4);
@@ -215,7 +215,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
 
             const uint8_t hm = uint8_t(1 << (iqs / 16));
 
-            const vec2 loadd = vec2(data_a[ib].d);
+            const vec2 loadd = vec2(data_a[ib].dm);
 
             const uint scidx0 = (is < 4) ? is : (is + 4);
             const uint scidx1 = (is < 4) ? is : (is - 4);
@@ -468,7 +468,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint ib = idx / 8;
             const uint iqs = (idx & 0x07) * 2;
 
-            const float d = e8m0_to_fp32(data_a[ib].e);
+            const float d = e8m0_to_fp32(data_a[ib].e) * 0.5;
             const uint vui = uint(data_a[ib].qs[iqs]);
             const uint vui2 = uint(data_a[ib].qs[iqs+1]);
 

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

@@ -0,0 +1,70 @@
+#ifdef MUL_MAT_ID
+shared u16vec2 row_ids[BN];
+uint _ne1;
+
+#ifdef MUL_MAT_ID_USE_SUBGROUPS
+shared uvec4 ballots_sh[NUM_WARPS];
+
+void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
+    _ne1 = 0;
+    uint num_elements = p.nei1 * p.nei0;
+    uint nei0shift = findLSB(p.nei0);
+
+    uint ids[16];
+    uint iter = 0;
+
+    for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
+        // prefetch up to 16 elements
+        if (iter == 0) {
+            [[unroll]] for (uint k = 0; k < 16; ++k) {
+                uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
+                bool in_range = i < num_elements;
+                uint ii1;
+                if (nei0_is_pow2) {
+                    ii1 = i >> nei0shift;
+                } else {
+                    ii1 = i / p.nei0;
+                }
+                uint ii0 = i - ii1 * p.nei0;
+                ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+            }
+        }
+        uint i = j + gl_LocalInvocationIndex;
+        bool in_range = i < num_elements;
+        uint ii1;
+        if (nei0_is_pow2) {
+            ii1 = i >> nei0shift;
+        } else {
+            ii1 = i / p.nei0;
+        }
+        uint ii0 = i - ii1 * p.nei0;
+        uint id = ids[iter++];
+        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
+
+        ballots_sh[gl_SubgroupID] = ballot;
+        barrier();
+
+        uint subgroup_base = 0;
+        uint total = 0;
+        for (uint k = 0; k < gl_NumSubgroups; ++k) {
+            if (k == gl_SubgroupID) {
+                subgroup_base = total;
+            }
+            total += subgroupBallotBitCount(ballots_sh[k]);
+        }
+        barrier();
+
+        uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
+        if (in_range && id == expert_idx && _ne1 + idx >= ic * BN && _ne1 + idx < (ic + 1) * BN) {
+            row_ids[_ne1 + idx - ic * BN] = u16vec2(ii0, ii1);
+        }
+        _ne1 += total;
+        iter &= 15;
+        if (_ne1 >= (ic + 1) * BN) {
+            break;
+        }
+    }
+    barrier();
+}
+#endif // MUL_MAT_ID_USE_SUBGROUPS
+#endif // MUL_MAT_ID