Add a workaround for compilation with ROCWMMA_FATTN and gfx9 (#19461 )

There is an upstream problem [1] with AMD's LLVM 22 fork and rocWMMA 2.2.0 causing compilation issues on devices without native fp16 support (CDNA devices). The specialized types aren't resolved properly: ``` /opt/rocm/include/rocwmma/internal/mfma_impl.hpp:2549:37: error: ambiguous partial specializations of 'amdgcn_mfma<__half, __half, __half, 16, 16, 16>' 2549 | using ARegsT = typename Impl::ARegsT; ``` Add a workaround to explicitly declare the types and cast when compiling with HIP and ROCWMMA_FATTN [2]. When this is actually fixed upstream some guards can be used to detect and wrap the version that has the fix to only apply when necessary. Link: https://github.com/ROCm/rocm-libraries/issues/4398 [1] Link: https://github.com/ggml-org/llama.cpp/issues/19269 [2] Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
server : fix typo in README.md for features list (#19510 )
2026-07-15 08:55:56 +02:00 · 2026-02-12 09:38:35 +01:00 · 2026-02-12 08:56:25 +01:00 · 2026-02-12 08:13:51 +01:00 · 2026-02-11 23:04:27 -08:00 · 2026-02-12 07:27:52 +01:00
220 changed files with 16553 additions and 6423 deletions
@@ -54,6 +54,7 @@ RUN apt-get update \
    build-essential \
    git \
    python3 \
+    python3-dev \
    python3-pip \
    python3-wheel \
    && pip install --break-system-packages --upgrade setuptools \
@@ -293,7 +293,9 @@ jobs:
          cmake -B build \
            -DLLAMA_FATAL_WARNINGS=ON \
            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+
          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)

      - name: Build (no OpenMP)
@@ -303,8 +305,10 @@ jobs:
          cmake -B build \
            -DLLAMA_FATAL_WARNINGS=ON \
            -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
+            -DGGML_SANITIZE_${{ matrix.sanitizer }}=ON \
            -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
            -DGGML_OPENMP=OFF
+
          cmake --build build --config ${{ matrix.build_type }} -j $(nproc)

      - name: Test
@@ -466,7 +470,7 @@ jobs:
          export GGML_VK_VISIBLE_DEVICES=0
          export GGML_VK_DISABLE_F16=1
          # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 4200
+          ctest -L main --verbose --timeout 4800

  ubuntu-24-cmake-webgpu:
    runs-on: ubuntu-24.04
@@ -1532,7 +1536,7 @@ jobs:
      - name: Test
        id: ggml-ci
        run: |
-          LLAMA_ARG_THREADS=$(nproc) bash ./ci/run.sh ./tmp/results ./tmp/mnt
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

  ggml-ci-arm64-cpu-high-perf:
    runs-on: ubuntu-22.04-arm
@@ -1558,7 +1562,7 @@ jobs:
      - name: Test
        id: ggml-ci
        run: |
-          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_NO_SVE=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+          LLAMA_ARG_THREADS=$(nproc) GG_BUILD_HIGH_PERF=1 GG_BUILD_NO_SVE=1 GG_BUILD_NO_BF16=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt

  ggml-ci-arm64-cpu-high-perf-sve:
    runs-on: ubuntu-22.04-arm
@@ -0,0 +1,73 @@
+name: Server-Metal
+
+on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
+      slow_tests:
+        description: 'Run slow tests'
+        required: true
+        type: boolean
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/server-metal.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'tools/server/**.*']
+
+env:
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_VERBOSITY: 10
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  server-metal:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    name: server-metal (${{ matrix.wf_name }})
+    strategy:
+      matrix:
+        build_type: [Release]
+        wf_name: ["GPUx1"]
+        include:
+          - build_type: Release
+            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
+            wf_name:    "GPUx1, backend-sampling"
+          - build_type: Release
+            extra_args: "GGML_METAL_DEVICES=2"
+            wf_name:    "GPUx2"
+          - build_type: Release
+            extra_args: "GGML_METAL_DEVICES=2 LLAMA_ARG_BACKEND_SAMPLING=1"
+            wf_name:    "GPUx2, backend-sampling"
+      fail-fast: false
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build -DGGML_SCHED_NO_REALLOC=ON
+          cmake --build build --config ${{ matrix.build_type }} -j $(sysctl -n hw.logicalcpu) --target llama-server
+
+      - name: Tests
+        id: server_integration_tests
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
+        run: |
+          cd tools/server/tests
+          python3 -m venv venv
+          source venv/bin/activate
+          pip install -r requirements.txt
+          export ${{ matrix.extra_args }}
+          pytest -v -x -m "not slow"
@@ -8,10 +8,6 @@ on:
        description: 'Commit SHA1 to build'
        required: false
        type: string
-      slow_tests:
-        description: 'Run slow tests'
-        required: true
-        type: boolean
  push:
    branches:
      - master
@@ -101,119 +97,3 @@ jobs:
        if: ${{ always() && steps.playwright.conclusion == 'success' }}
        run: npm run test:e2e
        working-directory: tools/server/webui
-
-  server-build:
-    runs-on: ubuntu-latest
-
-    strategy:
-      matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
-        build_type: [RelWithDebInfo]
-        include:
-          - build_type: Release
-            sanitizer: ""
-      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
-
-    steps:
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get -y install \
-            build-essential \
-            xxd \
-            git \
-            cmake \
-            curl \
-            wget \
-            language-pack-en \
-            libssl-dev
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
-
-      - name: Python setup
-        id: setup_python
-        uses: actions/setup-python@v6
-        with:
-          python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r tools/server/tests/requirements.txt
-
-      - name: Setup Node.js for WebUI
-        uses: actions/setup-node@v6
-        with:
-          node-version: "22"
-          cache: "npm"
-          cache-dependency-path: "tools/server/webui/package-lock.json"
-
-      - name: Install WebUI dependencies
-        run: npm ci
-        working-directory: tools/server/webui
-
-      - name: Build WebUI
-        run: npm run build
-        working-directory: tools/server/webui
-
-      - name: Build (no OpenMP)
-        id: cmake_build_no_openmp
-        if: ${{ matrix.sanitizer == 'THREAD' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \
-              -DGGML_OPENMP=OFF ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Build (sanitizers)
-        id: cmake_build_sanitizers
-        if: ${{ matrix.sanitizer != '' && matrix.sanitizer != 'THREAD' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Build (sanitizers)
-        id: cmake_build
-        if: ${{ matrix.sanitizer == '' }}
-        run: |
-          cmake -B build \
-              -DGGML_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ;
-          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server
-
-      - name: Tests
-        id: server_integration_tests
-        if: ${{ matrix.sanitizer == '' }}
-        env:
-          GITHUB_ACTIONS: "true"
-        run: |
-          cd tools/server/tests
-          ./tests.sh
-
-      - name: Tests (sanitizers)
-        id: server_integration_tests_sanitizers
-        if: ${{ matrix.sanitizer != '' }}
-        run: |
-          cd tools/server/tests
-          LLAMA_SANITIZE=1 ./tests.sh
-
-      - name: Slow tests
-        id: server_integration_tests_slow
-        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
-        run: |
-          cd tools/server/tests
-          SLOW_TESTS=1 ./tests.sh
@@ -36,7 +36,7 @@ jobs:

    strategy:
      matrix:
-        sanitizer: [ADDRESS, UNDEFINED] # THREAD is broken
+        sanitizer: [ADDRESS, UNDEFINED] # THREAD is very slow
        build_type: [RelWithDebInfo]
        include:
          - build_type: Release
@@ -45,7 +45,7 @@ jobs:
          - build_type: Release
            sanitizer: ""
            extra_args: "LLAMA_ARG_BACKEND_SAMPLING=1"
-      fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
+      fail-fast: false

    steps:
      - name: Dependencies
@@ -72,28 +72,40 @@ jobs:
      - name: Build
        id: cmake_build
        run: |
-          cmake -B build -DLLAMA_BUILD_BORINGSSL=ON -DGGML_SCHED_NO_REALLOC=ON
-          cmake --build build --config ${{ matrix.build_type }} -j ${env:NUMBER_OF_PROCESSORS} --target llama-server
+          cmake -B build \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_SCHED_NO_REALLOC=ON \
+            -DGGML_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
+            -DGGML_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
+            -DGGML_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }} \
+            -DLLAMA_SANITIZE_ADDRESS=${{ matrix.sanitizer == 'ADDRESS' }} \
+            -DLLAMA_SANITIZE_THREAD=${{ matrix.sanitizer == 'THREAD' }} \
+            -DLLAMA_SANITIZE_UNDEFINED=${{ matrix.sanitizer == 'UNDEFINED' }}
+          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server

      - name: Python setup
        id: setup_python
        uses: actions/setup-python@v6
        with:
          python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r tools/server/tests/requirements.txt
+          pip-install: -r tools/server/tests/requirements.txt

      - name: Tests
        id: server_integration_tests
-        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) && matrix.build_type == 'Release' }}
+        if: ${{ (!matrix.disabled_on_pr || !github.event.pull_request) }}
        run: |
          cd tools/server/tests
          export ${{ matrix.extra_args }}
          pytest -v -x -m "not slow"

+      - name: Slow tests
+        id: server_integration_tests_slow
+        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
+        run: |
+          cd tools/server/tests
+          export ${{ matrix.extra_args }}
+          SLOW_TESTS=1 pytest -v -x
+
  server-windows:
    runs-on: windows-2022

@@ -116,11 +128,7 @@ jobs:
        uses: actions/setup-python@v6
        with:
          python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r tools/server/tests/requirements.txt
+          pip-install: -r tools/server/tests/requirements.txt

      - name: Tests
        id: server_integration_tests
@@ -109,6 +109,7 @@ option(LLAMA_BUILD_TOOLS    "llama: build tools"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
 option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_DEFAULT})
+option(LLAMA_TESTS_INSTALL  "llama: install tests"        ON)

 # 3rd party libs
 option(LLAMA_HTTPLIB    "llama: httplib for downloading functionality" ON)
@@ -164,29 +165,6 @@ llama_option_depr(WARNING     LLAMA_SYCL                GGML_SYCL)
 llama_option_depr(WARNING     LLAMA_SYCL_F16            GGML_SYCL_F16)
 llama_option_depr(WARNING     LLAMA_CANN                GGML_CANN)

-if (NOT MSVC)
-    if (LLAMA_SANITIZE_THREAD)
-        message(STATUS "Using -fsanitize=thread")
-
-        add_compile_options(-fsanitize=thread)
-        link_libraries     (-fsanitize=thread)
-    endif()
-
-    if (LLAMA_SANITIZE_ADDRESS)
-        message(STATUS "Using -fsanitize=address")
-
-        add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-        link_libraries     (-fsanitize=address)
-    endif()
-
-    if (LLAMA_SANITIZE_UNDEFINED)
-        message(STATUS "Using -fsanitize=undefined")
-
-        add_compile_options(-fsanitize=undefined)
-        link_libraries     (-fsanitize=undefined)
-    endif()
-endif()
-
 include("cmake/license.cmake")
 license_add_file("llama.cpp" "LICENSE")

@@ -27,6 +27,7 @@
 /examples/batched.swift/                @ggerganov
 /examples/batched/                      @ggerganov
 /examples/convert-llama2c-to-ggml/      @ggerganov
+/examples/debug/                        @danbev @pwilkin
 /examples/deprecation-warning/          @ggerganov
 /examples/diffusion/                    @am17an
 /examples/embedding/                    @ggerganov
@@ -20,7 +20,7 @@ If AI is used to generate any portion of the code, contributors must adhere to t
 1. Explicitly disclose the manner in which AI was employed.
 2. Perform a comprehensive manual review prior to submitting the pull request.
 3. Be prepared to explain every line of code they submitted when asked about it by a maintainer.
-4. Using AI to write pull request descriptions or to respond to human reviewers is strictly prohibited.
+4. It is strictly prohibited to use AI to write your posts for you (bug reports, feature requests, pull request descriptions, Github discussions, responding to humans, ...).

 For more info, please refer to the [AGENTS.md](AGENTS.md) file.

@@ -288,6 +288,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [WebGPU [In Progress]](docs/build.md#webgpu) | All |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 | [Hexagon [In Progress]](docs/backend/hexagon/README.md) | Snapdragon |
+| [VirtGPU](docs/backend/VirtGPU.md) | VirtGPU APIR |

 ## Obtaining and quantizing models

@@ -8,7 +8,7 @@ g++ --version
 g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0

 nvidia-smi
-Sun Nov  2 10:43:25 2025
+Thu Feb  5 13:49:40 2026
 +-----------------------------------------------------------------------------------------+
 | NVIDIA-SMI 580.95.05              Driver Version: 580.95.05      CUDA Version: 13.0     |
 +-----------------------------------------+------------------------+----------------------+
@@ -17,7 +17,7 @@ Sun Nov  2 10:43:25 2025
 |                                         |                        |               MIG M. |
 |=========================================+========================+======================|
 |   0  NVIDIA GB10                    On  |   0000000F:01:00.0 Off |                  N/A |
-| N/A   35C    P8              4W /  N/A  | Not Supported          |      0%      Default |
+| N/A   47C    P0             13W /  N/A  | Not Supported          |      0%      Default |
 |                                         |                        |                  N/A |
 +-----------------------------------------+------------------------+----------------------+
 ```
@@ -29,46 +29,46 @@ Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
 - `llama-batched-bench`


-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20

 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.374 |  1369.01 |    0.383 |    83.64 |    0.757 |   719.01 |
-|   512 |     32 |    2 |   1088 |    0.274 |  3741.35 |    0.659 |    97.14 |    0.933 |  1166.66 |
-|   512 |     32 |    4 |   2176 |    0.526 |  3896.47 |    0.817 |   156.73 |    1.342 |  1621.08 |
-|   512 |     32 |    8 |   4352 |    1.044 |  3925.10 |    0.987 |   259.44 |    2.030 |  2143.56 |
-|   512 |     32 |   16 |   8704 |    2.076 |  3945.84 |    1.248 |   410.32 |    3.324 |  2618.60 |
-|   512 |     32 |   32 |  17408 |    4.170 |  3929.28 |    1.630 |   628.40 |    5.799 |  3001.76 |
-|  4096 |     32 |    1 |   4128 |    1.083 |  3782.66 |    0.394 |    81.21 |    1.477 |  2795.13 |
-|  4096 |     32 |    2 |   8256 |    2.166 |  3782.72 |    0.725 |    88.28 |    2.891 |  2856.14 |
-|  4096 |     32 |    4 |  16512 |    4.333 |  3780.88 |    0.896 |   142.82 |    5.230 |  3157.38 |
-|  4096 |     32 |    8 |  33024 |    8.618 |  3802.14 |    1.155 |   221.69 |    9.773 |  3379.08 |
-|  4096 |     32 |   16 |  66048 |   17.330 |  3781.73 |    1.598 |   320.34 |   18.928 |  3489.45 |
-|  4096 |     32 |   32 | 132096 |   34.671 |  3780.48 |    2.336 |   438.35 |   37.007 |  3569.51 |
-|  8192 |     32 |    1 |   8224 |    2.233 |  3668.56 |    0.438 |    72.98 |    2.671 |  3078.44 |
-|  8192 |     32 |    2 |  16448 |    4.425 |  3702.95 |    0.756 |    84.66 |    5.181 |  3174.95 |
-|  8192 |     32 |    4 |  32896 |    8.859 |  3698.64 |    0.967 |   132.38 |    9.826 |  3347.72 |
-|  8192 |     32 |    8 |  65792 |   17.714 |  3699.57 |    1.277 |   200.52 |   18.991 |  3464.35 |
-|  8192 |     32 |   16 | 131584 |   35.494 |  3692.84 |    1.841 |   278.12 |   37.335 |  3524.46 |
-|  8192 |     32 |   32 | 263168 |   70.949 |  3694.82 |    2.798 |   365.99 |   73.747 |  3568.53 |
+|   512 |     32 |    1 |    544 |    0.270 |  1895.57 |    0.399 |    80.13 |    0.669 |   812.60 |
+|   512 |     32 |    2 |   1088 |    0.230 |  4451.23 |    0.583 |   109.71 |    0.813 |  1337.56 |
+|   512 |     32 |    4 |   2176 |    0.437 |  4688.87 |    0.820 |   156.03 |    1.257 |  1730.91 |
+|   512 |     32 |    8 |   4352 |    0.863 |  4744.23 |    0.942 |   271.79 |    1.805 |  2410.73 |
+|   512 |     32 |   16 |   8704 |    1.725 |  4748.19 |    1.173 |   436.38 |    2.899 |  3002.85 |
+|   512 |     32 |   32 |  17408 |    3.437 |  4767.38 |    1.503 |   681.49 |    4.939 |  3524.40 |
+|  4096 |     32 |    1 |   4128 |    0.907 |  4513.91 |    0.407 |    78.54 |    1.315 |  3139.56 |
+|  4096 |     32 |    2 |   8256 |    1.796 |  4560.42 |    0.625 |   102.37 |    2.422 |  3409.45 |
+|  4096 |     32 |    4 |  16512 |    3.596 |  4555.66 |    0.888 |   144.11 |    4.485 |  3681.93 |
+|  4096 |     32 |    8 |  33024 |    7.184 |  4561.44 |    1.098 |   233.11 |    8.282 |  3987.51 |
+|  4096 |     32 |   16 |  66048 |   14.369 |  4560.82 |    1.503 |   340.74 |   15.872 |  4161.30 |
+|  4096 |     32 |   32 | 132096 |   28.760 |  4557.52 |    2.162 |   473.59 |   30.922 |  4271.95 |
+|  8192 |     32 |    1 |   8224 |    1.859 |  4405.59 |    0.430 |    74.36 |    2.290 |  3591.61 |
+|  8192 |     32 |    2 |  16448 |    3.698 |  4430.02 |    0.656 |    97.59 |    4.354 |  3777.47 |
+|  8192 |     32 |    4 |  32896 |    7.403 |  4426.10 |    0.957 |   133.82 |    8.360 |  3934.97 |
+|  8192 |     32 |    8 |  65792 |   14.802 |  4427.63 |    1.222 |   209.44 |   16.024 |  4105.87 |
+|  8192 |     32 |   16 | 131584 |   29.596 |  4428.67 |    1.741 |   294.13 |   31.337 |  4199.00 |
+|  8192 |     32 |   32 | 263168 |   59.169 |  4430.42 |    2.619 |   390.92 |   61.789 |  4259.17 |


 - `llama-bench`

-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |      3714.25 ± 20.36 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         86.58 ± 0.43 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |      3445.17 ± 17.85 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         81.72 ± 0.53 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |      3218.78 ± 11.34 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         74.86 ± 0.64 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       2732.83 ± 7.17 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         71.57 ± 0.51 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      2119.75 ± 12.81 |
-| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         62.33 ± 0.24 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      4505.82 ± 12.90 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         83.43 ± 0.59 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |      4158.34 ± 18.84 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         79.22 ± 0.60 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      3993.81 ± 17.55 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         75.22 ± 1.05 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |      3449.98 ± 12.13 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         70.36 ± 0.37 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      2689.42 ± 18.89 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         61.65 ± 0.30 |

-build: eeee367de (6989)
+build: 11fb327bf (7941)

 ## ggml-org/gpt-oss-120b-GGUF

@@ -77,46 +77,46 @@ Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
 - `llama-batched-bench`


-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20

 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.571 |   897.18 |    0.543 |    58.96 |    1.113 |   488.60 |
-|   512 |     32 |    2 |   1088 |    0.593 |  1725.37 |    1.041 |    61.45 |    1.635 |   665.48 |
-|   512 |     32 |    4 |   2176 |    1.043 |  1963.15 |    1.334 |    95.95 |    2.377 |   915.36 |
-|   512 |     32 |    8 |   4352 |    2.099 |  1951.63 |    1.717 |   149.07 |    3.816 |  1140.45 |
-|   512 |     32 |   16 |   8704 |    4.207 |  1947.12 |    2.311 |   221.56 |    6.518 |  1335.35 |
-|   512 |     32 |   32 |  17408 |    8.422 |  1945.36 |    3.298 |   310.46 |   11.720 |  1485.27 |
-|  4096 |     32 |    1 |   4128 |    2.138 |  1915.88 |    0.571 |    56.09 |    2.708 |  1524.12 |
-|  4096 |     32 |    2 |   8256 |    4.266 |  1920.25 |    1.137 |    56.27 |    5.404 |  1527.90 |
-|  4096 |     32 |    4 |  16512 |    8.564 |  1913.02 |    1.471 |    86.99 |   10.036 |  1645.29 |
-|  4096 |     32 |    8 |  33024 |   17.092 |  1917.19 |    1.979 |   129.33 |   19.071 |  1731.63 |
-|  4096 |     32 |   16 |  66048 |   34.211 |  1915.65 |    2.850 |   179.66 |   37.061 |  1782.15 |
-|  4096 |     32 |   32 | 132096 |   68.394 |  1916.44 |    4.381 |   233.72 |   72.775 |  1815.13 |
-|  8192 |     32 |    1 |   8224 |    4.349 |  1883.45 |    0.620 |    51.65 |    4.969 |  1655.04 |
-|  8192 |     32 |    2 |  16448 |    8.674 |  1888.83 |    1.178 |    54.33 |    9.852 |  1669.48 |
-|  8192 |     32 |    4 |  32896 |   17.351 |  1888.55 |    1.580 |    81.01 |   18.931 |  1737.68 |
-|  8192 |     32 |    8 |  65792 |   34.743 |  1886.31 |    2.173 |   117.80 |   36.916 |  1782.20 |
-|  8192 |     32 |   16 | 131584 |   69.413 |  1888.29 |    3.297 |   155.28 |   72.710 |  1809.70 |
-|  8192 |     32 |   32 | 263168 |  138.903 |  1887.24 |    5.004 |   204.63 |  143.907 |  1828.73 |
+|   512 |     32 |    1 |    544 |    0.445 |  1151.80 |    0.560 |    57.14 |    1.005 |   541.53 |
+|   512 |     32 |    2 |   1088 |    0.472 |  2169.85 |    0.874 |    73.27 |    1.345 |   808.65 |
+|   512 |     32 |    4 |   2176 |    0.826 |  2480.33 |    1.299 |    98.51 |    2.125 |  1023.94 |
+|   512 |     32 |    8 |   4352 |    1.644 |  2491.67 |    1.608 |   159.18 |    3.252 |  1338.20 |
+|   512 |     32 |   16 |   8704 |    3.292 |  2488.35 |    2.117 |   241.85 |    5.409 |  1609.13 |
+|   512 |     32 |   32 |  17408 |    6.604 |  2481.07 |    2.898 |   353.31 |    9.502 |  1832.04 |
+|  4096 |     32 |    1 |   4128 |    1.698 |  2412.65 |    0.580 |    55.21 |    2.277 |  1812.66 |
+|  4096 |     32 |    2 |   8256 |    3.399 |  2409.88 |    0.934 |    68.53 |    4.333 |  1905.27 |
+|  4096 |     32 |    4 |  16512 |    6.823 |  2401.21 |    1.411 |    90.72 |    8.234 |  2005.30 |
+|  4096 |     32 |    8 |  33024 |   13.574 |  2413.97 |    1.841 |   139.07 |   15.415 |  2142.31 |
+|  4096 |     32 |   16 |  66048 |   27.176 |  2411.52 |    2.609 |   196.26 |   29.785 |  2217.49 |
+|  4096 |     32 |   32 | 132096 |   54.359 |  2411.23 |    3.905 |   262.20 |   58.264 |  2267.19 |
+|  8192 |     32 |    1 |   8224 |    3.491 |  2346.81 |    0.613 |    52.23 |    4.103 |  2004.21 |
+|  8192 |     32 |    2 |  16448 |    6.939 |  2361.03 |    0.981 |    65.21 |    7.921 |  2076.56 |
+|  8192 |     32 |    4 |  32896 |   13.888 |  2359.40 |    1.511 |    84.71 |   15.399 |  2136.21 |
+|  8192 |     32 |    8 |  65792 |   27.756 |  2361.18 |    2.034 |   125.86 |   29.790 |  2208.56 |
+|  8192 |     32 |   16 | 131584 |   55.554 |  2359.34 |    3.021 |   169.49 |   58.575 |  2246.41 |
+|  8192 |     32 |   32 | 263168 |  111.036 |  2360.89 |    4.537 |   225.72 |  115.573 |  2277.08 |


 - `llama-bench`

-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       1919.36 ± 5.01 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         60.40 ± 0.30 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       1825.30 ± 6.37 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         56.94 ± 0.29 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       1739.19 ± 6.00 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         52.51 ± 0.42 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1536.75 ± 4.27 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         49.33 ± 0.27 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1255.85 ± 3.26 |
-| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         42.99 ± 0.18 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |       2443.91 ± 7.47 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         58.72 ± 0.20 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2309.84 ± 3.63 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         55.67 ± 0.35 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      2216.68 ± 10.16 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         52.87 ± 0.43 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1956.31 ± 6.39 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         49.45 ± 0.20 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      1567.08 ± 11.79 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         42.76 ± 0.14 |

-build: eeee367de (6989)
+build: 11fb327bf (7941)

 ## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF

@@ -125,46 +125,46 @@ Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
 - `llama-batched-bench`


-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20

 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.398 |  1285.90 |    0.530 |    60.41 |    0.928 |   586.27 |
-|   512 |     32 |    2 |   1088 |    0.386 |  2651.65 |    0.948 |    67.50 |    1.334 |   815.38 |
-|   512 |     32 |    4 |   2176 |    0.666 |  3076.37 |    1.209 |   105.87 |    1.875 |  1160.71 |
-|   512 |     32 |    8 |   4352 |    1.325 |  3091.39 |    1.610 |   158.98 |    2.935 |  1482.65 |
-|   512 |     32 |   16 |   8704 |    2.664 |  3075.58 |    2.150 |   238.19 |    4.813 |  1808.39 |
-|   512 |     32 |   32 |  17408 |    5.336 |  3070.31 |    2.904 |   352.59 |    8.240 |  2112.50 |
-|  4096 |     32 |    1 |   4128 |    1.444 |  2836.81 |    0.581 |    55.09 |    2.025 |  2038.81 |
-|  4096 |     32 |    2 |   8256 |    2.872 |  2852.14 |    1.084 |    59.06 |    3.956 |  2086.99 |
-|  4096 |     32 |    4 |  16512 |    5.744 |  2852.32 |    1.440 |    88.90 |    7.184 |  2298.47 |
-|  4096 |     32 |    8 |  33024 |   11.463 |  2858.68 |    2.068 |   123.78 |   13.531 |  2440.65 |
-|  4096 |     32 |   16 |  66048 |   22.915 |  2859.95 |    3.018 |   169.67 |   25.933 |  2546.90 |
-|  4096 |     32 |   32 | 132096 |   45.956 |  2852.10 |    4.609 |   222.18 |   50.565 |  2612.39 |
-|  8192 |     32 |    1 |   8224 |    3.063 |  2674.72 |    0.693 |    46.20 |    3.755 |  2189.92 |
-|  8192 |     32 |    2 |  16448 |    6.109 |  2681.87 |    1.214 |    52.71 |    7.323 |  2245.98 |
-|  8192 |     32 |    4 |  32896 |   12.197 |  2686.63 |    1.682 |    76.11 |   13.878 |  2370.30 |
-|  8192 |     32 |    8 |  65792 |   24.409 |  2684.94 |    2.556 |   100.17 |   26.965 |  2439.95 |
-|  8192 |     32 |   16 | 131584 |   48.753 |  2688.50 |    3.994 |   128.20 |   52.747 |  2494.64 |
-|  8192 |     32 |   32 | 263168 |   97.508 |  2688.42 |    6.528 |   156.86 |  104.037 |  2529.57 |
+|   512 |     32 |    1 |    544 |    0.393 |  1303.73 |    0.548 |    58.36 |    0.941 |   578.10 |
+|   512 |     32 |    2 |   1088 |    0.387 |  2648.68 |    0.910 |    70.35 |    1.296 |   839.27 |
+|   512 |     32 |    4 |   2176 |    0.659 |  3107.63 |    1.302 |    98.33 |    1.961 |  1109.77 |
+|   512 |     32 |    8 |   4352 |    1.322 |  3099.35 |    1.669 |   153.42 |    2.990 |  1455.43 |
+|   512 |     32 |   16 |   8704 |    2.639 |  3104.63 |    2.212 |   231.44 |    4.851 |  1794.32 |
+|   512 |     32 |   32 |  17408 |    5.284 |  3100.80 |    2.955 |   346.53 |    8.239 |  2112.93 |
+|  4096 |     32 |    1 |   4128 |    1.417 |  2890.36 |    0.598 |    53.51 |    2.015 |  2048.45 |
+|  4096 |     32 |    2 |   8256 |    2.829 |  2895.62 |    1.019 |    62.82 |    3.848 |  2145.60 |
+|  4096 |     32 |    4 |  16512 |    5.656 |  2896.96 |    1.528 |    83.79 |    7.183 |  2298.71 |
+|  4096 |     32 |    8 |  33024 |   11.338 |  2890.02 |    2.127 |   120.36 |   13.465 |  2452.53 |
+|  4096 |     32 |   16 |  66048 |   22.709 |  2885.96 |    3.104 |   164.97 |   25.812 |  2558.79 |
+|  4096 |     32 |   32 | 132096 |   45.301 |  2893.35 |    4.723 |   216.80 |   50.024 |  2640.63 |
+|  8192 |     32 |    1 |   8224 |    3.022 |  2711.09 |    0.678 |    47.20 |    3.700 |  2222.89 |
+|  8192 |     32 |    2 |  16448 |    6.039 |  2713.01 |    1.149 |    55.70 |    7.188 |  2288.21 |
+|  8192 |     32 |    4 |  32896 |   12.050 |  2719.35 |    1.785 |    71.69 |   13.835 |  2377.67 |
+|  8192 |     32 |    8 |  65792 |   24.113 |  2717.90 |    2.629 |    97.39 |   26.741 |  2460.31 |
+|  8192 |     32 |   16 | 131584 |   48.178 |  2720.58 |    4.099 |   124.91 |   52.277 |  2517.06 |
+|  8192 |     32 |   32 | 263168 |   96.401 |  2719.31 |    6.696 |   152.93 |  103.097 |  2552.63 |


 - `llama-bench`

-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       2925.55 ± 4.25 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         62.80 ± 0.27 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       2531.01 ± 6.79 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         55.86 ± 0.33 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       2244.39 ± 5.33 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         45.95 ± 0.33 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1783.17 ± 3.68 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         39.07 ± 0.10 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1241.90 ± 3.13 |
-| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         29.92 ± 0.06 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      2986.97 ± 18.87 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         61.06 ± 0.23 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2633.45 ± 6.26 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         54.77 ± 0.28 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |       2354.14 ± 3.84 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         48.02 ± 0.40 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1908.86 ± 4.25 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         40.23 ± 0.10 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |       1348.17 ± 2.00 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         30.21 ± 0.04 |

-build: eeee367de (6989)
+build: 11fb327bf (7941)

 ## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF

@@ -173,46 +173,46 @@ Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
 - `llama-batched-bench`


-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20

 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.211 |  2421.57 |    1.055 |    30.33 |    1.266 |   429.57 |
-|   512 |     32 |    2 |   1088 |    0.419 |  2441.34 |    1.130 |    56.65 |    1.549 |   702.32 |
-|   512 |     32 |    4 |   2176 |    0.873 |  2345.54 |    1.174 |   108.99 |    2.048 |  1062.74 |
-|   512 |     32 |    8 |   4352 |    1.727 |  2371.85 |    1.254 |   204.22 |    2.980 |  1460.19 |
-|   512 |     32 |   16 |   8704 |    3.452 |  2373.22 |    1.492 |   343.16 |    4.944 |  1760.56 |
-|   512 |     32 |   32 |  17408 |    6.916 |  2368.93 |    1.675 |   611.51 |    8.591 |  2026.36 |
-|  4096 |     32 |    1 |   4128 |    1.799 |  2277.26 |    1.084 |    29.51 |    2.883 |  1431.91 |
-|  4096 |     32 |    2 |   8256 |    3.577 |  2290.01 |    1.196 |    53.50 |    4.774 |  1729.51 |
-|  4096 |     32 |    4 |  16512 |    7.172 |  2284.36 |    1.313 |    97.50 |    8.485 |  1946.00 |
-|  4096 |     32 |    8 |  33024 |   14.341 |  2284.96 |    1.520 |   168.46 |   15.860 |  2082.18 |
-|  4096 |     32 |   16 |  66048 |   28.675 |  2285.44 |    1.983 |   258.21 |   30.658 |  2154.33 |
-|  4096 |     32 |   32 | 132096 |   57.354 |  2285.32 |    2.640 |   387.87 |   59.994 |  2201.82 |
-|  8192 |     32 |    1 |   8224 |    3.701 |  2213.75 |    1.119 |    28.59 |    4.820 |  1706.34 |
-|  8192 |     32 |    2 |  16448 |    7.410 |  2211.19 |    1.272 |    50.31 |    8.682 |  1894.56 |
-|  8192 |     32 |    4 |  32896 |   14.802 |  2213.83 |    1.460 |    87.68 |   16.261 |  2022.96 |
-|  8192 |     32 |    8 |  65792 |   29.609 |  2213.35 |    1.781 |   143.74 |   31.390 |  2095.93 |
-|  8192 |     32 |   16 | 131584 |   59.229 |  2212.96 |    2.495 |   205.17 |   61.725 |  2131.79 |
-|  8192 |     32 |   32 | 263168 |  118.449 |  2213.15 |    3.714 |   275.75 |  122.162 |  2154.25 |
+|   512 |     32 |    1 |    544 |    0.212 |  2420.12 |    1.100 |    29.10 |    1.311 |   414.85 |
+|   512 |     32 |    2 |   1088 |    0.428 |  2393.89 |    1.185 |    54.00 |    1.613 |   674.56 |
+|   512 |     32 |    4 |   2176 |    0.894 |  2290.41 |    1.229 |   104.17 |    2.123 |  1025.02 |
+|   512 |     32 |    8 |   4352 |    1.758 |  2330.36 |    1.319 |   194.15 |    3.076 |  1414.70 |
+|   512 |     32 |   16 |   8704 |    3.508 |  2335.21 |    1.543 |   331.90 |    5.051 |  1723.33 |
+|   512 |     32 |   32 |  17408 |    7.035 |  2328.93 |    1.738 |   589.21 |    8.773 |  1984.29 |
+|  4096 |     32 |    1 |   4128 |    1.831 |  2237.25 |    1.125 |    28.44 |    2.956 |  1396.42 |
+|  4096 |     32 |    2 |   8256 |    3.642 |  2249.48 |    1.253 |    51.07 |    4.895 |  1686.64 |
+|  4096 |     32 |    4 |  16512 |    7.274 |  2252.26 |    1.380 |    92.72 |    8.655 |  1907.81 |
+|  4096 |     32 |    8 |  33024 |   14.576 |  2248.09 |    1.617 |   158.29 |   16.193 |  2039.37 |
+|  4096 |     32 |   16 |  66048 |   29.138 |  2249.17 |    2.081 |   246.01 |   31.219 |  2115.63 |
+|  4096 |     32 |   32 | 132096 |   58.275 |  2249.19 |    2.814 |   363.87 |   61.089 |  2162.34 |
+|  8192 |     32 |    1 |   8224 |    3.757 |  2180.26 |    1.184 |    27.03 |    4.941 |  1664.37 |
+|  8192 |     32 |    2 |  16448 |    7.522 |  2178.05 |    1.341 |    47.73 |    8.863 |  1855.77 |
+|  8192 |     32 |    4 |  32896 |   15.043 |  2178.25 |    1.548 |    82.69 |   16.591 |  1982.74 |
+|  8192 |     32 |    8 |  65792 |   30.111 |  2176.49 |    1.937 |   132.13 |   32.048 |  2052.90 |
+|  8192 |     32 |   16 | 131584 |   60.405 |  2169.90 |    2.706 |   189.21 |   63.111 |  2084.97 |
+|  8192 |     32 |   32 | 263168 |  120.439 |  2176.58 |    3.993 |   256.46 |  124.432 |  2114.96 |


 - `llama-bench`

-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |       2272.74 ± 4.68 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         30.66 ± 0.02 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       2107.80 ± 9.55 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         29.71 ± 0.05 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |       1937.80 ± 6.75 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         28.86 ± 0.04 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |       1641.12 ± 1.78 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         27.24 ± 0.04 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |       1296.02 ± 2.67 |
-| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         23.78 ± 0.03 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |       2250.28 ± 6.41 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         29.43 ± 0.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |       2100.19 ± 8.96 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         28.61 ± 0.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |       2007.56 ± 4.16 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         27.38 ± 0.09 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |       1779.11 ± 6.42 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         25.72 ± 0.03 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |       1471.23 ± 1.71 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         22.51 ± 0.02 |

-build: eeee367de (6989)
+build: 11fb327bf (7941)

 ## ggml-org/gemma-3-4b-it-qat-GGUF

@@ -221,44 +221,91 @@ Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
 - `llama-batched-bench`


-main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20

 |    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
 |-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
-|   512 |     32 |    1 |    544 |    0.094 |  5434.73 |    0.394 |    81.21 |    0.488 |  1114.15 |
-|   512 |     32 |    2 |   1088 |    0.168 |  6091.68 |    0.498 |   128.52 |    0.666 |  1633.41 |
-|   512 |     32 |    4 |   2176 |    0.341 |  6010.68 |    0.542 |   236.37 |    0.882 |  2466.43 |
-|   512 |     32 |    8 |   4352 |    0.665 |  6161.46 |    0.678 |   377.74 |    1.342 |  3241.72 |
-|   512 |     32 |   16 |   8704 |    1.323 |  6193.19 |    0.902 |   567.41 |    2.225 |  3911.74 |
-|   512 |     32 |   32 |  17408 |    2.642 |  6202.03 |    1.231 |   832.03 |    3.872 |  4495.36 |
-|  4096 |     32 |    1 |   4128 |    0.701 |  5840.49 |    0.439 |    72.95 |    1.140 |  3621.23 |
-|  4096 |     32 |    2 |   8256 |    1.387 |  5906.82 |    0.574 |   111.48 |    1.961 |  4210.12 |
-|  4096 |     32 |    4 |  16512 |    2.758 |  5940.33 |    0.651 |   196.58 |    3.409 |  4843.33 |
-|  4096 |     32 |    8 |  33024 |    5.491 |  5967.56 |    0.876 |   292.40 |    6.367 |  5187.12 |
-|  4096 |     32 |   16 |  66048 |   10.978 |  5969.58 |    1.275 |   401.69 |   12.253 |  5390.38 |
-|  4096 |     32 |   32 | 132096 |   21.944 |  5972.93 |    1.992 |   514.16 |   23.936 |  5518.73 |
-|  8192 |     32 |    1 |   8224 |    1.402 |  5841.91 |    0.452 |    70.73 |    1.855 |  4434.12 |
-|  8192 |     32 |    2 |  16448 |    2.793 |  5865.34 |    0.637 |   100.55 |    3.430 |  4795.51 |
-|  8192 |     32 |    4 |  32896 |    5.564 |  5889.64 |    0.770 |   166.26 |    6.334 |  5193.95 |
-|  8192 |     32 |    8 |  65792 |   11.114 |  5896.44 |    1.122 |   228.07 |   12.237 |  5376.51 |
-|  8192 |     32 |   16 | 131584 |   22.210 |  5901.38 |    1.789 |   286.15 |   24.000 |  5482.74 |
-|  8192 |     32 |   32 | 263168 |   44.382 |  5906.56 |    3.044 |   336.38 |   47.426 |  5549.02 |
+|   512 |     32 |    1 |    544 |    0.092 |  5566.97 |    0.412 |    77.63 |    0.504 |  1078.95 |
+|   512 |     32 |    2 |   1088 |    0.161 |  6345.67 |    0.522 |   122.70 |    0.683 |  1593.06 |
+|   512 |     32 |    4 |   2176 |    0.325 |  6309.87 |    0.562 |   227.68 |    0.887 |  2453.87 |
+|   512 |     32 |    8 |   4352 |    0.643 |  6374.42 |    0.685 |   373.67 |    1.328 |  3277.94 |
+|   512 |     32 |   16 |   8704 |    1.277 |  6413.64 |    0.915 |   559.47 |    2.192 |  3970.01 |
+|   512 |     32 |   32 |  17408 |    2.518 |  6506.57 |    1.249 |   819.61 |    3.767 |  4620.64 |
+|  4096 |     32 |    1 |   4128 |    0.674 |  6079.68 |    0.453 |    70.60 |    1.127 |  3662.88 |
+|  4096 |     32 |    2 |   8256 |    1.335 |  6137.82 |    0.627 |   102.03 |    1.962 |  4208.11 |
+|  4096 |     32 |    4 |  16512 |    2.657 |  6167.35 |    0.749 |   170.92 |    3.405 |  4848.71 |
+|  4096 |     32 |    8 |  33024 |    5.307 |  6173.91 |    0.974 |   262.89 |    6.281 |  5257.53 |
+|  4096 |     32 |   16 |  66048 |   10.610 |  6176.96 |    1.379 |   371.42 |   11.988 |  5509.40 |
+|  4096 |     32 |   32 | 132096 |   21.213 |  6178.89 |    2.122 |   482.50 |   23.335 |  5660.82 |
+|  8192 |     32 |    1 |   8224 |    1.359 |  6027.34 |    0.467 |    68.52 |    1.826 |  4503.48 |
+|  8192 |     32 |    2 |  16448 |    2.699 |  6069.68 |    0.653 |    98.03 |    3.352 |  4906.68 |
+|  8192 |     32 |    4 |  32896 |    5.366 |  6106.74 |    0.818 |   156.55 |    6.184 |  5319.96 |
+|  8192 |     32 |    8 |  65792 |   10.755 |  6093.50 |    1.174 |   218.04 |   11.929 |  5515.22 |
+|  8192 |     32 |   16 | 131584 |   21.484 |  6100.82 |    1.829 |   279.90 |   23.314 |  5644.11 |
+|  8192 |     32 |   32 | 263168 |   42.950 |  6103.40 |    3.058 |   334.91 |   46.008 |  5720.05 |


 - `llama-bench`

-| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap |            test |                  t/s |
-| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --------------: | -------------------: |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |          pp2048 |      5810.04 ± 21.71 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |            tg32 |         84.54 ± 0.18 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d4096 |       5288.04 ± 3.54 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d4096 |         78.82 ± 1.37 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |  pp2048 @ d8192 |      4960.43 ± 16.64 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |    tg32 @ d8192 |         74.13 ± 0.30 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d16384 |      4495.92 ± 31.11 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d16384 |         72.37 ± 0.29 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 | pp2048 @ d32768 |      3746.90 ± 40.01 |
-| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   tg32 @ d32768 |         63.02 ± 0.20 |
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | mmap | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | ---: | --: | --------------: | -------------------: |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |          pp2048 |      5948.74 ± 10.61 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |            tg32 |         81.05 ± 0.20 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d4096 |      5652.69 ± 34.29 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d4096 |         76.37 ± 0.58 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |  pp2048 @ d8192 |      5509.57 ± 40.69 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |    tg32 @ d8192 |         71.61 ± 0.80 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d16384 |      5340.86 ± 36.92 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d16384 |         70.89 ± 0.34 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 | pp2048 @ d32768 |      5023.30 ± 13.52 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | CUDA       |  99 |     2048 |  1 |    0 |   1 |   tg32 @ d32768 |         62.28 ± 0.30 |

-build: eeee367de (6989)
+build: 11fb327bf (7941)

+## ggml-org/GLM-4.7-Flash-GGUF
+
+Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.433 |  1181.83 |    0.693 |    46.16 |    1.126 |   482.94 |
+|   512 |     32 |    2 |   1088 |    0.439 |  2334.46 |    1.034 |    61.89 |    1.473 |   738.75 |
+|   512 |     32 |    4 |   2176 |    0.772 |  2654.46 |    1.459 |    87.76 |    2.230 |   975.77 |
+|   512 |     32 |    8 |   4352 |    1.541 |  2658.78 |    2.043 |   125.31 |    3.583 |  1214.47 |
+|   512 |     32 |   16 |   8704 |    3.083 |  2656.91 |    2.675 |   191.42 |    5.758 |  1511.62 |
+|   512 |     32 |   32 |  17408 |    6.159 |  2660.12 |    3.615 |   283.24 |    9.774 |  1780.98 |
+|  4096 |     32 |    1 |   4128 |    1.915 |  2139.30 |    0.725 |    44.14 |    2.640 |  1563.83 |
+|  4096 |     32 |    2 |   8256 |    3.834 |  2136.40 |    1.119 |    57.21 |    4.953 |  1666.81 |
+|  4096 |     32 |    4 |  16512 |    7.636 |  2145.72 |    1.631 |    78.49 |    9.266 |  1781.93 |
+|  4096 |     32 |    8 |  33024 |   15.295 |  2142.40 |    2.344 |   109.21 |   17.639 |  1872.20 |
+|  4096 |     32 |   16 |  66048 |   30.573 |  2143.62 |    3.773 |   135.70 |   34.346 |  1923.04 |
+|  4096 |     32 |   32 | 132096 |   61.282 |  2138.82 |    5.795 |   176.71 |   67.077 |  1969.31 |
+|  8192 |     32 |    1 |   8224 |    4.510 |  1816.24 |    0.760 |    42.11 |    5.270 |  1560.44 |
+|  8192 |     32 |    2 |  16448 |    9.036 |  1813.19 |    1.206 |    53.06 |   10.242 |  1605.91 |
+|  8192 |     32 |    4 |  32896 |   18.070 |  1813.43 |    1.783 |    71.80 |   19.852 |  1657.03 |
+|  8192 |     32 |    8 |  65792 |   36.125 |  1814.15 |    2.635 |    97.14 |   38.760 |  1697.41 |
+|  8192 |     32 |   16 | 131584 |   72.367 |  1811.20 |    4.954 |   103.34 |   77.322 |  1701.77 |
+|  8192 |     32 |   32 | 263168 |  144.501 |  1814.13 |    8.103 |   126.37 |  152.604 |  1724.51 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | ngl | n_ubatch | fa | dio |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | -------: | -: | --: | --------------: | -------------------: |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |          pp2048 |      2364.18 ± 11.43 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |            tg32 |         48.68 ± 0.12 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |  pp2048 @ d4096 |       1684.13 ± 1.24 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |    tg32 @ d4096 |         44.62 ± 0.22 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |  pp2048 @ d8192 |       1314.68 ± 1.41 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |    tg32 @ d8192 |         42.59 ± 0.11 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 | pp2048 @ d16384 |        914.05 ± 3.32 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |   tg32 @ d16384 |         38.72 ± 0.13 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 | pp2048 @ d32768 |        567.20 ± 0.90 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | CUDA       |  99 |     2048 |  1 |   1 |   tg32 @ d32768 |         32.65 ± 0.09 |
+
+build: 11fb327bf (7941)
@@ -0,0 +1,298 @@
+## System info
+
+```bash
+uname -a
+Darwin gg-studio 25.2.0 Darwin Kernel Version 25.2.0: Tue Nov 18 21:07:05 PST 2025; root:xnu-12377.61.12~1/RELEASE_ARM64_T6020 arm64
+
+g++ --version
+Apple clang version 17.0.0 (clang-1700.3.19.1)
+Target: arm64-apple-darwin25.2.0
+```
+
+## ggml-org/gpt-oss-20b-GGUF
+
+Model: https://huggingface.co/ggml-org/gpt-oss-20b-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.215 |  2381.35 |    0.245 |   130.45 |    0.460 |  1181.81 |
+|   512 |     32 |    2 |   1088 |    0.379 |  2701.43 |    0.382 |   167.56 |    0.761 |  1429.67 |
+|   512 |     32 |    4 |   2176 |    0.721 |  2839.27 |    0.604 |   211.76 |    1.326 |  1641.32 |
+|   512 |     32 |    8 |   4352 |    1.433 |  2858.30 |    1.033 |   247.75 |    2.466 |  1764.57 |
+|   512 |     32 |   16 |   8704 |    2.853 |  2871.12 |    1.570 |   326.11 |    4.423 |  1967.77 |
+|   512 |     32 |   32 |  17408 |    5.699 |  2874.95 |    1.910 |   536.15 |    7.609 |  2287.88 |
+|  4096 |     32 |    1 |   4128 |    1.552 |  2638.56 |    0.334 |    95.72 |    1.887 |  2188.00 |
+|  4096 |     32 |    2 |   8256 |    3.084 |  2655.88 |    0.404 |   158.54 |    3.488 |  2366.86 |
+|  4096 |     32 |    4 |  16512 |    6.151 |  2663.78 |    0.652 |   196.39 |    6.802 |  2427.37 |
+|  4096 |     32 |    8 |  33024 |   12.288 |  2666.77 |    1.135 |   225.47 |   13.423 |  2460.27 |
+|  4096 |     32 |   16 |  66048 |   24.563 |  2668.12 |    1.762 |   290.55 |   26.325 |  2508.97 |
+|  4096 |     32 |   32 | 132096 |   49.114 |  2668.73 |    2.398 |   426.94 |   51.512 |  2564.35 |
+|  8192 |     32 |    1 |   8224 |    3.345 |  2448.78 |    0.275 |   116.46 |    3.620 |  2271.76 |
+|  8192 |     32 |    2 |  16448 |    6.665 |  2458.11 |    0.425 |   150.71 |    7.090 |  2319.91 |
+|  8192 |     32 |    4 |  32896 |   13.315 |  2460.92 |    0.691 |   185.21 |   14.006 |  2348.63 |
+|  8192 |     32 |    8 |  65792 |   26.611 |  2462.73 |    1.212 |   211.16 |   27.823 |  2364.62 |
+|  8192 |     32 |   16 | 131584 |   53.232 |  2462.27 |    1.919 |   266.83 |   55.151 |  2385.88 |
+|  8192 |     32 |   32 | 263168 |  110.455 |  2373.30 |    2.752 |   372.03 |  113.208 |  2324.64 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2713.40 ± 3.56 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |        129.97 ± 3.90 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       2324.59 ± 3.01 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |        123.38 ± 0.17 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |      1989.82 ± 30.11 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |        117.39 ± 0.33 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       1556.54 ± 6.22 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |        109.75 ± 0.42 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |       1122.63 ± 1.45 |
+| gpt-oss 20B MXFP4 MoE          |  11.27 GiB |    20.91 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         98.25 ± 0.08 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/gpt-oss-120b-GGUF
+
+Model: https://huggingface.co/ggml-org/gpt-oss-120b-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.426 |  1200.92 |    0.361 |    88.56 |    0.788 |   690.64 |
+|   512 |     32 |    2 |   1088 |    0.683 |  1500.14 |    0.545 |   117.35 |    1.228 |   886.02 |
+|   512 |     32 |    4 |   2176 |    1.204 |  1701.56 |    0.847 |   151.19 |    2.050 |  1061.34 |
+|   512 |     32 |    8 |   4352 |    2.402 |  1705.20 |    1.455 |   176.00 |    3.857 |  1128.45 |
+|   512 |     32 |   16 |   8704 |    4.802 |  1705.90 |    2.349 |   217.93 |    7.152 |  1217.08 |
+|   512 |     32 |   32 |  17408 |    9.593 |  1707.85 |    3.665 |   279.42 |   13.258 |  1313.01 |
+|  4096 |     32 |    1 |   4128 |    2.581 |  1587.08 |    0.390 |    82.12 |    2.970 |  1389.67 |
+|  4096 |     32 |    2 |   8256 |    5.124 |  1598.79 |    0.589 |   108.62 |    5.713 |  1445.10 |
+|  4096 |     32 |    4 |  16512 |   10.231 |  1601.47 |    0.928 |   137.98 |   11.158 |  1479.80 |
+|  4096 |     32 |    8 |  33024 |   20.468 |  1600.94 |    1.606 |   159.38 |   22.074 |  1496.04 |
+|  4096 |     32 |   16 |  66048 |   40.924 |  1601.42 |    2.639 |   193.99 |   43.563 |  1516.15 |
+|  4096 |     32 |   32 | 132096 |   81.819 |  1601.98 |    4.466 |   229.29 |   86.284 |  1530.94 |
+|  8192 |     32 |    1 |   8224 |    5.517 |  1484.74 |    0.409 |    78.16 |    5.927 |  1387.58 |
+|  8192 |     32 |    2 |  16448 |   11.008 |  1488.43 |    0.622 |   102.92 |   11.629 |  1414.34 |
+|  8192 |     32 |    4 |  32896 |   22.002 |  1489.29 |    0.987 |   129.66 |   22.990 |  1430.90 |
+|  8192 |     32 |    8 |  65792 |   46.051 |  1423.11 |    1.858 |   137.79 |   47.909 |  1373.27 |
+|  8192 |     32 |   16 | 131584 |   97.680 |  1341.85 |    2.872 |   178.28 |  100.552 |  1308.62 |
+|  8192 |     32 |   32 | 263168 |  176.407 |  1486.02 |    5.048 |   202.85 |  181.455 |  1450.32 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1648.69 ± 1.80 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         85.60 ± 0.52 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1429.86 ± 1.01 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         82.03 ± 0.12 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1257.90 ± 1.81 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         78.23 ± 0.33 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       1013.49 ± 0.70 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         73.20 ± 0.28 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        721.11 ± 0.58 |
+| gpt-oss 120B MXFP4 MoE         |  59.02 GiB |   116.83 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         65.52 ± 0.10 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
+
+Model: https://huggingface.co/ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.243 |  2109.23 |    0.419 |    76.34 |    0.662 |   821.84 |
+|   512 |     32 |    2 |   1088 |    0.406 |  2521.40 |    0.575 |   111.36 |    0.981 |  1109.27 |
+|   512 |     32 |    4 |   2176 |    0.744 |  2751.65 |    0.841 |   152.22 |    1.585 |  1372.71 |
+|   512 |     32 |    8 |   4352 |    1.479 |  2770.20 |    1.330 |   192.48 |    2.809 |  1549.53 |
+|   512 |     32 |   16 |   8704 |    2.951 |  2776.20 |    2.572 |   199.05 |    5.523 |  1575.93 |
+|   512 |     32 |   32 |  17408 |    5.899 |  2777.64 |    2.603 |   393.34 |    8.502 |  2047.54 |
+|  4096 |     32 |    1 |   4128 |    1.901 |  2154.15 |    0.474 |    67.58 |    2.375 |  1738.14 |
+|  4096 |     32 |    2 |   8256 |    3.788 |  2162.89 |    0.652 |    98.17 |    4.439 |  1859.69 |
+|  4096 |     32 |    4 |  16512 |    7.564 |  2166.18 |    0.990 |   129.24 |    8.554 |  1930.34 |
+|  4096 |     32 |    8 |  33024 |   15.121 |  2166.98 |    1.632 |   156.82 |   16.754 |  1971.12 |
+|  4096 |     32 |   16 |  66048 |   30.241 |  2167.09 |    3.166 |   161.72 |   33.407 |  1977.04 |
+|  4096 |     32 |   32 | 132096 |   60.474 |  2167.42 |    3.780 |   270.93 |   64.254 |  2055.86 |
+|  8192 |     32 |    1 |   8224 |    4.733 |  1730.92 |    0.483 |    66.29 |    5.215 |  1576.85 |
+|  8192 |     32 |    2 |  16448 |    9.459 |  1732.09 |    0.722 |    88.58 |   10.182 |  1615.46 |
+|  8192 |     32 |    4 |  32896 |   18.912 |  1732.65 |    1.120 |   114.26 |   20.032 |  1642.14 |
+|  8192 |     32 |    8 |  65792 |   37.797 |  1733.91 |    1.873 |   136.67 |   39.670 |  1658.49 |
+|  8192 |     32 |   16 | 131584 |   84.133 |  1557.92 |    3.718 |   137.72 |   87.850 |  1497.82 |
+|  8192 |     32 |   32 | 263168 |  157.550 |  1663.88 |    4.854 |   210.98 |  162.403 |  1620.46 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2453.11 ± 1.70 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         78.97 ± 0.46 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1569.46 ± 1.97 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         71.18 ± 0.37 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1145.51 ± 1.16 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         65.11 ± 0.36 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        741.04 ± 0.74 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         56.87 ± 0.14 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        431.31 ± 0.31 |
+| qwen3moe 30B.A3B Q8_0          |  30.25 GiB |    30.53 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         45.26 ± 0.11 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
+
+Model: https://huggingface.co/ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.339 |  1509.22 |    0.409 |    78.17 |    0.749 |   726.67 |
+|   512 |     32 |    2 |   1088 |    0.646 |  1584.93 |    0.483 |   132.45 |    1.129 |   963.45 |
+|   512 |     32 |    4 |   2176 |    1.258 |  1627.50 |    0.585 |   218.67 |    1.844 |  1180.21 |
+|   512 |     32 |    8 |   4352 |    2.506 |  1634.41 |    1.005 |   254.83 |    3.511 |  1239.64 |
+|   512 |     32 |   16 |   8704 |    5.007 |  1635.99 |    1.595 |   321.07 |    6.602 |  1318.38 |
+|   512 |     32 |   32 |  17408 |   10.007 |  1637.19 |    1.676 |   611.12 |   11.683 |  1490.03 |
+|  4096 |     32 |    1 |   4128 |    2.730 |  1500.46 |    0.431 |    74.31 |    3.160 |  1306.12 |
+|  4096 |     32 |    2 |   8256 |    5.446 |  1504.33 |    0.524 |   122.04 |    5.970 |  1382.91 |
+|  4096 |     32 |    4 |  16512 |   10.875 |  1506.59 |    0.662 |   193.45 |   11.537 |  1431.28 |
+|  4096 |     32 |    8 |  33024 |   21.749 |  1506.61 |    1.158 |   221.11 |   22.907 |  1441.64 |
+|  4096 |     32 |   16 |  66048 |   43.477 |  1507.36 |    1.901 |   269.32 |   45.378 |  1455.49 |
+|  4096 |     32 |   32 | 132096 |   86.954 |  1507.37 |    2.325 |   440.42 |   89.279 |  1479.59 |
+|  8192 |     32 |    1 |   8224 |    5.940 |  1379.21 |    0.449 |    71.20 |    6.389 |  1287.20 |
+|  8192 |     32 |    2 |  16448 |   11.865 |  1380.84 |    0.559 |   114.59 |   12.424 |  1323.92 |
+|  8192 |     32 |    4 |  32896 |   23.723 |  1381.25 |    0.728 |   175.80 |   24.452 |  1345.35 |
+|  8192 |     32 |    8 |  65792 |   47.434 |  1381.63 |    1.279 |   200.09 |   48.713 |  1350.60 |
+|  8192 |     32 |   16 | 131584 |   94.864 |  1381.69 |    2.198 |   232.97 |   97.061 |  1355.68 |
+|  8192 |     32 |   32 | 263168 |  189.743 |  1381.57 |    3.052 |   335.50 |  192.795 |  1365.01 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1565.91 ± 0.86 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         79.68 ± 0.39 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       1317.41 ± 1.02 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         74.70 ± 0.04 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       1134.65 ± 0.76 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         71.31 ± 0.12 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        886.46 ± 0.78 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         65.93 ± 0.06 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        612.21 ± 0.30 |
+| qwen2 7B Q8_0                  |   7.54 GiB |     7.62 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         56.83 ± 0.02 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/gemma-3-4b-it-qat-GGUF
+
+Model: https://huggingface.co/ggml-org/gemma-3-4b-it-qat-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.186 |  2748.06 |    0.235 |   136.28 |    0.421 |  1291.78 |
+|   512 |     32 |    2 |   1088 |    0.342 |  2990.95 |    0.312 |   204.99 |    0.655 |  1662.15 |
+|   512 |     32 |    4 |   2176 |    0.662 |  3092.69 |    0.404 |   316.97 |    1.066 |  2041.21 |
+|   512 |     32 |    8 |   4352 |    1.317 |  3110.41 |    0.579 |   441.80 |    1.896 |  2294.97 |
+|   512 |     32 |   16 |   8704 |    2.625 |  3120.23 |    1.207 |   424.08 |    3.833 |  2270.93 |
+|   512 |     32 |   32 |  17408 |    5.242 |  3125.34 |    1.299 |   788.23 |    6.541 |  2661.19 |
+|  4096 |     32 |    1 |   4128 |    1.408 |  2909.90 |    0.296 |   108.07 |    1.704 |  2422.95 |
+|  4096 |     32 |    2 |   8256 |    2.793 |  2933.40 |    0.325 |   197.00 |    3.118 |  2648.25 |
+|  4096 |     32 |    4 |  16512 |    5.567 |  2943.22 |    0.440 |   291.07 |    6.006 |  2749.05 |
+|  4096 |     32 |    8 |  33024 |   11.114 |  2948.23 |    0.640 |   400.26 |   11.754 |  2809.59 |
+|  4096 |     32 |   16 |  66048 |   22.217 |  2949.76 |    1.327 |   385.83 |   23.544 |  2805.26 |
+|  4096 |     32 |   32 | 132096 |   44.420 |  2950.77 |    1.553 |   659.30 |   45.973 |  2873.36 |
+|  8192 |     32 |    1 |   8224 |    2.860 |  2864.58 |    0.250 |   127.90 |    3.110 |  2644.42 |
+|  8192 |     32 |    2 |  16448 |    5.702 |  2873.63 |    0.335 |   191.07 |    6.036 |  2724.77 |
+|  8192 |     32 |    4 |  32896 |   11.383 |  2878.69 |    0.456 |   280.72 |   11.839 |  2778.63 |
+|  8192 |     32 |    8 |  65792 |   22.750 |  2880.75 |    0.671 |   381.48 |   23.421 |  2809.14 |
+|  8192 |     32 |   16 | 131584 |   45.484 |  2881.74 |    1.406 |   364.04 |   46.890 |  2806.22 |
+|  8192 |     32 |   32 | 263168 |   90.956 |  2882.10 |    1.793 |   570.98 |   92.749 |  2837.41 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       2923.59 ± 3.10 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |        134.28 ± 1.29 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |       2748.21 ± 3.05 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |        133.11 ± 0.08 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |       2641.45 ± 2.31 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |        125.85 ± 0.35 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |       2446.20 ± 2.94 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |        125.00 ± 0.12 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |       2129.18 ± 7.43 |
+| gemma3 4B Q4_0                 |   2.35 GiB |     3.88 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |        113.14 ± 0.10 |
+
+build: b828e18c7 (7948)
+
+## ggml-org/GLM-4.7-Flash-GGUF
+
+Model: https://huggingface.co/ggml-org/GLM-4.7-Flash-GGUF
+
+- `llama-batched-bench`
+
+
+main: n_kv_max = 270336, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = -1, n_threads = 16, n_threads_batch = 16
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    0.326 |  1568.69 |    0.522 |    61.28 |    0.849 |   641.09 |
+|   512 |     32 |    2 |   1088 |    0.528 |  1939.42 |    0.744 |    86.07 |    1.272 |   855.63 |
+|   512 |     32 |    4 |   2176 |    0.968 |  2114.85 |    1.105 |   115.85 |    2.073 |  1049.56 |
+|   512 |     32 |    8 |   4352 |    1.928 |  2124.62 |    1.684 |   151.99 |    3.612 |  1204.82 |
+|   512 |     32 |   16 |   8704 |    3.844 |  2131.34 |    3.141 |   162.99 |    6.985 |  1246.11 |
+|   512 |     32 |   32 |  17408 |    7.683 |  2132.38 |    3.924 |   260.95 |   11.608 |  1499.71 |
+|  4096 |     32 |    1 |   4128 |    3.280 |  1248.75 |    0.723 |    44.29 |    4.003 |  1031.33 |
+|  4096 |     32 |    2 |   8256 |    6.545 |  1251.63 |    0.930 |    68.85 |    7.475 |  1104.53 |
+|  4096 |     32 |    4 |  16512 |   13.080 |  1252.64 |    1.454 |    88.03 |   14.534 |  1136.12 |
+|  4096 |     32 |    8 |  33024 |   26.154 |  1252.90 |    2.388 |   107.20 |   28.542 |  1157.04 |
+|  4096 |     32 |   16 |  66048 |   52.297 |  1253.14 |    4.724 |   108.37 |   57.022 |  1158.30 |
+|  4096 |     32 |   32 | 132096 |  104.578 |  1253.34 |    7.266 |   140.93 |  111.844 |  1181.08 |
+|  8192 |     32 |    1 |   8224 |    9.623 |   851.31 |    0.767 |    41.72 |   10.390 |   791.54 |
+|  8192 |     32 |    2 |  16448 |   20.916 |   783.32 |    1.148 |    55.74 |   22.064 |   745.45 |
+|  8192 |     32 |    4 |  32896 |   43.509 |   753.14 |    1.833 |    69.82 |   45.342 |   725.51 |
+|  8192 |     32 |    8 |  65792 |   79.621 |   823.10 |    3.180 |    80.50 |   82.801 |   794.58 |
+|  8192 |     32 |   16 | 131584 |  153.770 |   852.39 |    6.502 |    78.74 |  160.272 |   821.00 |
+|  8192 |     32 |   32 | 263168 |  307.539 |   852.39 |   10.839 |    94.48 |  318.378 |   826.59 |
+
+
+- `llama-bench`
+
+| model                          |       size |     params | backend    | threads | n_ubatch | fa |            test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | ------: | -------: | -: | --------------: | -------------------: |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |          pp2048 |       1629.33 ± 0.27 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |            tg32 |         59.58 ± 0.13 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d4096 |        732.67 ± 0.42 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d4096 |         47.44 ± 0.15 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |  pp2048 @ d8192 |        474.33 ± 0.33 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |    tg32 @ d8192 |         40.20 ± 0.20 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d16384 |        277.46 ± 0.09 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d16384 |         31.50 ± 0.93 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 | pp2048 @ d32768 |        151.44 ± 0.05 |
+| deepseek2 30B.A3B Q8_0         |  29.65 GiB |    29.94 B | MTL,BLAS   |      16 |     2048 |  1 |   tg32 @ d32768 |         21.81 ± 0.01 |
+
+build: b828e18c7 (7948)
@@ -534,7 +534,7 @@ xcodebuild -create-xcframework \
    -framework $(pwd)/build-ios-device/framework/llama.framework \
    -debug-symbols $(pwd)/build-ios-device/dSYMs/llama.dSYM \
    -framework $(pwd)/build-macos/framework/llama.framework \
-    -debug-symbols $(pwd)/build-macos/dSYMS/llama.dSYM \
+    -debug-symbols $(pwd)/build-macos/dSYMs/llama.dSYM \
    -framework $(pwd)/build-visionos/framework/llama.framework \
    -debug-symbols $(pwd)/build-visionos/dSYMs/llama.dSYM \
    -framework $(pwd)/build-visionos-sim/framework/llama.framework \
@@ -635,6 +635,29 @@ function gg_check_build_requirements {
    fi
 }

+function gg_run_test_backend_ops_cpu {
+    cd ${SRC}
+
+    cd build-ci-release
+
+    set -e
+
+    (time ./bin/test-backend-ops -b CPU ) 2>&1 | tee -a $OUT/${ci}-test-backend-ops-cpu.log
+
+    set +e
+}
+
+function gg_sum_test_backend_ops_cpu {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs test-backend-ops for CPU backend\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-test-backend-ops-cpu.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
 ## main

 export LLAMA_LOG_PREFIX=1
@@ -663,6 +686,10 @@ ret=0
 test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release

+if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
+    test $ret -eq 0 && gg_run test_backend_ops_cpu
+fi
+
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
    test $ret -eq 0 && gg_run embd_bge_small
    test $ret -eq 0 && gg_run rerank_tiny
@@ -32,4 +32,27 @@ function(llama_add_compile_flags)
            set(CXX_FLAGS "" PARENT_SCOPE)
        endif()
    endif()
+
+    if (NOT MSVC)
+        if (LLAMA_SANITIZE_THREAD)
+            message(STATUS "Using -fsanitize=thread")
+
+            add_compile_options(-fsanitize=thread)
+            link_libraries     (-fsanitize=thread)
+        endif()
+
+        if (LLAMA_SANITIZE_ADDRESS)
+            message(STATUS "Using -fsanitize=address")
+
+            add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
+            link_libraries     (-fsanitize=address)
+        endif()
+
+        if (LLAMA_SANITIZE_UNDEFINED)
+            message(STATUS "Using -fsanitize=undefined")
+
+            add_compile_options(-fsanitize=undefined)
+            link_libraries     (-fsanitize=undefined)
+        endif()
+    endif()
 endfunction()
@@ -3437,16 +3437,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.speculative.ngram_size_m = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}));
-    add_opt(common_arg(
-        {"--spec-ngram-check-rate"}, "N",
-        string_format("ngram check rate for ngram-simple/ngram-map speculative decoding (default: %d)", params.speculative.ngram_check_rate),
-        [](common_params & params, int value) {
-            if (value < 1) {
-                throw std::invalid_argument("ngram check rate must be at least 1");
-            }
-            params.speculative.ngram_check_rate = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
    add_opt(common_arg(
        {"--spec-ngram-min-hits"}, "N",
        string_format("minimum hits for ngram-map speculative decoding (default: %d)", params.speculative.ngram_min_hits),
@@ -380,15 +380,46 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
    return msgs;
 }

-json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
+static json render_message_to_json(const std::vector<common_chat_msg> & msgs, const jinja::caps & c) {
+    if (!c.supports_string_content && !c.supports_typed_content) {
+        LOG_WRN("%s: Neither string content nor typed content is supported by the template. This is unexpected and may lead to issues.\n", __func__);
+    }
+
+    bool only_string_accepted =  c.supports_string_content && !c.supports_typed_content;
+    bool only_typed_accepted  = !c.supports_string_content &&  c.supports_typed_content;
+
    json messages = json::array();
    for (const auto & msg : msgs) {
-        json jmsg = msg.to_json_oaicompat(concat_typed_text);
-        messages.push_back(jmsg);
+        if (only_string_accepted) {
+            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ true);
+            messages.push_back(jmsg);
+        } else if (only_typed_accepted) {
+            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
+            if (jmsg.at("content").is_string()) {
+                jmsg["content"] = json::array({
+                    json{
+                        {"type", "text"},
+                        {"text", jmsg.at("content").get<std::string>()},
+                    }
+                });
+            }
+            messages.push_back(jmsg);
+        } else {
+            json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
+            messages.push_back(jmsg);
+        }
    }
    return messages;
 }

+// DEPRECATED: only used in tests
+json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text) {
+    jinja::caps c;
+    c.supports_string_content = true;
+    c.supports_typed_content = !concat_typed_text;
+    return render_message_to_json(msgs, c);
+}
+
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {
    std::vector<common_chat_tool> result;

@@ -3020,7 +3051,7 @@ static common_chat_params common_chat_templates_apply_jinja(
        : *tmpls->template_default;
    const auto & src = tmpl.source();
    const auto & caps = tmpl.original_caps();
-    params.messages = common_chat_msgs_to_json_oaicompat(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
+    params.messages = render_message_to_json(inputs.messages, tmpl.original_caps());
    params.add_generation_prompt = inputs.add_generation_prompt;
    params.tool_choice = inputs.tool_choice;
    params.reasoning_format = inputs.reasoning_format;
@@ -240,6 +240,8 @@ bool common_chat_templates_support_enable_thinking(const common_chat_templates *

 // Parses a JSON array of messages in OpenAI's chat completion API format.
 std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
+
+// DEPRECATED: only used in tests
 nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);

 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
@@ -1,7 +1,3 @@
-#if defined(_MSC_VER)
-#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
-#endif
-
 #include "ggml.h"
 #include "gguf.h"

@@ -9,12 +5,12 @@
 #include "log.h"
 #include "llama.h"
 #include "sampling.h"
+#include "unicode.h"

 #include <algorithm>
 #include <cinttypes>
 #include <climits>
 #include <cmath>
-#include <codecvt>
 #include <chrono>
 #include <cstdarg>
 #include <cstring>
@@ -706,45 +702,28 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
        return false;
    }

-    std::u32string filename_utf32;
-    try {
-#if defined(__clang__)
-        // disable C++17 deprecation warning for std::codecvt_utf8
-#    pragma clang diagnostic push
-#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic push
-#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
-#endif
+    size_t offset = 0;
+    while (offset < filename.size()) {
+        utf8_parse_result result = parse_utf8_codepoint(filename, offset);

-        std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
-
-#if defined(__clang__)
-#    pragma clang diagnostic pop
-#elif defined(__GNUC__)
-#    pragma GCC diagnostic pop
-#endif
-
-        filename_utf32 = converter.from_bytes(filename);
-
-        // If the reverse conversion mismatches, it means overlong UTF-8 sequences were used,
-        // or invalid encodings were encountered. Reject such attempts
-        std::string filename_reencoded = converter.to_bytes(filename_utf32);
-        if (filename_reencoded != filename) {
+        if (result.status != utf8_parse_result::SUCCESS) {
            return false;
        }
-    } catch (const std::exception &) {
-        return false;
-    }
+        uint32_t c = result.codepoint;

-    // Check for forbidden codepoints:
-    // - Control characters
-    // - Unicode equivalents of illegal characters
-    // - UTF-16 surrogate pairs
-    // - UTF-8 replacement character
-    // - Byte order mark (BOM)
-    // - Illegal characters: / \ : * ? " < > |
-    for (char32_t c : filename_utf32) {
+        if ((result.bytes_consumed == 2 && c < 0x80) ||
+            (result.bytes_consumed == 3 && c < 0x800) ||
+            (result.bytes_consumed == 4 && c < 0x10000)) {
+            return false;
+        }
+
+        // Check for forbidden codepoints:
+        // - Control characters
+        // - Unicode equivalents of illegal characters
+        // - UTF-16 surrogate pairs
+        // - UTF-8 replacement character
+        // - Byte order mark (BOM)
+        // - Illegal characters: / \ : * ? " < > |
        if (c <= 0x1F // Control characters (C0)
            || c == 0x7F // Control characters (DEL)
            || (c >= 0x80 && c <= 0x9F) // Control characters (C1)
@@ -752,6 +731,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
            || c == 0x2215 // Division Slash (forward slash equivalent)
            || c == 0x2216 // Set Minus (backslash equivalent)
            || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
+            || c > 0x10FFFF // Max Unicode limit
            || c == 0xFFFD // Replacement Character (UTF-8)
            || c == 0xFEFF // Byte Order Mark (BOM)
            || c == ':' || c == '*' // Illegal characters
@@ -762,6 +742,7 @@ bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
            // Subdirectories not allowed, reject path separators
            return false;
        }
+        offset += result.bytes_consumed;
    }

    // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
@@ -1469,66 +1450,6 @@ void common_batch_add(
    batch.n_tokens++;
 }

-//
-// Token utils
-//
-
-size_t common_lcp(const llama_tokens & a, const llama_tokens & b) {
-    size_t i;
-    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
-
-    return i;
-}
-
-size_t common_lcs(const llama_tokens & a, const llama_tokens & b) {
-    // check for empty sequences
-    if (a.empty() || b.empty()) {
-        return 0;
-    }
-
-    // get the lengths of the input sequences
-    size_t a_len = a.size();
-    size_t b_len = b.size();
-
-    // initialize the maximum length of the longest common subsequence (LCS)
-    size_t max_length = 0;
-
-    // use two rows instead of a 2D matrix to optimize space
-    std::vector<size_t> prev_row(b_len + 1, 0);
-    std::vector<size_t> curr_row(b_len + 1, 0);
-
-    // iterate through the elements of a
-    for (size_t i = 1; i <= a_len; i++) {
-        // iterate through the elements of b
-        for (size_t j = 1; j <= b_len; j++) {
-            // if elements at the current positions match
-            if (a[i - 1] == b[j - 1]) {
-                // if it's the first element of either sequences, set LCS length to 1
-                if (i == 1 || j == 1) {
-                    curr_row[j] = 1;
-                } else {
-                    // increment LCS length by 1 compared to the previous element
-                    curr_row[j] = prev_row[j - 1] + 1;
-                }
-
-                // update max_length if necessary
-                if (curr_row[j] > max_length) {
-                    max_length = curr_row[j];
-                }
-            } else {
-                // reset LCS length if elements don't match
-                curr_row[j] = 0;
-            }
-        }
-
-        // update the previous row for the next iteration
-        prev_row = curr_row;
-    }
-
-    // return the maximum length of the LCS
-    return max_length;
-}
-
 //
 // Vocab utils
 //
@@ -269,7 +269,6 @@ struct common_params_speculative {

    uint16_t ngram_size_n     = 12; // ngram size for lookup
    uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
-    uint16_t ngram_check_rate =  1; // check rate for ngram lookup
    uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed

    std::shared_ptr<common_ngram_mod> ngram_mod;
@@ -780,16 +779,6 @@ void common_batch_add(
    const std::vector<llama_seq_id> & seq_ids,
                               bool   logits);

-//
-// Token utils
-//
-
-// longest common prefix
-size_t common_lcp(const llama_tokens & a, const llama_tokens & b);
-
-// longet common subsequence
-size_t common_lcs(const llama_tokens & a, const llama_tokens & b);
-
 //
 // Vocab utils
 //
@@ -45,6 +45,8 @@ static float common_ggml_get_float_value(const uint8_t * data,
    return v;
 }

+#define INDENT "    "
+
 template <bool abort>
 void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
    GGML_ASSERT(n > 0);
@@ -60,41 +62,41 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
        }
    }
    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
-        LOG_ERR("                                     [\n");
+        LOG(INDENT "[\n");
        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
            if (i2 == n && ne[2] > 2 * n) {
-                LOG_ERR("                                      ..., \n");
+                LOG(INDENT INDENT "..., \n");
                i2 = ne[2] - n;
            }
-            LOG_ERR("                                      [\n");
+            LOG(INDENT INDENT "[\n");
            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
                if (i1 == n && ne[1] > 2 * n) {
-                    LOG_ERR("                                       ..., \n");
+                    LOG(INDENT INDENT INDENT "..., \n");
                    i1 = ne[1] - n;
                }
-                LOG_ERR("                                       [");
+                LOG(INDENT INDENT INDENT "[");
                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
                    if (i0 == n && ne[0] > 2 * n) {
-                        LOG_ERR("..., ");
+                        LOG("   ..., ");
                        i0 = ne[0] - n;
                    }
                    const float v = common_ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
-                    LOG_ERR("%12.4f", v);
+                    LOG("%12.4f", v);
                    if (i0 < ne[0] - 1) {
-                        LOG_ERR(", ");
+                        LOG(", ");
                    }
                }
-                LOG_ERR("],\n");
+                LOG("  ],\n");
            }
-            LOG_ERR("                                      ],\n");
+            LOG(INDENT INDENT "],\n");
        }
-        LOG_ERR("                                     ]\n");
-        LOG_ERR("                                     sum = %f\n", sum);
+        LOG(INDENT "]\n");
+        LOG(INDENT "sum = %f\n", sum);
    }

    if constexpr (abort) {
        if (std::isnan(sum)) {
-            LOG_ERR("encountered NaN - aborting\n");
+            LOG("encountered NaN - aborting\n");
            exit(0);
        }
    }
@@ -137,9 +139,9 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
    }

    if (matches_filter) {
-        LOG_ERR("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
-                ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
-                common_ggml_ne_string(t).c_str());
+        LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__, t->name, ggml_type_name(t->type),
+            ggml_op_desc(t), src0->name, common_ggml_ne_string(src0).c_str(), src1 ? src1_str : "",
+            common_ggml_ne_string(t).c_str());
    }

    const bool is_host = ggml_backend_buffer_is_host(t->buffer);
@@ -305,7 +305,10 @@ static bool common_pull_file(httplib::Client & cli,
    );

    if (!res) {
-        LOG_ERR("%s: error during download. Status: %d\n", __func__, res ? res->status : -1);
+        LOG_ERR("%s: download failed: %s (status: %d)\n",
+                __func__,
+                httplib::to_string(res.error()).c_str(),
+                res ? res->status : -1);
        return false;
    }

@@ -63,7 +63,8 @@ static void caps_print_stats(value & v, const std::string & path) {

 std::map<std::string, bool> caps::to_map() const {
    return {
-        {"requires_typed_content", requires_typed_content},
+        {"supports_string_content", supports_string_content},
+        {"supports_typed_content", supports_typed_content},
        {"supports_tools", supports_tools},
        {"supports_tool_calls", supports_tool_calls},
        {"supports_parallel_tool_calls", supports_parallel_tool_calls},
@@ -89,7 +90,7 @@ caps caps_get(jinja::program & prog) {
        return v->stats.ops.find(op_name) != v->stats.ops.end();
    };

-    // case: typed content requirement
+    // case: typed content support
    caps_try_execute(
        prog,
        [&]() {
@@ -105,12 +106,16 @@ caps caps_get(jinja::program & prog) {
            // tools
            return json{nullptr};
        },
-        [&](bool, value & messages, value &) {
+        [&](bool success, value & messages, value &) {
            auto & content = messages->at(0)->at("content");
            caps_print_stats(content, "messages[0].content");
            if (has_op(content, "selectattr") || has_op(content, "array_access")) {
                // accessed as an array
-                result.requires_typed_content = true;
+                result.supports_typed_content = true;
+            }
+            if (!success) {
+                // failed to execute with content as string
+                result.supports_string_content = false;
            }
        }
    );
@@ -14,7 +14,9 @@ struct caps {
    bool supports_parallel_tool_calls = true;
    bool supports_preserve_reasoning = false; // support assistant message with reasoning_content

-    bool requires_typed_content = false; // default: use string content
+    // one of the 2 content capabilities must be true
+    bool supports_string_content = true;
+    bool supports_typed_content = false;

    // for reporting on server
    std::map<std::string, bool> to_map() const;
@@ -144,6 +144,13 @@ value binary_expression::execute_impl(context & ctx) {
        return false;
    };

+    auto test_is_in = [&]() -> bool {
+        func_args args(ctx);
+        args.push_back(left_val);
+        args.push_back(right_val);
+        return global_builtins().at("test_is_in")(args)->as_bool();
+    };
+
    // Handle undefined and null values
    if (is_val<value_undefined>(left_val) || is_val<value_undefined>(right_val)) {
        if (is_val<value_undefined>(right_val) && (op.value == "in" || op.value == "not in")) {
@@ -223,19 +230,11 @@ value binary_expression::execute_impl(context & ctx) {
            return result;
        }
    } else if (is_val<value_array>(right_val)) {
-        auto & arr = right_val->as_array();
-        bool member = false;
-        for (const auto & item : arr) {
-            if (*left_val == *item) {
-                member = true;
-                break;
-            }
-        }
+        // case: 1 in [0, 1, 2]
+        bool member = test_is_in();
        if (op.value == "in") {
-            JJ_DEBUG("Checking membership: %s in Array is %d", left_val->type().c_str(), member);
            return mk_val<value_bool>(member);
        } else if (op.value == "not in") {
-            JJ_DEBUG("Checking non-membership: %s not in Array is %d", left_val->type().c_str(), !member);
            return mk_val<value_bool>(!member);
        }
    }
@@ -252,22 +251,23 @@ value binary_expression::execute_impl(context & ctx) {

    // String membership
    if (is_val<value_string>(left_val) && is_val<value_string>(right_val)) {
-        auto left_str = left_val->as_string().str();
-        auto right_str = right_val->as_string().str();
+        // case: "a" in "abc"
+        bool member = test_is_in();
        if (op.value == "in") {
-            return mk_val<value_bool>(right_str.find(left_str) != std::string::npos);
+            return mk_val<value_bool>(member);
        } else if (op.value == "not in") {
-            return mk_val<value_bool>(right_str.find(left_str) == std::string::npos);
+            return mk_val<value_bool>(!member);
        }
    }

    // Value key in object
    if (is_val<value_object>(right_val)) {
-        bool has_key = right_val->has_key(left_val);
+        // case: key in {key: value}
+        bool member = test_is_in();
        if (op.value == "in") {
-            return mk_val<value_bool>(has_key);
+            return mk_val<value_bool>(member);
        } else if (op.value == "not in") {
-            return mk_val<value_bool>(!has_key);
+            return mk_val<value_bool>(!member);
        }
    }

@@ -446,6 +446,12 @@ value for_statement::execute_impl(context & ctx) {

    value iterable_val = iter_expr->execute(scope);

+    // mark the variable being iterated as used for stats
+    if (ctx.is_get_stats) {
+        iterable_val->stats.used = true;
+        iterable_val->stats.ops.insert("array_access");
+    }
+
    if (iterable_val->is_undefined()) {
        JJ_DEBUG("%s", "For loop iterable is undefined, skipping loop");
        iterable_val = mk_val<value_array>();
@@ -393,6 +393,33 @@ const func_builtins & global_builtins() {
        {"test_is_lt", test_compare_fn<value_compare_op::lt>},
        {"test_is_lessthan", test_compare_fn<value_compare_op::lt>},
        {"test_is_ne", test_compare_fn<value_compare_op::ne>},
+        {"test_is_in", [](const func_args & args) -> value {
+            args.ensure_count(2);
+            auto needle   = args.get_pos(0);
+            auto haystack = args.get_pos(1);
+            if (is_val<value_undefined>(haystack)) {
+                return mk_val<value_bool>(false);
+            }
+            if (is_val<value_array>(haystack)) {
+                for (const auto & item : haystack->as_array()) {
+                    if (*needle == *item) {
+                        return mk_val<value_bool>(true);
+                    }
+                }
+                return mk_val<value_bool>(false);
+            }
+            if (is_val<value_string>(haystack)) {
+                if (!is_val<value_string>(needle)) {
+                    throw raised_exception("'in' test expects args[1] as string when args[0] is string, got args[1] as " + needle->type());
+                }
+                return mk_val<value_bool>(
+                    haystack->as_string().str().find(needle->as_string().str()) != std::string::npos);
+            }
+            if (is_val<value_object>(haystack)) {
+                return mk_val<value_bool>(haystack->has_key(needle));
+            }
+            throw raised_exception("'in' test expects iterable as first argument, got " + haystack->type());
+        }},
        {"test_is_test", [](const func_args & args) -> value {
            args.ensure_vals<value_string>();
            auto & builtins = global_builtins();
@@ -47,21 +47,15 @@ static std::string common_tokens_to_str(const llama_tokens & inp, size_t start,
 * @return Vector of draft tokens, empty if no matching pattern is found
 */
 llama_tokens common_ngram_simple_draft(
-        common_ngram_simple_state & state,
+        const common_ngram_simple_config & config,
        const llama_tokens & tokens, llama_token sampled) {

    // Simple implementation of self-speculative decoding without a draft model.
    //
    const size_t cur_len = tokens.size();
-    // Only check every check_rate tokens to save compute
-    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
-    if (state.idx_last_check + state.config.check_rate > cur_len) {
-        llama_tokens draft_tokens;
-        return draft_tokens;
-    }

-    size_t n_draft_min = state.config.size_ngram; // size of n-gram to lookup in token history
-    size_t n_draft_max = state.config.size_mgram; // the m-gram following the found n-gram is used for draft
+    const size_t n_draft_min = config.size_ngram; // size of n-gram to lookup in token history
+    const size_t n_draft_max = config.size_mgram; // the m-gram following the found n-gram is used for draft

    // vector for tokens we want to verify.
    // return empty vector if there is no match.
@@ -80,9 +74,6 @@ llama_tokens common_ngram_simple_draft(
    }
    pattern.push_back(sampled); // add the last token to the pattern

-    // We do a search in the token history.
-    state.idx_last_check = cur_len;
-
    size_t match_pos = 0; // we ignore position 0, position 0 == no match
                          // search backwards, but skip the current match (we are currently there)
    for (size_t j = cur_len - n_draft_min - 1; j > 0; --j) {
@@ -240,10 +231,9 @@ void common_ngram_map_draft(common_ngram_map & map,
        GGML_ABORT("%s: cur_len exceeds UINT32_MAX: %zu", __func__, cur_len);
    }

-    // Only check every check_rate tokens to save compute
-    // i.e., perform check if (cur_len - idx_last_check) >= check_rate
-    if (map.idx_last_check + map.check_rate > cur_len) {
-        return;
+    if (map.idx_last_check  > cur_len) {
+        // Should not happen because of common_ngram_map_begin().
+        GGML_ABORT("%s: map.idx_last_check > cur_len: %zu > %zu", __func__, map.idx_last_check, cur_len);
    }
    map.idx_last_check = cur_len;

@@ -471,7 +461,7 @@ void common_ngram_map_draft(common_ngram_map & map,
            slot_max = v;
        }
    }
-    // What is sum of the other occurences?
+    // What is sum of the other occurrences?
    uint32_t sum_occur = 0;
    for (int v = 0; v < COMMON_NGRAM_MAX_VALUES; ++v) {
        if (v == slot_max) {
@@ -24,26 +24,11 @@
 struct common_ngram_simple_config {
    uint16_t   size_ngram;      // size of n-grams to lookup in self-mode
    uint16_t   size_mgram;      // size of m-grams to draft in self-mode
-    uint16_t   check_rate;      // check for speculative decoding without draft model for each check_rate token
-};
-
-// current state (and config) of n-gram simple.
-struct common_ngram_simple_state {
-    common_ngram_simple_config config;
-
-    size_t idx_last_check = 0; // index of last check in context history (mutable)
-
-    common_ngram_simple_state(const common_ngram_simple_config & config)
-        : config(config) {}
 };

 // Searches for a n-gram in the history and checks whether a draft sequence should be generated.
-// state:              the ngram simple state to search in.
-// inp:                the tokens generated so far.
-// sampled:            the token that was just sampled.
-// draft:              vector to store the draft tokens, initially empty.
 llama_tokens common_ngram_simple_draft(
-        common_ngram_simple_state & state,
+        const common_ngram_simple_config & config,
        const llama_tokens & tokens, llama_token sampled);


@@ -59,7 +44,7 @@ llama_tokens common_ngram_simple_draft(
 // statistics of a m-gram after a known n-gram
 struct common_ngram_map_value {
    size_t   value_idx =  0;  // index of value m-gram in token-history (0 if unused)
-    uint16_t value_num =  0;  // number of occurences of this value m-gram after the key n-gram (0 in an unused values-slot)
+    uint16_t value_num =  0;  // number of occurrences of this value m-gram after the key n-gram (0 in an unused values-slot)
    int16_t n_accepted = -1;  // number of accepted tokens at last draft (-1 if unused)
 };

@@ -68,7 +53,7 @@ struct common_ngram_map_key {
    size_t   key_idx;   // index of key n-gram in token-history
    size_t   stat_idx;  // index of last token of stastistics computation (key_num, values)

-    uint16_t key_num;   // number of occurences of this key n-gram in token-history
+    uint16_t key_num;   // number of occurrences of this key n-gram in token-history
    common_ngram_map_value values[COMMON_NGRAM_MAX_VALUES]; // some known values after the key
 };

@@ -80,15 +65,14 @@ struct common_ngram_map {
    bool key_only;       // true if only key n-grams are used, no values.

    std::vector<common_ngram_map_key> keys; // key n-grams which occur several times in token-history
-    uint16_t check_rate; // check for speculative decoding without draft model for each check_rate token
    uint16_t min_hits;   // minimum number of key hits to consider a draft

-    bool     show_key_map_stats = false; // true, if statitics of the key_map should be printed.
+    bool     show_key_map_stats = false; // true, if statistics of the key_map should be printed.

    common_ngram_map(uint16_t sz_key, uint16_t sz_value, bool only_keys,
-                     uint16_t check_rate, uint16_t min_hits)
+                     uint16_t min_hits)
        : size_key(sz_key), size_value(sz_value), key_only(only_keys),
-          check_rate(check_rate), min_hits(min_hits) {
+          min_hits(min_hits) {
        key_map.resize(COMMON_NGRAM_HASH_MAP_SIZE); // 2^18 hash entries, 0 entries if key_map shouldn't be used
    }

@@ -113,13 +113,14 @@ static bool common_speculative_are_compatible(
 struct common_speculative_state {
    const enum common_speculative_type type;

-    // TODO: rename to n_call_draft, n_gen_drafts, n_acc_drafts, n_gen_tokens, n_acc_tokens
-    // TODO: add n_call_begin, n_call_accept
-    size_t drafts_call_count       = 0; // number of times this implementation was called.
-    size_t drafts_generated_count  = 0; // number of times a draft or part was generated by this implementation.
-    size_t drafts_accepted_count   = 0; // number of times a draft or part was accepted by the target model.
-    size_t drafts_generated_tokens = 0; // number of tokens generated by this implementation.
-    size_t drafts_accepted_tokens  = 0; // number of tokens accepted by the target model.
+    size_t n_call_begin  = 0; // number of times this implementation was called for refresh.
+    size_t n_call_draft  = 0; // number of times this implementation was called for generation.
+    size_t n_call_accept = 0; // number of times this implementation was called for accumulation.
+
+    size_t n_gen_drafts = 0; // number of times a draft or part was generated by this implementation.
+    size_t n_acc_drafts = 0; // number of times a draft or part was accepted by the target model.
+    size_t n_gen_tokens = 0; // number of tokens generated by this implementation.
+    size_t n_acc_tokens = 0; // number of tokens accepted by the target model.

    // TODO: track performance of most recent calls
    const bool gen_perf = true; // whether to generate performance stats.
@@ -463,12 +464,12 @@ struct common_speculative_state_eagle3 : public common_speculative_state {

 // state of self-speculation (simple implementation, not ngram-map)
 struct common_speculative_state_ngram_simple : public common_speculative_state {
-    common_ngram_simple_state state;
+    common_ngram_simple_config config;

    common_speculative_state_ngram_simple(
            enum common_speculative_type type,
-            common_ngram_simple_state state)
-        : common_speculative_state(type), state(state) {}
+            common_ngram_simple_config config)
+        : common_speculative_state(type), config(config) {}

    void begin(const llama_tokens & prompt) override {
        GGML_UNUSED(prompt);
@@ -479,7 +480,8 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
            const llama_tokens & prompt_tgt,
            llama_token id_last,
            llama_tokens & result) override {
-        result = common_ngram_simple_draft(state, prompt_tgt, id_last);
+
+        result = common_ngram_simple_draft(config, prompt_tgt, id_last);
        GGML_UNUSED(params);
    }

@@ -744,10 +746,9 @@ static common_ngram_map get_common_ngram_map(const common_speculative_config & c
    uint16_t size_key   = config.params.ngram_size_n;
    uint16_t size_value = config.params.ngram_size_m;
    bool     key_only   = (config.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
-    uint16_t check_rate = config.params.ngram_check_rate;
    uint16_t min_hits   = config.params.ngram_min_hits;

-    return common_ngram_map(size_key, size_value, key_only, check_rate, min_hits);
+    return common_ngram_map(size_key, size_value, key_only, min_hits);
 }

 static common_speculative_state_ngram_cache create_state_ngram_cache(
@@ -797,6 +798,42 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
    return it->second;
 }

+bool common_speculative_is_compat(llama_context * ctx_tgt) {
+    auto * mem = llama_get_memory(ctx_tgt);
+    if (mem == nullptr) {
+        return false;
+    }
+
+    bool res = true;
+
+    llama_memory_clear(mem, true);
+
+    // eval 2 tokens to check if the context is compatible
+    std::vector<llama_token> tmp;
+    tmp.push_back(0);
+    tmp.push_back(0);
+
+    int ret = llama_decode(ctx_tgt, llama_batch_get_one(tmp.data(), tmp.size()));
+    if (ret != 0) {
+        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
+        res = false;
+        goto done;
+    }
+
+    // try to remove the last tokens
+    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
+        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
+        res = false;
+        goto done;
+    }
+
+done:
+    llama_memory_clear(mem, true);
+    llama_synchronize(ctx_tgt);
+
+    return res;
+}
+
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(
@@ -887,16 +924,14 @@ common_speculative * common_speculative_init(

                uint16_t ngram_size_key   = ngram_map.size_key;
                uint16_t mgram_size_value = ngram_map.size_value;
-                uint16_t check_rate       = ngram_map.check_rate;

-                auto config_simple = common_ngram_simple_config{
+                auto config_simple = common_ngram_simple_config {
                    /* .size_ngram      = */ ngram_size_key,
-                    /* .size_mgram      = */ mgram_size_value,
-                    /* .check_rate      = */ check_rate
+                    /* .size_mgram      = */ mgram_size_value
                };
                auto state = std::make_unique<common_speculative_state_ngram_simple>(
                    /* .type            = */ config.type,
-                    /* .state           = */ common_ngram_simple_state(config_simple)
+                    /* .state           = */ config_simple
                );
                impls.push_back(std::move(state));
                break;
@@ -951,12 +986,9 @@ void common_speculative_begin(common_speculative * spec, const llama_tokens & pr
    }

    for (auto & impl : spec->impls) {
-        const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
-
+        common_time_meas tm(impl->t_begin_us, !impl->gen_perf);
        impl->begin(prompt);
-
-        const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
-        impl->t_begin_us += t_now_us - t_start_us; // accumulate duration for this refresh
+        impl->n_call_begin++;
    }
 }

@@ -971,24 +1003,19 @@ llama_tokens common_speculative_draft(

    for (auto & impl : spec->impls) {
        {
-            const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
-
+            common_time_meas tm(impl->t_draft_us, !impl->gen_perf);
            impl->draft(params, prompt_tgt, id_last, result);
-
-            const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
-
-            impl->drafts_call_count++;
-            impl->t_draft_us += t_now_us - t_start_us; // accumulate duration for this implementation
+            impl->n_call_draft++;
        }

        if (!result.empty()) {
            LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__,
                    common_speculative_type_to_str(impl.get()->type).c_str(), prompt_tgt.size(),
-                    impl.get()->drafts_call_count, result.size());
+                    impl.get()->n_call_draft, result.size());

            spec->curr_impl = impl.get(); // set current implementation for stats
-            impl->drafts_generated_count++;
-            impl->drafts_generated_tokens += result.size();
+            impl->n_gen_drafts++;
+            impl->n_gen_tokens += result.size();

            break; // We have a draft, so break out of the loop and return it.
        }
@@ -1006,15 +1033,16 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) {

    GGML_ASSERT(impl);

-    const int64_t t_start_us = impl->gen_perf ? ggml_time_us() : 0;
-    if (n_accepted > 0) {
-        impl->drafts_accepted_count++;
-        impl->drafts_accepted_tokens += n_accepted;
-    }
+    {
+        common_time_meas tm(impl->t_accept_us, !impl->gen_perf);
+        if (n_accepted > 0) {
+            impl->n_acc_drafts++;
+            impl->n_acc_tokens += n_accepted;
+        }

-    impl->accept(n_accepted);
-    const int64_t t_now_us = impl->gen_perf ? ggml_time_us() : 0;
-    impl->t_accept_us += t_now_us - t_start_us; // accumulate duration for this acculumulation
+        impl->accept(n_accepted);
+        impl->n_call_accept++;
+    }
 }

 void common_speculative_print_stats(const common_speculative * spec) {
@@ -1034,13 +1062,13 @@ void common_speculative_print_stats(const common_speculative * spec) {
            str_perf = "";
        }

-        LOG_INF("statistics %s: #calls = %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
+        LOG_INF("statistics %s: #calls(b,g,a) = %zu %zu %zu, #gen drafts = %zu, #acc drafts = %zu, #gen tokens = %zu, #acc tokens = %zu%s\n",
                common_speculative_type_to_str(impl->type).c_str(),
-                impl->drafts_call_count,
-                impl->drafts_generated_count,
-                impl->drafts_accepted_count,
-                impl->drafts_generated_tokens,
-                impl->drafts_accepted_tokens,
+                impl->n_call_begin, impl->n_call_draft, impl->n_call_accept,
+                impl->n_gen_drafts,
+                impl->n_acc_drafts,
+                impl->n_gen_tokens,
+                impl->n_acc_tokens,
                str_perf.c_str());
    }
 }
@@ -14,6 +14,10 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
 // convert type to string
 std::string common_speculative_type_to_str(enum common_speculative_type type);

+// check if the llama_context is compatible for speculative decoding
+// note: clears the memory of the context
+bool common_speculative_is_compat(llama_context * ctx_tgt);
+
 common_speculative * common_speculative_init(
        common_params_speculative & params,
        llama_context             * ctx_tgt);
@@ -160,8 +160,6 @@ class ModelBase:
                self.ftype = gguf.LlamaFileType.MOSTLY_F16
                logger.info("heuristics unable to detect tensor dtype, defaulting to --outtype f16")

-        self.dequant_model()
-
        # Configure GGUF Writer
        self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file,
                                           split_max_tensors=split_max_tensors, split_max_size=split_max_size, dry_run=dry_run, small_first_shard=small_first_shard)
@@ -527,6 +525,8 @@ class ModelBase:
        return ()

    def prepare_tensors(self):
+        self.dequant_model()
+
        # Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
        if self.tensor_map.mapping:
            max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
@@ -586,6 +586,10 @@ class ModelBase:
                            gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
                            gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
                            gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
+                            # Kimi KDA conv weights should be F32
+                            gguf.MODEL_TENSOR.SSM_CONV1D_Q,
+                            gguf.MODEL_TENSOR.SSM_CONV1D_K,
+                            gguf.MODEL_TENSOR.SSM_CONV1D_V,
                        )
                    )
                    or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
@@ -903,10 +907,10 @@ class TextModel(ModelBase):
        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
            self.gguf_writer.add_layer_norm_eps(f_norm_eps)
            logger.info(f"gguf: layer norm epsilon = {f_norm_eps}")
-        if (n_experts := self.hparams.get("num_local_experts")) is not None:
+        if (n_experts := self.find_hparam(["num_local_experts", "num_experts"], optional=True)) is not None:
            self.gguf_writer.add_expert_count(n_experts)
            logger.info(f"gguf: expert count = {n_experts}")
-        if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
+        if (n_experts_used := self.find_hparam(["num_experts_per_tok", "num_experts_per_token"], optional=True)) is not None:
            self.gguf_writer.add_expert_used_count(n_experts_used)
            logger.info(f"gguf: experts used count = {n_experts_used}")
        if (n_expert_groups := self.hparams.get("n_group")) is not None:
@@ -916,7 +920,7 @@ class TextModel(ModelBase):
            self.gguf_writer.add_expert_group_used_count(n_group_used)
            logger.info(f"gguf: expert groups used count = {n_group_used}")

-        if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func"], optional=True)) is not None:
+        if (score_func := self.find_hparam(["score_function", "scoring_func", "score_func", "moe_router_activation", "moe_router_activation_func"], optional=True)) is not None:
            if score_func == "sigmoid":
                self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
            elif score_func == "softmax":
@@ -1257,6 +1261,9 @@ class TextModel(ModelBase):
        if chkhsh == "6c81ce329e0802883b22eabab0d3fa48357337ef1ecb45443828bf1f6254833f":
            # ref: https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B
            res = "exaone-moe"
+        if chkhsh == "d30d75d9059f1aa2c19359de71047b3ae408c70875e8a3ccf8c5fba56c9d8af4":
+            # ref: https://huggingface.co/Qwen/Qwen3.5-9B-Instruct
+            res = "qwen35"

        if res is None:
            logger.warning("\n")
@@ -1808,7 +1815,7 @@ class MmprojModel(ModelBase):
    preprocessor_config: dict[str, Any]
    global_config: dict[str, Any]

-    n_block_keys = ["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth", "encoder_layers"]
+    n_block_keys = ["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth", "encoder_layers", "vt_num_hidden_layers"]

    has_vision_encoder: bool = True # by default
    has_audio_encoder: bool = False
@@ -1863,7 +1870,15 @@ class MmprojModel(ModelBase):
        preprocessor_config_path = self.dir_model / "preprocessor_config.json"
        if preprocessor_config_path.is_file():
            with open(preprocessor_config_path, "r", encoding="utf-8") as f:
-                self.preprocessor_config = json.load(f)
+                cfg = json.load(f)
+                # move media_proc_cfg to root level for compat
+                if "media_proc_cfg" in cfg:
+                    cfg = {
+                        **cfg,
+                        **cfg["media_proc_cfg"],
+                    }
+                # merge configs
+                self.preprocessor_config = {**self.preprocessor_config, **cfg}

        # prefer processor_config.json if possible
        processor_config_path = self.dir_model / "processor_config.json"
@@ -1912,10 +1927,10 @@ class MmprojModel(ModelBase):
            self.image_size = self.find_vparam(["image_size"])
            self.gguf_writer.add_vision_image_size(self.image_size)
            self.gguf_writer.add_vision_patch_size(self.find_vparam(["patch_size"]))
-            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_embedding_length(self.find_vparam(["hidden_size", "vt_hidden_size"]))
+            self.gguf_writer.add_vision_feed_forward_length(self.find_vparam(["intermediate_size", "vt_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", "num_heads"]))
+            self.gguf_writer.add_vision_head_count(self.find_vparam(["num_attention_heads", "num_heads", "vt_num_attention_heads"]))

            # preprocessor config
            image_mean = _MISTRAL_COMMON_DATASET_MEAN if self.is_mistral_format else self.preprocessor_config["image_mean"]
@@ -4105,37 +4120,29 @@ class Qwen2MoeModel(TextModel):
        # 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()
-            yield from super().modify_tensors(permuted, mapped, bid)
+            # HF: [n_expert, n_embd, n_ff] -> GGML: {n_ff, n_embd, n_expert}
+            yield from super().modify_tensors(data_torch, mapped, bid)
            return

        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:
+            if data_torch.ndim < 3 or data_torch.shape[-2] % 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()
-            yield from super().modify_tensors(perm_gate, mapped_gate, bid)
-            yield from super().modify_tensors(perm_up, mapped_up, bid)
+            # HF: [n_expert, 2*n_ff, n_embd] -> split on dim=-2
+            n_ff = data_torch.shape[-2] // 2
+            gate = data_torch[..., :n_ff, :].contiguous()
+            up = data_torch[..., n_ff:, :].contiguous()
+            # gate/up: [n_expert, n_ff, n_embd] -> GGML: {n_embd, n_ff, n_expert}
+            base_name = name.removesuffix(".weight").removesuffix(".gate_up_proj")
+            mapped_gate = f"{base_name}.gate_proj.weight"
+            mapped_up = f"{base_name}.up_proj.weight"
+            yield from super().modify_tensors(gate, mapped_gate, bid)
+            yield from super().modify_tensors(up, mapped_up, bid)
            return

        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:
            n_experts = self.hparams["num_experts"]
            assert bid is not None
@@ -4291,6 +4298,7 @@ class Qwen3NextModel(Qwen2MoeModel):
        self.gguf_writer.add_ssm_group_count(self.hparams["linear_num_key_heads"])
        self.gguf_writer.add_ssm_time_step_rank(self.hparams["linear_num_value_heads"])
        self.gguf_writer.add_ssm_inner_size(self.hparams["linear_value_head_dim"] * self.hparams["linear_num_value_heads"])
+        self.gguf_writer.add_full_attention_interval(self.hparams.get("full_attention_interval", 4))
        if (rope_dim := self.hparams.get("head_dim")) is None:
            rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.25)))
@@ -4355,7 +4363,7 @@ class RND1Model(Qwen2MoeModel):
            self.gguf_writer.add_mask_token_id(mask_token_id)


-@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration")
+@ModelBase.register("Qwen3VLForConditionalGeneration", "Qwen3VLMoeForConditionalGeneration", "Qwen3_5ForConditionalGeneration", "Qwen3_5MoeForConditionalGeneration")
 class Qwen3VLVisionModel(MmprojModel):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
@@ -4401,6 +4409,10 @@ class Qwen3VLVisionModel(MmprojModel):
        if name.startswith("model.language_model.") or name.startswith("lm_head."):
            return

+        # Skip MTP tensors
+        if name.startswith("mtp."):
+            return
+
        if name.startswith("model.visual."):
            name = name.replace("model.visual.", "visual.", 1)

@@ -4531,9 +4543,125 @@ class Qwen3VLMoeTextModel(Qwen3MoeModel):
        if name.startswith("model.visual."):
            return

+        # Qwen3VL has transposed packed tensors, so we treat it differently from general Qwen2MoE packed tensors
+        if name.endswith("mlp.experts.down_proj") or name.endswith("mlp.experts.down_proj.weight"):
+            name = name.replace("language_model.", "")
+            mapped = f"{name}.weight" if not name.endswith(".weight") else name
+            permuted = data_torch.permute(0, 2, 1).contiguous()
+            yield from ModelBase.modify_tensors(self, permuted, mapped, bid)
+            return
+
+        if name.endswith("mlp.experts.gate_up_proj") or name.endswith("mlp.experts.gate_up_proj.weight"):
+            name = name.replace("language_model.", "")
+            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()
+            yield from ModelBase.modify_tensors(self, perm_gate, mapped_gate, bid)
+            yield from ModelBase.modify_tensors(self, perm_up, mapped_up, bid)
+            return
+
        yield from super().modify_tensors(data_torch, name, bid)


+class _LinearAttentionVReorderBase(Qwen3NextModel):
+    model_arch = gguf.MODEL_ARCH.QWEN3NEXT  # overridden by subclasses
+    """reorders V heads from grouped to tiled order for ggml broadcast
+
+    see https://github.com/ggml-org/llama.cpp/pull/19468#discussion_r2786394306
+
+    Linear attention may has num_k_heads < num_v_heads. The HF weights store
+    V heads grouped by K head: [G0_v0..v{r-1}, G1_v0..v{r-1}, ...].
+    ggml binary ops use tiled broadcast: [K0, K1, ..., K0, K1, ...].
+    We reorder V heads to tiled order so ggml_repeat can replace the expensive
+    interleaved repeat: [G0_v0, G1_v0, ..., G0_v1, G1_v1, ...].
+    """
+
+    @staticmethod
+    def _reorder_v_heads(tensor: Tensor, dim: int, num_k_heads: int, num_v_per_k: int, head_dim: int) -> Tensor:
+        """Reorder V heads from grouped (by K head) to tiled order along the given dimension."""
+        shape = list(tensor.shape)
+        if dim < 0:
+            dim += len(shape)
+        new_shape = shape[:dim] + [num_k_heads, num_v_per_k, head_dim] + shape[dim + 1:]
+        tensor = tensor.reshape(*new_shape)
+        perm = list(range(len(new_shape)))
+        perm[dim], perm[dim + 1] = perm[dim + 1], perm[dim]
+        return tensor.permute(*perm).contiguous().reshape(*shape)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        num_k_heads = self.hparams.get("linear_num_key_heads", 0)
+        num_v_heads = self.hparams.get("linear_num_value_heads", 0)
+
+        if num_k_heads > 0 and num_v_heads > 0 and num_k_heads != num_v_heads and "linear_attn." in name:
+            head_k_dim = self.hparams["linear_key_head_dim"]
+            head_v_dim = self.hparams["linear_value_head_dim"]
+            num_v_per_k = num_v_heads // num_k_heads
+
+            if ".in_proj_qkv." in name:
+                # QKV weight: reorder only the V rows
+                q_dim = head_k_dim * num_k_heads
+                k_dim = head_k_dim * num_k_heads
+                q = data_torch[:q_dim]
+                k = data_torch[q_dim:q_dim + k_dim]
+                v = data_torch[q_dim + k_dim:]
+                v = self._reorder_v_heads(v, 0, num_k_heads, num_v_per_k, head_v_dim)
+                data_torch = torch.cat([q, k, v], dim=0)
+
+            elif ".in_proj_z." in name:
+                # Z gate weight: reorder rows (num_v_heads * head_v_dim)
+                data_torch = self._reorder_v_heads(data_torch, 0, num_k_heads, num_v_per_k, head_v_dim)
+
+            elif ".in_proj_b." in name or ".in_proj_a." in name:
+                # Beta/Alpha weight: reorder rows (num_v_heads, head_dim=1)
+                data_torch = self._reorder_v_heads(data_torch, 0, num_k_heads, num_v_per_k, 1)
+
+            elif ".A_log" in name or ".dt_bias" in name or ".dt_proj" in name:
+                # A_log / dt_bias: 1D parameters with num_v_heads elements
+                if data_torch.ndim == 1:
+                    data_torch = self._reorder_v_heads(
+                        data_torch.unsqueeze(-1), 0, num_k_heads, num_v_per_k, 1
+                    ).squeeze(-1)
+                else:
+                    data_torch = self._reorder_v_heads(data_torch, -1, num_k_heads, num_v_per_k, 1)
+
+            elif ".conv1d" in name:
+                # Conv1d kernel: reorder only the V channel portion
+                data = data_torch.squeeze()
+                qk_channels = head_k_dim * num_k_heads * 2
+                qk_part = data[:qk_channels]
+                v_part = data[qk_channels:]
+                v_part = self._reorder_v_heads(v_part, 0, num_k_heads, num_v_per_k, head_v_dim)
+                data_torch = torch.cat([qk_part, v_part], dim=0)
+
+            elif ".out_proj." in name:
+                # Out projection weight: reorder columns (input dimension)
+                data_torch = self._reorder_v_heads(data_torch, 1, num_k_heads, num_v_per_k, head_v_dim)
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Qwen3_5ForConditionalGeneration")
+class Qwen3_5TextModel(_LinearAttentionVReorderBase):
+    model_arch = gguf.MODEL_ARCH.QWEN35
+
+
+@ModelBase.register("Qwen3_5MoeForConditionalGeneration")
+class Qwen3_5MoeTextModel(_LinearAttentionVReorderBase):
+    model_arch = gguf.MODEL_ARCH.QWEN35MOE
+
+
@ModelBase.register("GPT2LMHeadModel")
 class GPT2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.GPT2
@@ -5013,6 +5141,221 @@ class CodeShellModel(TextModel):
        self.gguf_writer.add_rope_scaling_factor(1.0)


+@ModelBase.register("KimiLinearModel", "KimiLinearForCausalLM")
+class KimiLinearModel(TextModel):
+    """Kimi-Linear model with hybrid MLA+KDA architecture"""
+    model_arch = gguf.MODEL_ARCH.KIMI_LINEAR
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def set_vocab(self):
+        try:
+            self._set_vocab_gpt2()
+            return
+        except Exception:
+            pass
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        if tokpre == "kimi-k2":
+            # Build merges list using the approach similar to HunYuanMoE
+            merges = []
+            vocab = {}
+            mergeable_ranks = tokenizer.model._mergeable_ranks
+            for token, rank in mergeable_ranks.items():
+                vocab[QwenModel.token_bytes_to_string(token)] = rank
+                if len(token) == 1:
+                    continue
+                merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
+                if len(merged) == 2:
+                    merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
+            # Build token list
+            vocab_size = self.hparams["vocab_size"]
+            special_tokens = tokenizer.special_tokens
+            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
+            tokens: list[str] = []
+            toktypes: list[int] = []
+
+            for i in range(vocab_size):
+                if i not in reverse_vocab:
+                    tokens.append(f"[PAD{i}]")
+                    toktypes.append(gguf.TokenType.UNUSED)
+                else:
+                    token = reverse_vocab[i]
+                    tokens.append(token)
+                    if i in special_tokens.values():
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.NORMAL)
+
+            self.gguf_writer.add_tokenizer_model("gpt2")
+            self.gguf_writer.add_tokenizer_pre(tokpre)
+            self.gguf_writer.add_token_list(tokens)
+            self.gguf_writer.add_token_types(toktypes)
+            self.gguf_writer.add_token_merges(merges)
+
+            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+            special_vocab.add_to_gguf(self.gguf_writer)
+            # override eos id in config.json with tiktoken eos id
+            self.gguf_writer.add_eos_token_id(tokenizer.eos_id)
+        else:
+            raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")
+
+    def set_gguf_parameters(self):
+        # note: To enable MLA KV cache, attention needs to be converted into MQA (ie: GQA with 1 group)
+        self.hparams["num_key_value_heads"] = 1
+
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+
+        # KDA & MLA params
+        # Get ssm_d_conv from linear_attn_config.short_conv_kernel_size or ssm_d_conv
+        linear_attn_config = self.hparams["linear_attn_config"]
+        # n_head == 0 for KDA layers, n_head > 0 for MLA layers
+        # full_attention_layers list will be used to distingush layer type
+        _num_kv_heads = list()
+        _full_attn_layers = linear_attn_config["full_attn_layers"]
+        for il in range(self.hparams["num_hidden_layers"]):
+            if il + 1 in _full_attn_layers:
+                _num_kv_heads.append(self.hparams["num_key_value_heads"])
+            else:
+                _num_kv_heads.append(0)
+        assert len(_num_kv_heads) == self.hparams["num_hidden_layers"]
+        self.gguf_writer.add_head_count_kv(_num_kv_heads)
+
+        if (ssm_d_conv := linear_attn_config.get("short_conv_kernel_size")) is not None:
+            self.gguf_writer.add_ssm_conv_kernel(ssm_d_conv)
+        if (kda_head_dim := linear_attn_config.get("head_dim")) is not None:
+            self.gguf_writer.add_kda_head_dim(kda_head_dim)
+
+        # MLA params - use add_* methods that handle arch substitution
+        # Support both HuggingFace naming (q_lora_rank, kv_lora_rank) and internal naming (n_lora_q, n_lora_kv)
+        if (q_lora_rank := self.find_hparam(["q_lora_rank", "n_lora_q"], optional=True)) is not None:
+            self.gguf_writer.add_q_lora_rank(q_lora_rank)
+        # To enable MLA KV cache, MLA needs to be converted into MQA with larger heads, then decompresses to MHA
+        kv_lora_rank = self.find_hparam(["kv_lora_rank", "n_lora_kv"], optional=False)
+        self.gguf_writer.add_kv_lora_rank(kv_lora_rank)
+
+        # MLA head dimensions
+        # Support HuggingFace naming: qk_nope_head_dim, qk_rope_head_dim, v_head_dim
+        qk_nope_head_dim = self.hparams.get("qk_nope_head_dim")
+        # Rotation - use qk_rope_head_dim for Kimi
+        qk_rope_head_dim = self.find_hparam(["qk_rope_head_dim", "n_rot"], optional=False)
+        self.gguf_writer.add_rope_dimension_count(qk_rope_head_dim)
+        self.gguf_writer.add_key_length(kv_lora_rank + qk_rope_head_dim)
+        v_head_dim = self.hparams.get("v_head_dim")
+
+        # Calculate n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim
+        if (n_embd_head_k_mla := self.find_hparam(["n_embd_head_k_mla"], optional=True)) is not None:
+            self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)
+        elif qk_nope_head_dim is not None:
+            n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim
+            self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)
+
+        # n_embd_head_v_mla = v_head_dim
+        if (n_embd_head_v_mla := self.hparams.get("n_embd_head_v_mla")) is not None:
+            self.gguf_writer.add_value_length_mla(n_embd_head_v_mla)
+        elif v_head_dim is not None:
+            self.gguf_writer.add_value_length_mla(v_head_dim)
+
+        # moe_intermediate_size (1024 for Kimi)
+        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
+        # num_shared_experts (1 for Kimi)
+        self.gguf_writer.add_expert_shared_count(self.hparams["num_shared_experts"])
+        # first_k_dense_replace (1 for Kimi - first layer uses dense MLP)
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
+        # Routed scaling factor (expert_weights_scale = 2.446 for Kimi)
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        if self._experts is not None:
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        logger.info(f"Processing {name}: shape before = {tuple(data_torch.shape)}")
+
+        # Handle KDA conv1d weights
+        # HuggingFace/vLLM stores as [d_inner, d_conv] (2D), memory layout: conv_step changes fastest
+        # llama.cpp expects ggml ne = [d_conv, 1, d_inner, 1], memory layout: ne[0]=d_conv changes fastest
+        # GGUF reverses numpy shape when writing, so numpy (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]
+        # Memory layouts match: both have conv_step (d_conv) changing fastest
+        if name.endswith((".q_conv1d.weight", ".k_conv1d.weight", ".v_conv1d.weight")):
+            # HF shape: [d_inner, d_conv] e.g. [4096, 4]
+            # Target numpy shape: (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]
+            if data_torch.ndim == 2:
+                d_inner, d_conv = data_torch.shape
+                # Reshape to (1, d_inner, 1, d_conv) - memory layout preserved (d_conv fastest)
+                data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
+                logger.info(f"Reshaped conv1d weight {name}: [d_inner={d_inner}, d_conv={d_conv}] -> numpy {tuple(data_torch.shape)} -> ggml ne=[{d_conv}, 1, {d_inner}, 1]")
+            elif data_torch.ndim == 3:
+                # Already 3D [d_inner, 1, d_conv] from unsqueeze
+                d_inner, _, d_conv = data_torch.shape
+                data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
+                logger.info(f"Reshaped conv1d weight {name}: [d_inner={d_inner}, 1, d_conv={d_conv}] -> numpy {tuple(data_torch.shape)} -> ggml ne=[{d_conv}, 1, {d_inner}, 1]")
+
+        # Kimi specific bias
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        # Handle A_log: iHF stores as [1, 1, num_heads, 1]
+        # llama.cpp expects ggml ne = [1, num_heads, 1, 1]
+        # GGUF reverses numpy shape: numpy (1, 1, num_heads, 1) -> ggml ne = [1, num_heads, 1, 1]
+        if name.endswith(".A_log"):
+            data_torch = -torch.exp(data_torch)
+        if name.endswith(".dt_bias"):
+            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
+            logger.info("Changed dt_bias to dt_proj.bias")
+
+        # process the experts separately
+        if name.find("block_sparse_moe.experts") != -1:
+            n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=False)
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                # merge the experts into a single 3d tensor
+                # w1: gate, w2: down, w3: up
+                for wid, tname in [("w1", gguf.MODEL_TENSOR.FFN_GATE_EXP),
+                                   ("w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP),
+                                   ("w3", gguf.MODEL_TENSOR.FFN_UP_EXP)]:
+                    datas: list[Tensor] = []
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+                    data_torch = torch.stack(datas, dim=0)
+                    new_name = self.format_tensor_name(tname, bid)
+                    yield from super().modify_tensors(data_torch, new_name, bid)
+            return
+
+        # note: MLA with the absorption optimization, needs these two split and k_b_proj transposed
+        if name.endswith("kv_b_proj.weight"):
+            name_kb = name.replace("kv_b_proj", "k_b_proj")
+            name_vb = name.replace("kv_b_proj", "v_b_proj")
+            n_head_kv = self.hparams["num_key_value_heads"]
+            v_head_dim = self.find_hparam(["n_embd_head_v_mla", "v_head_dim"], optional=False)
+            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+            logger.info("Split kv_b n_head_kv %d\n" % n_head_kv)
+            assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)
+            kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
+            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
+            k_b = k_b.transpose(1, 2)
+            yield from super().modify_tensors(k_b, name_kb, bid)
+            yield from super().modify_tensors(v_b, name_vb, bid)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
@ModelBase.register("InternLM2ForCausalLM")
 class InternLM2Model(TextModel):
    model_arch = gguf.MODEL_ARCH.INTERNLM2
@@ -7360,6 +7703,7 @@ class DeepseekModel(TextModel):
    "DeepseekV2ForCausalLM",
    "DeepseekV3ForCausalLM",
    "KimiVLForConditionalGeneration",
+    "KimiK25ForConditionalGeneration",
    "YoutuForCausalLM",
    "YoutuVLForConditionalGeneration",
 )
@@ -7478,8 +7822,8 @@ class DeepseekV2Model(TextModel):
    _experts: list[dict[str, Tensor]] | None = None

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # skip vision tensors and remove "language_model." for Kimi-VL
-        if "vision_tower" in name or "multi_modal_projector" in name:
+        # skip vision tensors and remove "language_model." for Kimi-VL and Kimi-K2.5
+        if "vision_tower" in name or "multi_modal_projector" in name or "mm_projector" in name:
            return
        if name.startswith("siglip2.") or name.startswith("merger."):
            return
@@ -7693,6 +8037,135 @@ class MimoV2Model(TextModel):
                raise ValueError(f"Unprocessed experts: {experts}")


+@ModelBase.register("Step3p5ForCausalLM")
+class Step35Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.STEP35
+
+    def set_gguf_parameters(self):
+        rope_theta = self.hparams.get("rope_theta")
+        if isinstance(rope_theta, list):
+            self.hparams["rope_theta"] = float(rope_theta[0])
+            self.hparams["local_rope_theta"] = float(rope_theta[1])
+            self.rope_parameters["rope_theta"] = self.hparams["rope_theta"]
+            self.rope_parameters["sliding_attention"] = {"rope_theta": self.hparams["local_rope_theta"]}
+
+        super().set_gguf_parameters()
+
+        layer_types = self.hparams.get("layer_types") or []
+        partial_rotary_factors = self.hparams.get("partial_rotary_factors") or []
+        attn_other = self.hparams.get("attention_other_setting") or {}
+
+        n_head_base = self.hparams["num_attention_heads"]
+        n_kv_base = self.hparams["num_attention_groups"]
+
+        n_head_swa = attn_other.get("num_attention_heads", n_head_base)
+        n_kv_swa = attn_other.get("num_attention_groups", n_kv_base)
+
+        layer_types = layer_types[: self.block_count]
+        partial_rotary_factors = partial_rotary_factors[: self.block_count]
+        assert [1.0 if lt == "sliding_attention" else 0.5 for lt in layer_types] == partial_rotary_factors
+        head_arr = [n_head_swa if lt == "sliding_attention" else n_head_base for lt in layer_types]
+        kv_arr = [n_kv_swa if lt == "sliding_attention" else n_kv_base for lt in layer_types]
+        swa_pat = [lt == "sliding_attention" for lt in layer_types]
+
+        self.gguf_writer.add_head_count(head_arr)
+        self.gguf_writer.add_head_count_kv(kv_arr)
+
+        self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
+        self.gguf_writer.add_sliding_window_pattern(swa_pat)
+
+        self.gguf_writer.add_value_length(self.hparams["head_dim"])
+
+        # MoE params
+        self.gguf_writer.add_expert_count(self.hparams["moe_num_experts"])
+        self.gguf_writer.add_expert_used_count(self.hparams["moe_top_k"])
+        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
+        self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["share_expert_dim"])
+
+        if (moe_router_scaling_factor := self.hparams.get("moe_router_scaling_factor")) is not None:
+            self.gguf_writer.add_expert_weights_scale(moe_router_scaling_factor)
+        if (norm_expert_weight := self.hparams.get("norm_expert_weight")) is not None:
+            self.gguf_writer.add_expert_weights_norm(norm_expert_weight)
+
+        # leading dense blocks
+        leading_dense = 0
+        moe_layers_enum = self.hparams.get("moe_layers_enum")
+        if isinstance(moe_layers_enum, str) and moe_layers_enum.strip():
+            moe_layers = sorted(int(i) for i in moe_layers_enum.strip().split(","))
+            if moe_layers:
+                leading_dense = max(0, moe_layers[0])
+        self.gguf_writer.add_leading_dense_block_count(leading_dense)
+        self.gguf_writer.add_moe_every_n_layers(int(self.hparams.get("moe_every_n_layer", 1)))
+
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-5))
+
+        # Optional per-layer SwiGLU clamps.
+        if (limits := self.hparams.get("swiglu_limits")) is not None:
+            limits_f = [0.0 if v is None else float(v) for v in limits[: self.block_count]]
+            self.gguf_writer.add_swiglu_clamp_exp(limits_f)
+        if (limits_shared := self.hparams.get("swiglu_limits_shared")) is not None:
+            limits_shared_f = [0.0 if v is None else float(v) for v in limits_shared[: self.block_count]]
+            self.gguf_writer.add_swiglu_clamp_shexp(limits_shared_f)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        # remove mtp layers
+        if (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None:
+            il = int(m.group(1))
+            n_main = int(self.hparams.get("num_hidden_layers", self.block_count))
+            if il >= n_main:
+                return
+        if name.endswith("norm.weight"):
+            data_torch += 1.0
+        # Map router bias (expert selection bias) to a GGUF bias tensor
+        if name.endswith(".moe.router_bias"):
+            name += ".bias"
+
+        if name.endswith((".self_attn.g_proj.weight", ".moe.gate.weight", ".moe.up_proj.weight", ".moe.gate_proj.weight", ".moe.down_proj.weight")):
+            data_torch = data_torch.squeeze().contiguous()
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # Step35 can optionally use Llama-3 style RoPE scaling (HF: rope_scaling.rope_type == "llama3").
+        # llama.cpp represents this via a single extra tensor: "rope_freqs.weight" (aka MODEL_TENSOR.ROPE_FREQS).
+        rope_params = self.rope_parameters.get("full_attention", self.rope_parameters)
+        rope_type = rope_params.get("rope_type") or ""
+        if rope_type.lower() != "llama3":
+            return
+
+        # Step35 configs can carry per-layer rope_theta as a list; for llama3 rope factors we use the base value.
+        rope_theta = self.hparams.get("rope_theta", 10000.0)
+        if isinstance(rope_theta, list):
+            rope_theta = rope_theta[0]
+        base = float(rope_theta)
+        if (dim := self.hparams.get("head_dim")) is None:
+            dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+        dim = int(dim)
+
+        freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+
+        factor = float(rope_params.get("factor", 8.0))
+        low_freq_factor = float(rope_params.get("low_freq_factor", 1.0))
+        high_freq_factor = float(rope_params.get("high_freq_factor", 4.0))
+        old_context_len = int(rope_params.get("original_max_position_embeddings", self.hparams.get("original_max_position_embeddings", 8192)))
+
+        low_freq_wavelen = old_context_len / low_freq_factor
+        high_freq_wavelen = old_context_len / high_freq_factor
+
+        rope_factors: list[float] = []
+        for freq in freqs:
+            wavelen = 2 * math.pi / float(freq)
+            if wavelen < high_freq_wavelen:
+                rope_factors.append(1.0)
+            elif wavelen > low_freq_wavelen:
+                rope_factors.append(factor)
+            else:
+                smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
+                rope_factors.append(1.0 / ((1.0 - smooth) / factor + smooth))
+
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
+
+
@ModelBase.register("PanguEmbeddedForCausalLM")
 class PanguEmbeddedModel(TextModel):
    model_arch = gguf.MODEL_ARCH.PANGU_EMBED
@@ -10712,6 +11185,103 @@ class KimiVLModel(MmprojModel):
                yield from super().modify_tensors(data_torch, name, bid)


+@ModelBase.register("KimiK25ForConditionalGeneration")
+class KimiK25Model(MmprojModel):
+    """Kimi-K2.5 with MoonViT3d vision encoder"""
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        assert self.hparams_vision is not None, "Kimi-K2.5 requires vision_config in model config"
+
+        self.merge_kernel_size = tuple(self.hparams_vision.get("merge_kernel_size", [2, 2]))
+        self.patch_size = self.hparams_vision.get("patch_size", 14)
+
+        # Set image_size for compatibility with base class
+        # Use position embedding dimensions as image_size reference
+        pos_emb_h = self.hparams_vision.get("init_pos_emb_height", 64)
+        self.hparams_vision["image_size"] = pos_emb_h * self.patch_size
+
+    def set_gguf_parameters(self):
+        # Base class MmprojModel.set_gguf_parameters() already writes:
+        # - vision_block_count, vision_head_count, vision_embedding_length
+        # - vision_feed_forward_length, vision_patch_size, image_mean, image_std
+        # via find_vparam() which handles the vt_* prefixed keys in Kimi-K2.5's config
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.KIMIK25)
+
+        # Position embedding parameters (for interpolation)
+        self.gguf_writer.add_uint32("vision.pos_emb_height", self.hparams_vision.get("init_pos_emb_height", 64))
+        self.gguf_writer.add_uint32("vision.pos_emb_width", self.hparams_vision.get("init_pos_emb_width", 64))
+        self.gguf_writer.add_uint32("vision.pos_emb_time", self.hparams_vision.get("init_pos_emb_time", 4))
+
+        # Projector parameters
+        self.gguf_writer.add_vision_use_gelu(self.hparams_vision.get("projector_hidden_act", "gelu") == "gelu")
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("projector_ln_eps", 1e-5))
+        self.gguf_writer.add_vision_projector_scale_factor(self.merge_kernel_size[0])
+
+        # Image size limits
+        # Note: in_patch_limit is for images, in_patch_limit_each_frame is for video (not supported yet)
+        in_patch_limit = self.preprocessor_config.get("in_patch_limit", 16384)
+        min_patches = 8  # reasonable minimum
+        pixels_per_patch = self.patch_size ** 2
+        self.gguf_writer.add_vision_min_pixels(min_patches * pixels_per_patch)
+        self.gguf_writer.add_vision_max_pixels(in_patch_limit * pixels_per_patch)
+
+    @staticmethod
+    def permute(weights: Tensor, n_head: int) -> Tensor:
+        out_dim, in_dim = weights.shape
+        head_dim = out_dim // n_head
+        w = weights.reshape(n_head, head_dim // 4, 2, 2, in_dim)
+        w = w.permute(0, 2, 1, 3, 4)
+        return w.reshape(out_dim, in_dim)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Only process vision and projector tensors
+        is_vision = any(x in name for x in ["vision_tower", "mm_projector"])
+
+        if not is_vision:
+            return
+
+        assert self.hparams_vision is not None
+        n_head = self.hparams_vision.get("num_attention_heads", 16)
+
+        # Permute Q/K weights/biases from interleaved to split RoPE format
+        # This allows using build_rope_2d at runtime without post-permutation.
+        if "wqkv" in name:
+            out_dim = data_torch.shape[0]
+            qkv_dim = out_dim // 3
+            head_dim = qkv_dim // n_head
+
+            if "weight" in name:
+                wq, wk, wv = data_torch[:qkv_dim, :], data_torch[qkv_dim:2 * qkv_dim, :], data_torch[2 * qkv_dim:, :]
+                wq = self.permute(wq, n_head)
+                wk = self.permute(wk, n_head)
+                data_torch = torch.cat([wq, wk, wv], dim=0)
+            elif "bias" in name:
+                bq, bk, bv = data_torch[:qkv_dim], data_torch[qkv_dim:2 * qkv_dim], data_torch[2 * qkv_dim:]
+                bq = bq.reshape(n_head, head_dim // 4, 2, 2).permute(0, 2, 1, 3).reshape(-1)
+                bk = bk.reshape(n_head, head_dim // 4, 2, 2).permute(0, 2, 1, 3).reshape(-1)
+                data_torch = torch.cat([bq, bk, bv], dim=0)
+
+        # Temporal embeddings: (T, 1, C) → (T, C)
+        if "pos_emb.time_weight" in name:
+            T, _, C = data_torch.shape
+            data_torch = data_torch.reshape(T, C)
+
+        # PatchMergerMLP tensor name mapping
+        # proj.0.weight → proj.linear_1.weight
+        # proj.2.weight → proj.linear_2.weight
+        if "mm_projector.proj.0." in name:
+            name = name.replace(".proj.0.", ".proj.linear_1.")
+        elif "mm_projector.proj.2." in name:
+            name = name.replace(".proj.2.", ".proj.linear_2.")
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
@ModelBase.register("CogVLMForCausalLM")
 class CogVLMVisionModel(MmprojModel):

@@ -148,6 +148,7 @@ models = [
    {"name": "youtu",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Youtu-LLM-2B", },
    {"name": "solar-open",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/upstage/Solar-Open-100B", },
    {"name": "exaone-moe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B", },
+    {"name": "qwen35",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3.5-9B-Instruct", }
 ]

 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -22,12 +22,11 @@
 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
 - **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*.
 - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over Intel iGPUs and dGPUs.
- **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.

 ### Llama.cpp + SYCL

 The llama.cpp SYCL backend is primarily designed for **Intel GPUs**.
-SYCL cross-platform capabilities enable support for Nvidia GPUs as well, with limited support for AMD.
+SYCL cross-platform capabilities enable support for other vendor GPUs as well.

 ## Recommended Release

@@ -42,6 +41,9 @@ The following releases are verified and recommended:

 ## News

+- 2026.02
+  - Remove support for Nvidia & AMD GPU, because the oneAPI plugin for Nvidia & AMD GPU is unavailable: download/installation channels are out of work. User can't build up the software for Nvidia & AMD GPU.
+
 - 2025.11
  - Support malloc memory on device more than 4GB.

@@ -111,8 +113,8 @@ On older Intel GPUs, you may try [OpenCL](/docs/backend/OPENCL.md) although the
 |-------------------------------|---------|---------------------------------------|
 | Intel Data Center Max Series  | Support | Max 1550, 1100                        |
 | Intel Data Center Flex Series | Support | Flex 170                              |
-| Intel Arc A-Series              | Support | Arc A770, Arc A730M, Arc A750         |
-| Intel Arc B-Series              | Support | Arc B580         |
+| Intel Arc A-Series            | Support | Arc A770, Arc A730M, Arc A750         |
+| Intel Arc B-Series            | Support | Arc B580                              |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake, Arrow Lake, Lunar Lake |
 | Intel iGPU                    | Support | iGPU in 13700k, 13400, i5-1250P, i7-1260P, i7-1165G7  |

@@ -127,20 +129,7 @@ On older Intel GPUs, you may try [OpenCL](/docs/backend/OPENCL.md) although the

 ### Other Vendor GPU

-**Verified devices**
-
-| Nvidia GPU               | Status    | Verified Model |
-|--------------------------|-----------|----------------|
-| Ampere Series            | Supported | A100, A4000    |
-| Ampere Series *(Mobile)* | Supported | RTX 40 Series  |
-
-| AMD GPU                  | Status       | Verified Model |
-|--------------------------|--------------|----------------|
-| Radeon Pro               | Experimental | W6800          |
-| Radeon RX                | Experimental | 6700 XT        |
-
-Note: AMD GPU support is highly experimental and is incompatible with F16.
-Additionally, it only supports GPUs with a sub_group_size (warp size) of 32.
+NA

 ## Docker

@@ -149,11 +138,11 @@ The docker build option is currently limited to *Intel GPU* targets.
 ### Build image

 ```sh
-# Using FP16
-docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
-
 # Using FP32
 docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=OFF" --target light -f .devops/intel.Dockerfile .
+
+# Using FP16
+docker build -t llama-cpp-sycl --build-arg="GGML_SYCL_F16=ON" --target light -f .devops/intel.Dockerfile .
 ```

 *Notes*:
@@ -212,14 +201,6 @@ Platform #0: Intel(R) OpenCL HD Graphics
 `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
 ```

- **Nvidia GPU**
-
-In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed.
-
- **AMD GPU**
-
-To target AMD GPUs with SYCL, the ROCm stack must be installed first.
-
 2. **Install Intel® oneAPI Base toolkit**

 SYCL backend depends on:
@@ -248,23 +229,6 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
 |2025.1|
 |2024.1|

- **Adding support to Nvidia GPUs**
-
-**oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
-
-**oneDNN**: The current oneDNN releases *(shipped with the oneAPI base-toolkit)* do not include the NVIDIA backend. Therefore, oneDNN must be compiled from source to enable the NVIDIA target:
-
-```sh
-git clone https://github.com/oneapi-src/oneDNN.git
-cd oneDNN
-cmake -GNinja -Bbuild-nvidia -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP -DDNNL_GPU_VENDOR=NVIDIA -DONEDNN_BUILD_GRAPH=OFF -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-cmake --build build-nvidia --config Release
-```
-
- **Adding support to AMD GPUs**
-
-**oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
-
 3. **Verify installation and environment**

 In order to check the available SYCL devices on the machine, please use the `sycl-ls` command.
@@ -285,25 +249,6 @@ When targeting an intel GPU, the user should expect one or more devices among th
 [opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) UHD Graphics 730 OpenCL 3.0 NEO  [24.39.31294]
 ```

- **Nvidia GPU**
-
-Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`cuda:gpu`] as below:
-
-```
-[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.12.0.12_195853.xmain-hotfix]
-[opencl:cpu][opencl:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
-[cuda:gpu][cuda:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.5]
-```
-
- **AMD GPU**
-
-For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]:
-
-```
-[opencl:cpu][opencl:0] Intel(R) OpenCL, 12th Gen Intel(R) Core(TM) i9-12900K OpenCL 3.0 (Build 0) [2024.18.6.0.02_160000]
-[hip:gpu][hip:0] AMD HIP BACKEND, AMD Radeon PRO W6800 gfx1030 [HIP 60140.9]
-```
-
 ### II. Build llama.cpp

 #### Intel GPU
@@ -332,47 +277,6 @@ It is possible to come across some precision issues when running tests that stem
 instructions, which can be circumvented by setting the environment variable `SYCL_PROGRAM_COMPILE_OPTIONS`
 as `-cl-fp32-correctly-rounded-divide-sqrt`

-#### Nvidia GPU
-
-The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
-By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
-
-```sh
-# Build LLAMA with Nvidia BLAS acceleration through SYCL
-# Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance
-GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture
-
-# Option 1: Use FP32 (recommended for better performance in most cases)
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
-
-# Option 2: Use FP16
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl
-
-# build all binary
-cmake --build build --config Release -j -v
-```
-
-It is possible to come across some precision issues when running tests that stem from using faster
-instructions, which can be circumvented by passing the `-fno-fast-math` flag to the compiler.
-
-#### AMD GPU
-
-The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices.
-By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`.
-
-```sh
-# Build LLAMA with rocBLAS acceleration through SYCL
-
-## AMD
-# Use FP32, FP16 is not supported
-# Find your GGML_SYCL_DEVICE_ARCH with rocminfo, under the key 'Name:'
-GGML_SYCL_DEVICE_ARCH=gfx90a # Example architecture
-cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-# build all binary
-cmake --build build --config Release -j -v
-```
-
 ### III. Run the inference

 #### Retrieve and prepare model
@@ -766,15 +670,15 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | Name               | Value                                 | Function                                    |
 |--------------------|---------------------------------------|---------------------------------------------|
 | GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.           |
-| GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
-| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)             | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
+| GGML_SYCL_TARGET   | INTEL *(default)*                     | Set the SYCL target device type.            |
+| GGML_SYCL_DEVICE_ARCH | Optional                           | Set the SYCL device architecture. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path. (1.) |
-| GGML_SYCL_GRAPH    | ON *(default)* \|OFF *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
+| GGML_SYCL_GRAPH    | OFF *(default)* \|ON *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |

-1. FP16 is recommended for better prompt processing performance on quantized models. Performance is equivalent in text generation but set `GGML_SYCL_F16=OFF` if you are experiencing issues with FP16 builds.
+1. FP32 or FP16 have different performance impact to LLM. Recommended to test them for better prompt processing performance on your models. You need to rebuild the code after change `GGML_SYCL_F16=OFF/ON`.

 #### Runtime

@@ -782,7 +686,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
 | GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features for Intel GPUs. (Recommended to 1 for intel devices older than Gen 10) |
-| GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
+| GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Support malloc device memory more than 4GB.|
@@ -0,0 +1,180 @@
+# GGML-VirtGPU Backend
+
+The GGML-VirtGPU backend enables GGML applications to run machine
+learning computations on host hardware while the application itself
+runs inside a virtual machine.  It uses host-guest shared memory to
+efficiently share data buffers between the two sides.
+
+This backend relies on the virtio-gpu, and VirglRenderer API Remoting
+(APIR) component. The backend is split into two libraries:
+- a GGML implementation (the "remoting frontend"), running in the
+  guest and interacting with the virtgpu device
+- a VirglRenderer APIR compatible library (the "remoting backend"),
+  running in the host and interacting with Virglrenderer and an actual
+  GGML device backend.
+
+## OS support
+
+| OS       | Status            | Backend     | CI testing  | Notes
+| -------- | ----------------- | ----------- | ----------- | -----
+| MacOS 14 | Supported         | ggml-metal  | X           | Working when compiled on MacOS 14
+| MacOS 15 | Supported         | ggml-metal  | X           | Working when compiled on MacOS 14 or MacOS 15
+| MacOS 26 | Not tested        |             |             |
+| Linux    | Under development | ggml-vulkan | not working | Working locally, CI running into deadlocks
+
+
+## Architecture Overview
+
+The GGML-VirtGPU backend consists of three main components:
+
+```mermaid
+graph TD
+    %% Nodes
+
+ subgraph GuestVM ["Guest VM - Frontend"]
+        App([GGML Application<br/>llama.cpp, etc.])
+
+        direction TB
+        Interface[GGML Backend Interface]
+        Comm["GGML-VirtGPU<br/>(hypercalls + shared mem)"]
+
+        App --> Interface
+        Interface --> Comm
+    end
+
+    API[virtio-gpu / virglrenderer API]
+
+    subgraph HostSystem [Host System - Backend]
+        direction TB
+        Dispatcher[GGML-VirtGPU-Backend]
+        BackendLib[GGML Backend library<br/>Metal / Vulkan / CPU / ...]
+
+        Dispatcher --> BackendLib
+    end
+
+    %% Connections
+    Comm --> API
+    API --> HostSystem
+```
+
+### Key Components
+
+1. **Guest-side Frontend** (`ggml-virtgpu/`): Implements the GGML backend interface and forwards operations to the host
+2. **Host-side Backend** (`ggml-virtgpu/backend/`): Receives forwarded operations and executes them on actual hardware backends
+3. **Communication Layer**: Uses virtio-gpu hypercalls and shared memory for efficient data transfer
+
+## Features
+
+- **Dynamic backend loading** on the host side (CPU, CUDA, Metal, etc.)
+- **Zero-copy data transfer** via host-guest shared memory pages
+
+## Communication Protocol
+
+### Hypercalls and Shared Memory
+
+The backend uses two primary communication mechanisms:
+
+1. **Hypercalls (`DRM_IOCTL_VIRTGPU_EXECBUFFER`)**: Trigger remote execution from guest to host
+2. **Shared Memory Pages**: Zero-copy data transfer for tensors and parameters
+
+#### Shared Memory Layout
+
+Each connection uses two shared memory buffers:
+
+- **Data Buffer** (24 MiB): For command/response data and tensor transfers
+- **Reply Buffer** (16 KiB): For command replies and status information
+- **Data Buffers**: Dynamically allocated host-guest shared buffers
+  served as GGML buffers.
+
+### APIR Protocol
+
+The Virglrender API Remoting protocol defines three command types:
+
+- `HANDSHAKE`: Protocol version negotiation and capability discovery
+- `LOADLIBRARY`: Dynamic loading of backend libraries on the host
+- `FORWARD`: API function call forwarding
+
+### Binary Serialization
+
+Commands and data are serialized using a custom binary protocol with:
+
+- Fixed-size encoding for basic types
+- Variable-length arrays with size prefixes
+- Buffer bounds checking
+- Error recovery mechanisms
+
+## Supported Operations
+
+### Device Operations
+- Device enumeration and capability queries
+- Memory information (total/free)
+- Backend type detection
+
+### Buffer Operations
+- Buffer allocation and deallocation
+- Tensor data transfer (host ↔ guest)
+- Memory copying and clearing
+
+### Computation Operations
+- Graph execution forwarding
+
+## Build Requirements
+
+### Guest-side Dependencies
+- `libdrm` for DRM/virtio-gpu communication
+- C++20 compatible compiler
+- CMake 3.14+
+
+### Host-side Dependencies
+- virglrenderer with APIR support (pending upstream review)
+- Target backend libraries (libggml-metal, libggml-vulkan, etc.)
+
+## Configuration
+
+### Environment Variables
+
+- `GGML_VIRTGPU_BACKEND_LIBRARY`: Path to the host-side backend library
+- `GGML_VIRTGPU_DEBUG`: Enable debug logging
+
+### Build Options
+
+- `GGML_VIRTGPU`: Enable the VirtGPU backend (`ON` or `OFF`, default: `OFF`)
+- `GGML_VIRTGPU_BACKEND`: Build the host-side backend component (`ON`, `OFF` or `ONLY`, default: `OFF`)
+
+### System Requirements
+
+- VM with virtio-gpu support
+- VirglRenderer with APIR patches
+- Compatible backend libraries on host
+
+## Limitations
+
+- **VM-specific**: Only works in virtual machines with virtio-gpu support
+- **Host dependency**: Requires properly configured host-side backend
+- **Latency**: Small overhead from VM escaping for each operation
+
+
+* This work is pending upstream changes in the VirglRenderer
+  project.
+  * The backend can be tested with Virglrenderer compiled from source
+  using this PR:
+  https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590
+* This work is pending changes in the VMM/hypervisor running the
+  virtual machine, which need to know how to route the newly
+  introduced APIR capset.
+  * The environment variable `VIRGL_ROUTE_VENUS_TO_APIR=1` allows
+    using the Venus capset, until the relevant hypervisors have been
+    patched. However, setting this flag breaks the Vulkan/Venus normal
+    behavior.
+  * The environment variable `GGML_REMOTING_USE_APIR_CAPSET` tells the
+    `ggml-virtgpu` backend to use the APIR capset. This will become
+    the default when the relevant hypervisors have been patched.
+
+* This work focused on improving the performance of llama.cpp running
+  on MacOS containers, and is mainly tested on this platform. The
+  linux support (via `krun`) is in progress.
+
+## See Also
+
+- [Development and Testing](VirtGPU/development.md)
+- [Backend configuration](VirtGPU/configuration.md)
@@ -0,0 +1,174 @@
+# GGML-VirtGPU Backend Configuration
+
+This document describes the environment variables used by the ggml-virtgpu backend system, covering both the frontend (guest-side) and backend (host-side) components.
+
+## Environment Variables Overview
+
+The ggml-virtgpu backend uses environment variables for configuration across three main components:
+- **Frontend (Guest)**: GGML applications running in VMs
+- **Hypervisor**: Virglrenderer/APIR system
+- **Backend (Host)**: Host-side GGML backend integration
+
+## Frontend (Guest-side) Configuration
+
+### GGML_REMOTING_USE_APIR_CAPSET
+- **Location**: `ggml/src/ggml-virtgpu/virtgpu.cpp`
+- **Type**: Boolean flag (presence-based)
+- **Purpose**: Controls which virtio-gpu capability set to use for communication
+- **Values**:
+  - Set (any value): Use the APIR capset (long-term setup)
+  - Unset: Use the Venus capset (easier for testing with an unmodified hypervisor)
+- **Default**: Unset (Venus capset)
+- **Usage**:
+  ```bash
+  export GGML_REMOTING_USE_APIR_CAPSET=1  # Use APIR capset
+  # or leave unset for Venus capset
+  ```
+
+## Hypervisor (Virglrenderer/APIR) Configuration
+
+These environment variables are used during the transition phase for
+running with an unmodified hypervisor (not supporting the
+VirglRenderer APIR component). They will be removed in the future, and
+the hypervisor will instead configure VirglRenderer with the APIR
+_Configuration Key_.
+
+### VIRGL_APIR_BACKEND_LIBRARY
+- **Location**: `virglrenderer/src/apir/apir-context.c`
+- **Configuration Key**: `apir.load_library.path`
+- **Type**: File path string
+- **Purpose**: Path to the APIR backend library that virglrenderer should dynamically load
+- **Required**: Yes
+- **Example**:
+  ```bash
+  export VIRGL_APIR_BACKEND_LIBRARY="/path/to/libggml-remotingbackend.so"
+  ```
+
+### VIRGL_ROUTE_VENUS_TO_APIR
+- **Location**: `virglrenderer/src/apir/apir-renderer.h`
+- **Type**: Boolean flag (presence-based)
+- **Purpose**: Temporary workaround to route Venus capset calls to APIR during hypervisor transition period
+- **Status**: will be removed once hypervisors support APIR natively
+- **Warning**: Breaks normal Vulkan/Venus functionality
+- **Usage**:
+  ```bash
+  export VIRGL_ROUTE_VENUS_TO_APIR=1  # For testing with an unmodified hypervisor
+  ```
+
+### VIRGL_APIR_LOG_TO_FILE
+- **Location**: `virglrenderer/src/apir/apir-renderer.c`
+- **Environment Variable**: `VIRGL_APIR_LOG_TO_FILE`
+- **Type**: File path string
+- **Purpose**: Enable debug logging from the VirglRenderer APIR component to specified file
+- **Required**: No (optional debugging)
+- **Default**: Logging to `stderr`
+- **Usage**:
+  ```bash
+  export VIRGL_APIR_LOG_TO_FILE="/tmp/apir-debug.log"
+  ```
+
+## Backend (Host-side) Configuration
+
+These environment variables are used during the transition phase for
+running with an unmodified hypervisor (not supporting the
+VirglRenderer APIR component). They will be removed in the future, and
+the hypervisor will instead configure VirglRenderer with the APIR
+_Configuration Key_.
+
+### APIR_LLAMA_CPP_GGML_LIBRARY_PATH
+- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp`
+- **Environment Variable**: `APIR_LLAMA_CPP_GGML_LIBRARY_PATH`
+- **Configuration Key**: `ggml.library.path`
+- **Type**: File path string
+- **Purpose**: Path to the actual GGML backend library (Metal, CUDA, Vulkan, etc.)
+- **Required**: **Yes** - backend initialization fails without this
+- **Examples**:
+  ```bash
+  # macOS with Metal backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-metal.dylib"
+
+  # Linux with CUDA backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-cuda.so"
+
+  # macOS or Linux with Vulkan backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-vulkan.so"
+  ```
+
+### APIR_LLAMA_CPP_GGML_LIBRARY_REG
+- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp`
+- **Environment Variable**: `APIR_LLAMA_CPP_GGML_LIBRARY_REG`
+- **Configuration Key**: `ggml.library.reg`
+- **Type**: Function symbol name string
+- **Purpose**: Name of the backend registration function to call after loading the library
+- **Required**: No (defaults to `ggml_backend_init`)
+- **Default**: `ggml_backend_init`
+- **Examples**:
+  ```bash
+  # Metal backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_metal_reg"
+
+  # CUDA backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_cuda_reg"
+
+  # Vulkan backend
+  export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_vulkan_reg"
+
+  # Generic fallback (default)
+  # export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_init"
+  ```
+
+### APIR_LLAMA_CPP_LOG_TO_FILE
+- **Location**: `ggml/src/ggml-virtgpu/backend/backend.cpp:62`
+- **Environment Variable**: `APIR_LLAMA_CPP_LOG_TO_FILE`
+- **Type**: File path string
+- **Purpose**: Enable debug logging from the GGML backend to specified file
+- **Required**: No (optional debugging)
+- **Usage**:
+  ```bash
+  export APIR_LLAMA_CPP_LOG_TO_FILE="/tmp/ggml-backend-debug.log"
+  ```
+
+## Configuration Flow
+
+The configuration system works as follows:
+
+1. **Hypervisor Setup**: Virglrenderer loads the APIR backend library specified by `VIRGL_APIR_BACKEND_LIBRARY`
+
+2. **Context Creation**: When an APIR context is created, it populates a configuration table with environment variables:
+   - `apir.load_library.path` ← `VIRGL_APIR_BACKEND_LIBRARY`
+   - `ggml.library.path` ← `APIR_LLAMA_CPP_GGML_LIBRARY_PATH`
+   - `ggml.library.reg` ← `APIR_LLAMA_CPP_GGML_LIBRARY_REG`
+   - this step will eventually be performed by the hypervisor itself, with command-line arguments instead of environment variables.
+
+3. **Backend Initialization**: The backend queries the configuration via callbacks:
+   - `virgl_cbs->get_config(ctx_id, "ggml.library.path")` returns the library path
+   - `virgl_cbs->get_config(ctx_id, "ggml.library.reg")` returns the registration function
+
+4. **Library Loading**: The backend dynamically loads and initializes the specified GGML library
+
+## Error Messages
+
+Common error scenarios and their messages:
+
+- **Missing library path**: `"cannot open the GGML library: env var 'APIR_LLAMA_CPP_GGML_LIBRARY_PATH' not defined"`
+- **Missing registration function**: `"cannot register the GGML library: env var 'APIR_LLAMA_CPP_GGML_LIBRARY_REG' not defined"`
+
+## Example Complete Configuration
+
+Here's an example configuration for a macOS host with Metal backend:
+
+```bash
+# Hypervisor environment
+export VIRGL_APIR_BACKEND_LIBRARY="/opt/llama.cpp/lib/libggml-virtgpu-backend.dylib"
+
+# Backend configuration
+export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="/opt/llama.cpp/lib/libggml-metal.dylib"
+export APIR_LLAMA_CPP_GGML_LIBRARY_REG="ggml_backend_metal_reg"
+
+# Optional logging
+export VIRGL_APIR_LOG_TO_FILE="/tmp/apir.log"
+export APIR_LLAMA_CPP_LOG_TO_FILE="/tmp/ggml.log"
+
+# Guest configuration
+export GGML_REMOTING_USE_APIR_CAPSET=1
+```
@@ -0,0 +1,220 @@
+# Development and Testing
+
+## Development
+
+### Code Generation
+
+The backend uses code generation from YAML configuration:
+
+```bash
+# Regenerate protocol code
+cd ggml-virtgpu/
+python regenerate_remoting.py
+```
+
+### Adding New Operations
+
+1. Add function definition to `ggmlremoting_functions.yaml`
+2. Regenerate code with `regenerate_remoting.py`
+3. Implement guest-side forwarding in `virtgpu-forward-*.cpp`
+4. Implement host-side handling in `backend-dispatched-*.cpp`
+
+## Testing
+
+This document provides instructions for building and testing the GGML-VirtGPU backend on macOS with containers.
+
+### Prerequisites
+
+The testing setup requires:
+
+- macOS host system
+- Container runtime with `libkrun` provider (podman machine)
+- Access to development patchset for VirglRenderer
+
+### Required Patchsets
+
+The backend requires patches that are currently under review:
+
+- **Virglrenderer APIR upstream PR**: https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1590 (for reference)
+- **MacOS Virglrenderer (for krunkit)**: https://gitlab.freedesktop.org/kpouget/virglrenderer/-/tree/main-macos
+- **Linux Virglrenderer (for krun)**: https://gitlab.freedesktop.org/kpouget/virglrenderer/-/tree/main-linux
+
+### Build Instructions
+
+#### 1. Build ggml-virtgpu-backend (Host-side, macOS)
+
+```bash
+# Build the backend that runs natively on macOS
+mkdir llama.cpp
+cd llama.cpp
+git clone https://github.com/ggml-org/llama.cpp.git src
+cd src
+
+LLAMA_MAC_BUILD=$PWD/build/ggml-virtgpu-backend
+
+cmake -S . -B $LLAMA_MAC_BUILD \
+      -DGGML_NATIVE=OFF \
+      -DLLAMA_CURL=ON \
+      -DGGML_REMOTINGBACKEND=ONLY \
+      -DGGML_METAL=ON
+
+TARGETS="ggml-metal"
+cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $TARGETS
+
+# Build additional tools for native benchmarking
+EXTRA_TARGETS="llama-run llama-bench"
+cmake --build $LLAMA_MAC_BUILD --parallel 8 --target $EXTRA_TARGETS
+```
+
+#### 2. Build virglrenderer (Host-side, macOS)
+
+```bash
+# Build virglrenderer with APIR support
+mkdir virglrenderer
+git clone https://gitlab.freedesktop.org/kpouget/virglrenderer -b main-macos src
+cd src
+
+VIRGL_BUILD_DIR=$PWD/build
+
+# -Dvenus=true and VIRGL_ROUTE_VENUS_TO_APIR=1 route the APIR requests via the Venus backend, for easier testing without a patched hypervisor
+
+meson setup $VIRGL_BUILD_DIR \
+      -Dvenus=true \
+      -Dapir=true
+
+ninja -C $VIRGL_BUILD_DIR
+```
+
+#### 3. Build ggml-virtgpu (Guest-side, Linux)
+
+Option A: Build from a script:
+
+```bash
+# Inside a Linux container
+mkdir llama.cpp
+git clone https://github.com/ggml-org/llama.cpp.git src
+cd src
+
+LLAMA_LINUX_BUILD=$PWD//build-virtgpu
+
+cmake -S . -B $LLAMA_LINUX_BUILD \
+      -DGGML_VIRTGPU=ON
+
+ninja -C $LLAMA_LINUX_BUILD
+```
+
+Option B: Build container image with frontend:
+
+```bash
+cat << EOF > remoting.containerfile
+FROM quay.io/fedora/fedora:43
+USER 0
+
+WORKDIR /app/remoting
+
+ARG LLAMA_CPP_REPO="https://github.com/ggml-org/llama.cpp.git"
+ARG LLAMA_CPP_VERSION="master"
+ARG LLAMA_CPP_CMAKE_FLAGS="-DGGML_VIRTGPU=ON"
+ARG LLAMA_CPP_CMAKE_BUILD_FLAGS="--parallel 4"
+
+RUN dnf install -y git cmake gcc gcc-c++ libcurl-devel libdrm-devel
+
+RUN git clone "\${LLAMA_CPP_REPO}" src \\
+ && git -C src fetch origin \${LLAMA_CPP_VERSION} \\
+ && git -C src reset --hard FETCH_HEAD
+
+RUN mkdir -p build \\
+ && cd src \\
+ && set -o pipefail \\
+ && cmake -S . -B ../build \${LLAMA_CPP_CMAKE_FLAGS} \\
+ && cmake --build ../build/ \${LLAMA_CPP_CMAKE_BUILD_FLAGS}
+
+ENTRYPOINT ["/app/remoting/src/build/bin/llama-server"]
+EOF
+
+mkdir -p empty_dir
+podman build -f remoting.containerfile ./empty_dir -t localhost/llama-cpp.virtgpu
+```
+
+### Environment Setup
+
+#### Set krunkit Environment Variables
+
+```bash
+# Define the base directories (adapt these paths to your system)
+VIRGL_BUILD_DIR=$HOME/remoting/virglrenderer/build
+LLAMA_MAC_BUILD=$HOME/remoting/llama.cpp/build-backend
+
+# For krunkit to load the custom virglrenderer library
+export DYLD_LIBRARY_PATH=$VIRGL_BUILD_DIR/src
+
+# For Virglrenderer to load the ggml-remotingbackend library
+export VIRGL_APIR_BACKEND_LIBRARY="$LLAMA_MAC_BUILD/bin/libggml-virtgpu-backend.dylib"
+
+# For llama.cpp remotingbackend to load the ggml-metal backend
+export APIR_LLAMA_CPP_GGML_LIBRARY_PATH="$LLAMA_MAC_BUILD/bin/libggml-metal.dylib"
+export APIR_LLAMA_CPP_GGML_LIBRARY_REG=ggml_backend_metal_reg
+```
+
+#### Launch Container Environment
+
+```bash
+# Set container provider to libkrun
+export CONTAINERS_MACHINE_PROVIDER=libkrun
+podman machine start
+```
+
+#### Verify Environment
+
+Confirm that krunkit is using the correct virglrenderer library:
+
+```bash
+lsof -c krunkit | grep virglrenderer
+# Expected output:
+# krunkit 50574 user  txt  REG  1,14  2273912  10849442 ($VIRGL_BUILD_DIR/src)/libvirglrenderer.1.dylib
+```
+
+### Running Tests
+
+#### Launch Test Container
+
+```bash
+# Optional model caching
+mkdir -p models
+PODMAN_CACHE_ARGS="-v models:/models --user root:root --cgroupns host --security-opt label=disable -w /models"
+
+podman run $PODMAN_CACHE_ARGS -it --rm --device /dev/dri localhost/llama-cpp.virtgpu
+```
+
+#### Test llama.cpp in Container
+
+```bash
+
+# Run performance benchmark
+/app/remoting/build/bin/llama-bench -m ./llama3.2
+```
+
+Expected output (performance may vary):
+```
+| model                          |       size |     params | backend    | ngl |          test |                  t/s |
+| ------------------------------ | ---------: | ---------: | ---------- | --: | ------------: | -------------------: |
+| llama 3B Q4_K - Medium         |   1.87 GiB |     3.21 B | ggml-virtgpu |  99 |         pp512 |        991.30 ± 0.66 |
+| llama 3B Q4_K - Medium         |   1.87 GiB |     3.21 B | ggml-virtgpu |  99 |         tg128 |         85.71 ± 0.11 |
+```
+
+### Troubleshooting
+
+#### SSH Environment Variable Issues
+
+⚠️ **Warning**: Setting `DYLD_LIBRARY_PATH` from SSH doesn't work on macOS. Here is a workaround:
+
+**Workaround 1: Replace system library**
+```bash
+VIRGL_BUILD_DIR=$HOME/remoting/virglrenderer/build  # ⚠️ adapt to your system
+BREW_VIRGL_DIR=/opt/homebrew/Cellar/virglrenderer/0.10.4d/lib
+VIRGL_LIB=libvirglrenderer.1.dylib
+
+cd $BREW_VIRGL_DIR
+mv $VIRGL_LIB ${VIRGL_LIB}.orig
+ln -s $VIRGL_BUILD_DIR/src/$VIRGL_LIB
+```
@@ -35,7 +35,7 @@ Adapt below build commands accordingly.
 Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:

 ```
-[d]/workspace> cp docs/backend/hexagon/CMakeUserPresets.json .
+[d]/workspace> cp docs/backend/snapdragon/CMakeUserPresets.json .

 [d]/workspace> cmake --preset arm64-android-snapdragon-release -B build-snapdragon
 Preset CMake variables:
@@ -252,9 +252,7 @@ CUDA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.ggu

 The environment variable [`CUDA_SCALE_LAUNCH_QUEUES`](https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/environment-variables.html#cuda-scale-launch-queues) controls the size of CUDA's command buffer, which determines how many GPU operations can be queued before the CPU must wait for the GPU to catch up. A larger buffer reduces CPU-side stalls and allows more work to be queued on a GPU.

-**Default behavior:** llama.cpp automatically sets `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.
-
-See PR [#19042](https://github.com/ggml-org/llama.cpp/pull/19042) for performance benchmarks and technical details.
+Consider setting `CUDA_SCALE_LAUNCH_QUEUES=4x`, which increases the CUDA command buffer to 4 times its default size. This optimization is particularly beneficial for **Multi-GPU setups with pipeline parallelism**, where it significantly improves prompt processing throughput by allowing more operations to be enqueued across GPUs.

 ### Unified Memory

@@ -22,7 +22,7 @@ Legend:
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
@@ -113,7 +113,7 @@ Legend:
 |                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ |
+|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ |
@@ -77,8 +77,8 @@
 "SYCL0","GELU_ERF","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","FLOOR","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","FLOOR","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
+"SYCL0","CEIL","type=f16,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
+"SYCL0","CEIL","type=f16,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","ROUND","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","ROUND","type=f16,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
 "SYCL0","TRUNC","type=f16,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
@@ -161,8 +161,8 @@
 "SYCL0","GELU_ERF","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","FLOOR","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","FLOOR","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
-"SYCL0","CEIL","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
+"SYCL0","CEIL","type=f32,ne_a=[128,2,2,2],v=1","support","1","yes","SYCL"
+"SYCL0","CEIL","type=f32,ne_a=[5,7,11,13],v=1","support","1","yes","SYCL"
 "SYCL0","ROUND","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
 "SYCL0","ROUND","type=f32,ne_a=[5,7,11,13],v=1","support","0","no","SYCL"
 "SYCL0","TRUNC","type=f32,ne_a=[128,2,2,2],v=1","support","0","no","SYCL"
@@ -9677,168 +9677,168 @@
 "SYCL0","ARGSORT","type=f32,ne=[2048,2,1,3],order=1","support","1","yes","SYCL"
 "SYCL0","ARGSORT","type=f32,ne=[2049,2,1,3],order=1","support","1","yes","SYCL"
 "SYCL0","ARGSORT","type=f32,ne=[2,8,8192,1],order=1","support","1","yes","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[12,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[13,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[13,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[12,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[13,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[13,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[15,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[19,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[27,1,2,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[43,1,2,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[64,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[75,1,2,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[128,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[139,1,2,1],k=100,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[256,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[267,1,2,1],k=100,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[512,1,1,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[523,1,2,1],k=500,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1024,1,1,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[1035,1,2,1],k=1023,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2048,1,1,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[2059,1,2,1],k=1023,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4096,1,1,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[4107,1,2,1],k=1023,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=500,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8192,1,1,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[8203,1,2,1],k=1023,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9847,16 +9847,16 @@
 "SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[16395,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9865,16 +9865,16 @@
 "SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[32768,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[32779,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9883,16 +9883,16 @@
 "SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[65536,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[65547,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9901,16 +9901,16 @@
 "SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[131072,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[131083,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9919,16 +9919,16 @@
 "SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[262144,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[262155,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=100,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=500,ties=0","support","0","no","SYCL"
@@ -9937,51 +9937,51 @@
 "SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=1023,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[524288,1,1,1],k=9999,ties=0","support","0","no","SYCL"
 "SYCL0","TOP_K","type=f32,ne=[524299,1,2,1],k=9999,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=1,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=2,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=3,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=7,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=15,ties=0","support","0","no","SYCL"
-"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=15,ties=0","support","0","no","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=1,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=2,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=3,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=7,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16,10,10,10],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[60,10,10,10],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1023,2,1,3],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1024,2,1,3],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[1025,2,1,3],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2047,2,1,3],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2048,2,1,3],k=15,ties=0","support","1","yes","SYCL"
+"SYCL0","TOP_K","type=f32,ne=[2049,2,1,3],k=15,ties=0","support","1","yes","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=0","support","1","yes","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","1","yes","SYCL"
 "SYCL0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest","support","1","yes","SYCL"
@@ -119,8 +119,6 @@ If a draft model is combined with a draftless decoding the draftless decoding ha
                                        of lookup n-gram (default: 12)
 --spec-ngram-size-m N                   ngram size M for ngram-simple/ngram-map speculative decoding, length
                                        of draft m-gram (default: 48)
--spec-ngram-check-rate N               ngram check rate for ngram-simple/ngram-map speculative decoding
-                                        (default: 1)
 --spec-ngram-min-hits N                 minimum hits for ngram-map speculative decoding (default: 1)
 ```

@@ -153,10 +151,6 @@ Sets the size M of the draft m-gram for n-gram map based speculative decoding.
 The m-gram size determines how many tokens to draft when a match is found.
 Larger values can provide more speedup but may reduce acceptance rate.

-### `--spec-ngram-check-rate R`
-
-This option aims at performance if the n-gram lookup in history is to costly. A lookup will be executed at every R tokens (default is 1, every token).
-
 ### `--spec-ngram-min-hits H`

 This option defines how often a key has to appear in the token history to be used as a draft (default is 1).
@@ -175,7 +169,12 @@ draft acceptance rate = 0.70312 (   90 accepted /   128 generated)
 statistics ngram_mod: #calls = 810, #gen drafts = 15, #acc drafts = 15, #gen tokens = 960, #acc tokens = 730, dur(b,g,a) = 0.149, 0.347, 0.005 ms
 ```

- `#calls`: number of calls of this implementations
+```
+statistics ngram_map_k: #calls(b,g,a) = 6 1690 26, #gen drafts = 26, #acc drafts = 26, #gen tokens = 1248, #acc tokens = 968, dur(b,g,a) = 2.234, 1.427, 0.016 ms
+```
+
+
+- `#calls(b,g,a)`: number of calls of begin (new prompt), generation and accumulation of this implementations
 - `#gen drafts`: number of drafts generated by this implementation
 - `#acc drafts`: number of drafts accepted (partially) by the main model
 - `#gen tokens`: number of tokens generated by this implementation (including rejected tokens)
@@ -33,11 +33,14 @@ DEVICE ?= auto
 causal-convert-model-bf16: OUTTYPE=bf16
 causal-convert-model-bf16: causal-convert-model

+causal-convert-model-debug: DEBUG=--debug
+causal-convert-model-debug: causal-convert-model
+
 causal-convert-model:
 	$(call validate_model_path,causal-convert-model)
 	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(MODEL_PATH)" \
 	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
-	./scripts/causal/convert-model.sh
+	./scripts/causal/convert-model.sh $(DEBUG)

 causal-convert-mm-model-bf16: OUTTYPE=bf16
 causal-convert-mm-model-bf16: MM_OUTTYPE=f16
@@ -4,12 +4,17 @@ set -e

 # Parse command line arguments
 MMPROJ=""
+DEBUG=""
 while [[ $# -gt 0 ]]; do
    case $1 in
        --mmproj)
            MMPROJ="--mmproj"
            shift
            ;;
+        --debug)
+            DEBUG="1"
+            shift
+            ;;
        *)
            shift
            ;;
@@ -28,7 +33,12 @@ echo "Data  type: ${TYPE}"
 echo "Converted model path:: ${CONVERTED_MODEL}"
 echo "Metadata override: ${METADATA_OVERRIDE}"

-CMD_ARGS=("python" "../../convert_hf_to_gguf.py" "--verbose")
+if [[ -n "$DEBUG" ]]; then
+    CMD_ARGS=("python" "-m" "pdb")
+else
+    CMD_ARGS=("python")
+fi
+CMD_ARGS+=("../../convert_hf_to_gguf.py" "--verbose")
 CMD_ARGS+=("${MODEL_PATH}")
 CMD_ARGS+=("--outfile" "${CONVERTED_MODEL}")
 CMD_ARGS+=("--outtype" "${TYPE}")
@@ -0,0 +1,159 @@
+#!/usr/bin/env python3
+
+import argparse
+import json
+import os
+import re
+import sys
+from pathlib import Path
+from typing import Optional
+from safetensors import safe_open
+
+
+MODEL_SAFETENSORS_FILE = "model.safetensors"
+MODEL_SAFETENSORS_INDEX = "model.safetensors.index.json"
+
+
+def get_weight_map(model_path: Path) -> Optional[dict[str, str]]:
+    index_file = model_path / MODEL_SAFETENSORS_INDEX
+
+    if index_file.exists():
+        with open(index_file, 'r') as f:
+            index = json.load(f)
+            return index.get("weight_map", {})
+
+    return None
+
+
+def get_all_tensor_names(model_path: Path) -> list[str]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        return list(weight_map.keys())
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        try:
+            with safe_open(single_file, framework="pt", device="cpu") as f:
+                return list(f.keys())
+        except Exception as e:
+            print(f"Error reading {single_file}: {e}")
+            sys.exit(1)
+
+    print(f"Error: No safetensors files found in {model_path}")
+    sys.exit(1)
+
+
+def find_tensor_file(model_path: Path, tensor_name: str) -> Optional[str]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        return weight_map.get(tensor_name)
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        return single_file.name
+
+    return None
+
+
+def normalize_tensor_name(tensor_name: str) -> str:
+    normalized = re.sub(r'\.\d+\.', '.#.', tensor_name)
+    normalized = re.sub(r'\.\d+$', '.#', normalized)
+    return normalized
+
+
+def list_all_tensors(model_path: Path, unique: bool = False):
+    tensor_names = get_all_tensor_names(model_path)
+
+    if unique:
+        seen = set()
+        for tensor_name in sorted(tensor_names):
+            normalized = normalize_tensor_name(tensor_name)
+            if normalized not in seen:
+                seen.add(normalized)
+                print(normalized)
+    else:
+        for tensor_name in sorted(tensor_names):
+            print(tensor_name)
+
+
+def print_tensor_info(model_path: Path, tensor_name: str):
+    tensor_file = find_tensor_file(model_path, tensor_name)
+
+    if tensor_file is None:
+        print(f"Error: Could not find tensor '{tensor_name}' in model index")
+        print(f"Model path: {model_path}")
+        sys.exit(1)
+
+    file_path = model_path / tensor_file
+
+    try:
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            if tensor_name in f.keys():
+                tensor_slice = f.get_slice(tensor_name)
+                shape = tensor_slice.get_shape()
+                print(f"Tensor: {tensor_name}")
+                print(f"File:   {tensor_file}")
+                print(f"Shape:  {shape}")
+            else:
+                print(f"Error: Tensor '{tensor_name}' not found in {tensor_file}")
+                sys.exit(1)
+
+    except FileNotFoundError:
+        print(f"Error: The file '{file_path}' was not found.")
+        sys.exit(1)
+    except Exception as e:
+        print(f"An error occurred: {e}")
+        sys.exit(1)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Print tensor information from a safetensors model"
+    )
+    parser.add_argument(
+        "tensor_name",
+        nargs="?",  # optional (if --list is used for example)
+        help="Name of the tensor to inspect"
+    )
+    parser.add_argument(
+        "-m", "--model-path",
+        type=Path,
+        help="Path to the model directory (default: MODEL_PATH environment variable)"
+    )
+    parser.add_argument(
+        "-l", "--list",
+        action="store_true",
+        help="List unique tensor patterns in the model (layer numbers replaced with #)"
+    )
+
+    args = parser.parse_args()
+
+    model_path = args.model_path
+    if model_path is None:
+        model_path_str = os.environ.get("MODEL_PATH")
+        if model_path_str is None:
+            print("Error: --model-path not provided and MODEL_PATH environment variable not set")
+            sys.exit(1)
+        model_path = Path(model_path_str)
+
+    if not model_path.exists():
+        print(f"Error: Model path does not exist: {model_path}")
+        sys.exit(1)
+
+    if not model_path.is_dir():
+        print(f"Error: Model path is not a directory: {model_path}")
+        sys.exit(1)
+
+    if args.list:
+        list_all_tensors(model_path, unique=True)
+    else:
+        if args.tensor_name is None:
+            print("Error: tensor_name is required when not using --list")
+            sys.exit(1)
+        print_tensor_info(model_path, args.tensor_name)
+
+
+if __name__ == "__main__":
+    main()
@@ -19,6 +19,9 @@ extern "C" {
        // abort ggml_graph_compute when true
        ggml_abort_callback abort_callback;
        void *              abort_callback_data;
+
+        // use only reference implementations
+        bool use_ref;
    };

    // numa strategies
@@ -132,6 +135,8 @@ extern "C" {
    GGML_BACKEND_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
    GGML_BACKEND_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);

+    GGML_BACKEND_API void ggml_backend_cpu_set_use_ref(ggml_backend_t backend_cpu, bool use_ref);
+
    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);

    GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);
@@ -7,8 +7,6 @@
 extern "C" {
 #endif

-#define GGML_REMOTING_FRONTEND_NAME "RemotingFrontend"
-
 GGML_BACKEND_API ggml_backend_reg_t ggml_backend_virtgpu_reg();

 #ifdef  __cplusplus
@@ -471,9 +471,10 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,

    int best_score = 0;
    fs::path best_path;
+    std::error_code ec;

    for (const auto & search_path : search_paths) {
-        if (std::error_code ec; !fs::exists(search_path, ec)) {
+        if (!fs::exists(search_path, ec)) {
            if (ec) {
                GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(search_path).c_str(), ec.message().c_str());
            } else {
@@ -483,7 +484,7 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
        }
        fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
        for (const auto & entry : dir_it) {
-            if (entry.is_regular_file()) {
+            if (entry.is_regular_file(ec)) {
                auto filename = entry.path().filename();
                auto ext = entry.path().extension();
                if (filename.native().find(file_prefix) == 0 && ext == file_extension) {
@@ -258,6 +258,7 @@ void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor *
    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");

    if (backend->iface.set_tensor_async == NULL) {
+        ggml_backend_synchronize(backend);
        ggml_backend_tensor_set(tensor, data, offset, size);
    } else {
        backend->iface.set_tensor_async(backend, tensor, data, offset, size);
@@ -271,6 +272,7 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");

    if (backend->iface.get_tensor_async == NULL) {
+        ggml_backend_synchronize(backend);
        ggml_backend_tensor_get(tensor, data, offset, size);
    } else {
        backend->iface.get_tensor_async(backend, tensor, data, offset, size);
@@ -3286,130 +3286,223 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context & ctx, ggml_tensor
 }

 /**
- * @brief Performs expert-specific matrix multiplication (MoE) with
- * quantized precision using the CANN backend.
+ * @brief Performs quantized matrix multiplication for Mixture of Experts (MoE)
+ * models using the CANN backend.
 *
- * This function executes a matrix multiplication operation tailored for
- * Mixture of Experts (MoE) models, where the input tensor is multiplied
- * with expert-specific quantized weight matrices. It leverages the CANN
- * backend to perform efficient low-precision computations and stores the
- * quantized result in the destination tensor `dst`.
+ * This function implements MUL_MAT_ID operation for quantized weight matrices
+ * (Q4_0 and Q8_0 formats). It selects expert-specific weight matrices based on
+ * the provided expert indices, and computes matrix multiplication using CANN's
+ * WeightQuantBatchMatmulV2 operator.
 *
- * Quantization techniques reduce memory footprint and improve performance
- * by using lower-bit representations (e.g., int8) instead of floating-point.
- * This function is designed to work with such formats and may incorporate
- * optimizations like identity-based fast paths or routing masks for sparse
- * expert selection.
+ * The function performs the following steps:
+ * 1. Converts input/output tensors to F16 format if necessary
+ * 2. Uses IndexSelect to extract expert-specific weights and scales based on indices
+ * 3. Performs quantized matrix multiplication for each expert using WeightQuantBatchMatmulV2
+ * 4. Converts output back to the target type if needed
 *
- * @param ctx The context for executing CANN backend operations.
- * @param dst The destination tensor where the quantized MoE multiplication result
- * will be stored.
+ * Tensor shapes:
+ * - dst:  [M, K, N, 1] - output tensor
+ * - src0: [D, M, A, 1] - quantized weight matrices (Q4_0 or Q8_0)
+ * - src1: [D, B, N, 1] - input activations (B = K for per-expert input, or B = 1 for broadcast)
+ * - ids:  [K, N] - expert indices for routing
 *
- * @note This function assumes quantized data types and is designed for
- * MoE architectures with potential sparse expert routing.
+ * @param ctx The CANN backend context for operation execution.
+ * @param dst The destination tensor where the multiplication result will be stored.
+ *
+ * @note Only Q4_0 and Q8_0 quantization formats are supported.
+ * @note The function handles automatic type conversion to/from F16 as needed by the hardware.
 */
 static void ggml_cann_mul_mat_id_quant(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    // TODO: Use aclnnGroupedMatMul
-    //dst   [M, K, N, 1]
-    ggml_tensor * src0 = dst->src[0];  //src0	[D, M, A, 1]
-    ggml_tensor * src1 = dst->src[1];  //src1	[D, B, N, 1], B = K or B = 1
-    ggml_tensor * ids  = dst->src[2];  //ids	[K, N]
+    // dst:  [M, K, N, 1]
+    // src0: [D, M, A, 1] - quantized weights
+    // src1: [D, B, N, 1] - input activations, B = K or B = 1
+    // ids:  [K, N] - expert indices
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src1 = dst->src[1];
+    ggml_tensor * ids  = dst->src[2];

-    GGML_TENSOR_BINARY_OP_LOCALS
+    GGML_ASSERT(src0->ne[3] == 1);
+    GGML_ASSERT(src1->ne[3] == 1);
+    GGML_ASSERT(dst->ne[3] == 1);
+    GGML_ASSERT(src1->ne[2] == ids->ne[1]);

-    // copy index from npu to cpu
-    int64_t n_as  = ne02;        // A
-    int64_t n_ids = ids->ne[0];  // K
+    const int64_t        n_batches        = ids->ne[1];
+    const int64_t        n_select_experts = ids->ne[0];
+    const enum ggml_type type             = src0->type;

-    std::vector<char> ids_host(ggml_nbytes(ids));
-    ACL_CHECK(aclrtMemcpyAsync(ids_host.data(), ggml_nbytes(ids), ids->data, ggml_nbytes(ids),
-                               ACL_MEMCPY_DEVICE_TO_HOST, ctx.stream()));
-    ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
+    const int32_t group_size = QK8_0;  // Both Q4_0 and Q8_0 use group size of 32
+    GGML_ASSERT(group_size == QK4_0);

-    char * src0_original = (char *) src0->data;
-    char * src1_original = (char *) src1->data;
-    char * dst_original  = (char *) dst->data;
+    // Calculate element size for quantized weights
+    const float weight_elem_size =
+        (type == GGML_TYPE_Q4_0) ? 0.5f :
+        (type == GGML_TYPE_Q8_0) ? 1.0f :
+                                   (GGML_ABORT("MUL_MAT_ID only supports Q4_0 and Q8_0"), 0.0f);

-    ggml_tensor src0_row = *src0;
-    ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row  = *dst;
+    // Calculate scale offset in memory
+    const size_t weight_size     = src0->ne[0] * src0->ne[1] * src0->ne[2] * weight_elem_size;
+    const size_t scale_elem_size = sizeof(uint16_t);
+    char *       scale_data      = (char *) src0->data + weight_size;

-    const enum ggml_type type = dst->src[0]->type;
-    float                weight_elem_size;
-    if (type == GGML_TYPE_Q4_0) {
-        weight_elem_size = float(sizeof(uint8_t)) / 2;
-    } else if (type == GGML_TYPE_Q8_0) {
-        weight_elem_size = float(sizeof(uint8_t));
-    } else {
-        GGML_ABORT("MUL_MAT_ID only support quant type Q4_0 and Q8_0 ");
-    }
+    // Allocate buffers for selected expert weights and scales
+    const size_t         selected_weight_size = src0->ne[0] * src0->ne[1] * n_select_experts * weight_elem_size;
+    ggml_cann_pool_alloc selected_weight_alloc(ctx.pool(), selected_weight_size);
+    void *               selected_weight_buffer = selected_weight_alloc.get();

-    // src0_row [D, M, 1, 1] weight without permute
-    src0_row.ne[2]       = 1;
-    src0_row.ne[3]       = 1;
-    src0_row.nb[0]       = weight_elem_size;
-    src0_row.nb[1]       = weight_elem_size * ne00;
-    src0_row.nb[2]       = weight_elem_size * ne00;
-    src0_row.nb[3]       = weight_elem_size * ne00;
-    size_t weight_stride = ne00 * ne01 * weight_elem_size;
-    size_t weight_size   = weight_stride * ne02 * ne03;
+    const size_t selected_scale_size = (src0->ne[0] / group_size) * src0->ne[1] * n_select_experts * scale_elem_size;
+    ggml_cann_pool_alloc selected_scale_alloc(ctx.pool(), selected_scale_size);
+    void *               selected_scale_buffer = selected_scale_alloc.get();

-    // scale [D, M, 1, 1] -> scale && permute
-    size_t scale_elem_size = sizeof(uint16_t);
-    size_t scale_stride    = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
+    // Helper lambda to allocate and cast tensor to F16 if needed
+    constexpr size_t f16_elem_size      = sizeof(uint16_t);
+    auto             prepare_f16_buffer = [&](ggml_tensor * tensor, ggml_cann_pool_alloc & allocator,
+                                  bool need_cast = false) -> void * {
+        if (tensor->type == GGML_TYPE_F16) {
+            return tensor->data;
+        }

-    // src1_row [D, 1, 1, 1] -> input
-    src1_row.ne[1] = 1;
-    src1_row.ne[2] = 1;
-    src1_row.ne[3] = 1;
-    src1_row.nb[2] = nb11;
-    src1_row.nb[3] = nb11;
+        size_t total_size = f16_elem_size;
+        for (int i = 0; i < GGML_MAX_DIMS; i++) {
+            total_size *= tensor->ne[i];
+        }
+        void * buffer = allocator.alloc(total_size);

-    // dst_row [M, 1, 1, 1] -> out
-    dst_row.ne[1] = 1;
-    dst_row.ne[2] = 1;
-    dst_row.ne[3] = 1;
-    dst_row.nb[2] = nb1;
-    dst_row.nb[3] = nb1;
+        if (need_cast == false) {
+            return buffer;
+        }

-    //create weight for one row
-    ggml_cann_pool_alloc weight_allocator(ctx.pool());
-    void *               weight_buffer = weight_allocator.alloc(nb02);
-    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-        for (int64_t id = 0; id < n_ids; id++) {
-            // expert index
-            int32_t i02 = *(int32_t *) (ids_host.data() + iid1 * ids->nb[1] + id * ids->nb[0]);
-            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+        int64_t ne[GGML_MAX_DIMS];
+        size_t  nb[GGML_MAX_DIMS] = { f16_elem_size };
+        for (int i = 0; i < GGML_MAX_DIMS; i++) {
+            ne[i] = tensor->ne[i];
+            if (i > 0) {
+                nb[i] = nb[i - 1] * ne[i - 1];
+            }
+        }

-            // If B = 1 (broadcast), always use 0; otherwise, use id.
-            int64_t i11 = (ne11 == 1 ? 0 : id);
-            int64_t i12 = iid1;
+        acl_tensor_ptr src_tensor = ggml_cann_create_tensor(tensor);
+        acl_tensor_ptr f16_tensor = ggml_cann_create_tensor(buffer, ACL_FLOAT16, f16_elem_size, ne, nb, GGML_MAX_DIMS);
+        aclnn_cast(ctx, src_tensor.get(), f16_tensor.get(), ACL_FLOAT16);

-            int64_t i1 = id;
-            int64_t i2 = i12;
+        return buffer;
+    };

-            void * src0_tmp_ptr  = src0_original + i02 * weight_stride;
-            void * scale_tmp_ptr = src0_original + weight_size + i02 * scale_stride;
-            void * src1_tmp_ptr  = src1_original + i11 * nb11 + i12 * nb12;
-            void * dst_tmp_ptr   = dst_original + i1 * nb1 + i2 * nb2;
+    // Prepare input and output buffers
+    ggml_cann_pool_alloc input_alloc(ctx.pool());
+    void *               input_buffer = prepare_f16_buffer(src1, input_alloc, true);

-            // mem cpy
-            ACL_CHECK(aclrtMemcpyAsync(weight_buffer, weight_stride, src0_tmp_ptr, weight_stride,
-                                       ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
-            void * scale_buffer = (char *) weight_buffer + weight_stride;
-            ACL_CHECK(aclrtMemcpyAsync(scale_buffer, scale_stride, scale_tmp_ptr, scale_stride,
-                                       ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+    ggml_cann_pool_alloc output_alloc(ctx.pool());
+    void *               output_buffer = prepare_f16_buffer(dst, output_alloc, false);

-            src0_row.data  = weight_buffer;
-            src1_row.data  = src1_tmp_ptr;
-            dst_row.data   = dst_tmp_ptr;
-            dst_row.src[0] = &src0_row;
-            dst_row.src[1] = &src1_row;
+    // Process each batch
+    for (int64_t batch_idx = 0; batch_idx < n_batches; batch_idx++) {
+        // Create index tensor for current batch
+        const size_t   index_offset  = batch_idx * ids->nb[1];
+        acl_tensor_ptr batch_indices = ggml_cann_create_tensor(ids, ids->ne, ids->nb, 1, ACL_FORMAT_ND, index_offset);

-            ggml_cann_mul_mat(ctx, &dst_row);
+        // Select quantized weights using expert indices
+        // Q4_0 stores 2 values per byte, Q8_0 stores 1 value per byte
+        const int64_t weight_d         = (type == GGML_TYPE_Q4_0) ? src0->ne[0] / 2 : src0->ne[0];
+        const int64_t weight_m         = src0->ne[1];
+        const int64_t weight_n_experts = src0->ne[2];
+
+        int64_t weight_ne[3] = { weight_d, weight_m, weight_n_experts };
+        size_t  weight_nb[3] = { sizeof(int8_t), weight_d * sizeof(int8_t), weight_d * weight_m * sizeof(int8_t) };
+
+        acl_tensor_ptr all_weights =
+            ggml_cann_create_tensor(src0->data, ACL_INT8, sizeof(int8_t), weight_ne, weight_nb, 3);
+
+        int64_t selected_weight_ne[3] = { weight_d, weight_m, n_select_experts };
+        size_t  selected_weight_nb[3] = { sizeof(int8_t), weight_d * sizeof(int8_t),
+                                          weight_d * weight_m * sizeof(int8_t) };
+
+        acl_tensor_ptr selected_weights = ggml_cann_create_tensor(selected_weight_buffer, ACL_INT8, sizeof(int8_t),
+                                                                  selected_weight_ne, selected_weight_nb, 3);
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, all_weights.get(), 0, batch_indices.get(), selected_weights.get());
+
+        // Select scales using the same expert indices
+        const int64_t scale_d     = src0->ne[0] / group_size;
+        int64_t       scale_ne[3] = { scale_d, weight_m, weight_n_experts };
+        size_t scale_nb[3] = { scale_elem_size, scale_d * scale_elem_size, scale_d * weight_m * scale_elem_size };
+
+        acl_tensor_ptr all_scales =
+            ggml_cann_create_tensor(scale_data, ACL_FLOAT16, scale_elem_size, scale_ne, scale_nb, 3);
+
+        int64_t selected_scale_ne[3] = { scale_d, weight_m, n_select_experts };
+        size_t  selected_scale_nb[3] = { scale_elem_size, scale_d * scale_elem_size,
+                                         scale_d * weight_m * scale_elem_size };
+
+        acl_tensor_ptr selected_scales = ggml_cann_create_tensor(selected_scale_buffer, ACL_FLOAT16, scale_elem_size,
+                                                                 selected_scale_ne, selected_scale_nb, 3);
+
+        GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, all_scales.get(), 0, batch_indices.get(), selected_scales.get());
+
+        // Process each expert for current batch
+        // IndexSelect output layout: [D, M, K] in contiguous format
+        // WeightQuantBatchMatmulV2 expects: [M, D] with row-major stride
+        for (int64_t expert_idx = 0; expert_idx < n_select_experts; expert_idx++) {
+            // Determine input offset: broadcast if src1->ne[1]==1, otherwise use per-expert input
+            const size_t input_offset =
+                (batch_idx * src1->ne[1] + (src1->ne[1] == 1 ? 0 : expert_idx)) * src1->ne[0] * f16_elem_size;
+            const size_t output_offset = (batch_idx * dst->ne[1] + expert_idx) * dst->ne[0] * f16_elem_size;
+
+            // Create weight view for current expert: [D, M, K] -> [M, D]
+            int64_t      weight_view_ne[2]  = { weight_m, src0->ne[0] };
+            float        weight_view_nb[2]  = { src0->ne[0] * weight_elem_size, weight_elem_size };
+            const size_t weight_view_offset = expert_idx * selected_weight_nb[2];
+
+            acl_tensor_ptr weight_view =
+                ggml_cann_create_tensor(selected_weight_buffer, ggml_cann_type_mapping(type), weight_elem_size,
+                                        weight_view_ne, weight_view_nb, 2, ACL_FORMAT_ND, weight_view_offset);
+
+            // Create scale view for current expert: [D, M, K] -> [M, D]
+            int64_t      scale_view_ne[2]  = { weight_m, scale_d };
+            size_t       scale_view_nb[2]  = { selected_scale_nb[1], selected_scale_nb[0] };
+            const size_t scale_view_offset = expert_idx * selected_scale_nb[2];
+
+            acl_tensor_ptr scale_view =
+                ggml_cann_create_tensor(selected_scale_buffer, ACL_FLOAT16, scale_elem_size, scale_view_ne,
+                                        scale_view_nb, 2, ACL_FORMAT_ND, scale_view_offset);
+
+            // Create input activation tensor [D, 1]
+            int64_t input_ne[2] = { src1->ne[0], 1 };
+            size_t  input_nb[2] = { f16_elem_size, src1->ne[0] * f16_elem_size };
+
+            acl_tensor_ptr input_tensor = ggml_cann_create_tensor(input_buffer, ACL_FLOAT16, f16_elem_size, input_ne,
+                                                                  input_nb, 2, ACL_FORMAT_ND, input_offset);
+
+            // Create output tensor [M, 1]
+            int64_t output_ne[2] = { dst->ne[0], 1 };
+            size_t  output_nb[2] = { f16_elem_size, dst->ne[0] * f16_elem_size };
+
+            acl_tensor_ptr output_tensor = ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, f16_elem_size, output_ne,
+                                                                   output_nb, 2, ACL_FORMAT_ND, output_offset);
+
+            // Perform quantized matrix multiplication
+            GGML_CANN_CALL_ACLNN_OP(ctx, WeightQuantBatchMatmulV2, input_tensor.get(), weight_view.get(),
+                                    scale_view.get(), nullptr, nullptr, nullptr, nullptr, group_size,
+                                    output_tensor.get());
        }
    }
-    return;
+
+    // Cast output back to original type if we used a temporary F16 buffer
+    if (dst->type != GGML_TYPE_F16) {
+        int64_t ne[GGML_MAX_DIMS];
+        size_t  nb[GGML_MAX_DIMS] = { f16_elem_size };
+        for (int i = 0; i < GGML_MAX_DIMS; i++) {
+            ne[i] = dst->ne[i];
+            if (i > 0) {
+                nb[i] = nb[i - 1] * ne[i - 1];
+            }
+        }
+
+        acl_tensor_ptr f16_output =
+            ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, f16_elem_size, ne, nb, GGML_MAX_DIMS);
+        acl_tensor_ptr dst_tensor = ggml_cann_create_tensor(dst);
+
+        aclnn_cast(ctx, f16_output.get(), dst_tensor.get(), ggml_cann_type_mapping(dst->type));
+    }
 }

 void ggml_cann_mul_mat_id(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
@@ -794,19 +794,44 @@ struct ggml_backend_cann_buffer_context {
    ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
 };

+// cann buffer type
 /**
- * @brief Check if a buffer is a CANN buffer.
- *
- * This function checks if a given buffer is a CANN buffer by comparing its
- * `get_name` function pointer to `ggml_backend_cann_buffer_get_name`.
- *
- * @param buffer The buffer to check.
- * @return true if the buffer is a CANN buffer, false otherwise.
+ * @brief Structure representing context information for a specific backend
+ * buffer type.
 */
-static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft);
+struct ggml_backend_cann_buffer_type_context {
+    int32_t     device; /**< Device identifier associated with the buffer context. */
+    std::string name;   /**< Name associated with the buffer context. */
+};

-static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
-    return ggml_backend_buft_is_cann(buffer->buft);
+/**
+ * @brief Retrieves the name associated with a CANN buffer type.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN buffer type context.
+ *
+ * @param buft Pointer to the buffer type context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+static const char * ggml_backend_cann_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
+
+    return buft_ctx->name.c_str();
+}
+
+/**
+ * @brief Checks if the backend buffer type is associated with the CANN backend.
+ *
+ * This function checks whether the provided backend buffer type is associated
+ * with the CANN backend based on the comparison of its name retrieval function
+ * pointer.
+ *
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the buffer type is associated with the CANN
+ * backend, otherwise false.
+ */
+static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
 }

 /**
@@ -1271,7 +1296,7 @@ static void ggml_backend_cann_buffer_get_tensor(ggml_backend_buffer_t buffer,
 static bool ggml_backend_cann_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
                                                const ggml_tensor *   src,
                                                ggml_tensor *         dst) {
-    if (ggml_backend_buffer_is_cann(src->buffer)) {
+    if (ggml_backend_buft_is_cann(src->buffer->buft)) {
        ggml_backend_cann_buffer_context * src_ctx = (ggml_backend_cann_buffer_context *) src->buffer->context;
        ggml_backend_cann_buffer_context * dst_ctx = (ggml_backend_cann_buffer_context *) buffer->context;

@@ -1335,31 +1360,6 @@ static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
    /* .reset           = */ NULL,
 };

-// cann buffer type
-/**
- * @brief Structure representing context information for a specific backend
- * buffer type.
- */
-struct ggml_backend_cann_buffer_type_context {
-    int32_t     device; /**< Device identifier associated with the buffer context. */
-    std::string name;   /**< Name associated with the buffer context. */
-};
-
-/**
- * @brief Retrieves the name associated with a CANN buffer type.
- *
- * This function returns the descriptive name associated with the specified
- * CANN buffer type context.
- *
- * @param buft Pointer to the buffer type context.
- * @return Const pointer to the C-style string containing the name.
- */
-static const char * ggml_backend_cann_buffer_type_name(ggml_backend_buffer_type_t buft) {
-    ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
-
-    return buft_ctx->name.c_str();
-}
-
 /**
 * @brief Allocates a new CANN buffer of the specified type and size.
 *
@@ -1997,7 +1997,7 @@ static bool ggml_backend_cann_cpy_tensor_async(ggml_backend_t      backend_src,

    GGML_ASSERT(!is_matmul_weight((const ggml_tensor *) src));

-    if (!ggml_backend_buffer_is_cann(src->buffer) || !ggml_backend_buffer_is_cann(dst->buffer)) {
+    if (!ggml_backend_buft_is_cann(src->buffer->buft) || !ggml_backend_buft_is_cann(dst->buffer->buft)) {
        return false;
    }

@@ -2523,21 +2523,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
    GGML_UNUSED(dev);
 }

-/**
- * @brief Checks if the backend buffer type is associated with the CANN backend.
- *
- * This function checks whether the provided backend buffer type is associated
- * with the CANN backend based on the comparison of its name retrieval function
- * pointer.
- *
- * @param buft Pointer to the backend buffer type to check.
- * @return bool Returns true if the buffer type is associated with the CANN
- * backend, otherwise false.
- */
-static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
-}
-
 /**
 * @brief Records an event on the CANN backend stream.
 *
@@ -43,6 +43,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -55,7 +56,8 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
-#    define ggml_gemm_q6_K_8x8_q8_K_generic   ggml_gemm_q6_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic   ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -76,6 +78,7 @@
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
@@ -84,6 +87,7 @@
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -107,6 +111,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -119,6 +124,7 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
@@ -143,6 +149,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -155,6 +162,7 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
@@ -186,6 +194,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -197,6 +206,7 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
@@ -227,6 +237,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -239,6 +250,7 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
@@ -271,6 +283,7 @@
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
 #define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
@@ -283,6 +296,7 @@
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
@@ -1072,6 +1072,195 @@ void ggml_gemv_q5_K_8x8_q8_K(int                        n,
    ggml_gemv_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }

+void ggml_gemv_q6_K_8x4_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_groups = ncols_interleaved / 4;
+    const uint8x16_t m4b        = vdupq_n_u8(0x0f);
+    const uint8x16_t mask_lo    = vdupq_n_u8(0x03);
+    const uint8x16_t mask_hi    = vdupq_n_u8(0x30);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[2];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q6_Kx8 * GGML_RESTRICT q6_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < col_groups; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q6_d_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q6_d_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d       = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0 = vmulq_f32(q6_d_0, q8_d);
+            float32x4_t sb_scale_1 = vmulq_f32(q6_d_1, q8_d);
+
+            int32x4_t acc[col_groups];
+            for (int i = 0; i < col_groups; i++) {
+                acc[i] = vdupq_n_s32(0);
+            }
+
+            // Load all 16 scales once and widen to int16 (Q6_K has 16 scales per block)
+            // Reused for bias and dequantization later
+            int16_t q6_scales[16 * 8];
+            for (int i = 0; i < 16; i++) {
+                int16x8_t scales = vmovl_s8(vld1_s8(q6_ptr[b].scales + i * 8));
+                vst1q_s16(q6_scales + i * 8, scales);
+            }
+
+            // Compute bias per column using q8 bsums and preloaded scales to skip the -32 shift
+            int32x4_t bias_lo = vdupq_n_s32(0);
+            int32x4_t bias_hi = vdupq_n_s32(0);
+
+            // Load bsums in chunks of 4 to process with vectorized operations
+            for (int i = 0; i < 16; i += 4) {
+                int16x4_t bsums_vec   = vld1_s16(q8_ptr[b].bsums + i);
+                int16x4_t scales_lo_0 = vld1_s16(q6_scales + (i + 0) * 8);
+                int16x4_t scales_hi_0 = vld1_s16(q6_scales + (i + 0) * 8 + 4);
+                int16x4_t scales_lo_1 = vld1_s16(q6_scales + (i + 1) * 8);
+                int16x4_t scales_hi_1 = vld1_s16(q6_scales + (i + 1) * 8 + 4);
+                int16x4_t scales_lo_2 = vld1_s16(q6_scales + (i + 2) * 8);
+                int16x4_t scales_hi_2 = vld1_s16(q6_scales + (i + 2) * 8 + 4);
+                int16x4_t scales_lo_3 = vld1_s16(q6_scales + (i + 3) * 8);
+                int16x4_t scales_hi_3 = vld1_s16(q6_scales + (i + 3) * 8 + 4);
+
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_0, bsums_vec, 0);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_0, bsums_vec, 0);
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_1, bsums_vec, 1);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_1, bsums_vec, 1);
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_2, bsums_vec, 2);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_2, bsums_vec, 2);
+                bias_lo = vmlal_lane_s16(bias_lo, scales_lo_3, bsums_vec, 3);
+                bias_hi = vmlal_lane_s16(bias_hi, scales_hi_3, bsums_vec, 3);
+            }
+            bias_lo = vshlq_n_s32(bias_lo, 5);
+            bias_hi = vshlq_n_s32(bias_hi, 5);
+
+            // Process two 128-value halves per superblock
+            for (int half = 0; half < 2; half++) {
+                const uint8_t * ql_base = q6_ptr[b].ql + half * 512;
+                const uint8_t * qh_base = q6_ptr[b].qh + half * 256;
+
+                // A subblock (sb) is a set of weights that share the scale
+                // Since q6_K scales are per 16 elements
+                // num sbs -> 256 elements / (16 elements/scale * 2 elements/byte * 2 halves)
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    const int8_t * q8_base_l = q8_ptr[b].qs + half * 128 + sb * 16;
+                    const int8_t * q8_base_h = q8_base_l + 64;
+
+                    // Load and duplicate q8 values (each register covers four interleaved columns of q6)
+                    int8x16_t q8_l[4];
+                    int8x16_t q8_h[4];
+                    for (int i = 0; i < 4; i++) {
+                        q8_l[i] = (int8x16_t) vld1q_dup_s32((const int32_t *) (q8_base_l + i * 4));
+                        q8_h[i] = (int8x16_t) vld1q_dup_s32((const int32_t *) (q8_base_h + i * 4));
+                    }
+
+                    const int ql_off_base = sb * QK_K / 2;
+                    const int qh_off_base = ql_off_base & 255;  // wraps after 256 bytes
+
+                    // Load 4 vectors at once (64 bytes each for ql_0, ql_1, qh_0, qh_1)
+                    uint8x16x4_t q6_ql_0 = vld1q_u8_x4(ql_base + ql_off_base);
+                    uint8x16x4_t q6_ql_1 = vld1q_u8_x4(ql_base + ql_off_base + 64);
+                    uint8x16x4_t q6_qh_0 = vld1q_u8_x4(qh_base + qh_off_base);
+                    uint8x16x4_t q6_qh_1 = vld1q_u8_x4(qh_base + qh_off_base + 64);
+
+                    // Adjust qh for subblocks 2 and 3 (shift right by 2)
+                    if (sb > 1) {
+                        q6_qh_0.val[0] = vshrq_n_u8(q6_qh_0.val[0], 2);
+                        q6_qh_0.val[1] = vshrq_n_u8(q6_qh_0.val[1], 2);
+                        q6_qh_0.val[2] = vshrq_n_u8(q6_qh_0.val[2], 2);
+                        q6_qh_0.val[3] = vshrq_n_u8(q6_qh_0.val[3], 2);
+                        q6_qh_1.val[0] = vshrq_n_u8(q6_qh_1.val[0], 2);
+                        q6_qh_1.val[1] = vshrq_n_u8(q6_qh_1.val[1], 2);
+                        q6_qh_1.val[2] = vshrq_n_u8(q6_qh_1.val[2], 2);
+                        q6_qh_1.val[3] = vshrq_n_u8(q6_qh_1.val[3], 2);
+                    }
+
+                    const uint8x16_t q6_ql[8] = { q6_ql_0.val[0], q6_ql_0.val[1], q6_ql_0.val[2], q6_ql_0.val[3],
+                                                  q6_ql_1.val[0], q6_ql_1.val[1], q6_ql_1.val[2], q6_ql_1.val[3] };
+                    const uint8x16_t q6_qh[8] = { q6_qh_0.val[0], q6_qh_0.val[1], q6_qh_0.val[2], q6_qh_0.val[3],
+                                                  q6_qh_1.val[0], q6_qh_1.val[1], q6_qh_1.val[2], q6_qh_1.val[3] };
+
+                    // Process column groups (0-3, 4-7)
+                    for (int g = 0; g < col_groups; g++) {
+                        int32x4_t sb_acc_l = vdupq_n_s32(0);
+                        int32x4_t sb_acc_h = vdupq_n_s32(0);
+
+                        for (int chunk = 0; chunk < 4; chunk++) {
+                            const int idx = chunk * 2 + g;
+
+                            const uint8x16_t q6_qs_l = q6_ql[idx];
+                            const uint8x16_t q6_qs_h = q6_qh[idx];
+
+                            // Extract high 2 bits for upper nibble reconstruction
+                            const uint8x16_t q6_qs_hh = vandq_u8(q6_qs_h, mask_hi);
+
+                            // q6 = (low4 | high2<<4), without -32 bias (handled via bsums)
+                            const int8x16_t q6_l =
+                                vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_qs_l, m4b), vandq_u8(q6_qs_h, mask_lo), 4));
+                            const int8x16_t q6_h = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_qs_l, 4), q6_qs_hh));
+
+                            sb_acc_l = vdotq_s32(sb_acc_l, q6_l, q8_l[chunk]);
+                            sb_acc_h = vdotq_s32(sb_acc_h, q6_h, q8_h[chunk]);
+                        }
+
+                        const int scale_idx_l = half * 8 + sb;
+                        const int scale_idx_h = half * 8 + sb + 4;
+
+                        const int32x4_t scale_vec_l = vmovl_s16(vld1_s16(q6_scales + scale_idx_l * 8 + g * 4));
+                        const int32x4_t scale_vec_h = vmovl_s16(vld1_s16(q6_scales + scale_idx_h * 8 + g * 4));
+
+                        acc[g] = vmlaq_s32(acc[g], sb_acc_l, scale_vec_l);
+                        acc[g] = vmlaq_s32(acc[g], sb_acc_h, scale_vec_h);
+                    }
+                }
+            }  // for half
+
+            // Bias correction
+            acc[0] = vsubq_s32(acc[0], bias_lo);
+            acc[1] = vsubq_s32(acc[1], bias_hi);
+
+            // Apply superblock scale (no mins for q6_K)
+            // acc[g] has [c0, c1, c2, c3]
+            float32x4_t w_0123 = vmulq_f32(vcvtq_f32_s32(acc[0]), sb_scale_0);
+            float32x4_t w_4567 = vmulq_f32(vcvtq_f32_s32(acc[1]), sb_scale_1);
+
+            acc_f32[0] = vaddq_f32(acc_f32[0], w_0123);
+            acc_f32[1] = vaddq_f32(acc_f32[1], w_4567);
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q6_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemv_q6_K_8x8_q8_K(int                        n,
                             float * GGML_RESTRICT      s,
                             size_t                     bs,
@@ -1177,15 +1366,14 @@ void ggml_gemv_q6_K_8x8_q8_K(int                        n,
                        q8_h[i] = (int8x16_t) vld1q_dup_s64((const int64_t *) (q8_base_h + i * 8));
                    }

-                    // TODO: Test other qh repack patterns to reduce loads
                    const int ql_off_base = sb * QK_K / 2;
                    const int qh_off_base = ql_off_base & 255;  // wraps after 256 bytes

                    // Load 4 vectors at once (64 bytes each for ql_0, ql_1, qh_0, qh_1)
-                    ggml_uint8x16x4_t q6_ql_0 = ggml_vld1q_u8_x4(ql_base + ql_off_base);
-                    ggml_uint8x16x4_t q6_ql_1 = ggml_vld1q_u8_x4(ql_base + ql_off_base + 64);
-                    ggml_uint8x16x4_t q6_qh_0 = ggml_vld1q_u8_x4(qh_base + qh_off_base);
-                    ggml_uint8x16x4_t q6_qh_1 = ggml_vld1q_u8_x4(qh_base + qh_off_base + 64);
+                    uint8x16x4_t q6_ql_0 = vld1q_u8_x4(ql_base + ql_off_base);
+                    uint8x16x4_t q6_ql_1 = vld1q_u8_x4(ql_base + ql_off_base + 64);
+                    uint8x16x4_t q6_qh_0 = vld1q_u8_x4(qh_base + qh_off_base);
+                    uint8x16x4_t q6_qh_1 = vld1q_u8_x4(qh_base + qh_off_base + 64);

                    // Adjust qh for subblocks 2 and 3 (shift right by 2)
                    if (sb > 1) {
@@ -3474,6 +3662,208 @@ void ggml_gemm_q5_K_8x8_q8_K(int                        n,
    ggml_gemm_q5_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }

+void ggml_gemm_q6_K_8x4_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    q8_k_blocklen = 4;
+    constexpr int    col_groups    = ncols_interleaved / 4;
+    constexpr int    acc_size      = q8_k_blocklen * col_groups;  // 4 rows, 2 column groups
+    const uint8x16_t m4b           = vdupq_n_u8(0x0f);
+    const uint8x16_t mask_lo       = vdupq_n_u8(0x03);
+    const uint8x16_t mask_hi       = vdupq_n_u8(0x30);
+    const int8x16_t  m32s          = vdupq_n_s8(32);
+
+    float32x4_t acc_f32[acc_size];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q6_Kx8 * GGML_RESTRICT q6_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < acc_size; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                float32x4_t q6_d_0123 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d));
+                float32x4_t q6_d_4567 = vcvt_f32_f16(vld1_f16((const __fp16 *) q6_ptr[b].d + 4));
+                float32x4_t q8_d_0123 = vld1q_f32(q8_ptr[b].d);
+
+                float32x4_t sbd_scale_0123[q8_k_blocklen];
+                float32x4_t sbd_scale_4567[q8_k_blocklen];
+
+                sbd_scale_0123[0] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 0);
+                sbd_scale_4567[0] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 0);
+                sbd_scale_0123[1] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 1);
+                sbd_scale_4567[1] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 1);
+                sbd_scale_0123[2] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 2);
+                sbd_scale_4567[2] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 2);
+                sbd_scale_0123[3] = vmulq_laneq_f32(q6_d_0123, q8_d_0123, 3);
+                sbd_scale_4567[3] = vmulq_laneq_f32(q6_d_4567, q8_d_0123, 3);
+
+                int32x4_t acc_s32[acc_size];
+                for (int i = 0; i < acc_size; i++) {
+                    acc_s32[i] = vdupq_n_s32(0);
+                }
+
+                int16_t q6_scales[8 * 16];
+                for (int i = 0; i < 16; i++) {
+                    int16x8_t scales = vmovl_s8(vld1_s8(q6_ptr[b].scales + i * 8));
+                    vst1q_s16(q6_scales + i * 8, scales);
+                }
+
+                for (int half = 0; half < 2; half++) {
+                    const uint8_t * ql_base = q6_ptr[b].ql + half * 512;
+                    const uint8_t * qh_base = q6_ptr[b].qh + half * 256;
+
+                    for (int sb = 0; sb < QK_K / 64; sb++) {
+                        int32x4_t acc_lo[acc_size];
+                        int32x4_t acc_hi[acc_size];
+                        for (int i = 0; i < acc_size; i++) {
+                            acc_lo[i] = vdupq_n_s32(0);
+                            acc_hi[i] = vdupq_n_s32(0);
+                        }
+
+                        const int8_t * q8_base_l = q8_ptr[b].qs + half * 512 + sb * 64;
+                        const int8_t * q8_base_h = q8_ptr[b].qs + half * 512 + 256 + sb * 64;
+
+                        // 4 rows * 16 elements per scale
+                        // 4 reads of 16 bytes each
+                        constexpr int reads_per_sb = 4;
+                        int8x16_t     q8_l[reads_per_sb];
+                        int8x16_t     q8_h[reads_per_sb];
+                        for (int k = 0; k < reads_per_sb; k++) {
+                            q8_l[k] = vld1q_s8(q8_base_l + 16 * k);
+                            q8_h[k] = vld1q_s8(q8_base_h + 16 * k);
+                        }
+
+                        const int ql_off_base = sb * QK_K / 2;
+                        const int qh_off_base = ql_off_base & 255;
+
+                        uint8x16_t q6_ql_0123[reads_per_sb];
+                        uint8x16_t q6_ql_4567[reads_per_sb];
+                        uint8x16_t q6_qh_0123[reads_per_sb];
+                        uint8x16_t q6_qh_4567[reads_per_sb];
+
+                        for (int k = 0; k < reads_per_sb; k++) {
+                            q6_ql_0123[k] = vld1q_u8(ql_base + ql_off_base + k * 32);
+                            q6_ql_4567[k] = vld1q_u8(ql_base + ql_off_base + k * 32 + 16);
+                            q6_qh_0123[k] = vld1q_u8(qh_base + qh_off_base + k * 32);
+                            q6_qh_4567[k] = vld1q_u8(qh_base + qh_off_base + k * 32 + 16);
+                        }
+
+                        if (sb > 1) {
+                            for (int k = 0; k < reads_per_sb; k++) {
+                                q6_qh_0123[k] = vshrq_n_u8(q6_qh_0123[k], 2);
+                                q6_qh_4567[k] = vshrq_n_u8(q6_qh_4567[k], 2);
+                            }
+                        }
+
+                        for (int k = 0; k < reads_per_sb; k++) {
+                            // q = (ql | qh) - 32
+                            const uint8x16_t hbit_lo_0123 = vandq_u8(q6_qh_0123[k], mask_lo);
+                            const uint8x16_t hbit_hi_0123 = vandq_u8(q6_qh_0123[k], mask_hi);
+                            const uint8x16_t hbit_lo_4567 = vandq_u8(q6_qh_4567[k], mask_lo);
+                            const uint8x16_t hbit_hi_4567 = vandq_u8(q6_qh_4567[k], mask_hi);
+
+                            const int8x16_t q6_0123_lo = vsubq_s8(
+                                vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_0123[k], m4b), hbit_lo_0123, 4)), m32s);
+                            const int8x16_t q6_0123_hi = vsubq_s8(
+                                vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_0123[k], 4), hbit_hi_0123)), m32s);
+
+                            acc_lo[0] = vdotq_laneq_s32(acc_lo[0], q6_0123_lo, q8_l[k], 0);  //  0..3  r0 c0123
+                            acc_lo[1] = vdotq_laneq_s32(acc_lo[1], q6_0123_lo, q8_l[k], 1);  //  0..3  r1 c0123
+                            acc_lo[2] = vdotq_laneq_s32(acc_lo[2], q6_0123_lo, q8_l[k], 2);  //  0..3  r2 c0123
+                            acc_lo[3] = vdotq_laneq_s32(acc_lo[3], q6_0123_lo, q8_l[k], 3);  //  0..3  r3 c0123
+
+                            acc_hi[0] = vdotq_laneq_s32(acc_hi[0], q6_0123_hi, q8_h[k], 0);  // 64..67 r0 c0123
+                            acc_hi[1] = vdotq_laneq_s32(acc_hi[1], q6_0123_hi, q8_h[k], 1);  // 64..67 r1 c0123
+                            acc_hi[2] = vdotq_laneq_s32(acc_hi[2], q6_0123_hi, q8_h[k], 2);  // 64..67 r2 c0123
+                            acc_hi[3] = vdotq_laneq_s32(acc_hi[3], q6_0123_hi, q8_h[k], 3);  // 64..67 r3 c0123
+
+                            const int8x16_t q6_4567_lo = vsubq_s8(
+                                vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q6_ql_4567[k], m4b), hbit_lo_4567, 4)), m32s);
+                            const int8x16_t q6_4567_hi = vsubq_s8(
+                                vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6_ql_4567[k], 4), hbit_hi_4567)), m32s);
+
+                            acc_lo[4] = vdotq_laneq_s32(acc_lo[4], q6_4567_lo, q8_l[k], 0);  //  0..3  r0 c4567
+                            acc_lo[5] = vdotq_laneq_s32(acc_lo[5], q6_4567_lo, q8_l[k], 1);  //  0..3  r1 c4567
+                            acc_lo[6] = vdotq_laneq_s32(acc_lo[6], q6_4567_lo, q8_l[k], 2);  //  0..3  r2 c4567
+                            acc_lo[7] = vdotq_laneq_s32(acc_lo[7], q6_4567_lo, q8_l[k], 3);  //  0..3  r3 c4567
+
+                            acc_hi[4] = vdotq_laneq_s32(acc_hi[4], q6_4567_hi, q8_h[k], 0);  // 64..67 r0 c4567
+                            acc_hi[5] = vdotq_laneq_s32(acc_hi[5], q6_4567_hi, q8_h[k], 1);  // 64..67 r1 c4567
+                            acc_hi[6] = vdotq_laneq_s32(acc_hi[6], q6_4567_hi, q8_h[k], 2);  // 64..67 r2 c4567
+                            acc_hi[7] = vdotq_laneq_s32(acc_hi[7], q6_4567_hi, q8_h[k], 3);  // 64..67 r3 c4567
+                        }
+
+                        // Scale and bias
+                        const int scale_idx_l = half * 8 + sb;
+                        const int scale_idx_h = half * 8 + sb + 4;
+
+                        for (int g = 0; g < col_groups; g++) {
+                            const int16x4_t scales_l16  = vld1_s16(q6_scales + scale_idx_l * 8 + g * 4);
+                            const int16x4_t scales_h16  = vld1_s16(q6_scales + scale_idx_h * 8 + g * 4);
+                            const int32x4_t scale_vec_l = vmovl_s16(scales_l16);
+                            const int32x4_t scale_vec_h = vmovl_s16(scales_h16);
+                            const int       acc_offset  = g * q8_k_blocklen;
+
+                            for (int row = 0; row < q8_k_blocklen; row++) {
+                                const int idx = row * 2 + g;
+                                acc_s32[idx]  = vmlaq_s32(acc_s32[idx], acc_lo[acc_offset + row], scale_vec_l);
+                                acc_s32[idx]  = vmlaq_s32(acc_s32[idx], acc_hi[acc_offset + row], scale_vec_h);
+                            }
+                        }
+                    }
+                }
+
+                // Finally we apply the superblock scales
+                for (int row = 0; row < q8_k_blocklen; row++) {
+                    const int       idx0     = 2 * row;
+                    const int       idx1     = 2 * row + 1;
+                    const int32x4_t acc_0123 = acc_s32[idx0];
+                    const int32x4_t acc_4567 = acc_s32[idx1];
+
+                    acc_f32[idx0] = vmlaq_f32(acc_f32[idx0], vcvtq_f32_s32(acc_0123), sbd_scale_0123[row]);
+                    acc_f32[idx1] = vmlaq_f32(acc_f32[idx1], vcvtq_f32_s32(acc_4567), sbd_scale_4567[row]);
+                }
+            }  // for b
+
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemm_q6_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemm_q6_K_8x8_q8_K(int                        n,
                             float * GGML_RESTRICT      s,
                             size_t                     bs,
@@ -268,9 +268,9 @@ static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const
                           _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
 }

-static inline __m256 quad_mx_delta_float(const int8_t x0, const float y0, const int8_t x1, const float y1) {
-    return _mm256_set_m128(_mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
-                           _mm_set1_ps(GGML_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
+static inline __m256 quad_mx_delta_float(const uint8_t x0, const float y0, const uint8_t x1, const float y1) {
+    return _mm256_set_m128(_mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
+                           _mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
 }
 #endif
 #elif defined(__SSSE3__)
@@ -782,6 +782,7 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo

    __m256 accum1 = _mm256_setzero_ps();
    __m256 accum2 = _mm256_setzero_ps();
+
    for (; ib + 1 < nb; ib += 2) {
        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
@@ -795,10 +796,10 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
        const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
        const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
-        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 0].e)),
-                _mm256_cvtepi32_ps(p_1), accum1);
-        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_E8M0_TO_FP32_HALF(x[ib + 1].e)),
-                _mm256_cvtepi32_ps(p_2), accum2);
+        const __m256 scale0 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 0].e));
+        const __m256 scale1 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 1].e));
+        accum1 = _mm256_fmadd_ps(scale0, _mm256_cvtepi32_ps(p_1), accum1);
+        accum2 = _mm256_fmadd_ps(scale1, _mm256_cvtepi32_ps(p_2), accum2);
    }

    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
@@ -830,7 +831,7 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo

 #endif
    for (; ib < nb; ++ib) {
-        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib].e);
        int sumi1 = 0;
        int sumi2 = 0;
        for (int j = 0; j < QK_MXFP4/2; ++j) {
@@ -3817,4 +3818,3 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
    ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
 }
-
@@ -59,11 +59,7 @@ static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * ds
    GGML_ASSERT(nb00 == sizeof(src0_t));

    const auto [ir0, ir1] = get_thread_range(params, src0);
-    const bool is_src1_contiguous = (nb10 == sizeof(src1_t));
-
-    if (!is_src1_contiguous) { // broadcast not implemented yet for non-contiguous
-        GGML_ASSERT(ggml_are_same_shape(src0, src1));
-    }
+    const bool is_src1_contiguous_rows = ggml_is_contiguous_rows(src1);

 #ifdef GGML_USE_ACCELERATE
    vDSP_fn_t vDSP_op = nullptr;
@@ -94,7 +90,7 @@ static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * ds
        const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
        const src1_t * src1_ptr = (const src1_t *) ((const char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);

-        if (is_src1_contiguous) {
+        if (is_src1_contiguous_rows) {
            // src1 is broadcastable across src0 and dst in i1, i2, i3
            const int64_t nr0 = ne00 / ne10;

@@ -24,6 +24,9 @@ struct ggml_compute_params {
    void * wdata;

    struct ggml_threadpool * threadpool;
+
+    // use reference implementation
+    bool use_ref;
 };


@@ -5,7 +5,6 @@
 #include "ggml-backend.h"
 #include "traits.h"
 #include "ggml-cpu-impl.h"
-#include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "quants.h"
 #include "ggml-threading.h"
@@ -76,6 +75,9 @@
 // precomputed f32 table for f16 (256 KB) (simd-mappings.h)
 float ggml_table_f32_f16[1 << 16];

+// precomputed f32 table for e8m0 half (1 KB) (simd-mappings.h)
+float ggml_table_f32_e8m0_half[1 << 8];
+
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
    int sve_cnt;
@@ -2867,12 +2869,20 @@ struct ggml_cplan ggml_graph_plan(
                    } break;
                case GGML_OP_FLASH_ATTN_EXT:
                    {
+                        const int64_t neq2 = node->src[0]->ne[2]; // number of query heads
                        const int64_t DK = node->src[1]->ne[0];
                        const int64_t DV = node->src[2]->ne[0];

                        // Tiled flash attention scratch (tile sizes defined in common.h)
                        // Per-thread: Q_q + KQ + mask + VKQ32 + V32 + padding
-                        cur = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV)*n_tasks;
+                        size_t prefill  = sizeof(float)*(GGML_FA_TILE_Q*DK + 2*GGML_FA_TILE_Q*GGML_FA_TILE_KV + GGML_FA_TILE_Q*DV + GGML_FA_TILE_KV*DV)*n_tasks;
+
+                        // Decode path: n_kv_chunks = n_tasks (one chunk per thread)
+                        // Per-thread: VKQ accmulator (DV), partial M, partial S + intra-thread scratch for V, Q and VKQ
+                        size_t n_chunks = n_tasks;
+                        size_t decode   = sizeof(float)*(neq2*n_chunks*(2+DV) + n_tasks*(DK + 2*DV));
+
+                        cur += MAX(prefill, decode);
                    } break;
                case GGML_OP_FLASH_ATTN_BACK:
                    {
@@ -2929,11 +2939,12 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
    set_numa_thread_affinity(state->ith);

    struct ggml_compute_params params = {
-        /*.ith       =*/ state->ith,
-        /*.nth       =*/ atomic_load_explicit(&tp->n_graph, memory_order_relaxed) & GGML_THREADPOOL_N_THREADS_MASK,
-        /*.wsize     =*/ cplan->work_size,
-        /*.wdata     =*/ cplan->work_data,
-        /*.threadpool=*/ tp,
+        /*.ith        =*/ state->ith,
+        /*.nth        =*/ atomic_load_explicit(&tp->n_graph, memory_order_relaxed) & GGML_THREADPOOL_N_THREADS_MASK,
+        /*.wsize      =*/ cplan->work_size,
+        /*.wdata      =*/ cplan->work_data,
+        /*.threadpool =*/ tp,
+        /*.use_ref    =*/ cplan->use_ref,
    };

    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d \n", state->ith, cplan, state->last_graph);
@@ -3673,6 +3684,11 @@ void ggml_cpu_init(void) {
                ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
            }

+            // initialize E8M0 half table (256 entries)
+            for (int i = 0; i < (1 << 8); ++i) {
+                ggml_table_f32_e8m0_half[i] = GGML_E8M0_TO_FP32_HALF(i);
+            }
+
            const uint64_t t_end = ggml_time_us(); UNUSED(t_end);

            GGML_PRINT_DEBUG("%s: GELU, Quick GELU, SILU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0);
@@ -105,6 +105,8 @@ struct ggml_backend_cpu_context {

    ggml_abort_callback abort_callback;
    void *              abort_callback_data;
+
+    bool                use_ref;  // use reference implementation
 };

 static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
@@ -143,6 +145,7 @@ static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend

    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+    cpu_plan->cplan.use_ref             = cpu_ctx->use_ref;

    return cpu_plan;
 }
@@ -182,6 +185,7 @@ static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, s

    cplan.abort_callback      = cpu_ctx->abort_callback;
    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+    cplan.use_ref             = cpu_ctx->use_ref;

    return ggml_graph_compute(cgraph, &cplan);
 }
@@ -223,6 +227,7 @@ ggml_backend_t ggml_backend_cpu_init(void) {
    ctx->work_size           = 0;
    ctx->abort_callback      = NULL;
    ctx->abort_callback_data = NULL;
+    ctx->use_ref             = false;

    ggml_backend_t cpu_backend = new ggml_backend {
        /* .guid    = */ ggml_backend_cpu_guid(),
@@ -270,6 +275,13 @@ void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_
    ctx->abort_callback_data = abort_callback_data;
 }

+void ggml_backend_cpu_set_use_ref(ggml_backend_t backend_cpu, bool use_ref) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->use_ref = use_ref;
+}
+
 // CPU backend - device

 struct ggml_backend_cpu_device_context {
@@ -646,6 +658,9 @@ static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const ch
    if (strcmp(name, "ggml_backend_cpu_is_numa") == 0) {
        return (void *)ggml_is_numa;
    }
+    if (strcmp(name, "ggml_backend_cpu_set_use_ref") == 0) {
+        return (void *)ggml_backend_cpu_set_use_ref;
+    }

    // threadpool - TODO:  move to ggml-base
    if (strcmp(name, "ggml_threadpool_new") == 0) {
@@ -2096,10 +2096,14 @@ static void ggml_compute_forward_gelu_f32(

    const ggml_tensor * src0 = dst->src[0];

-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
    assert(ggml_are_same_shape(src0, dst));

+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    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;

@@ -2113,10 +2117,14 @@ static void ggml_compute_forward_gelu_f32(
    const int ir0 = dr*ith;
    const int ir1 = MIN(ir0 + dr, nr);

-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
        ggml_vec_gelu_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));

 #ifndef NDEBUG
        for (int k = 0; k < nc; k++) {
@@ -2135,10 +2143,14 @@ static void ggml_compute_forward_gelu_f16(

    const ggml_tensor * src0 = dst->src[0];

-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
    assert(ggml_are_same_shape(src0, dst));

+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    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;

@@ -2152,10 +2164,14 @@ static void ggml_compute_forward_gelu_f16(
    const int ir0 = dr*ith;
    const int ir1 = MIN(ir0 + dr, nr);

-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
        ggml_vec_gelu_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));

 #ifndef NDEBUG
        for (int k = 0; k < nc; k++) {
@@ -2276,10 +2292,14 @@ static void ggml_compute_forward_gelu_erf_f32(

    const ggml_tensor * src0 = dst->src[0];

-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
    assert(ggml_are_same_shape(src0, dst));

+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    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;

@@ -2293,10 +2313,14 @@ static void ggml_compute_forward_gelu_erf_f32(
    const int ir0 = dr*ith;
    const int ir1 = MIN(ir0 + dr, nr);

-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
        ggml_vec_gelu_erf_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));

 #ifndef NDEBUG
        for (int k = 0; k < nc; k++) {
@@ -2315,10 +2339,14 @@ static void ggml_compute_forward_gelu_erf_f16(

    const ggml_tensor * src0 = dst->src[0];

-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
    assert(ggml_are_same_shape(src0, dst));

+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    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;

@@ -2332,10 +2360,14 @@ static void ggml_compute_forward_gelu_erf_f16(
    const int ir0 = dr*ith;
    const int ir1 = MIN(ir0 + dr, nr);

-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
        ggml_vec_gelu_erf_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));

 #ifndef NDEBUG
        for (int k = 0; k < nc; k++) {
@@ -2379,10 +2411,14 @@ static void ggml_compute_forward_gelu_quick_f32(

    const ggml_tensor * src0 = dst->src[0];

-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
    assert(ggml_are_same_shape(src0, dst));

+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    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;

@@ -2396,10 +2432,14 @@ static void ggml_compute_forward_gelu_quick_f32(
    const int ir0 = dr*ith;
    const int ir1 = MIN(ir0 + dr, nr);

-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
        ggml_vec_gelu_quick_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));

 #ifndef NDEBUG
        for (int k = 0; k < nc; k++) {
@@ -2418,10 +2458,14 @@ static void ggml_compute_forward_gelu_quick_f16(

    const ggml_tensor * src0 = dst->src[0];

-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
    assert(ggml_are_same_shape(src0, dst));

+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    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;

@@ -2435,10 +2479,14 @@ static void ggml_compute_forward_gelu_quick_f16(
    const int ir0 = dr*ith;
    const int ir1 = MIN(ir0 + dr, nr);

-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
        ggml_vec_gelu_quick_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));

 #ifndef NDEBUG
        for (int k = 0; k < nc; k++) {
@@ -2482,10 +2530,14 @@ static void ggml_compute_forward_silu_f32(

    const ggml_tensor * src0 = dst->src[0];

-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
    assert(ggml_are_same_shape(src0, dst));

+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    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;

@@ -2499,10 +2551,14 @@ static void ggml_compute_forward_silu_f32(
    const int ir0 = dr*ith;
    const int ir1 = MIN(ir0 + dr, nr);

-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
        ggml_vec_silu_f32(nc,
-                (float *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (float *) ((char *) src0->data + i1*(src0->nb[1])));
+                (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));

 #ifndef NDEBUG
        for (int k = 0; k < nc; k++) {
@@ -2521,10 +2577,14 @@ static void ggml_compute_forward_silu_f16(

    const ggml_tensor * src0 = dst->src[0];

-    assert(ggml_is_contiguous_1(src0));
-    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_is_contiguous_rows(src0));
    assert(ggml_are_same_shape(src0, dst));

+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+    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;

@@ -2538,10 +2598,14 @@ static void ggml_compute_forward_silu_f16(
    const int ir0 = dr*ith;
    const int ir1 = MIN(ir0 + dr, nr);

-    for (int i1 = ir0; i1 < ir1; i1++) {
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int i3 = ir/(ne02*ne01);
+        const int i2 = (ir - i3*ne02*ne01)/ne01;
+        const int i1 = (ir - i3*ne02*ne01 - i2*ne01);
+
        ggml_vec_silu_f16(nc,
-                (ggml_fp16_t *) ((char *) dst->data  + i1*( dst->nb[1])),
-                (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1])));
+                (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1),
+                (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));

 #ifndef NDEBUG
        for (int k = 0; k < nc; k++) {
@@ -7629,8 +7693,7 @@ static void ggml_compute_forward_pad_f32(

    const ggml_tensor * src0 = dst->src[0];

-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-    GGML_ASSERT( dst->nb[0] == sizeof(float));
+    assert(dst->nb[0] == sizeof(float));

    const int ith = params->ith;
    const int nth = params->nth;
@@ -8042,12 +8105,14 @@ void ggml_compute_forward_top_k(
    }
 }

-// ggml_compute_forward_flash_attn_ext
-
 static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
        const ggml_compute_params * params,
        ggml_tensor * dst,
-        int ir0, int ir1) {
+        int ir0, int ir1,
+        int64_t ic_start, int64_t ic_end,
+        float * partials, int64_t partial_stride) {
+
+    const bool write_partials = (partials != nullptr);
    const ggml_tensor * q     = dst->src[0];
    const ggml_tensor * k     = dst->src[1];
    const ggml_tensor * v     = dst->src[2];
@@ -8124,7 +8189,6 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(

    int ith = params->ith;

-    // loop over n_batch and n_head
    for (int ir = ir0; ir < ir1; ++ir) {
        // q indices
        const int iq3 = ir/(neq2*neq1);
@@ -8165,7 +8229,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
        // loop over n_kv and n_head_kv
        // ref: https://arxiv.org/pdf/2112.05682.pdf

-        for (int64_t ic = 0; ic < nek1; ++ic) {
+        for (int64_t ic = ic_start; ic < ic_end; ++ic) {
            const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f;
            if (mv == -INFINITY) {
                continue;
@@ -8238,8 +8302,8 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
            }
        }

-        // sinks
-        if (sinks) {
+        // sinks - apply only on the first kv-chunk
+        if (sinks && ic_start == 0) {
            const float s = ((float *)((char *) sinks->data))[h];

            float ms = 1.0f;
@@ -8247,6 +8311,7 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(

            if (s > M) {
                ms = expf(M - s);
+                M = s;
                ggml_vec_scale_f32(DV, VKQ32, ms);
            } else {
                vs = expf(s - M);
@@ -8255,20 +8320,26 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(
            S = S*ms + vs;
        }

-        // V /= S
-        const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
-        ggml_vec_scale_f32(DV, VKQ32, S_inv);
+        if (write_partials) {
+            // Write M, S, VKQ to partials for later reduction
+            // partials layout: [M, S, VKQ[DV]] per query head
+            float * partial = partials + ir * partial_stride;
+            partial[0] = M;
+            partial[1] = S;
+            memcpy(partial + 2, VKQ32, DV * sizeof(float));
+        } else {
+            // V /= S
+            const float S_inv = S == 0.0f ? 0.0f : 1.0f/S;
+            ggml_vec_scale_f32(DV, VKQ32, S_inv);

-        // dst indices
-        const int i1 = iq1;
-        const int i2 = iq2;
-        const int i3 = iq3;
+            // dst indices
+            const int i1 = iq1;
+            const int i2 = iq2;
+            const int i3 = iq3;

-        // original
-        //memcpy((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3), V, nev0*sizeof(float));
-
-        // permute(0, 2, 1, 3)
-        memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1);
+            // permute(0, 2, 1, 3)
+            memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1);
+        }
    }
 }

@@ -8546,6 +8617,78 @@ static void ggml_compute_forward_flash_attn_ext_tiled(
    }
 }

+// Reduction function: combines partial results across KV chunks
+// Partials layout in wdata: [n_q_heads][n_chunks][2 + DV]
+static void ggml_flash_attn_ext_reduce_partials(
+        const ggml_compute_params * params,
+        ggml_tensor * dst,
+        const int64_t n_chunks,
+        const int64_t chunk_size) {
+
+    const ggml_tensor * q = dst->src[0];
+    const ggml_tensor * k = dst->src[1];
+    const ggml_tensor * v = dst->src[2];
+
+    const int64_t DK        = k->ne[0];
+    const int64_t DV        = v->ne[0];
+    const int64_t nek1      = k->ne[1];
+    const int64_t n_q_heads = q->ne[2];
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t wdata_per_thread = DK + 2*DV + CACHE_LINE_SIZE_F32;
+    float *       thread_wdata     = (float *) params->wdata + ith * wdata_per_thread;
+
+    const int64_t partials_offset  = nth * (DK + 2*DV + CACHE_LINE_SIZE_F32);
+    const int64_t partial_size     = 2 + DV;
+    const float * partials_base    = (const float *) params->wdata + partials_offset;
+
+    // Output layout
+    const int64_t ne1 = dst->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    const size_t  nb1 = dst->nb[1];
+
+    // Each thread reduces a subset of query heads
+    for (int64_t q_head = ith; q_head < n_q_heads; q_head += nth) {
+        float   M_final   = -INFINITY;
+        float   S_final   = 0.0f;
+        float * VKQ_final = thread_wdata;
+        memset(VKQ_final, 0, DV * sizeof(float));
+
+        // Combine partials from all chunks
+        for (int64_t chunk_idx = 0; chunk_idx < n_chunks; ++chunk_idx) {
+            const int64_t ic_start = chunk_idx * chunk_size;
+            if (ic_start >= nek1) continue;
+
+            const float * partial   = partials_base + (q_head * n_chunks + chunk_idx) * partial_size;
+            const float   M_chunk   = partial[0];
+            const float   S_chunk   = partial[1];
+            const float * VKQ_chunk = partial + 2;
+
+            if (S_chunk == 0.0f) continue;
+
+            const float M_new     = fmaxf(M_final, M_chunk);
+            const float scale_old = expf(M_final - M_new);
+            const float scale_new = expf(M_chunk - M_new);
+
+            for (int64_t d = 0; d < DV; ++d) {
+                VKQ_final[d] = VKQ_final[d] * scale_old + VKQ_chunk[d] * scale_new;
+            }
+            S_final = S_final * scale_old + S_chunk * scale_new;
+            M_final = M_new;
+        }
+
+        // Normalize and write to output
+        if (S_final != 0.0f) {
+            const float S_inv = 1.0f / S_final;
+            ggml_vec_scale_f32(DV, VKQ_final, S_inv);
+        }
+        // iq1=0, iq3=0 for decode
+        memcpy((char *) dst->data + (0*ne2*ne1 + q_head + 0*ne1)*nb1, VKQ_final, nb1);
+    }
+}
+
 static void ggml_compute_forward_flash_attn_ext_f16(
        const ggml_compute_params * params,
        ggml_tensor * dst) {
@@ -8567,6 +8710,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
    const int64_t DV = nev0;
    const int64_t N  = neq1;

+
    GGML_ASSERT(ne0 == DV);
    GGML_ASSERT(ne2 == N);

@@ -8587,60 +8731,92 @@ static void ggml_compute_forward_flash_attn_ext_f16(
    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();
+    // When use_ref is set, force the vec-only reference implementation (no tiling, no KV-chunking)
+    const bool use_ref = params->use_ref;

-    // 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;
-
-    static constexpr int64_t KV_TILE_SZ = ggml_fa_tile_config::KV;
-    static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
    const bool kv_is_f32_or_f16 = (k->type == GGML_TYPE_F32 || k->type == GGML_TYPE_F16);
-    const bool use_tiled = (q->type == GGML_TYPE_F32 &&
-                            kv_is_f32_or_f16 &&
-                            k->type == v->type &&
-                            nek1 % KV_TILE_SZ == 0 &&
-                            neq1 >= Q_TILE_SZ);  // Only use tiled for batch >= tile size
+    const bool use_split_kv_path = !use_ref && (neq1 == 1 && neq3 == 1) && kv_is_f32_or_f16 && (k->type == v->type) && q->type == GGML_TYPE_F32 && nek1 >= 512;

-    // The first chunk comes from our thread_id, the rest will get auto-assigned.
-    int current_chunk = ith;
+    if (use_split_kv_path) {
+        const int64_t chunk_size = (nek1 + nth - 1) / nth;

-    while (current_chunk < nchunk) {
-        const int64_t ir0 = dr * current_chunk;
-        const int64_t ir1 = MIN(ir0 + dr, nr);
+        // Partials buffer layout: [q_head][kv_chunk][M, S, VKQ]
+        const int64_t partial_size  = 2 + DV;
+        float *       partials_base = (float *) params->wdata + nth * (DK + 2*DV + CACHE_LINE_SIZE_F32);

-        if (use_tiled) {
-            ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1);
+        const int64_t ic_start = ith * chunk_size;
+        const int64_t ic_end   = std::min(ic_start + chunk_size, nek1);
+
+        const int64_t partial_stride = nth * partial_size;
+        float *       chunk_partials = partials_base + ith * partial_size;
+
+        if (ic_start < nek1) {
+            for (int64_t q_head = 0; q_head < neq2; q_head++) {
+                ggml_compute_forward_flash_attn_ext_f16_one_chunk(
+                    params, dst, q_head, q_head + 1, ic_start, ic_end,
+                    chunk_partials, partial_stride);
+            }
        } else {
-            ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1);
+            for (int64_t q_head = 0; q_head < neq2; q_head++) {
+                float * q_partials = chunk_partials + q_head * partial_stride;
+                q_partials[0] = -INFINITY;  // M
+                q_partials[1] = 0.0f;       // S
+            }
        }

-        current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+        ggml_barrier(params->threadpool);
+        ggml_flash_attn_ext_reduce_partials(params, dst, nth, chunk_size);
+    } else {
+
+        // total rows in q
+        const int64_t nr = neq1*neq2*neq3;
+
+        // 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) {
+            ggml_threadpool_chunk_set(params->threadpool, nth);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        const int64_t dr = (nr + nchunk - 1) / nchunk;
+
+        static constexpr int64_t KV_TILE_SZ = ggml_fa_tile_config::KV;
+        static constexpr int64_t Q_TILE_SZ  = ggml_fa_tile_config::Q;
+        const bool use_tiled = !use_ref &&
+                               (q->type == GGML_TYPE_F32 &&
+                                kv_is_f32_or_f16 &&
+                                k->type == v->type &&
+                                nek1 % KV_TILE_SZ == 0 &&
+                                neq1 >= Q_TILE_SZ);
+
+        int current_chunk = ith;
+
+        while (current_chunk < nchunk) {
+            const int64_t ir0 = dr * current_chunk;
+            const int64_t ir1 = MIN(ir0 + dr, nr);
+
+            if (use_tiled) {
+                ggml_compute_forward_flash_attn_ext_tiled(params, dst, ir0, ir1);
+            } else {
+                ggml_compute_forward_flash_attn_ext_f16_one_chunk(params, dst, ir0, ir1, 0, nek1, nullptr, 0);
+            }
+
+            current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1);
+        }
    }
 }

@@ -256,6 +256,200 @@ template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTR
    ggml_quantize_mat_q8_K_4x8(x, vy, n_per_row);
 }

+template <int M, int N>
+static void ggml_gemv_q6_K_NxM_q8_K_generic_impl(int                        n,
+                                                 float * GGML_RESTRICT      s,
+                                                 size_t                     bs,
+                                                 const void * GGML_RESTRICT vx,
+                                                 const void * GGML_RESTRICT vy,
+                                                 int                        nr,
+                                                 int                        nc) {
+    constexpr int blocklen          = M;
+    constexpr int ncols_interleaved = N;
+    const int     qk                = QK_K;
+    const int     nb                = n / qk;
+    const int     blocks_per_half   = 64 / blocklen;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j] = 0.0f;
+        }
+
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen;
+                const int base_h = base_l + 64;
+
+                const int scale_idx_l = base_l / 16;
+                const int scale_idx_h = base_h / 16;
+
+                const int qh_shift_l = ((base_l % 128) / 32) * 2;
+                const int qh_shift_h = ((base_h % 128) / 32) * 2;
+
+                const int qh_half_l = (base_l / 128) * 32;
+                const int qh_half_h = (base_h / 128) * 32;
+
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j];
+                    const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j];
+
+                    int sumi_l = 0;
+                    int sumi_h = 0;
+
+                    for (int i = 0; i < blocklen; i++) {
+                        const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + i;
+                        const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
+                        const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
+
+                        const int qh_idx_l    = qh_half_l + ((base_l + i) % 32);
+                        const int qh_chunk_l  = qh_idx_l / blocklen;
+                        const int qh_pos_l    = qh_idx_l % blocklen;
+                        const int qh_offset_l = qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l;
+                        const int hi_2_l      = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
+
+                        const int qh_idx_h    = qh_half_h + ((base_h + i) % 32);
+                        const int qh_chunk_h  = qh_idx_h / blocklen;
+                        const int qh_pos_h    = qh_idx_h % blocklen;
+                        const int qh_offset_h = qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h;
+                        const int hi_2_h      = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
+
+                        const int q_l = ((hi_2_l << 4) | l_4) - 32;
+                        const int q_h = ((hi_2_h << 4) | hi_4) - 32;
+
+                        const int8_t a_l = a_ptr[l].qs[base_l + i];
+                        const int8_t a_h = a_ptr[l].qs[base_h + i];
+
+                        sumi_l += q_l * a_l;
+                        sumi_h += q_h * a_h;
+                    }
+
+                    sumf[j] +=
+                        (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+        }
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+template <int M, int N>
+static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(int                        n,
+                                                 float * GGML_RESTRICT      s,
+                                                 size_t                     bs,
+                                                 const void * GGML_RESTRICT vx,
+                                                 const void * GGML_RESTRICT vy,
+                                                 int                        nr,
+                                                 int                        nc) {
+    constexpr int blocklen          = M;
+    constexpr int ncols_interleaved = N;
+    const int     qk                = QK_K;
+    const int     nb                = n / qk;
+    const int     blocks_per_half   = 64 / blocklen;
+    const int     q8_half_stride    = 512;
+    const int     q8_low_high_step  = 256;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+
+    float sumf[4][8];
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
+
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j] = 0.0f;
+                }
+            }
+
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen;
+                    const int base_h = base_l + 64;
+
+                    const int scale_idx_l = base_l / 16;
+                    const int scale_idx_h = base_h / 16;
+
+                    const int qh_shift_l = ((base_l % 128) / 32) * 2;
+                    const int qh_shift_h = ((base_h % 128) / 32) * 2;
+
+                    const int qh_half_l = (base_l / 128) * 32;
+                    const int qh_half_h = (base_h / 128) * 32;
+
+                    const int q8_base = (k / blocks_per_half) * q8_half_stride + (k % blocks_per_half) * (blocklen * 4);
+
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j];
+                            const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j];
+
+                            int sumi_l = 0;
+                            int sumi_h = 0;
+
+                            for (int i = 0; i < blocklen; i++) {
+                                const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + i;
+                                const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
+                                const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
+
+                                const int qh_idx_l   = qh_half_l + ((base_l + i) % 32);
+                                const int qh_chunk_l = qh_idx_l / blocklen;
+                                const int qh_pos_l   = qh_idx_l % blocklen;
+                                const int qh_offset_l =
+                                    qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l;
+                                const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
+
+                                const int qh_idx_h   = qh_half_h + ((base_h + i) % 32);
+                                const int qh_chunk_h = qh_idx_h / blocklen;
+                                const int qh_pos_h   = qh_idx_h % blocklen;
+                                const int qh_offset_h =
+                                    qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h;
+                                const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
+
+                                const int q_l = ((hi_2_l << 4) | l_4) - 32;
+                                const int q_h = ((hi_2_h << 4) | hi_4) - 32;
+
+                                const int8_t q8_l = a_ptr[l].qs[q8_base + m * blocklen + i];
+                                const int8_t q8_h = a_ptr[l].qs[q8_base + m * blocklen + i + q8_low_high_step];
+
+                                sumi_l += q_l * q8_l;
+                                sumi_h += q_h * q8_h;
+                            }
+
+                            sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) *
+                                          a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+
 extern "C" {

 void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -704,94 +898,12 @@ void ggml_gemv_q5_K_8x8_q8_K_generic(int                        n,
 }


+void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemv_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
-    constexpr int qk = QK_K;
-    const int nb = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen = 8;
-
-    assert(n % qk == 0);
-    assert(nc % ncols_interleaved == 0);
-
-    UNUSED(bs);
-    UNUSED(nr);
-
-    float sumf[8];
-
-    const block_q8_K * a_ptr = (const block_q8_K *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) {
-            sumf[j] = 0.0f;
-        }
-
-        for (int l = 0; l < nb; l++) {
-
-
-            for (int k = 0; k < 16; k++) {
-                // k = 0.. 7 weights 0-63 low, 64-127 high
-                // k = 8..15 weights 128-191 low, 192-255 high
-                const int base_l = (k / 8) * 128 + (k % 8) * 8;
-                const int base_h = base_l + 64;
-
-                const int scale_idx_l = base_l / 16;
-                const int scale_idx_h = base_h / 16;
-
-                // Bit shift cycles 0,2,4,6 for each 32-value group within a 128-value half
-                const int qh_shift_l = ((base_l % 128) / 32) * 2;
-                const int qh_shift_h = ((base_h % 128) / 32) * 2;
-
-                // qh_half: offset to the correct 32-byte half (0 or 32)
-                const int qh_half_l = (base_l / 128) * 32;
-                const int qh_half_h = (base_h / 128) * 32;
-
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    // Interleaved scales
-                    const int8_t scale_l = b_ptr[l].scales[scale_idx_l * 8 + j];
-                    const int8_t scale_h = b_ptr[l].scales[scale_idx_h * 8 + j];
-
-                    int sumi_l = 0;
-                    int sumi_h = 0;
-
-                    for (int i = 0; i < blocklen; i++) {
-                        const int ql_pos = k * 64 + j * 8 + i;
-                        const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
-                        const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
-
-                        // qh indexing with 8-byte interleaving (like q5_K)
-                        const int qh_byte_l   = qh_half_l + ((base_l + i) % 32);
-                        const int qh_chunk_l  = qh_byte_l / 8;
-                        const int qh_pos_l    = qh_byte_l % 8;
-                        const int qh_offset_l = qh_chunk_l * 64 + j * 8 + qh_pos_l;
-                        const int hi_2_l      = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
-
-                        const int qh_byte_h   = qh_half_h + ((base_h + i) % 32);
-                        const int qh_chunk_h  = qh_byte_h / 8;
-                        const int qh_pos_h    = qh_byte_h % 8;
-                        const int qh_offset_h = qh_chunk_h * 64 + j * 8 + qh_pos_h;
-                        const int hi_2_h      = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
-
-                        const int q_l = ((hi_2_l << 4) | l_4) - 32;
-                        const int q_h = ((hi_2_h << 4) | hi_4) - 32;
-
-                        const int8_t a_l = a_ptr[l].qs[base_l + i];
-                        const int8_t a_h = a_ptr[l].qs[base_h + i];
-
-                        sumi_l += q_l * a_l;
-                        sumi_h += q_h * a_h;
-                    }
-
-                    sumf[j] +=
-                        (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
-                }
-            }
-        }
-
-        for (int j = 0; j < ncols_interleaved; j++) {
-            s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
+    ggml_gemv_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
 }

 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -1485,109 +1597,12 @@ void ggml_gemm_q5_K_8x8_q8_K_generic(int                        n,
    }
 }

-void ggml_gemm_q6_K_8x8_q8_K_generic(int                        n,
-                                     float * GGML_RESTRICT      s,
-                                     size_t                     bs,
-                                     const void * GGML_RESTRICT vx,
-                                     const void * GGML_RESTRICT vy,
-                                     int                        nr,
-                                     int                        nc) {
-    const int qk                = QK_K;
-    const int nb                = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen          = 8;
+void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemm_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
+}

-    assert(n % qk == 0);
-    assert(nr % 4 == 0);
-    assert(nc % ncols_interleaved == 0);
-
-    UNUSED(bs);
-
-    float sumf[4][8];
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb);
-
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumf[m][j] = 0.0f;
-                }
-            }
-
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < 16; k++) {
-                    // k = 0.. 7 weights 0-63 low, 64-127 high
-                    // k = 8..15 weights 128-191 low, 192-255 high
-                    const int base_l = (k / 8) * 128 + (k % 8) * 8;
-                    const int base_h = base_l + 64;
-
-                    const int scale_idx_l = base_l / 16;
-                    const int scale_idx_h = base_h / 16;
-
-                    // Bit shift cycles 0,2,4,6 for each 32-value group within a 128-value half
-                    const int qh_shift_l = ((base_l % 128) / 32) * 2;
-                    const int qh_shift_h = ((base_h % 128) / 32) * 2;
-
-                    // qh_half: offset to the correct 32-byte half (0 or 32)
-                    const int qh_half_l = (base_l / 128) * 32;
-                    const int qh_half_h = (base_h / 128) * 32;
-
-                    // Activation base indices for q8_Kx4 interleaved format
-                    // Layout: 128-value halves (k/8), then 8-value sub-blocks (k%8) with stride 32
-                    const int q8_base = (k / 8) * 512 + (k % 8) * 32;
-
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            // Interleaved scales
-                            const int8_t scale_l = b_ptr[l].scales[scale_idx_l * 8 + j];
-                            const int8_t scale_h = b_ptr[l].scales[scale_idx_h * 8 + j];
-
-                            int sumi_l = 0;
-                            int sumi_h = 0;
-
-                            for (int i = 0; i < blocklen; i++) {
-                                const int ql_pos = k * 64 + j * 8 + i;
-                                const int l_4    = b_ptr[l].ql[ql_pos] & 0xF;
-                                const int hi_4   = (b_ptr[l].ql[ql_pos] >> 4) & 0xF;
-
-                                const int qh_idx_l    = qh_half_l + ((base_l + i) % 32);
-                                const int qh_chunk_l  = qh_idx_l / 8;
-                                const int qh_pos_l    = qh_idx_l % 8;
-                                const int qh_offset_l = qh_chunk_l * 64 + j * 8 + qh_pos_l;
-                                const int hi_2_l      = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3;
-
-                                const int qh_idx_h    = qh_half_h + ((base_h + i) % 32);
-                                const int qh_chunk_h  = qh_idx_h / 8;
-                                const int qh_pos_h    = qh_idx_h % 8;
-                                const int qh_offset_h = qh_chunk_h * 64 + j * 8 + qh_pos_h;
-                                const int hi_2_h      = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3;
-
-                                const int q_l = ((hi_2_l << 4) | l_4) - 32;
-                                const int q_h = ((hi_2_h << 4) | hi_4) - 32;
-
-                                const int8_t q8_l = a_ptr[l].qs[q8_base + m * 8 + i];
-                                const int8_t q8_h = a_ptr[l].qs[q8_base + m * 8 + i + 256];
-
-                                sumi_l += q_l * q8_l;
-                                sumi_h += q_h * q8_h;
-                            }
-
-                            sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) *
-                                          a_ptr[l].d[m];
-                        }
-                    }
-                }
-            }
-
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
-                }
-            }
-        }
-    }
+void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+   ggml_gemm_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
 }

 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -2097,18 +2112,18 @@ static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_in
    }

    const int end_ls = QK_K * 4 / blck_size_interleave;
-    // Interleave Q6_K quants by taking 8 bytes at a time
+    // Interleave Q6_K quants by taking blck_size_interleave bytes at a time
    for (int i = 0; i < end_ls; ++i) {
        int src_id     = i % n_blocks;
        int src_offset = (i / n_blocks) * blck_size_interleave;
        int dst_offset = i * blck_size_interleave;

        uint64_t elem_ls;
-        memcpy(&elem_ls, &in[src_id].ql[src_offset], sizeof(uint64_t));
-        memcpy(&out.ql[dst_offset], &elem_ls, sizeof(uint64_t));
+        memcpy(&elem_ls, &in[src_id].ql[src_offset], blck_size_interleave);
+        memcpy(&out.ql[dst_offset], &elem_ls, blck_size_interleave);
    }

-    // Interleave high bits using same 8-byte pattern as low bits
+    // Interleave high bits using same chunk size as low bits
    const int end_hs = end_ls / 2;
    for (int i = 0; i < end_hs; ++i) {
        int src_id     = i % n_blocks;
@@ -2116,8 +2131,8 @@ static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_in
        int dst_offset = i * blck_size_interleave;

        uint64_t elem_hs;
-        memcpy(&elem_hs, &in[src_id].qh[src_offset], sizeof(uint64_t));
-        memcpy(&out.qh[dst_offset], &elem_hs, sizeof(uint64_t));
+        memcpy(&elem_hs, &in[src_id].qh[src_offset], blck_size_interleave);
+        memcpy(&out.qh[dst_offset], &elem_hs, blck_size_interleave);
    }

    // The below logic is designed so as to unpack and rearrange scales in Q6_K
@@ -2262,7 +2277,7 @@ static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor *       t,

 static int repack_q6_K_to_q6_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
    GGML_ASSERT(t->type == GGML_TYPE_Q6_K);
-    GGML_ASSERT(interleave_block == 8);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
    constexpr int nrows_interleaved = 8;

    block_q6_Kx8 * dst = (block_q6_Kx8 *)t->data;
@@ -2511,6 +2526,10 @@ template <> int repack<block_q5_K, 8, 8>(struct ggml_tensor * t, const void * da
    return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size);
 }

+template <> int repack<block_q6_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q6_K_to_q6_K_8_bl(t, 4, data, data_size);
+}
+
 template <> int repack<block_q6_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
    return repack_q6_K_to_q6_K_8_bl(t, 8, data, data_size);
 }
@@ -2575,6 +2594,10 @@ template <> void gemv<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
    ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }

+template <> void gemv<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
    ggml_gemv_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -2634,6 +2657,10 @@ template <> void gemm<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
    ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }

+template <> void gemm<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemm<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
    ggml_gemm_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -3043,6 +3070,7 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
    static const ggml::cpu::repack::tensor_traits<block_q5_K, 8, 8, GGML_TYPE_Q8_K> q5_K_8x8_q8_K;

    // instance for Q6_K
+    static const ggml::cpu::repack::tensor_traits<block_q6_K, 4, 8, GGML_TYPE_Q8_K> q6_K_8x4_q8_K;
    static const ggml::cpu::repack::tensor_traits<block_q6_K, 8, 8, GGML_TYPE_Q8_K> q6_K_8x8_q8_K;

    // instance for Q2
@@ -3107,6 +3135,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                return &q6_K_8x8_q8_K;
            }
        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q6_K_8x4_q8_K;
+            }
+        }
    } else if (cur->type == GGML_TYPE_IQ4_NL) {
        if (ggml_cpu_has_avx2()) {
            if (cur->ne[1] % 8 == 0) {
@@ -112,6 +112,7 @@ void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -122,6 +123,7 @@ void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -142,6 +144,7 @@ void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -152,6 +155,7 @@ void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -116,6 +116,17 @@ extern "C" {
 // defined in ggml-cpu.c, initialized in ggml_cpu_init()
 extern float ggml_table_f32_f16[1 << 16];

+// precomputed f32 table for e8m0 half (1 KB)
+// defined in ggml-cpu.c, initialized in ggml_cpu_init()
+extern float ggml_table_f32_e8m0_half[1 << 8];
+
+// Use lookup table for E8M0 on x86 (faster than bit manipulation)
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+#define GGML_CPU_E8M0_TO_FP32_HALF(x) ggml_table_f32_e8m0_half[(uint8_t)(x)]
+#else
+#define GGML_CPU_E8M0_TO_FP32_HALF(x) GGML_E8M0_TO_FP32_HALF(x)
+#endif
+
 // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
 // so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
 // This is also true for POWER9.
@@ -111,7 +111,7 @@ template <float (*op)(float), typename src0_t, typename dst_t>
 static void apply_unary_op(const ggml_compute_params * params, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];

-    GGML_ASSERT(ggml_is_contiguous_1(src0) && ggml_is_contiguous_1(dst) && ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_is_contiguous_rows(src0) && ggml_is_contiguous_rows(dst) && ggml_are_same_shape(src0, dst));

    GGML_TENSOR_UNARY_OP_LOCALS

@@ -64,7 +64,7 @@ if (CUDAToolkit_FOUND)
        FetchContent_Declare(
            CCCL
            GIT_REPOSITORY https://github.com/nvidia/cccl.git
-            GIT_TAG        v3.2.0-rc2
+            GIT_TAG        v3.2.0
            GIT_SHALLOW    TRUE
        )

@@ -39,13 +39,16 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
                                   const uint3            ne11,
                                   const uint3            ne12,
                                   const uint3            ne13,
-                                   /*int s0, */ const int s1,
+                                 /*const int              s0,*/
+                                   const int              s1,
                                   const int              s2,
                                   const int              s3,
-                                   /*int s00,*/ const int s01,
+                                   const int              s00,
+                                   const int              s01,
                                   const int              s02,
                                   const int              s03,
-                                   /*int s10,*/ const int s11,
+                                   const int              s10,
+                                   const int              s11,
                                   const int              s12,
                                   const int              s13,
                                   src1_ptrs... src1s) {
@@ -72,11 +75,11 @@ static __global__ void k_bin_bcast(const src0_t *         src0,
    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x * gridDim.x) {
        const uint32_t i10 = fastmodulo(i0, ne10);

-        float result = src0_row ? (float) src0_row[i0] : 0.0f;
+        float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
        if constexpr (sizeof...(src1_ptrs) > 0) {
-            result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10])));
+            result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
        } else {
-            result = bin_op(result, (float)src1[i_src1 + i10]);
+            result = bin_op(result, (float)src1[i_src1 + i10*s10]);
        }

        dst_row[i0] = (dst_t) result;
@@ -101,13 +104,16 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,
                                           const uint3            ne11,
                                           const uint3            ne12,
                                           const uint3            ne13,
-                                           /*int s0, */ const int s1,
+                                         /*const int              s0,*/
+                                           const int              s1,
                                           const int              s2,
                                           const int              s3,
-                                           /*int s00,*/ const int s01,
+                                           const int              s00,
+                                           const int              s01,
                                           const int              s02,
                                           const int              s03,
-                                           /*int s10,*/ const int s11,
+                                           const int              s10,
+                                           const int              s11,
                                           const int              s12,
                                           const int              s13,
                                           src1_ptrs... src1s) {
@@ -135,11 +141,11 @@ static __global__ void k_bin_bcast_unravel(const src0_t *         src0,

    const int i10 = fastmodulo(i0, ne10);

-    float result = src0_row ? (float) src0_row[i0] : 0.0f;
+    float result = src0_row ? (float) src0_row[i0*s00] : 0.0f;
    if constexpr (sizeof...(src1_ptrs) > 0) {
-        result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10])));
+        result = (..., (result = bin_op(result, (float)src1s[i_src1 + i10*s10])));
    } else {
-        result = bin_op(result, (float)src1[i_src1 + i10]);
+        result = bin_op(result, (float)src1[i_src1 + i10*s10]);
    }

    dst_row[i0] = (dst_t) result;
@@ -179,7 +185,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
        cnb[3] *= cne[3];
    };

-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
        for (int i = 0; i < 4; i++) {
            if (nr[i] != 1) {
                break;
@@ -221,7 +227,7 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
        size_t nb12 = cnb1[2];
        size_t nb13 = cnb1[3];

-        size_t s0 = nb0 / sizeof(dst_t);
+      //size_t s0 = nb0 / sizeof(dst_t);
        size_t s1 = nb1 / sizeof(dst_t);
        size_t s2 = nb2 / sizeof(dst_t);
        size_t s3 = nb3 / sizeof(dst_t);
@@ -251,10 +257,6 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
        GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
        GGML_ASSERT(nb13 % sizeof(src1_t) == 0);

-        GGML_ASSERT(s0 == 1);
-        GGML_ASSERT(s00 == 1);
-        GGML_ASSERT(s10 == 1);
-
        const int block_size = 128;

        int64_t hne0 = std::max(ne0 / 2LL, 1LL);
@@ -284,31 +286,31 @@ static void launch_bin_bcast_pack(const ggml_tensor * src0, const ggml_tensor *
                k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t><<<block_num, block_size, 0, stream>>>(
                    src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv, ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11,
                    ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00,*/ s01, s02, s03,
-                    /* s10,*/ s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
+                  /*s0,*/ s1,  s2,  s3,
+                    s00, s01, s02, s03,
+                    s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
            } else {
                k_bin_bcast_unravel<bin_op, src0_t, src1_t, dst_t>
                    <<<block_num, block_size, 0, stream>>>(src0_dd, src1_dd, dst_dd, ne0_fastdiv, ne1_fastdiv,
                                                           ne2_fastdiv, ne3, prod_012, prod_01, ne10, ne11, ne12, ne13,
-                                                           /* s0, */ s1, s2, s3,
-                                                           /* s00,*/ s01, s02, s03,
-                                                           /* s10,*/ s11, s12, s13);
+                                                         /*s0,*/ s1,  s2,  s3,
+                                                           s00, s01, s02, s03,
+                                                           s10, s11, s12, s13);
            }
        } else {
            const uint3 ne3_fastdiv = init_fastdiv_values((uint32_t) ne3);
            if constexpr (sizeof...(I) > 0) {
                k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
                    src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00,*/ s01, s02, s03,
-                    /* s10,*/ s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
+                  /*s0,*/ s1, s2,  s3,
+                    s00 ,s01, s02, s03,
+                    s10, s11, s12, s13, (const src1_t *) dst->src[I + 1]->data...);
            } else {
                k_bin_bcast<bin_op, src0_t, src1_t, dst_t><<<block_nums, block_dims, 0, stream>>>(
                    src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3_fastdiv, ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s00,*/ s01, s02, s03,
-                    /* s10,*/ s11, s12, s13);
+                  /*s0,*/ s1,  s2,  s3,
+                    s00, s01, s02, s03,
+                    s10, s11, s12, s13);
            }
        }
    }
@@ -63,11 +63,19 @@ static __global__ void flash_attn_ext_f16(
    constexpr int frag_m = ncols == 8 ? 32 : 16;
    constexpr int frag_n = ncols == 8 ?  8 : 16;
    static_assert(D % frag_m == 0, "If ncols == 8 then D % frag_m must be 0.");
+#if defined(GGML_USE_HIP)
+    typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, _Float16, wmma::row_major> frag_a_K;
+    typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, _Float16, wmma::col_major> frag_a_V;
+    typedef wmma::fragment<wmma::matrix_b,    frag_m, frag_n, 16, _Float16, wmma::col_major> frag_b;
+    typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, KQ_acc_t>                      frag_c_KQ;
+    typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, _Float16>                          frag_c_VKQ;
+#else
    typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, half, wmma::row_major> frag_a_K;
    typedef wmma::fragment<wmma::matrix_a,    frag_m, frag_n, 16, half, wmma::col_major> frag_a_V;
    typedef wmma::fragment<wmma::matrix_b,    frag_m, frag_n, 16, half, wmma::col_major> frag_b;
    typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, KQ_acc_t>                      frag_c_KQ;
    typedef wmma::fragment<wmma::accumulator, frag_m, frag_n, 16, half>                          frag_c_VKQ;
+#endif

    constexpr int KQ_stride_tc  = nwarps*frag_m; // Number of KQ rows calculated in parallel.
    constexpr int VKQ_ratio = KQ_stride_tc/VKQ_stride; // Number of parallel VKQ accumulators needed to keep all warps busy.
@@ -126,6 +134,19 @@ static __global__ void flash_attn_ext_f16(

    __shared__ half VKQ[ncols*D_padded]; // Accumulator for final VKQ slice.
    half2 * VKQ2 = (half2 *) VKQ;
+
+#if defined(GGML_USE_HIP)
+    const _Float16 * K_h_f16  = reinterpret_cast<const _Float16 *>(K_h);
+    const _Float16 * V_h_f16  = reinterpret_cast<const _Float16 *>(V_h);
+    _Float16       * KQ_f16   = reinterpret_cast<_Float16 *>(KQ);
+    _Float16       * VKQ_f16  = reinterpret_cast<_Float16 *>(VKQ);
+#else
+    const half * K_h_f16  = K_h;
+    const half * V_h_f16  = V_h;
+    half       * KQ_f16   = KQ;
+    half       * VKQ_f16  = VKQ;
+#endif
+
 #pragma unroll
    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
        const int j = j0 + threadIdx.y;
@@ -160,7 +181,7 @@ static __global__ void flash_attn_ext_f16(
    for (int i0 = 0; i0 < D; i0 += 16) {
 #pragma unroll
        for (int j0 = 0; j0 < ncols; j0 += frag_n) {
-            wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ + j0*D_padded + i0, D_padded);
+            wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ_f16 + j0*D_padded + i0, D_padded);
        }
    }

@@ -180,7 +201,7 @@ static __global__ void flash_attn_ext_f16(
 #pragma unroll
            for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 16) {
                frag_a_K K_a;
-                wmma::load_matrix_sync(K_a, K_h + int64_t(k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
+                wmma::load_matrix_sync(K_a, K_h_f16 + int64_t(k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
 #pragma unroll
                for (int j = 0; j < ncols/frag_n; ++j) {
                    wmma::mma_sync(KQ_c[j], K_a, Q_b[k_KQ_0/16][j], KQ_c[j]);
@@ -310,7 +331,7 @@ static __global__ void flash_attn_ext_f16(
                const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
                wmma::load_matrix_sync(
                    KQ_b[k0/(VKQ_ratio*16)][j0/frag_n],
-                    KQ + j0*(kqar*kqs_padded) + k,
+                    KQ_f16 + j0*(kqar*kqs_padded) + k,
                    kqar*kqs_padded);
            }
        }
@@ -328,7 +349,7 @@ static __global__ void flash_attn_ext_f16(
                const int k = k0 + (threadIdx.y % VKQ_ratio)*16;

                frag_a_V v_a;
-                wmma::load_matrix_sync(v_a, V_h + int64_t(k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
+                wmma::load_matrix_sync(v_a, V_h_f16 + int64_t(k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
 #pragma unroll
                for (int j = 0; j < ncols/frag_n; ++j) {
                    wmma::mma_sync(VKQ_c[i_VKQ_0/VKQ_stride][j], v_a, KQ_b[k0/(VKQ_ratio*16)][j], VKQ_c[i_VKQ_0/VKQ_stride][j]);
@@ -344,7 +365,7 @@ static __global__ void flash_attn_ext_f16(
 #pragma unroll
            for (int j0 = 0; j0 < ncols; j0 += frag_n) {
                wmma::store_matrix_sync(
-                    KQ + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio),
+                    KQ_f16 + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio),
                    VKQ_c[i_KQ_0/VKQ_stride][j0/frag_n],
                    D_padded, wmma::mem_col_major);
            }
@@ -2279,13 +2279,19 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;

    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        if (ne2 == 1) {
+        static_assert(MMVQ_MAX_BATCH_SIZE == MMVF_MAX_BATCH_SIZE);
+        if (ne2 <= MMVQ_MAX_BATCH_SIZE) {
            if (ggml_is_quantized(src0->type)) {
-                ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+                if (ne2 <= 4) {
+                    ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+                    return;
+                }
            } else {
-                ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
+                if (GGML_CUDA_CC_IS_AMD(cc)) {
+                    ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
+                    return;
+                }
            }
-            return;
        }

        if (ggml_cuda_should_use_mmq(src0->type, cc, ne12, /*n_experts=*/ne02)) {
@@ -2973,8 +2979,7 @@ static bool ggml_cuda_graph_node_properties_match(ggml_tensor * node, ggml_cuda_
        }
    }

-    if ((node->op == GGML_OP_SCALE || node->op == GGML_OP_GLU) &&
-        memcmp(props->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
+    if (memcmp(props->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
        return false;
    }

@@ -4829,8 +4834,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
        case GGML_OP_SUM_ROWS:
        case GGML_OP_MEAN:
        case GGML_OP_GROUP_NORM:
-        case GGML_OP_PAD:
            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_PAD:
+            return true;
        case GGML_OP_UPSCALE:
        case GGML_OP_PAD_REFLECT_1D:
        case GGML_OP_ARANGE:
@@ -5049,16 +5055,6 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
        static std::mutex mutex;
        std::lock_guard<std::mutex> lock(mutex);
        if (!initialized) {
-            // Set CUDA_SCALE_LAUNCH_QUEUES before any CUDA API call to improve multi-GPU pipeline parallelism performance
-            // PR: https://github.com/ggml-org/llama.cpp/pull/19042
-            if (getenv("CUDA_SCALE_LAUNCH_QUEUES") == nullptr) {
-#ifdef _WIN32
-                _putenv_s("CUDA_SCALE_LAUNCH_QUEUES", "4x");
-#else
-                setenv("CUDA_SCALE_LAUNCH_QUEUES", "4x", 0); // don't overwrite if already set
-#endif // _WIN32
-            }
-
            ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
            const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;

@@ -3697,13 +3697,20 @@ static __global__ void mul_mat_q(
         tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 }

-
 template <ggml_type type, int mmq_x, bool need_check>
-static __global__ void mul_mat_q_stream_k_fixup(
-        const int32_t * ids_dst, const int32_t * expert_bounds, float * __restrict__ dst, const float * __restrict__ tmp_last_tile,
-        const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_col_dst,
-        const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst,
-        const int ncols_max) {
+static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
+                                                const int32_t * expert_bounds,
+                                                float * __restrict__ dst,
+                                                const float * __restrict__ tmp_last_tile,
+                                                const int    ncols_x,
+                                                const int    nrows_x,
+                                                const int    ncols_dst,
+                                                const size_t stride_col_dst,
+                                                const int    nchannels_y,
+                                                const size_t stride_channel_dst,
+                                                const int    nsamples_y,
+                                                const size_t stride_sample_dst,
+                                                const int    ncols_max) {
    constexpr int     mmq_y           = get_mmq_y_device();
    constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
    constexpr int     ITER_K          = get_iter_k(type);
@@ -4,26 +4,48 @@
 #include "mmvf.cuh"
 #include "convert.cuh"

-template <typename T, typename type_acc, int ncols_dst, int block_size, bool has_fusion = false>
+template <typename T, typename type_acc, int ncols_dst, int block_size, bool has_fusion = false, bool is_multi_token_id = false>
 static __global__ void mul_mat_vec_f(
        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
-        const int ncols2, const int nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
+        const int ncols2, const uint3 nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
        const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
-        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const int ids_stride) {
    const int row         = blockIdx.x;
+    // for MUL_MAT_ID - blockIdx.y = n_expert_used, blockIdx.z = ncols_dst (tokens)
    const int channel_dst = blockIdx.y;
-    const int channel_x   = ids ? ids[channel_dst]          : fastdiv((uint32_t) channel_dst, channel_ratio);
-    const int channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
-    const int sample_dst  = blockIdx.z;
+    const int tid         = threadIdx.x;
+
+    int token_idx;
+    int channel_x;
+    int channel_y;
+    int sample_dst;
+
+    if constexpr (is_multi_token_id) {
+        // Multi-token MUL_MAT_ID path, adding these in the normal path causes a perf regression for n_tokens=1 case
+        token_idx  = blockIdx.z;
+        channel_x  = ids[channel_dst + token_idx * ids_stride];
+        channel_y  = fastmodulo(channel_dst, nchannels_y);
+        sample_dst = 0;
+    } else {
+        token_idx  = ids ? blockIdx.z                                          : 0;
+        channel_x  = ids ? ids[blockIdx.y + token_idx * ids_stride]            : fastdiv((uint32_t) channel_dst, channel_ratio);
+        channel_y  = ids ? fastmodulo(blockIdx.y, nchannels_y)                 : channel_dst;
+        sample_dst = ids ? 0                                                   : blockIdx.z;
+    }
+
    const int sample_x    = fastdiv((uint32_t) sample_dst, sample_ratio);
    const int sample_y    = sample_dst;
-    const int tid         = threadIdx.x;

    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();

    x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
    y   += int64_t(sample_y)  *stride_sample_y   + channel_y  *stride_channel_y;
    dst += int64_t(sample_dst)*stride_sample_dst + channel_dst*stride_channel_dst;
+    if constexpr (is_multi_token_id) {
+        y   += token_idx*stride_col_y2*2;
+        dst += token_idx*stride_col_dst;
+    }

    bool use_gate = false;
    bool use_bias = false;
@@ -56,8 +78,10 @@ static __global__ void mul_mat_vec_f(
    if (use_gate) {
        gate_x += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
    }
+
+    const int channel_bias = ids ? channel_x : channel_dst;
+
    if constexpr (has_fusion) {
-        const int channel_bias = ids ? channel_x : channel_dst;
        if (use_bias) {
            x_bias += int64_t(sample_dst)*stride_sample_dst + channel_bias*stride_channel_dst;
        }
@@ -349,36 +373,36 @@ static __global__ void mul_mat_vec_f(
    }
 }

-template<typename T, typename type_acc, int ncols_dst, int block_size>
+template<typename T, typename type_acc, int ncols_dst, int block_size, bool is_multi_token_id = false>
 static void mul_mat_vec_f_switch_fusion(
        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
-        const int64_t ncols, const int64_t nrows,
+        const int64_t ncols, const uint3 nchannels_y,
        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
        const uint3 channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
        const uint3 sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        const dim3 & block_dims, const dim3 & block_nums, const int nbytes_shared, const cudaStream_t stream) {
+        const dim3 & block_dims, const dim3 & block_nums, const int nbytes_shared, const int ids_stride, const cudaStream_t stream) {

    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
    if constexpr (ncols_dst == 1) {
        if (has_fusion) {
-            mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+            mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, y, ids, fusion, dst, ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
                channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
            return;
       }
    }

    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");

-    mul_mat_vec_f<T, type_acc, ncols_dst, block_size><<<block_nums, block_dims, nbytes_shared, stream>>>
-        (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+    mul_mat_vec_f<T, type_acc, ncols_dst, block_size, false, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
+        (x, y, ids, fusion, dst, ncols, nchannels_y, stride_row, stride_col_y, stride_col_dst,
        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);

 }

-template <typename T, typename type_acc, int ncols_dst>
+template <typename T, typename type_acc, int ncols_dst, bool is_multi_token_id = false>
 void launch_mul_mat_vec_f_cuda(
        const T * x, const float * y, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
        const int64_t ncols, const int64_t nrows,
@@ -386,12 +410,13 @@ void launch_mul_mat_vec_f_cuda(
        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
        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 int64_t nsamples_or_ntokens, const int64_t ids_stride, cudaStream_t stream) {
    GGML_ASSERT(ncols        % 2 == 0);
    GGML_ASSERT(stride_row   % 2 == 0);
    GGML_ASSERT(stride_col_y % 2 == 0);
    GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
    GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
+    const uint3 nchannels_y_fd   = ids ? init_fastdiv_values(nchannels_y) : make_uint3(0, 0, 0);
    const uint3 channel_ratio_fd = ids ? make_uint3(0, 0, 0) : init_fastdiv_values(nchannels_dst / nchannels_x);
    const uint3 sample_ratio_fd  = init_fastdiv_values(nsamples_dst  / nsamples_x);

@@ -415,56 +440,56 @@ void launch_mul_mat_vec_f_cuda(
    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;

    const int nbytes_shared = warp_size*sizeof(float) + (has_fusion ? warp_size*sizeof(float) : 0);
-    const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
+    const dim3 block_nums(nrows, nchannels_dst, nsamples_or_ntokens);
    const dim3 block_dims(block_size_best, 1, 1);
    switch (block_size_best) {
        case   32: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 32>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 32, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
        } break;
        case   64: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 64>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 64, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
        } break;
        case   96: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 96>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 96, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
        } break;
        case  128: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 128>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 128, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
        } break;
        case  160: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 160>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 160, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
        } break;
        case  192: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 192>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 192, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
        } break;
        case  224: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 224>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 224, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
        } break;
        case  256: {
-            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 256>
-                (x, y, ids, fusion, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+            mul_mat_vec_f_switch_fusion<T, type_acc, ncols_dst, 256, is_multi_token_id>
+                (x, y, ids, fusion, dst, ncols/2, nchannels_y_fd, stride_row, stride_col_y/2, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, stream);
+                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst, block_dims, block_nums, nbytes_shared, ids_stride, stream);
        } break;
        default: {
            GGML_ABORT("fatal error");
@@ -480,55 +505,88 @@ static void mul_mat_vec_f_cuda_switch_ncols_dst(
        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
        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 int64_t ids_stride, cudaStream_t stream) {
+
+    const bool has_ids = ids != nullptr;
+
+    if (has_ids && ncols_dst > 1) {
+        // Multi-token MUL_MAT_ID path only - single-token goes through regular path below
+        constexpr int c_ncols_dst = 1;
+        launch_mul_mat_vec_f_cuda<T, type_acc, c_ncols_dst, true>
+            (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+             nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+             stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+             ncols_dst, ids_stride, stream);
+        return;
+    }
+
+    if (has_ids) {
+        // Single-token MUL_MAT_ID path
+        constexpr int c_ncols_dst = 1;
+        launch_mul_mat_vec_f_cuda<T, type_acc, c_ncols_dst>
+            (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+             nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+             stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+             ncols_dst, ids_stride, stream);
+        return;
+    }
+
    switch (ncols_dst) {
        case 1:
            launch_mul_mat_vec_f_cuda<T, type_acc, 1>
                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
            break;
        case 2:
            launch_mul_mat_vec_f_cuda<T, type_acc, 2>
                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
            break;
        case 3:
            launch_mul_mat_vec_f_cuda<T, type_acc, 3>
                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
            break;
        case 4:
            launch_mul_mat_vec_f_cuda<T, type_acc, 4>
                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
            break;
        case 5:
            launch_mul_mat_vec_f_cuda<T, type_acc, 5>
                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
            break;
        case 6:
            launch_mul_mat_vec_f_cuda<T, type_acc, 6>
                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
            break;
        case 7:
            launch_mul_mat_vec_f_cuda<T, type_acc, 7>
                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
            break;
        case 8:
            launch_mul_mat_vec_f_cuda<T, type_acc, 8>
                (x, y, ids, fusion, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 nsamples_dst, ids_stride, stream);
            break;
        default:
            GGML_ABORT("fatal error");
@@ -544,21 +602,21 @@ static void mul_mat_vec_f_cuda(
        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
        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,
-        enum ggml_prec prec, cudaStream_t stream) {
+        const int64_t ids_stride, enum ggml_prec prec, cudaStream_t stream) {

    if constexpr(std::is_same_v<T, half>) {
        if (prec == GGML_PREC_DEFAULT) {
            mul_mat_vec_f_cuda_switch_ncols_dst<T, half>
                (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-                stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            return;
        }
    }
    mul_mat_vec_f_cuda_switch_ncols_dst<T, float>
        (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
        nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
-        stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+        stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
 }

 void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
@@ -573,7 +631,7 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
    const size_t ts_src1 = ggml_type_size(src1->type);
    const size_t ts_dst  = ggml_type_size(dst->type);

-    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
+    GGML_ASSERT(!ids || ne12 <= MMVF_MAX_BATCH_SIZE);
    GGML_ASSERT(ne13 == ne3);

    GGML_ASSERT(        nb00       == ts_src0);
@@ -626,29 +684,31 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
    const int64_t ncols_dst          = ids ? ne2  : ne1;
    const int64_t nchannels_y        = ids ? ne11 : ne12;
    const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_col_dst     = ids ? s2   : s1;
+    const int64_t stride_col_y       = ids ? s12  : s11;
    const int64_t stride_channel_dst = ids ? s1   : s2;
    const int64_t stride_channel_y   = ids ? s11  : s12;

-    GGML_ASSERT(!ids || ncols_dst == 1);
+    const int64_t ids_stride = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;

    switch (src0->type) {
        case GGML_TYPE_F32: {
            const float * src0_d = (const float *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+                ne03,              ne3,           s03, s13,              s3,                 ids_stride, prec, ctx.stream());
        } break;
        case GGML_TYPE_F16: {
            const half * src0_d = (const half *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+                ne03,              ne3,           s03, s13,              s3,                 ids_stride, prec, ctx.stream());
        } break;
        case GGML_TYPE_BF16: {
            const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
-            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+                ne03,              ne3,           s03, s13,              s3,                 ids_stride, prec, ctx.stream());
        } break;
        default:
            GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
@@ -695,19 +755,19 @@ void ggml_cuda_op_mul_mat_vec_f(
            const float * src0_d = (const float *) src0_dd_i;
            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, 0, prec, stream);
        } break;
        case GGML_TYPE_F16: {
            const half * src0_d = (const half *) src0_dd_i;
            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, 0, prec, stream);
        } break;
        case GGML_TYPE_BF16: {
            const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
            mul_mat_vec_f_cuda(src0_d, src1_ddf_i, nullptr, empty, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, 0, prec, stream);
        } break;
        default:
            GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
@@ -1,5 +1,7 @@
 #include "common.cuh"

+#define MMVF_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVF kernels.
+
 void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst,
    const ggml_cuda_mm_fusion_args_host * fusion = nullptr);

@@ -137,15 +137,15 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
    return 1;
 }

-// tell the compiler to use as many registers as it wants, see nwarps definition below
-template <ggml_type type, int ncols_dst, bool has_fusion>
+template <ggml_type type, int ncols_dst, bool has_fusion, bool is_multi_token_id = false>
 __launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
        const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
        const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
-        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst) {
+        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
+        const uint32_t ids_stride) {

    constexpr int qk  = ggml_cuda_type_traits<type>::qk;
    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
@@ -162,11 +162,25 @@ static __global__ void mul_mat_vec_q(
    const     int blocks_per_row_x = ncols_x / qk;
    constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;

-    // The MUL_MAT_ID code path with ids != nullptr is only implemented for ncols_dst == 1.
    const uint32_t channel_dst = blockIdx.y;
-    const uint32_t channel_x   = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
-    const uint32_t channel_y   = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
-    const uint32_t sample_dst  = blockIdx.z;
+
+    uint32_t token_idx = 0;
+    uint32_t channel_x;
+    uint32_t channel_y;
+    uint32_t sample_dst;
+
+    if constexpr (is_multi_token_id) {
+        // Multi-token MUL_MAT_ID path, adding these in the normal path causes a perf regression for n_tokens=1 case
+        token_idx  = blockIdx.z;
+        channel_x  = ids[channel_dst + token_idx * ids_stride];
+        channel_y  = fastmodulo(channel_dst, nchannels_y);
+        sample_dst = 0;
+    } else {
+        channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
+        channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
+        sample_dst = blockIdx.z;
+    }
+
    const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
    const uint32_t sample_y    = sample_dst;

@@ -188,11 +202,11 @@ static __global__ void mul_mat_vec_q(
        active_glu    = fusion.glu_op;
    }

-    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) {
+        const uint32_t channel_bias = ids ? channel_x : channel_dst;
        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
@@ -222,6 +236,9 @@ static __global__ void mul_mat_vec_q(
    float tmp_gate[ncols_dst][rows_per_cuda_block] = {{0.0f}};

    const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
+    if constexpr (is_multi_token_id) {
+        y += token_idx*stride_col_y;
+    }
    const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;

    for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
@@ -275,6 +292,10 @@ static __global__ void mul_mat_vec_q(

    dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;

+    if constexpr (is_multi_token_id) {
+        dst += token_idx*stride_col_dst;
+    }
+
    // sum up partial sums and write back result
 #pragma unroll
    for (int j = 0; j < ncols_dst; ++j) {
@@ -335,40 +356,41 @@ static __global__ void mul_mat_vec_q(
 }

 static std::pair<dim3, dim3> calc_launch_params(
-        const int ncols_dst, const int nrows_x, const int nchannels_y, const int nsamples_y,
+        const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
        const int warp_size, const mmvq_parameter_table_id table_id) {
    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
-    const dim3 block_nums(nblocks, nchannels_y, nsamples_y);
+    const dim3 block_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
    const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
    return {block_nums, block_dims};
 }

-template<ggml_type type, int c_ncols_dst>
+template<ggml_type type, int c_ncols_dst, bool is_multi_token_id = false>
 static void mul_mat_vec_q_switch_fusion(
        const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
        const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
        const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
-        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared, cudaStream_t stream) {
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared,
+        const uint32_t ids_stride, cudaStream_t stream) {

    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
    if constexpr (c_ncols_dst == 1) {
        if (has_fusion) {
-            mul_mat_vec_q<type, c_ncols_dst, true><<<block_nums, block_dims, nbytes_shared, stream>>>
+            mul_mat_vec_q<type, c_ncols_dst, true, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
                (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
            return;
        }
    }

    GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");

-    mul_mat_vec_q<type, c_ncols_dst, false><<<block_nums, block_dims, nbytes_shared, stream>>>
+    mul_mat_vec_q<type, c_ncols_dst, false, is_multi_token_id><<<block_nums, block_dims, nbytes_shared, stream>>>
        (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
        channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
 }

 template <ggml_type type>
@@ -379,7 +401,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
        const int nchannels_x, const int nchannels_y, const int nchannels_dst,
        const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
        const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        cudaStream_t stream) {
+        const int ids_stride, cudaStream_t stream) {

    GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
    GGML_ASSERT(ncols_dst <= MMVQ_MAX_BATCH_SIZE);
@@ -393,8 +415,19 @@ static void mul_mat_vec_q_switch_ncols_dst(
    const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);

    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    const bool has_ids = ids != nullptr;
+
+    if (has_ids && ncols_dst > 1) {
+        // Multi-token MUL_MAT_ID path only - single-token goes through regular path below
+        constexpr int c_ncols_dst = 1;
+        std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
+        mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+             channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
+             sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
+             dims.first, dims.second, 0, ids_stride, stream);
+        return;
+    }

-    GGML_ASSERT(!ids || ncols_dst == 1);
    switch (ncols_dst) {
        case 1: {
            constexpr int c_ncols_dst = 1;
@@ -402,7 +435,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
        } break;
        case 2: {
            constexpr int c_ncols_dst = 2;
@@ -410,7 +443,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
        } break;
        case 3: {
            constexpr int c_ncols_dst = 3;
@@ -418,7 +451,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
        } break;
        case 4: {
            constexpr int c_ncols_dst = 4;
@@ -426,7 +459,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
        } break;
        case 5: {
            constexpr int c_ncols_dst = 5;
@@ -434,7 +467,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
        } break;
        case 6: {
            constexpr int c_ncols_dst = 6;
@@ -442,7 +475,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
        } break;
        case 7: {
            constexpr int c_ncols_dst = 7;
@@ -450,7 +483,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
        } break;
        case 8: {
            constexpr int c_ncols_dst = 8;
@@ -458,7 +491,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
        } break;
        default:
            GGML_ABORT("fatal error");
@@ -474,127 +507,127 @@ static void mul_mat_vec_q_switch_type(
        const int nchannels_x, const int nchannels_y, const int nchannels_dst,
        const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
        const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        cudaStream_t stream) {
+        const int ids_stride, cudaStream_t stream) {
    switch (type_x) {
        case GGML_TYPE_Q4_0:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_Q4_1:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_Q5_0:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_Q5_1:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_Q8_0:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_MXFP4:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_MXFP4>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_Q2_K:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_Q3_K:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_Q4_K:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_Q5_K:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_Q6_K:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_IQ2_XXS:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_IQ2_XS:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_IQ2_S:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_IQ3_XXS:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_IQ1_S:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_IQ1_M:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_IQ4_NL:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_IQ4_XS:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        case GGML_TYPE_IQ3_S:
            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
            break;
        default:
            GGML_ABORT("fatal error");
@@ -622,7 +655,7 @@ void ggml_cuda_mul_mat_vec_q(
    GGML_ASSERT(        nb0        == ts_dst);
    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));

-    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
+    GGML_ASSERT(!ids || ne12 <= MMVQ_MAX_BATCH_SIZE);

    const float   * src1_d =       (const float   *) src1->data;
    const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
@@ -693,11 +726,13 @@ void ggml_cuda_mul_mat_vec_q(
    const int64_t stride_channel_dst = ids ? s1   : s2;
    const int64_t stride_channel_y   = ids ? s11  : s12;

+    const int64_t ids_stride = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
+
    mul_mat_vec_q_switch_type(
        src0->data, src0->type, src1_q8_1.get(), ids_d, fusion_local, dst_d, ne00,
        ne01,              ncols_dst,     s01, stride_col_y,     stride_col_dst,
        ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-        ne03,              ne3,           s03, s13,              s3,               stream);
+        ne03,              ne3,           s03, s13,              s3,               ids_stride, stream);
 }

 void ggml_cuda_op_mul_mat_vec_q(
@@ -726,7 +761,7 @@ void ggml_cuda_op_mul_mat_vec_q(
    ggml_cuda_mm_fusion_args_device fusion_local{};
    mul_mat_vec_q_switch_type(
        src0_dd_i, src0->type, src1_ddq_i, nullptr, fusion_local, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
-        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
+        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, stream);

    GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);
 }
@@ -7,7 +7,7 @@ __device__ __forceinline__ int64_t wrap_around(int64_t coord, int64_t size) {
    return (coord + size) % size;
 }

-static __global__ void pad_f32(const float * src, float * dst,
+static __global__ void pad_f32(const float * src, size_t s00, size_t s01, size_t s02, size_t s03, float * dst,
                               const int lp0, const int rp0, const int lp1, const int rp1,
                               const int lp2, const int rp2, const int lp3, const int rp3,
                               const int ne0, const int ne1, const int ne2, const int ne3,
@@ -34,11 +34,8 @@ static __global__ void pad_f32(const float * src, float * dst,
            const int64_t i01  = i1 - lp1;
            const int64_t i02  = i2 - lp2;
            const int64_t i03  = i3 - lp3;
-            const int64_t ne02 = ne2 - lp2 - rp2;
-            const int64_t ne01 = ne1 - lp1 - rp1;
-            const int64_t ne00 = ne0 - lp0 - rp0;

-            const int64_t src_idx = i03 * (ne00 * ne01 * ne02) + i02 * (ne00 * ne01) + i01 * ne00 + i00;
+            const int64_t src_idx = i03 * s03 + i02 * s02 + i01 * s01 + i00 * s00;

            dst[dst_idx] = src[src_idx];
        } else {
@@ -57,21 +54,21 @@ static __global__ void pad_f32(const float * src, float * dst,
        const int64_t i02 = wrap_around(i2 - lp2, ne02);
        const int64_t i03 = wrap_around(i3 - lp3, ne03);

-        const int64_t src_idx = i03 * (ne00 * ne01 * ne02) + i02 * (ne00 * ne01) + i01 * ne00 + i00;
+        const int64_t src_idx = i03 * s03 + i02 * s02 + i01 * s01 + i00 * s00;

        dst[dst_idx] = src[src_idx];
    }
 }


-static void pad_f32_cuda(const float * src, float * dst,
+static void pad_f32_cuda(const float * src, size_t s00, size_t s01, size_t s02, size_t s03, float * dst,
    const int lp0, const int rp0, const int lp1, const int rp1,
    const int lp2, const int rp2, const int lp3, const int rp3,
    const int ne0, const int ne1, const int ne2, const int ne3,
    const bool circular, cudaStream_t stream) {
    int  num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
    dim3 gridDim(num_blocks, ne1, ne2 * ne3);
-    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, dst,
+    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, s00, s01, s02, s03, dst,
                                                         lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
                                                         ne0, ne1, ne2, ne3, circular);
 }
@@ -82,9 +79,10 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    float *             dst_d  = (float *) dst->data;
    cudaStream_t        stream = ctx.stream();

+    GGML_TENSOR_UNARY_OP_LOCALS;
+
    GGML_ASSERT(src0->type == GGML_TYPE_F32);
    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(ggml_is_contiguous(src0));

    const int32_t lp0      = ((const int32_t *) (dst->op_params))[0];
    const int32_t rp0      = ((const int32_t *) (dst->op_params))[1];
@@ -96,7 +94,12 @@ void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const int32_t rp3      = ((const int32_t *) (dst->op_params))[7];
    const int32_t circular = ((const int32_t *) (dst->op_params))[8];

-    pad_f32_cuda(src0_d, dst_d,
+    const size_t s00 = nb00 / ggml_type_size(src0->type);
+    const size_t s01 = nb01 / ggml_type_size(src0->type);
+    const size_t s02 = nb02 / ggml_type_size(src0->type);
+    const size_t s03 = nb03 / ggml_type_size(src0->type);
+
+    pad_f32_cuda(src0_d, s00, s01, s02, s03, dst_d,
                 lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
                 dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
                 (bool) circular, stream);
@@ -43,10 +43,15 @@ static __device__ void rope_yarn(
 template <bool forward, bool has_ff, typename T, typename D>
 static __global__ void rope_norm(const T *            x,
                                 D *                  dst,
-                                 const int            ne0,
-                                 const int            ne1,
+                                 const int            ne00,
+                                 const int            ne01,
+                                 const int            ne02,
+                                 const int            s01,
+                                 const int            s02,
+                                 const int            s03,
                                 const int            s1,
                                 const int            s2,
+                                 const int            s3,
                                 const int            n_dims,
                                 const int32_t *      pos,
                                 const float          freq_scale,
@@ -59,23 +64,23 @@ static __global__ void rope_norm(const T *            x,
                                 const int            set_rows_stride) {
    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

-    if (i0 >= ne0) {
+    if (i0 >= ne00) {
        return;
    }

    const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;

-    const int row_x     = row_dst % ne1;
-    const int channel_x = row_dst / ne1;
-
-    int       idst = row_dst * ne0 + i0;
-    const int ix   = channel_x*s2 + row_x*s1 + i0;
+    const uint32_t i3 = row_dst / (ne01 * ne02);
+    const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
+    const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;

+    int       idst = i0 + i1 * s1  + i2 * s2  + i3 * s3;
+    const int ix   = i0 + i1 * s01 + i2 * s02 + i3 * s03;
    // Fusion optimization: ROPE + VIEW + SET_ROWS.
    // The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
    if (set_rows_stride != 0) {
-        idst = row_x * ne0 + i0;
-        idst += row_indices[channel_x] * set_rows_stride;
+        idst = i1 * s1 + i0;
+        idst += row_indices[i2] * set_rows_stride;
    }

    const auto & store_coaelsced = [&](float x0, float x1) {
@@ -92,7 +97,7 @@ static __global__ void rope_norm(const T *            x,
        return;
    }

-    const float theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+    const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);

    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;

@@ -110,10 +115,15 @@ static __global__ void rope_norm(const T *            x,
 template <bool forward, bool has_ff, typename T, typename D>
 static __global__ void rope_neox(const T *            x,
                                 D *                  dst,
-                                 const int            ne0,
-                                 const int            ne1,
+                                 const int            ne00,
+                                 const int            ne01,
+                                 const int            ne02,
+                                 const int            s01,
+                                 const int            s02,
+                                 const int            s03,
                                 const int            s1,
                                 const int            s2,
+                                 const int            s3,
                                 const int            n_dims,
                                 const int32_t *      pos,
                                 const float          freq_scale,
@@ -126,23 +136,24 @@ static __global__ void rope_neox(const T *            x,
                                 const int            set_rows_stride) {
    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

-    if (i0 >= ne0) {
+    if (i0 >= ne00) {
        return;
    }

    const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;

-    const int row_x     = row_dst % ne1;
-    const int channel_x = row_dst / ne1;
+    const uint32_t i3 = row_dst / (ne01 * ne02);
+    const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
+    const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;

-    int       idst = row_dst * ne0 + i0 / 2;
-    const int ix   = channel_x*s2 + row_x*s1 + i0/2;
+    int       idst = i0 / 2 + i1 * s1  + i2 * s2  + i3 * s3;
+    const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;

    // Fusion optimization: ROPE + VIEW + SET_ROWS.
    // The rope output is viewed as a 1D tensor and offset based on a row index in row_indices.
    if (set_rows_stride != 0) {
-        idst = row_x * ne0 + i0 / 2;
-        idst += row_indices[channel_x] * set_rows_stride;
+        idst = i1 * s1 + i0 / 2;
+        idst += row_indices[i2] * set_rows_stride;
    }

    if (i0 >= n_dims) {
@@ -152,7 +163,7 @@ static __global__ void rope_neox(const T *            x,
        return;
    }

-    const float theta_base = pos[channel_x]*powf(theta_scale, i0/2.0f);
+    const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);

    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;

@@ -168,24 +179,42 @@ static __global__ void rope_neox(const T *            x,
    dst[idst + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }

-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 bool is_imrope) {
-    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+template <bool forward, bool has_ff, typename T>
+static __global__ void rope_multi(const T *            x,
+                                  T *                  dst,
+                                  const int            ne00,
+                                  const int            ne01,
+                                  const int            ne02,
+                                  const int            s01,
+                                  const int            s02,
+                                  const int            s03,
+                                  const int            s1,
+                                  const int            s2,
+                                  const int            s3,
+                                  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 bool           is_imrope) {
+    const int i0 = 2 * (blockDim.y * blockIdx.y + threadIdx.y);

-    if (i0 >= ne0) {
+    if (i0 >= ne00) {
        return;
    }

    const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;

-    const int row_x     = row_dst % ne1;
-    const int channel_x = row_dst / ne1;
+    const uint32_t i3 = row_dst / (ne01 * ne02);
+    const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
+    const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;

-    const int idst = row_dst*ne0 + i0/2;
-    const int ix   = channel_x*s2 + row_x*s1 + i0/2;
+    int       idst = i0 / 2 + i1 * s1  + i2 * s2  + i3 * s3;
+    const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;

    if (i0 >= n_dims) {
        dst[idst + i0/2 + 0] = x[ix + i0/2 + 0];
@@ -200,27 +229,24 @@ static __global__ void rope_multi(

    float theta_base = 0.0;
    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);
+        if (sector % 3 == 1 && sector < 3 * sections.v[1]) {         // h
+            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
+        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) {  // w
+            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
+        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) {  // t
+            theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
        } else {
-            theta_base = pos[channel_x + ne2 * 3]*powf(theta_scale, i0/2.0f);
+            theta_base = pos[i2 + ne02 * 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);
+            theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
+        } else if (sector >= sections.v[0] && sector < sec_w) {
+            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
+        } else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
+        } else if (sector >= sec_w + sections.v[2]) {
+            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
        }
    }

@@ -238,37 +264,53 @@ static __global__ void rope_multi(
    dst[idst + n_dims/2] = x0*sin_theta + x1*cos_theta;
 }

-template<bool forward, bool has_ff, typename T>
-static __global__ void rope_vision(
-        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) {
+template <bool forward, bool has_ff, typename T>
+static __global__ void rope_vision(const T *            x,
+                                   T *                  dst,
+                                   const int            ne00,
+                                   const int            ne01,
+                                   const int            ne02,
+                                   const int            s01,
+                                   const int            s02,
+                                   const int            s03,
+                                   const int            s1,
+                                   const int            s2,
+                                   const int            s3,
+                                   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 int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

-    if (i0 >= ne0) {
+    if (i0 >= ne00) {
        return;
    }

    const int row_dst = blockDim.x*blockIdx.x + threadIdx.x;

-    const int row_x     = row_dst % ne1;
-    const int channel_x = row_dst / ne1;
+    const uint32_t i3 = row_dst / (ne01 * ne02);
+    const uint32_t i2 = (row_dst - i3 * ne01 * ne02) / ne01;
+    const uint32_t i1 = row_dst - i3 * ne01 * ne02 - i2 * ne01;

-    const int idst = row_dst*ne0 + i0/2;
-    const int ix   = channel_x*s2 + row_x*s1 + i0/2;
+    int       idst = i0 / 2 + i1 * s1  + i2 * s2  + i3 * s3;
+    const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;

    const int sect_dims = sections.v[0] + sections.v[1];
-    const int sec_w = sections.v[1] + sections.v[0];
-    const int sector = (i0 / 2) % sect_dims;
+    const int sec_w     = sections.v[1] + sections.v[0];
+    const int sector    = (i0 / 2) % sect_dims;

    float theta_base = 0.0;
    if (sector < sections.v[0]) {
        const int p = sector;
-        theta_base = pos[channel_x]*powf(theta_scale, p);
-    }
-    else if (sector >= sections.v[0] && sector < sec_w) {
+        theta_base  = pos[i2] * powf(theta_scale, p);
+    } else if (sector >= sections.v[0] && sector < sec_w) {
        const int p = sector - sections.v[0];
-        theta_base = pos[channel_x + ne2]*powf(theta_scale, p);
+        theta_base  = pos[i2 + ne02] * powf(theta_scale, p);
    }

    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
@@ -288,10 +330,15 @@ static __global__ void rope_vision(
 template <bool forward, typename T, typename D>
 static void rope_norm_cuda(const T *            x,
                           D *                  dst,
-                           const int            ne0,
-                           const int            ne1,
+                           const int            ne00,
+                           const int            ne01,
+                           const int            ne02,
+                           const int            s01,
+                           const int            s02,
+                           const int            s03,
                           const int            s1,
                           const int            s2,
+                           const int            s3,
                           const int            n_dims,
                           const int            nr,
                           const int32_t *      pos,
@@ -304,31 +351,36 @@ static void rope_norm_cuda(const T *            x,
                           const int64_t *      row_indices,
                           const int            set_rows_stride,
                           cudaStream_t         stream) {
-    GGML_ASSERT(ne0 % 2 == 0);
+    GGML_ASSERT(ne00 % 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);
+    const int  n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
    const dim3 block_nums(nr, n_blocks_x, 1);

-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);

    if (freq_factors == nullptr) {
        rope_norm<forward, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
-            freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
    } else {
        rope_norm<forward, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
-            freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
    }
 }

 template <bool forward, typename T, typename D>
 static void rope_neox_cuda(const T *            x,
                           D *                  dst,
-                           const int            ne0,
-                           const int            ne1,
+                           const int            ne00,
+                           const int            ne01,
+                           const int            ne02,
+                           const int            s01,
+                           const int            s02,
+                           const int            s03,
                           const int            s1,
                           const int            s2,
+                           const int            s3,
                           const int            n_dims,
                           const int            nr,
                           const int32_t *      pos,
@@ -341,55 +393,92 @@ static void rope_neox_cuda(const T *            x,
                           const int64_t *      row_indices,
                           const int            set_rows_stride,
                           cudaStream_t         stream) {
-    GGML_ASSERT(ne0 % 2 == 0);
+    GGML_ASSERT(ne00 % 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);
+    const int  n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
    const dim3 block_nums(nr, n_blocks_x, 1);

-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);

    if (freq_factors == nullptr) {
        rope_neox<forward, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
-            freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
    } else {
        rope_neox<forward, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims, theta_scale,
-            freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
    }
 }

-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, const bool is_imrope, cudaStream_t stream) {
-    GGML_ASSERT(ne0 % 2 == 0);
+template <bool forward, typename T>
+static void rope_multi_cuda(const T *            x,
+                            T *                  dst,
+                            const int            ne00,
+                            const int            ne01,
+                            const int            ne02,
+                            const int            s01,
+                            const int            s02,
+                            const int            s03,
+                            const int            s1,
+                            const int            s2,
+                            const int            s3,
+                            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,
+                            const bool           is_imrope,
+                            cudaStream_t         stream) {
+    GGML_ASSERT(ne00 % 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);
+    const int  n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
    const dim3 block_nums(nr, n_blocks_x, 1);

-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);

    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,
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
            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,
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
    }
 }

-template<bool forward, typename T>
-static void rope_vision_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) {
-    GGML_ASSERT(ne0 % 2 == 0);
+template <bool forward, typename T>
+static void rope_vision_cuda(const T *            x,
+                             T *                  dst,
+                             const int            ne00,
+                             const int            ne01,
+                             const int            ne02,
+                             const int            s01,
+                             const int            s02,
+                             const int            s03,
+                             const int            s1,
+                             const int            s2,
+                             const int            s3,
+                             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) {
+    GGML_ASSERT(ne00 % 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);
+    const int  n_blocks_x = (ne00 + 2 * CUDA_ROPE_BLOCK_SIZE - 1) / (2 * CUDA_ROPE_BLOCK_SIZE);
    const dim3 block_nums(nr, n_blocks_x, 1);
    // break down (head_dim, heads, seq) into (CUDA_ROPE_BLOCK_SIZE, x, heads * seq)
    // where x ~= ceil(head_dim / CUDA_ROPE_BLOCK_SIZE);
@@ -398,11 +487,11 @@ static void rope_vision_cuda(

    if (freq_factors == nullptr) {
        rope_vision<forward, false, T><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
            attn_factor, corr_dims, theta_scale, freq_factors, sections);
    } else {
        rope_vision<forward, true, T><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor,
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
            attn_factor, corr_dims, theta_scale, freq_factors, sections);
    }
 }
@@ -445,6 +534,11 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,

    const size_t s01 = src0->nb[1] / ggml_type_size(src0->type);
    const size_t s02 = src0->nb[2] / ggml_type_size(src0->type);
+    const size_t s03 = src0->nb[3] / ggml_type_size(src0->type);
+
+    const size_t s1 = dst->nb[1] / ggml_type_size(dst->type);
+    const size_t s2 = dst->nb[2] / ggml_type_size(dst->type);
+    const size_t s3 = dst->nb[3] / ggml_type_size(dst->type);

    //const int n_past     = ((int32_t *) dst->op_params)[0];
    const int n_dims     = ((int32_t *) dst->op_params)[1];
@@ -495,57 +589,63 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
    // compute
    if (is_neox) {
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
-            rope_neox_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims,
-                                                  nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                  freq_factors, row_indices, set_rows_stride, stream);
+            rope_neox_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                  s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                  ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                  set_rows_stride, stream);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
-            rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims,
-                                                 nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                 freq_factors, row_indices, set_rows_stride, stream);
+            rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                 s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                 ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                 set_rows_stride, stream);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
-            rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr,
-                                                pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                freq_factors, row_indices, set_rows_stride, stream);
+            rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                set_rows_stride, stream);
        } else {
            GGML_ABORT("fatal error");
        }
    } else if (is_mrope && !is_vision) {
        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, is_imrope, stream);
+            rope_multi_cuda<forward>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
+                                     s2, s3, n_dims, nr, pos, freq_scale, 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, is_imrope, stream);
+            rope_multi_cuda<forward>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
+                                     s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                     corr_dims, freq_factors, sections, is_imrope, stream);
        } else {
            GGML_ABORT("fatal error");
        }
    } else if (is_vision) {
        if (src0->type == GGML_TYPE_F32) {
-            rope_vision_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);
+            rope_vision_cuda<forward>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
+                                      s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                      corr_dims, freq_factors, sections, stream);
        } else if (src0->type == GGML_TYPE_F16) {
-            rope_vision_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);
+            rope_vision_cuda<forward>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
+                                      s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                      corr_dims, freq_factors, sections, stream);
        } else {
            GGML_ABORT("fatal error");
        }
    } else {
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
-            rope_norm_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, s01, s02, n_dims,
-                                                  nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                  freq_factors, row_indices, set_rows_stride, stream);
+            rope_norm_cuda<forward, float, float>((const float *) src0_d, (float *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                  s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                  ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                  set_rows_stride, stream);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
-            rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims,
-                                                 nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                 freq_factors, row_indices, set_rows_stride, stream);
+            rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                 s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                 ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                 set_rows_stride, stream);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
-            rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, s01, s02, n_dims, nr,
-                                                pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
-                                                freq_factors, row_indices, set_rows_stride, stream);
+            rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
+                                                s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
+                                                set_rows_stride, stream);
        } else {
            GGML_ABORT("fatal error");
        }
@@ -1935,11 +1935,6 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
        return false;
    }

-    // TODO: add support for non-contigiuos tensors
-    if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1) || !ggml_is_contiguous(dst)) {
-        return false;
-    }
-
    return true;
 }

@@ -1991,6 +1986,25 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
    return true;
 }

+static bool ggml_hexagon_supported_sum_rows(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * dst  = op;
+
+    if (!hex_supported_src0_type(src0->type)) {
+        return false;
+    }
+    if (!hex_supported_dst_type(dst->type)) {
+        return false;
+    }
+
+    // TODO: add support for non-contigiuos tensors
+    if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(dst)) {
+        return false;
+    }
+
+    return true;
+}
+
 static bool ggml_hexagon_supported_activations(const struct ggml_hexagon_session * sess,
                                               const struct ggml_tensor *          op) {
    const struct ggml_tensor * src0 = op->src[0];
@@ -2111,6 +2125,26 @@ static bool ggml_hexagon_supported_get_rows(const struct ggml_hexagon_session *
    return true;
 }

+static bool ggml_hexagon_supported_argsort(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0]; // values
+    const struct ggml_tensor * dst  = op;         // indices
+
+    if (src0->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    if (dst->type != GGML_TYPE_I32) {
+        return false;
+    }
+
+    if (src0->ne[0] > (16*1024)) {
+        // reject tensors with huge rows for now
+        return false;
+    }
+
+    return true;
+}
+
 static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
    const int32_t * op_params = &op->op_params[0];

@@ -2278,6 +2312,9 @@ static inline size_t init_binary_req(htp_general_req * req, dspqueue_buffer * bu
        case GGML_OP_SUB:
            req->op = HTP_OP_SUB;
            break;
+        case GGML_OP_DIV:
+            req->op = HTP_OP_DIV;
+            break;
        default:
            GGML_ABORT("ggml-hex: binary : unsupported op: %d\n", t->op);
            break;
@@ -2316,6 +2353,17 @@ static inline size_t init_get_rows_req(htp_general_req * req, dspqueue_buffer *
    return n_bufs;
 }

+static inline size_t init_argsort_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
+    req->op = HTP_OP_ARGSORT;
+    memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
+
+    size_t n_bufs = 0;
+    n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
+    n_bufs += htp_req_buff_init(&req->dst,  &bufs[n_bufs], t,         DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
+
+    return n_bufs;
+}
+
 template <bool _is_src0_constant>
 static inline size_t init_binary_id_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
    switch (t->op) {
@@ -2370,6 +2418,16 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
            supported = true;
            break;

+        case GGML_OP_SQR:
+            req->op   = HTP_OP_SQR;
+            supported = true;
+            break;
+
+        case GGML_OP_SQRT:
+            req->op   = HTP_OP_SQRT;
+            supported = true;
+            break;
+
        case GGML_OP_UNARY:
            if (ggml_get_unary_op(t) == GGML_UNARY_OP_SILU) {
                req->op   = HTP_OP_UNARY_SILU;
@@ -2387,6 +2445,9 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
            } else if (ggml_get_glu_op(t) == GGML_GLU_OP_SWIGLU_OAI) {
                req->op   = HTP_OP_GLU_SWIGLU_OAI;
                supported = true;
+            } else if (ggml_get_glu_op(t) == GGML_GLU_OP_GEGLU) {
+                req->op   = HTP_OP_GLU_GEGLU;
+                supported = true;
            }
            break;

@@ -2411,6 +2472,17 @@ static inline size_t init_unary_req(htp_general_req * req, dspqueue_buffer * buf
    return n_bufs;
 }

+static inline size_t init_sum_rows_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
+    memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
+    req->op = HTP_OP_SUM_ROWS;
+
+    size_t n_bufs = 0;
+    n_bufs += htp_req_buff_init(&req->src0, &bufs[n_bufs], t->src[0], DSPQBUF_TYPE_CPU_WRITE_DSP_READ);
+    n_bufs += htp_req_buff_init(&req->dst,  &bufs[n_bufs], t,         DSPQBUF_TYPE_DSP_WRITE_CPU_READ);
+
+    return n_bufs;
+}
+
 static inline size_t init_rope_req(htp_general_req * req, dspqueue_buffer * bufs, const ggml_tensor * t) {
    memcpy(&req->op_params, &t->op_params, sizeof(t->op_params));
    req->op = HTP_OP_ROPE;
@@ -2519,6 +2591,7 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
            case GGML_OP_MUL:
            case GGML_OP_ADD:
            case GGML_OP_SUB:
+            case GGML_OP_DIV:
                ggml_hexagon_dispatch_op<init_binary_req<false>>(sess, node, flags);
                break;
            case GGML_OP_ADD_ID:
@@ -2528,6 +2601,13 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
            case GGML_OP_SCALE:
                ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
                break;
+            case GGML_OP_SQR:
+            case GGML_OP_SQRT:
+                ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
+                break;
+            case GGML_OP_SUM_ROWS:
+                ggml_hexagon_dispatch_op<init_sum_rows_req>(sess, node, flags);
+                break;
            case GGML_OP_UNARY:
                if ((ggml_get_unary_op(node) == GGML_UNARY_OP_SILU) ||
                        (ggml_get_unary_op(node) == GGML_UNARY_OP_GELU)) {
@@ -2536,7 +2616,8 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                break;
            case GGML_OP_GLU:
                if ((ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU) ||
-                        (ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI)) {
+                        (ggml_get_glu_op(node) == GGML_GLU_OP_SWIGLU_OAI) ||
+                        (ggml_get_glu_op(node) == GGML_GLU_OP_GEGLU)) {
                    ggml_hexagon_dispatch_op<init_unary_req>(sess, node, flags);
                }
                break;
@@ -2564,6 +2645,10 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
                ggml_hexagon_dispatch_op<init_cpy_req>(sess, node, flags);
                break;

+            case GGML_OP_ARGSORT:
+                ggml_hexagon_dispatch_op<init_argsort_req>(sess, node, flags);
+                break;
+
            default:
                GGML_ABORT("\nggml-hex: graph-compute %s is not supported\n", ggml_op_desc(node));
        }
@@ -2916,6 +3001,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
        case GGML_OP_MUL:
        case GGML_OP_ADD:
        case GGML_OP_SUB:
+        case GGML_OP_DIV:
            supp = ggml_hexagon_supported_binary(sess, op);
            break;

@@ -2928,6 +3014,15 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
            supp = ggml_hexagon_supported_unary(sess, op);
            break;

+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+            supp = ggml_hexagon_supported_unary(sess, op);
+            break;
+
+        case GGML_OP_SUM_ROWS:
+            supp = ggml_hexagon_supported_sum_rows(sess, op);
+            break;
+
        case GGML_OP_SOFT_MAX:
            supp = ggml_hexagon_supported_softmax(sess, op);
            break;
@@ -2943,7 +3038,7 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
        case GGML_OP_GLU:
            {
                const auto glu_op = ggml_get_glu_op(op);
-                if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI)) {
+                if ((glu_op == GGML_GLU_OP_SWIGLU) || (glu_op == GGML_GLU_OP_SWIGLU_OAI) || (glu_op == GGML_GLU_OP_GEGLU)) {
                    supp = ggml_hexagon_supported_activations(sess, op);
                }
                break;
@@ -2968,6 +3063,10 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
            supp = ggml_hexagon_supported_cpy(sess, op);
            break;

+        case GGML_OP_ARGSORT:
+            supp = ggml_hexagon_supported_argsort(sess, op);
+            break;
+
        default:
            break;
    }
@@ -6,6 +6,7 @@ include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_fun.cmake)
 include_directories(
    ${HEXAGON_SDK_ROOT}/incs
    ${HEXAGON_SDK_ROOT}/incs/stddef
+    ${CMAKE_CURRENT_SOURCE_DIR}/../../../include
    ${CMAKE_CURRENT_SOURCE_DIR}/../..
    ${CMAKE_CURRENT_SOURCE_DIR}/..
    ${CMAKE_CURRENT_SOURCE_DIR}
@@ -21,6 +22,7 @@ add_library(${HTP_LIB} SHARED
    matmul-ops.c
    binary-ops.c
    unary-ops.c
+    sum-rows-ops.c
    softmax-ops.c
    act-ops.c
    rope-ops.c
@@ -28,6 +30,7 @@ add_library(${HTP_LIB} SHARED
    set-rows-ops.c
    get-rows-ops.c
    cpy-ops.c
+    argsort-ops.c
 )

 target_compile_definitions(${HTP_LIB} PRIVATE
@@ -410,7 +410,7 @@ static void unary_gelu_f32_per_thread(const struct htp_tensor * src0,
            // gelu = x * sigmoid(1.702 * x) // current implementation
            hvx_mul_scalar_f32((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (float) 1.702, ne0);
            hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
-            hvx_mul_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
+            hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
        }

        dma_queue_push_vtcm_to_ddr(dma_queue,
@@ -516,7 +516,7 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,

            // silu = x * sigmoid(x)
            hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, ne0);
-            hvx_mul_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
+            hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
        }

        dma_queue_push_vtcm_to_ddr(dma_queue,
@@ -541,6 +541,143 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
         ne03, src0_start_row, src0_end_row, ne0, ne1, ne2, ne3, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
 }

+static const float GELU_COEF_A     = 0.044715f;
+static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+
+static void glu_geglu_f32_per_thread(const struct htp_tensor * src0,
+                                       const struct htp_tensor * src1,
+                                       struct htp_tensor *       dst,
+                                       const int32_t *           op_params,
+                                       struct htp_spad *         src0_spad,
+                                       struct htp_spad *         src1_spad,
+                                       struct htp_spad *         dst_spad,
+                                       uint32_t                  nth,
+                                       uint32_t                  ith,
+                                       uint32_t                  src0_nrows_per_thread,
+                                       dma_queue *               dma_queue) {
+    htp_act_preamble3;
+
+    size_t src0_row_size = nb01;
+    size_t src1_row_size = nb11;
+    size_t dst_row_size  = nb1;
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return;
+    }
+
+    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
+    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
+    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+
+    const bool src1_valid = src1->ne[0];
+    const int  nc         = (src1_valid) ? ne00 : ne00 / 2;
+    if (!src1_valid) {
+        const int32_t swapped = op_params[1];
+        data_src1             = data_src0;
+        src1_row_size         = src0_row_size;
+
+        const size_t nc_in_bytes = nc * SIZEOF_FP32;
+        data_src0 += swapped ? nc_in_bytes : 0;
+        data_src1 += swapped ? 0 : nc_in_bytes;
+    }
+
+    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
+    const size_t src1_row_size_aligned = hex_round_up(src1_row_size, VLEN);
+    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+
+    uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread);
+    uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_spad->size_per_thread);
+    uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_spad->size_per_thread);
+
+    // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0
+    size_t src0_spad_half_size = src0_spad->size_per_thread / 2;
+    size_t src1_spad_half_size = src1_spad->size_per_thread / 2;
+    size_t dst_spad_half_size  = dst_spad->size_per_thread / 2;
+
+    const int BLOCK = src0_spad_half_size / src0_row_size_aligned;  // How many rows can we process in one block
+    if (BLOCK == 0) {
+        FARF(ERROR,
+             "geglu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
+             src0_spad->size_per_thread, src0_row_size_aligned);
+        return;
+    }
+
+    // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379
+    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
+
+        // Dummy DMA transation for sequencing (interleaving dst,src,dst,...)
+        dma_queue_push_vtcm_to_ddr(dma_queue,
+            dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)),
+            dst_row_size, dst_row_size_aligned, 0);
+
+        dma_queue_push_ddr_to_vtcm(dma_queue,
+            dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src0 + (ir * src0_row_size)),
+            src0_row_size_aligned, src0_row_size, block_size);
+        dma_queue_push_ddr_to_vtcm(dma_queue,
+            dma_make_ptr(src1_spad_data + (spad_idx * src1_spad_half_size), data_src1 + (ir * src1_row_size)),
+            src1_row_size_aligned, src1_row_size, block_size);
+    }
+
+    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
+
+        float * dst_spad  = (float *) dma_queue_pop(dma_queue).src;
+        float * src0_spad = (float *) dma_queue_pop(dma_queue).dst;
+        float * src1_spad = (float *) dma_queue_pop(dma_queue).dst;
+
+        for (uint32_t ib = 0; ib < block_size; ib++) {
+            const uint8_t * src0_spad_ptr = (const uint8_t *)(src0_spad + ib * (src0_row_size_aligned / sizeof(float)));
+            const uint8_t * src1_spad_ptr = (const uint8_t *)(src1_spad + ib * (src1_row_size_aligned / sizeof(float)));
+            uint8_t *       dst_spad_ptr  = (uint8_t *)(dst_spad + ib * (dst_row_size_aligned / sizeof(float)));
+
+            // geglu tanh implementation
+            // geglu(x, g) = gelu(x) * g
+            // gelu(x) = 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)))
+            hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, src0_spad_ptr, nc);                       // res = x*x
+            hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, GELU_COEF_A, nc);   // res = res * GELU_COEF_A
+            hvx_add_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, 1.0f, nc);          // res = res + 1.0f
+            hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, (const uint8_t *)dst_spad_ptr, nc);       // res = res * x
+            hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t*)dst_spad_ptr, SQRT_2_OVER_PI, nc); // res = result * SQRT_2_OVER_PI
+            hvx_tanh_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) dst_spad_ptr, nc);         // res = tanh(res)
+            hvx_add_scalar_f32_aa(dst_spad_ptr, (const uint8_t*)dst_spad_ptr, 1.0f, nc);           // res = res + 1.0f
+            hvx_mul_f32_aaa(dst_spad_ptr, src0_spad_ptr, (const uint8_t *)dst_spad_ptr, nc);       // res = res * x
+            hvx_mul_scalar_f32_aa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, 0.5f, nc);          // res = res + 0.5f
+            hvx_mul_f32_aaa(dst_spad_ptr, (const uint8_t *)dst_spad_ptr, src1_spad_ptr, nc);       // res = res * g
+        }
+
+        dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad), dst_row_size,
+                                   dst_row_size_aligned, block_size);
+
+        // prefetch N+2 loop iteration if any
+        const uint32_t pref_block = (ir + BLOCK * 2);
+        if (pref_block < src0_end_row) {
+            const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block);
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src0_spad, data_src0 + (pref_block * src0_row_size)),
+                                       src0_row_size_aligned, src0_row_size, pref_block_size);
+            dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src1_spad, data_src1 + (pref_block * src1_row_size)),
+                                       src1_row_size_aligned, src1_row_size, pref_block_size);
+        }
+    }
+
+    dma_queue_flush(dma_queue);
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "geglu-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+         ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3,
+         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
 static void unary_silu_f32(unsigned int n, unsigned int i, void * data) {
    struct htp_ops_context * octx = (struct htp_ops_context *) data;
    unary_silu_f32_per_thread(&octx->src0, &octx->dst, octx->op_params, &octx->src0_spad, &octx->dst_spad, n, i,
@@ -559,6 +696,12 @@ static void glu_swiglu_oai_f32(unsigned int n, unsigned int i, void * data) {
                                   &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
 }

+static void glu_geglu_f32(unsigned int n, unsigned int i, void * data) {
+    struct htp_ops_context * octx = (struct htp_ops_context *) data;
+    glu_geglu_f32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad,
+                               &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
+}
+
 static int execute_op_activations_f32(struct htp_ops_context * octx) {
    int err = HTP_STATUS_OK;

@@ -593,6 +736,11 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
            act_op_func = unary_gelu_f32;
            op_type     = "gelu-f32";
            break;
+
+        case HTP_OP_GLU_GEGLU:
+            act_op_func = glu_geglu_f32;
+            op_type     = "geglu-f32";
+            break;
        default:
            FARF(ERROR, "Unsupported activations Op %u\n", octx->op);
            return HTP_STATUS_NO_SUPPORT;
@@ -0,0 +1,281 @@
+#include <string.h>
+#include <stdlib.h>
+#include <math.h>
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "ggml.h"
+
+#include "hvx-utils.h"
+#include "hex-dma.h"
+
+#include "htp-ctx.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+
+#ifndef MIN
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#endif
+
+struct htp_argsort_context {
+    struct htp_ops_context * octx;
+    uint32_t                 nrows_per_thread;
+};
+
+static inline bool all_greater_f32(HVX_Vector x, HVX_Vector y)
+{
+    const HVX_Vector one  = Q6_V_vsplat_R(1);
+    const HVX_Vector zero = Q6_V_vzero();
+
+    HVX_VectorPred pred = Q6_Q_vcmp_gt_VsfVsf(x, y);
+    HVX_Vector matches = Q6_V_vmux_QVV(pred, one, zero);
+    HVX_Vector sum = hvx_vec_reduce_sum_i32(matches);
+    return hvx_vec_get_i32(sum) == 32;
+}
+
+// Sorts values and mirrors swaps to indices.
+static void quicksort_values_indices_asc(float * values, int32_t * indices, int left, int right) {
+    if (left >= right) return;
+
+    int pivot_idx = (left + right) / 2;
+    float pivot = values[pivot_idx];
+    int i = left;
+    int j = right;
+
+    HVX_Vector pivot_vec = hvx_vec_splat_f32(pivot);
+    while (i <= j) {
+        // Vectorized scan for i
+        while (i <= j) {
+            // Check if we have at least one full vector
+            if (i + 32 <= j) {
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + i);
+                if (all_greater_f32(pivot_vec, vals_vec)) {
+                    // If all elements are < pivot, we can skip this whole block
+                    i += 32;
+                    continue;
+                }
+            }
+
+            // Scalar fallback / cleanup
+            if (values[i] < pivot) {
+                i++;
+            } else {
+                break;
+            }
+        }
+
+        // Vectorized scan for j
+        while (i <= j) {
+            if (j - 32 >= i) {
+                // Load 32 elements ending at j.
+                // Since we want `values[j] > pivot`, let's load from j-31 to j.
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + j - 31);
+                if (all_greater_f32(vals_vec, pivot_vec)) {
+                    j -= 32;
+                    continue;
+                }
+            }
+
+            if (values[j] > pivot) {
+                j--;
+            } else {
+                break;
+            }
+        }
+
+        if (i <= j) {
+            float tmp_val = values[i];
+            values[i] = values[j];
+            values[j] = tmp_val;
+
+            int32_t tmp_idx = indices[i];
+            indices[i] = indices[j];
+            indices[j] = tmp_idx;
+            i++;
+            j--;
+        }
+    }
+
+    if (left < j) quicksort_values_indices_asc(values, indices, left, j);
+    if (i < right) quicksort_values_indices_asc(values, indices, i, right);
+}
+
+static void quicksort_values_indices_desc(float * values, int32_t * indices, int left, int right) {
+    if (left >= right) return;
+
+    int pivot_idx = (left + right) / 2;
+    float pivot = values[pivot_idx];
+    int i = left;
+    int j = right;
+
+    HVX_Vector pivot_vec = hvx_vec_splat_f32(pivot);
+
+    while (i <= j) {
+        // Vectorized scan for i (values[i] > pivot)
+        while (i <= j) {
+            if (i + 32 <= j) {
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + i);
+                if (all_greater_f32(vals_vec, pivot_vec)) {
+                    i += 32;
+                    continue;
+                }
+            }
+
+            if (values[i] > pivot) {
+                i++;
+            } else {
+                break;
+            }
+        }
+
+        // Vectorized scan for j (values[j] < pivot)
+        while (i <= j) {
+            if (j - 32 >= i) {
+                HVX_Vector vals_vec = *(HVX_UVector *)(values + j - 31);
+                if (all_greater_f32(pivot_vec, vals_vec)) {
+                    j -= 32;
+                    continue;
+                }
+            }
+
+            if (values[j] < pivot) {
+                j--;
+            } else {
+                break;
+            }
+        }
+
+        if (i <= j) {
+            float tmp_val = values[i];
+            values[i] = values[j];
+            values[j] = tmp_val;
+
+            int32_t tmp_idx = indices[i];
+            indices[i] = indices[j];
+            indices[j] = tmp_idx;
+            i++;
+            j--;
+        }
+    }
+
+    if (left < j) quicksort_values_indices_desc(values, indices, left, j);
+    if (i < right) quicksort_values_indices_desc(values, indices, i, right);
+}
+
+static void htp_argsort_f32(unsigned int n, unsigned int i, void * data) {
+    struct htp_argsort_context * actx = (struct htp_argsort_context *)data;
+    struct htp_ops_context * octx = actx->octx;
+
+    // Unpack context
+    const struct htp_tensor * src0 = &octx->src0;
+    const struct htp_tensor * dst = &octx->dst;
+
+    // Scratchpad memory
+    uint8_t * spad = octx->src0_spad.data + octx->src0_spad.size_per_thread * i;
+
+    // Dimensions
+    uint32_t ne00 = src0->ne[0];
+    uint32_t ne01 = src0->ne[1];
+    uint32_t ne02 = src0->ne[2];
+    uint32_t ne03 = src0->ne[3];
+
+    uint32_t nb01 = src0->nb[1];
+    //uint32_t nb02 = src0->nb[2];
+    //uint32_t nb03 = src0->nb[3];
+
+    uint32_t nb1 = dst->nb[1];
+    //uint32_t nb2 = dst->nb[2];
+    //uint32_t nb3 = dst->nb[3];
+
+    // Sort order
+    enum ggml_sort_order order = (enum ggml_sort_order) octx->op_params[0];
+
+    // Rows to process
+    uint32_t total_rows = ne01 * ne02 * ne03;
+    uint32_t rows_per_thread = actx->nrows_per_thread;
+    uint32_t start_row = rows_per_thread * i;
+    uint32_t end_row = MIN(start_row + rows_per_thread, total_rows);
+
+    // Scratchpad layout:
+    // We need space for one row of float data (values) and one row of int32 indices.
+    // values: ne00 * sizeof(float)
+    // indices: ne00 * sizeof(int32_t)
+    // Padded to 128 bytes.
+
+    size_t values_size = hex_round_up(ne00 * sizeof(float), 128);
+    float * values_buf = (float *) spad;
+    int32_t * indices_buf = (int32_t *) (spad + values_size);
+
+    for (uint32_t r = start_row; r < end_row; r++) {
+        uint32_t src_offset = r * nb01;
+        uint32_t dst_offset = r * nb1;
+
+        uint8_t * src_ptr = (uint8_t *) src0->data + src_offset;
+        uint8_t * dst_ptr = (uint8_t *) dst->data  + dst_offset;
+
+        hex_l2fetch(src_ptr, ne00 * sizeof(float), ne00 * sizeof(float), 1);
+        hvx_copy_f32_au((uint8_t*)values_buf, src_ptr, ne00);
+
+        // Initialize indices
+        for (uint32_t j = 0; j < ne00; j++) {
+            indices_buf[j] = j;
+        }
+
+        // Sort values and mirror swaps to indices
+        if (order == GGML_SORT_ORDER_ASC) {
+            quicksort_values_indices_asc(values_buf, indices_buf, 0, ne00 - 1);
+        } else {
+            quicksort_values_indices_desc(values_buf, indices_buf, 0, ne00 - 1);
+        }
+
+        // Copy indices back to DDR
+        hvx_copy_f32_ua(dst_ptr, (const uint8_t *) indices_buf, ne00);
+    }
+}
+
+int op_argsort(struct htp_ops_context * octx) {
+    // Check supported types
+    if (octx->src0.type != HTP_TYPE_F32) {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    // Allocate scratchpad
+    // We need 1 row of float + 1 row of int32 per thread.
+    uint32_t ne00 = octx->src0.ne[0];
+    size_t values_size  = hex_round_up(ne00 * sizeof(float), 128);
+    size_t indices_size = hex_round_up(ne00 * sizeof(int32_t), 128);
+    size_t spad_per_thread = values_size + indices_size;
+
+    // Make sure we round up to 256 for alignment requirements
+    spad_per_thread = hex_round_up(spad_per_thread, 256);
+
+    size_t total_spad_size = spad_per_thread * octx->n_threads;
+
+    if (octx->ctx->vtcm_size < total_spad_size) {
+        FARF(ERROR, "argsort: VTCM size too small. Needed %zu, have %zu", total_spad_size, octx->ctx->vtcm_size);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->src0_spad.size = total_spad_size;
+    octx->src0_spad.size_per_thread = spad_per_thread;
+
+    FARF(HIGH, "argsort: %ux%ux%ux%u -> %ux%ux%ux%u (0x%x, 0x%x)",
+         octx->src0.ne[0], octx->src0.ne[1], octx->src0.ne[2], octx->src0.ne[3],
+         octx->dst.ne[0], octx->dst.ne[1], octx->dst.ne[2], octx->dst.ne[3],
+         octx->src0.data, octx->dst.data);
+
+    uint32_t total_rows = octx->src0.ne[1] * octx->src0.ne[2] * octx->src0.ne[3];
+    uint32_t n_jobs = MIN(total_rows, octx->n_threads);
+
+    struct htp_argsort_context actx;
+    actx.octx = octx;
+    actx.nrows_per_thread = (total_rows + n_jobs - 1) / n_jobs;
+
+    // Run jobs
+    worker_pool_run_func(octx->ctx->worker_pool, htp_argsort_f32, &actx, n_jobs);
+
+    return HTP_STATUS_OK;
+}
@@ -42,32 +42,36 @@ enum htp_data_type {
    HTP_TYPE_COUNT
 };

-// These values are manually translated over to HTP
-// !!!! DO NOT ALTER THE ORDER OF THE FIRST FOUR ENUMS !!!!
+// Do not reorder first 4 (used as an index)
 enum htp_op {
-    HTP_OP_MUL            = 0,
-    HTP_OP_ADD            = 1,
-    HTP_OP_SUB            = 2,
-    HTP_OP_DIV            = 3,
-    HTP_OP_MUL_MAT        = 4,
-    HTP_OP_MUL_MAT_ID     = 5,
-    HTP_OP_RMS_NORM       = 6,
-    HTP_OP_UNARY_SILU     = 7,
-    HTP_OP_UNARY_GELU     = 8,
-    HTP_OP_GLU_SWIGLU     = 9,
-    HTP_OP_GLU_SWIGLU_OAI = 10,
-    HTP_OP_SOFTMAX        = 11,
-    HTP_OP_ADD_ID         = 12,
-    HTP_OP_ROPE           = 13,
-    HTP_OP_FLASH_ATTN_EXT = 14,
-    HTP_OP_SET_ROWS       = 15,
-    HTP_OP_SCALE          = 16,
-    HTP_OP_GET_ROWS       = 17,
-    HTP_OP_CPY            = 18,
+    HTP_OP_MUL = 0,
+    HTP_OP_ADD = 1,
+    HTP_OP_SUB = 2,
+    HTP_OP_DIV = 3,
+    HTP_OP_MUL_MAT,
+    HTP_OP_MUL_MAT_ID,
+    HTP_OP_RMS_NORM,
+    HTP_OP_UNARY_SILU,
+    HTP_OP_UNARY_GELU,
+    HTP_OP_GLU_SWIGLU,
+    HTP_OP_GLU_SWIGLU_OAI,
+    HTP_OP_GLU_GEGLU,
+    HTP_OP_SOFTMAX,
+    HTP_OP_ADD_ID,
+    HTP_OP_ROPE,
+    HTP_OP_FLASH_ATTN_EXT,
+    HTP_OP_SET_ROWS,
+    HTP_OP_GET_ROWS,
+    HTP_OP_SCALE,
+    HTP_OP_CPY,
+    HTP_OP_ARGSORT,
+    HTP_OP_SQR,
+    HTP_OP_SQRT,
+    HTP_OP_SUM_ROWS,
    INVALID
 };

-static inline size_t htp_type_block_size(uint32_t t) {
+static inline size_t htp_t_block_size(uint32_t t) {
    switch (t) {
        case HTP_TYPE_F32:
            return 1;
@@ -103,22 +107,6 @@ static inline size_t htp_type_nbytes(uint32_t t) {
    return 0;
 }

-static const char * htp_type_name(uint32_t t) {
-    switch (t) {
-        case HTP_TYPE_F32:
-            return "fp32";
-        case HTP_TYPE_F16:
-            return "fp16";
-        case HTP_TYPE_Q4_0:
-            return "q4_0";
-        case HTP_TYPE_Q8_0:
-            return "q8_0";
-        case HTP_TYPE_MXFP4:
-            return "mxfp4";
-    }
-    return 0;
-}
-
 // Internal types
 #define QK_Q4_0x4x2  256  // 4x Q4_0 blocks packed with next 4x Q4_0 blocks (size in bytes 128)
 #define QK_Q8_0x4x2  256  // 4x Q8_0 blocks concat with next 4x Q8_0 blocks
@@ -64,25 +64,12 @@ struct htp_ops_context {
    struct fastdiv_values broadcast_rv2;
    struct fastdiv_values broadcast_rv3;

-    struct fastdiv_values mm_div_ne12_ne1; // fastdiv values for ne12 * ne1
-    struct fastdiv_values mm_div_ne1;      // fastdiv values for ne1
-    struct fastdiv_values mm_div_r2;       // fastdiv values for ne12 / ne02
-    struct fastdiv_values mm_div_r3;       // fastdiv values for ne13 / ne03
-
    struct fastdiv_values set_rows_div_ne12; // fastdiv values for ne12
    struct fastdiv_values set_rows_div_ne11; // fastdiv values for ne11

    struct fastdiv_values get_rows_div_ne10;      // fastdiv values for ne10
    struct fastdiv_values get_rows_div_ne10_ne11; // fastdiv values for ne10 * ne11

-    struct fastdiv_values cpy_div_ne01; // fastdiv values for ne01
-    struct fastdiv_values cpy_div_ne02; // fastdiv values for ne02
-    struct fastdiv_values cpy_div_ne03; // fastdiv values for ne03
-
-    struct fastdiv_values cpy_rshp_div_n0;       // fastdiv values for ne00
-    struct fastdiv_values cpy_rshp_div_n1n0;     // fastdiv values for ne00*ne01
-    struct fastdiv_values cpy_rshp_div_n2n1n0;   // fastdiv values for ne00*ne01*ne02
-
    uint32_t flags;
 };

@@ -90,6 +77,7 @@ int op_matmul(struct htp_ops_context * octx);
 int op_matmul_id(struct htp_ops_context * octx);
 int op_binary(struct htp_ops_context * octx);
 int op_unary(struct htp_ops_context * octx);
+int op_sum_rows(struct htp_ops_context * octx);
 int op_activations(struct htp_ops_context * octx);
 int op_softmax(struct htp_ops_context * octx);
 int op_add_id(struct htp_ops_context * octx);
@@ -98,5 +86,6 @@ int op_flash_attn_ext(struct htp_ops_context * octx);
 int op_set_rows(struct htp_ops_context * octx);
 int op_get_rows(struct htp_ops_context * octx);
 int op_cpy(struct htp_ops_context * octx);
+int op_argsort(struct htp_ops_context * octx);

 #endif /* HTP_OPS_H */
@@ -46,127 +46,76 @@
 #define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
 #endif

-// ADD variants
+// Generic macro to define alignment permutations for an op
+#define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO) \
+static inline void OP_NAME##_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src0 % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src0 % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) dst % 128 == 0); \
+    hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a, OP_MACRO); \
+} \
+static inline void OP_NAME##_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src0 % 128 == 0); \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src0 % 128 == 0); \
+    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    assert((uintptr_t) src1 % 128 == 0); \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u, OP_MACRO); \
+} \
+static inline void OP_NAME##_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, OP_MACRO); \
+} \

-static inline void hvx_add_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_ADD);
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_add_f32, HVX_OP_ADD)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_sub_f32, HVX_OP_SUB)
+DEFINE_HVX_BINARY_OP_VARIANTS(hvx_mul_f32, HVX_OP_MUL)
+
+// Dispatcher logic
+#define HVX_BINARY_DISPATCHER(OP_NAME) \
+static inline void OP_NAME(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) { \
+    if (hex_is_aligned((void *) dst, 128)) { \
+        if (hex_is_aligned((void *) src0, 128)) { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aaa(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_aau(dst, src0, src1, num_elems); \
+        } else { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_aua(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_auu(dst, src0, src1, num_elems); \
+        } \
+    } else { \
+        if (hex_is_aligned((void *) src0, 128)) { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uaa(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_uau(dst, src0, src1, num_elems); \
+        } else { \
+            if (hex_is_aligned((void *) src1, 128)) OP_NAME##_uua(dst, src0, src1, num_elems); \
+            else                                    OP_NAME##_uuu(dst, src0, src1, num_elems); \
+        } \
+    } \
 }

-static inline void hvx_add_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_ADD);
-}
-
-static inline void hvx_add_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_ADD);
-}
-
-static inline void hvx_add_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_ADD);
-}
-
-// SUB variants
-
-static inline void hvx_sub_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_SUB);
-}
-
-static inline void hvx_sub_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_SUB);
-}
-
-static inline void hvx_sub_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_SUB);
-}
-
-static inline void hvx_sub_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_SUB);
-}
-
-// MUL variants
-
-static inline void hvx_mul_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a, HVX_OP_MUL);
-}
-
-static inline void hvx_mul_f32_au(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) dst % 128 == 0);
-    assert((unsigned long) src0 % 128 == 0);
-    hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a, HVX_OP_MUL);
-}
-
-static inline void hvx_mul_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    assert((unsigned long) src0 % 128 == 0);
-    assert((unsigned long) src1 % 128 == 0);
-    hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u, HVX_OP_MUL);
-}
-
-static inline void hvx_mul_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
-    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u, HVX_OP_MUL);
-}
-
-// Dispatchers
-
-static inline void hvx_add_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
-        if (hex_is_aligned((void *) src1, 128)) {
-            hvx_add_f32_aa(dst, src0, src1, num_elems);
-        } else {
-            hvx_add_f32_au(dst, src0, src1, num_elems);
-        }
-    } else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
-        hvx_add_f32_ua(dst, src0, src1, num_elems);
-    } else {
-        hvx_add_f32_uu(dst, src0, src1, num_elems);
-    }
-}
-
-static inline void hvx_sub_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
-        if (hex_is_aligned((void *) src1, 128)) {
-            hvx_sub_f32_aa(dst, src0, src1, num_elems);
-        } else {
-            hvx_sub_f32_au(dst, src0, src1, num_elems);
-        }
-    } else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
-        hvx_sub_f32_ua(dst, src0, src1, num_elems);
-    } else {
-        hvx_sub_f32_uu(dst, src0, src1, num_elems);
-    }
-}
-
-static inline void hvx_mul_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
-    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src0, 128)) {
-        if (hex_is_aligned((void *) src1, 128)) {
-            hvx_mul_f32_aa(dst, src0, src1, num_elems);
-        } else {
-            hvx_mul_f32_au(dst, src0, src1, num_elems);
-        }
-    } else if (hex_is_aligned((void *) src0, 128) && hex_is_aligned((void *) src1, 128)) {
-        hvx_mul_f32_ua(dst, src0, src1, num_elems);
-    } else {
-        hvx_mul_f32_uu(dst, src0, src1, num_elems);
-    }
-}
+HVX_BINARY_DISPATCHER(hvx_add_f32)
+HVX_BINARY_DISPATCHER(hvx_sub_f32)
+HVX_BINARY_DISPATCHER(hvx_mul_f32)

 // Mul-Mul Optimized
-
 static inline void hvx_mul_mul_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint8_t * restrict src2, const uint32_t num_elems) {
    assert((unsigned long) dst % 128 == 0);
    assert((unsigned long) src0 % 128 == 0);
@@ -443,6 +392,68 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
    }
 }

+//
+// Square
+//
+
+#define hvx_sqr_loop_body(dst_type, src_type, vec_store)           \
+    do {                                                                   \
+        dst_type * restrict vdst  = (dst_type *) dst;                      \
+        src_type * restrict vsrc = (src_type *) src;                       \
+                                                                           \
+        const uint32_t elem_size = sizeof(float);                          \
+        const uint32_t epv  = 128 / elem_size;                             \
+        const uint32_t nvec = n / epv;                                     \
+        const uint32_t nloe = n % epv;                                     \
+                                                                           \
+        uint32_t i = 0;                                                    \
+                                                                           \
+        _Pragma("unroll(4)")                                               \
+        for (; i < nvec; i++) {                                            \
+            vdst[i] = HVX_OP_MUL(vsrc[i], vsrc[i]);                        \
+        }                                                                  \
+        if (nloe) {                                                        \
+            HVX_Vector v = HVX_OP_MUL(vsrc[i], vsrc[i]);                   \
+            vec_store((void *) &vdst[i], nloe * elem_size, v);             \
+        }                                                                  \
+    } while(0)
+
+static inline void hvx_sqr_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqr_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_sqr_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqr_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqr_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqr_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_sqr_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) {
+    if (hex_is_aligned((void *) dst, 128)) {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_sqr_f32_aa(dst, src, num_elems);
+        } else {
+            hvx_sqr_f32_au(dst, src, num_elems);
+        }
+    } else {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_sqr_f32_ua(dst, src, num_elems);
+        } else {
+            hvx_sqr_f32_uu(dst, src, num_elems);
+        }
+    }
+}
+
 #undef HVX_OP_ADD
 #undef HVX_OP_SUB
 #undef HVX_OP_MUL
@@ -453,5 +464,7 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
 #undef hvx_scalar_loop_body
 #undef HVX_OP_MIN_SCALAR
 #undef HVX_OP_CLAMP_SCALAR
+#undef DEFINE_HVX_BINARY_OP_VARIANTS
+#undef HVX_BINARY_DISPATCHER

 #endif // HVX_ARITH_H
@@ -66,6 +66,12 @@ static inline float hvx_vec_get_f32(HVX_Vector v) {
    return x;
 }

+static inline int32_t hvx_vec_get_i32(HVX_Vector v) {
+    int32_t __attribute__((aligned(128))) x;
+    hvx_vec_store_a(&x, 4, v);
+    return x;
+}
+
 static inline HVX_Vector hvx_vec_abs_f16(HVX_Vector v) {
    // abs by clearing the fp16 sign bit
    HVX_Vector mask = Q6_Vh_vsplat_R(0x7fff);
@@ -136,8 +136,6 @@ static inline void hvx_copy_f32_uu(uint8_t * restrict dst, const uint8_t * restr
        dst_type * restrict vdst = (dst_type *) dst;                                \
        src_type * restrict vsrc = (src_type *) src;                                \
                                                                                    \
-        const HVX_Vector zero = Q6_V_vsplat_R(0);                                   \
-                                                                                    \
        const uint32_t elem_size = sizeof(__fp16);                                  \
        const uint32_t epv  = 128 / elem_size;                                      \
        const uint32_t nvec = n / epv;                                              \
@@ -0,0 +1,116 @@
+#ifndef HVX_DIV_H
+#define HVX_DIV_H
+
+#include <HAP_farf.h>
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "hvx-base.h"
+#include "hex-utils.h"
+#include "hvx-inverse.h"
+#include "hvx-arith.h"
+
+#if __HVX_ARCH__ < 79
+#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+#else
+#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+#endif
+
+#define hvx_div_f32_loop_body(dst_type, src0_type, src1_type, vec_store)             \
+    do {                                                                             \
+        dst_type * restrict vdst = (dst_type *) dst;                                 \
+        src0_type * restrict vsrc0 = (src0_type *) src0;                             \
+        src1_type * restrict vsrc1 = (src1_type *) src1;                             \
+                                                                                     \
+        const HVX_Vector nan_inf_mask = Q6_V_vsplat_R(0x7f800000);                   \
+                                                                                     \
+        const uint32_t nvec = n / VLEN_FP32;                                         \
+        const uint32_t nloe = n % VLEN_FP32;                                         \
+                                                                                     \
+        uint32_t i = 0;                                                              \
+                                                                                     \
+        _Pragma("unroll(4)")                                                         \
+        for (; i < nvec; i++) {                                                      \
+            HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
+            HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1);                         \
+            vdst[i] = res;                                                           \
+        }                                                                            \
+        if (nloe) {                                                                  \
+            HVX_Vector inv_src1 = hvx_vec_inverse_f32_guard(vsrc1[i], nan_inf_mask); \
+            HVX_Vector res = HVX_OP_MUL(vsrc0[i], inv_src1);                         \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP32, res);                   \
+        }                                                                            \
+    } while(0)
+
+// 3-letter suffix variants
+static inline void hvx_div_f32_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src0 % 128 == 0);
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src0 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) dst % 128 == 0);
+    hvx_div_f32_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_div_f32_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) src0 % 128 == 0);
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) src0 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    assert((uintptr_t) src1 % 128 == 0);
+    hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
+    hvx_div_f32_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_div_f32(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, const uint32_t num_elems) {
+    if (hex_is_aligned((void *) dst, 128)) {
+        if (hex_is_aligned((void *) src0, 128)) {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aaa(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_aau(dst, src0, src1, num_elems);
+        } else {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_aua(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_auu(dst, src0, src1, num_elems);
+        }
+    } else {
+        if (hex_is_aligned((void *) src0, 128)) {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uaa(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_uau(dst, src0, src1, num_elems);
+        } else {
+            if (hex_is_aligned((void *) src1, 128)) hvx_div_f32_uua(dst, src0, src1, num_elems);
+            else                                    hvx_div_f32_uuu(dst, src0, src1, num_elems);
+        }
+    }
+}
+
+#undef HVX_OP_MUL
+
+#endif // HVX_DIV_H
@@ -91,6 +91,27 @@ static inline HVX_Vector hvx_vec_tanh_f32(HVX_Vector x) {
        }                                                       \
    } while(0)

+#define hvx_tanh_loop_body(dst_type, src_type, vec_store)       \
+    do {                                                        \
+        dst_type * restrict vdst = (dst_type *) dst;            \
+        src_type * restrict vsrc = (src_type *) src;            \
+                                                                \
+        const uint32_t epv  = 128 / sizeof(float);              \
+        const uint32_t nvec = n / epv;                          \
+        const uint32_t nloe = n % epv;                          \
+                                                                \
+        uint32_t i = 0;                                         \
+                                                                \
+        _Pragma("unroll(4)")                                    \
+        for (; i < nvec; i++) {                                 \
+             vdst[i] = hvx_vec_tanh_f32(vsrc[i]);               \
+        }                                                       \
+        if (nloe) {                                             \
+             HVX_Vector tmp = hvx_vec_tanh_f32(vsrc[i]);        \
+             vec_store((void *) &vdst[i], nloe * sizeof(float), tmp); \
+        }                                                       \
+    } while(0)
+
 static inline void hvx_sigmoid_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
    assert((unsigned long) dst % 128 == 0);
    assert((unsigned long) src % 128 == 0);
@@ -111,4 +132,10 @@ static inline void hvx_sigmoid_f32_uu(uint8_t * restrict dst, const uint8_t * re
    hvx_sigmoid_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
 }

+static inline void hvx_tanh_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_tanh_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
 #endif /* HVX_SIGMOID_H */
@@ -12,11 +12,17 @@
 #define RSQRT_ONE_HALF     0x3f000000  // 0.5
 #define RSQRT_THREE_HALVES 0x3fc00000  // 1.5

+#if __HVX_ARCH__ < 79
+#define HVX_OP_MUL(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+#else
+#define HVX_OP_MUL(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+#endif
+
 static inline HVX_Vector hvx_vec_rsqrt_f32(HVX_Vector in_vec) {
    //Algorithm :
    //  x2 = input*0.5
    //  y  = * (long *) &input
-    //  y  = 0x5f3759df - (y>>2)
+    //  y  = 0x5f3759df - (y>>1)
    //  y  = y*(threehalfs - x2*y*y)

    HVX_Vector rsqrtconst = Q6_V_vsplat_R(RSQRT_CONST);
@@ -57,4 +63,64 @@ static inline HVX_Vector hvx_vec_rsqrt_f32(HVX_Vector in_vec) {
    return Q6_Vsf_equals_Vqf32(temp);
 }

+// Compute sqrt(x) as x*inv_sqrt(x)
+#define hvx_sqrt_f32_loop_body(dst_type, src_type, vec_store)                \
+    do {                                                                     \
+        dst_type * restrict vdst = (dst_type *) dst;                         \
+        src_type * restrict vsrc = (src_type *) src;                         \
+                                                                             \
+        const uint32_t nvec = n / VLEN_FP32;                                 \
+        const uint32_t nloe = n % VLEN_FP32;                                 \
+                                                                             \
+        uint32_t i = 0;                                                      \
+                                                                             \
+        _Pragma("unroll(4)")                                                 \
+        for (; i < nvec; i++) {                                              \
+            HVX_Vector inv_sqrt = hvx_vec_rsqrt_f32(vsrc[i]);                \
+            HVX_Vector sqrt_res = HVX_OP_MUL(inv_sqrt, vsrc[i]);             \
+            vdst[i] = sqrt_res;                                              \
+        }                                                                    \
+        if (nloe) {                                                          \
+            HVX_Vector inv_sqrt = hvx_vec_rsqrt_f32(vsrc[i]);                \
+            HVX_Vector sqrt_res = HVX_OP_MUL(inv_sqrt, vsrc[i]);             \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP32, sqrt_res);      \
+        }                                                                    \
+    } while(0)
+
+static inline void hvx_sqrt_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqrt_f32_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqrt_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_sqrt_f32_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqrt_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqrt_f32_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqrt_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_sqrt_f32_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqrt_f32(uint8_t * restrict dst, const uint8_t * restrict src, const int num_elems) {
+    if ((unsigned long) dst % 128 == 0) {
+        if ((unsigned long) src % 128 == 0) {
+            hvx_sqrt_f32_aa(dst, src, num_elems);
+        } else {
+            hvx_sqrt_f32_au(dst, src, num_elems);
+        }
+    } else {
+        if ((unsigned long) src % 128 == 0) {
+            hvx_sqrt_f32_ua(dst, src, num_elems);
+        } else {
+            hvx_sqrt_f32_uu(dst, src, num_elems);
+        }
+    }
+}
+
 #endif /* HVX_SQRT_H */
@@ -12,6 +12,7 @@
 #include "hvx-sigmoid.h"
 #include "hvx-sqrt.h"
 #include "hvx-arith.h"
+#include "hvx-div.h"
 #include "hvx-base.h"

 #endif /* HVX_UTILS_H */
@@ -440,6 +440,45 @@ static void proc_matmul_req(struct htp_context *     ctx,
    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }

+static void proc_argsort_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
+    struct dspqueue_buffer rsp_bufs[1];
+
+    // We had written to the output buffer, we'd also need to flush it
+    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
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.dst.data  = (uint32_t) bufs[1].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_argsort(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
+}
+
 static void proc_cpy_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
    struct dspqueue_buffer rsp_bufs[1];

@@ -679,6 +718,45 @@ static void proc_unary_req(struct htp_context * ctx, struct htp_general_req * re
    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
 }

+static void proc_sum_rows_req(struct htp_context * ctx, struct htp_general_req * req, struct dspqueue_buffer * bufs) {
+    struct dspqueue_buffer rsp_bufs[HTP_MAX_PACKET_BUFFERS];
+
+    // We had written to the output buffer, we'd also need to flush it
+    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
+    struct htp_ops_context octx = { 0 };
+    octx.ctx                    = ctx;
+    octx.src0                   = req->src0;
+    octx.dst                    = req->dst;
+    octx.flags                  = req->flags;
+    octx.op                     = req->op;
+
+    memcpy(octx.op_params, req->op_params, sizeof(octx.op_params));
+
+    // Update data pointers
+    octx.src0.data = (uint32_t) bufs[0].ptr;
+    octx.dst.data  = (uint32_t) bufs[1].ptr;
+    octx.n_threads = ctx->n_threads;
+
+    struct profile_data prof;
+    profile_start(&prof);
+
+    uint32_t rsp_status = HTP_STATUS_INTERNAL_ERR;
+    if (vtcm_acquire(ctx) == AEE_SUCCESS) {
+        rsp_status = op_sum_rows(&octx);
+        vtcm_release(ctx);
+    }
+
+    profile_stop(&prof);
+    send_htp_rsp(ctx, req->op, rsp_status, rsp_bufs, 1, &prof);
+}
+
 static void proc_activations_req(struct htp_context *     ctx,
                                 struct htp_general_req * req,
                                 struct dspqueue_buffer * bufs,
@@ -951,6 +1029,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
            case HTP_OP_MUL:
            case HTP_OP_ADD:
            case HTP_OP_SUB:
+            case HTP_OP_DIV:
                if (n_bufs != 3) {
                    FARF(ERROR, "Bad binary-req buffer list");
                    continue;
@@ -968,6 +1047,25 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
                proc_unary_req(ctx, &req, bufs);
                break;

+            case HTP_OP_SQR:
+            case HTP_OP_SQRT:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad unary-req buffer list");
+                    continue;
+                }
+
+                proc_unary_req(ctx, &req, bufs);
+                break;
+
+            case HTP_OP_SUM_ROWS:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad unary-req buffer list");
+                    continue;
+                }
+
+                proc_sum_rows_req(ctx, &req, bufs);
+                break;
+
            case HTP_OP_UNARY_SILU:
            case HTP_OP_UNARY_GELU:
                if (n_bufs != 2) {
@@ -980,6 +1078,7 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
            case HTP_OP_GLU_SWIGLU:
            case HTP_OP_GLU_SWIGLU_OAI:
            case HTP_OP_SOFTMAX:
+            case HTP_OP_GLU_GEGLU:
                if ((n_bufs != 2) && (n_bufs != 3)) {
                    FARF(ERROR, "Bad act-req buffer list");
                    continue;
@@ -1035,6 +1134,14 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
                proc_cpy_req(ctx, &req, bufs);
                break;

+            case HTP_OP_ARGSORT:
+                if (n_bufs != 2) {
+                    FARF(ERROR, "Bad argsort-req buffer list");
+                    continue;
+                }
+                proc_argsort_req(ctx, &req, bufs);
+                break;
+
            default:
                FARF(ERROR, "Unknown Op %u", req.op);
                break;
@@ -0,0 +1,115 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#include <string.h>
+#include <math.h>
+
+#include "hex-dma.h"
+#include "hvx-utils.h"
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+
+
+#define sum_rows_preamble                       \
+    struct htp_tensor *src0 =  &octx->src0;\
+    struct htp_tensor *dst  = &octx->dst;  \
+                                           \
+    const uint32_t ne00 = src0->ne[0];     \
+    const uint32_t ne01 = src0->ne[1];     \
+    const uint32_t ne02 = src0->ne[2];     \
+    const uint32_t ne03 = src0->ne[3];     \
+                                           \
+    const uint32_t nb00 = src0->nb[0];     \
+    const uint32_t nb01 = src0->nb[1];     \
+    const uint32_t nb02 = src0->nb[2];     \
+    const uint32_t nb03 = src0->nb[3];     \
+                                           \
+    const uint32_t  ne0 = dst->ne[0];      \
+    const uint32_t  ne1 = dst->ne[1];      \
+    const uint32_t  ne2 = dst->ne[2];      \
+    const uint32_t  ne3 = dst->ne[3];      \
+                                           \
+    const uint32_t  nb0 = dst->nb[0];      \
+    const uint32_t  nb1 = dst->nb[1];      \
+    const uint32_t  nb2 = dst->nb[2];      \
+    const uint32_t  nb3 = dst->nb[3];      \
+
+static int sum_rows_thread_f32(struct htp_ops_context * octx, const int nth, const int ith) {
+    sum_rows_preamble;
+
+    const uint32_t src0_nrows_per_thread  = octx->src0_nrows_per_thread;
+    const size_t src0_row_size = nb01;
+    const size_t dst_row_size  = nb1;
+
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+
+    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
+    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+
+    // no work for this thread
+    if (src0_start_row >= src0_end_row) {
+        return HTP_STATUS_OK;
+    }
+
+    int opt_path   = 0;
+    if ((0 == hex_is_aligned((void *) src0->data, VLEN)) && !(nb01 & (VLEN - 1))) {
+        opt_path = 1;
+    }
+
+    const uint8_t * restrict data_src = (const uint8_t *) src0->data;
+    uint8_t * restrict data_dst       = (uint8_t *) dst->data;
+
+    const float * restrict src_th = (float *) (data_src + (src0_start_row * src0_row_size));
+    float * restrict dst_th       = (float *) (data_dst + (src0_start_row * dst_row_size));
+
+    for (uint32_t ir = 0; ir < src0_nrows_per_thread; ir++) {
+        const float * restrict src_local = src_th + (ir * ne00);
+
+        if (ir + 1 < src0_nrows_per_thread) {
+            hex_l2fetch(src_local + ne00, src0_row_size, src0_row_size, 1);
+        }
+
+        if (1 == opt_path) {
+            dst_th[ir] = hvx_reduce_sum_f32_a((const uint8_t *) src_local, ne00);
+        } else {
+            dst_th[ir] = hvx_reduce_sum_f32((const uint8_t *) src_local, ne00);
+        }
+    }
+
+    return HTP_STATUS_OK;
+}
+
+static void sum_rows_work_f32(unsigned int n, unsigned int i, void *data) {
+    sum_rows_thread_f32((struct htp_ops_context *) data, n, i);
+}
+
+int op_sum_rows(struct htp_ops_context * octx) {
+    sum_rows_preamble;
+
+    if (octx->src0.type != HTP_TYPE_F32) {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
+
+    const int      n_threads  = octx->n_threads;
+    const uint32_t src0_nrows = ne01 * ne02 * ne03;
+
+    uint32_t n_jobs = MIN(n_threads, src0_nrows);
+    octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+
+    worker_pool_run_func(octx->ctx->worker_pool, sum_rows_work_f32, octx, n_jobs);
+
+    return HTP_STATUS_OK;
+}
+
@@ -132,6 +132,56 @@ static void rms_norm_htp_f32(const float * restrict src,
    }
 }

+static void sqr_htp_f32(const float * restrict src,
+                          float * restrict dst,
+                          uint8_t * restrict spad,
+                          const uint32_t num_rows,
+                          const uint32_t row_elems,
+                          const size_t   row_size,
+                          int32_t *      op_params,
+                          int            opt_path) {
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const float * restrict src_local = src + (ir * row_elems);
+        float * restrict dst_local       = dst + (ir * row_elems);
+
+        if (ir + 1 < num_rows) {
+            hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
+        }
+
+        if (1 == opt_path) {
+            hvx_sqr_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        } else {
+            hvx_sqr_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        }
+    }
+}
+
+static void sqrt_htp_f32(const float * restrict src,
+                          float * restrict dst,
+                          uint8_t * restrict spad,
+                          const uint32_t num_rows,
+                          const uint32_t row_elems,
+                          const size_t   row_size,
+                          int32_t *      op_params,
+                          int            opt_path) {
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const float * restrict src_local = src + (ir * row_elems);
+        float * restrict dst_local       = dst + (ir * row_elems);
+
+        if (ir + 1 < num_rows) {
+            hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
+        }
+
+        if (1 == opt_path) {
+            hvx_sqrt_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        } else {
+            hvx_sqrt_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
+        }
+    }
+}
+
 static void unary_job_f32_per_thread(const struct htp_tensor * src,
                                     struct htp_tensor *       dst,
                                     uint8_t *                 spad,
@@ -181,6 +231,12 @@ static void unary_job_f32_per_thread(const struct htp_tensor * src,
        case HTP_OP_SCALE:
            scale_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
            break;
+        case HTP_OP_SQR:
+            sqr_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
+            break;
+        case HTP_OP_SQRT:
+            sqrt_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
+            break;

        default:
            break;
@@ -218,6 +274,14 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
            unary_op_func = unary_job_dispatcher_f32;
            op_type       = "scale-f32";
            break;
+        case HTP_OP_SQR:
+            unary_op_func = unary_job_dispatcher_f32;
+            op_type       = "sqr-f32";
+            break;
+        case HTP_OP_SQRT:
+            unary_op_func = unary_job_dispatcher_f32;
+            op_type       = "sqrt-f32";
+            break;

        default:
            FARF(ERROR, "Unsupported unary Op %u\n", octx->op);
@@ -15,14 +15,22 @@ typedef struct ggml_metal * ggml_metal_t;
 ggml_metal_t ggml_metal_init(ggml_metal_device_t dev);
 void ggml_metal_free(ggml_metal_t ctx);

+const char * ggml_metal_get_name(ggml_metal_t ctx);
+
 void ggml_metal_synchronize(ggml_metal_t ctx);

 void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
 void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
+bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);

 enum ggml_status ggml_metal_graph_compute (ggml_metal_t ctx, struct ggml_cgraph * gf);
 void             ggml_metal_graph_optimize(ggml_metal_t ctx, struct ggml_cgraph * gf);

+void ggml_metal_event_record(ggml_metal_t ctx, ggml_metal_event_t ev);
+void ggml_metal_event_wait  (ggml_metal_t ctx, ggml_metal_event_t ev);
+
+ggml_metal_event_t ggml_metal_get_ev_cpy(ggml_metal_t ctx);
+
 void ggml_metal_set_n_cb            (ggml_metal_t ctx, int n_cb);
 void ggml_metal_set_abort_callback  (ggml_metal_t ctx, ggml_abort_callback abort_callback, void * user_data);
 bool ggml_metal_supports_family     (ggml_metal_t ctx, int family);
@@ -24,9 +24,13 @@ struct ggml_metal_command_buffer {
 };

 struct ggml_metal {
+    char name[128];
+
    ggml_metal_device_t  dev;
    ggml_metal_library_t lib;

+    ggml_metal_event_t ev_cpy; // for async copies
+
    dispatch_queue_t d_queue;

    // additional, inference-time compiled pipelines
@@ -117,7 +121,11 @@ 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);
+    res->ev_cpy = ggml_metal_device_event_init(dev);
+
+    const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
+
+    snprintf(res->name, sizeof(res->name), "%s", props_dev->name);

    res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);

@@ -206,9 +214,15 @@ void ggml_metal_free(ggml_metal_t ctx) {

    dispatch_release(ctx->d_queue);

+    ggml_metal_device_event_free(ctx->dev, ctx->ev_cpy);
+
    free(ctx);
 }

+const char * ggml_metal_get_name(ggml_metal_t ctx) {
+    return ctx->name;
+}
+
 void ggml_metal_synchronize(ggml_metal_t ctx) {
    // wait for any backend operations to finish
    if (ctx->cmd_buf_last) {
@@ -273,8 +287,8 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
        // wrap the source data into a Metal buffer
        id<MTLDevice> device = ggml_metal_device_get_obj(ctx->dev);
        id<MTLBuffer> buf_src = [device newBufferWithBytes:data
-                                                         length:size
-                                                        options:MTLResourceStorageModeShared];
+                                                    length:size
+                                                   options:MTLResourceStorageModeShared];

        GGML_ASSERT(buf_src);

@@ -316,9 +330,9 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
    @autoreleasepool {
        id<MTLDevice> device = ggml_metal_device_get_obj(ctx->dev);
        id<MTLBuffer> buf_dst = [device newBufferWithBytesNoCopy:data
-                                                               length:size
-                                                              options:MTLResourceStorageModeShared
-                                                          deallocator:nil];
+                                                          length:size
+                                                         options:MTLResourceStorageModeShared
+                                                     deallocator:nil];

        GGML_ASSERT(buf_dst);

@@ -356,9 +370,52 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
    }
 }

+bool ggml_metal_cpy_tensor_async(ggml_metal_t ctx_src, ggml_metal_t ctx_dst, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    @autoreleasepool {
+        struct ggml_metal_buffer_id bid_src = ggml_metal_get_buffer_id(src);
+        struct ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(dst);
+
+        if (bid_src.metal == nil || bid_dst.metal == nil) {
+            return false;
+        }
+
+        // queue the copy operation into the Metal context
+        // this will be queued at the end, after any currently ongoing GPU operations
+        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(ctx_src->dev);
+        id<MTLCommandBuffer> cmd_buf = [queue commandBuffer];
+        id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
+
+        [encoder copyFromBuffer:bid_src.metal
+                   sourceOffset:bid_src.offs
+                       toBuffer:bid_dst.metal
+              destinationOffset:bid_dst.offs
+                           size:ggml_nbytes(src)];
+
+        [encoder endEncoding];
+
+        ggml_metal_event_t ev_cpy = ggml_metal_get_ev_cpy(ctx_src);
+        ggml_metal_event_encode_signal(ev_cpy, cmd_buf);
+
+        [cmd_buf commit];
+
+        // do not wait here for completion
+        //[cmd_buf waitUntilCompleted];
+
+        // instead, remember a reference to the command buffer and wait for it later if needed
+        [ctx_src->cmd_bufs_ext addObject:cmd_buf];
+        ctx_src->cmd_buf_last = cmd_buf;
+
+        [cmd_buf retain];
+
+        ggml_metal_event_wait(ctx_dst, ev_cpy);
+
+        return true;
+    }
+}
+
 enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph * gf) {
    // number of nodes encoded by the main thread (empirically determined)
-    const int n_main = 64;
+    const int n_main = MAX(64, 0.1*gf->n_nodes);

    // number of threads in addition to the main thread
    const int n_cb = ctx->n_cb;
@@ -530,6 +587,42 @@ void ggml_metal_graph_optimize(ggml_metal_t ctx, struct ggml_cgraph * gf) {
    //printf("%s: graph optimize took %.3f ms\n", __func__, (ggml_time_us() - t_start) / 1000.0);
 }

+void ggml_metal_event_record(ggml_metal_t ctx, ggml_metal_event_t ev) {
+    @autoreleasepool {
+        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(ctx->dev);
+        id<MTLCommandBuffer> cmd_buf = [queue commandBuffer];
+
+        ggml_metal_event_encode_signal(ev, cmd_buf);
+
+        [cmd_buf commit];
+
+        [ctx->cmd_bufs_ext addObject:cmd_buf];
+        ctx->cmd_buf_last = cmd_buf;
+
+        [cmd_buf retain];
+    }
+}
+
+void ggml_metal_event_wait(ggml_metal_t ctx, ggml_metal_event_t ev) {
+    @autoreleasepool {
+        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(ctx->dev);
+        id<MTLCommandBuffer> cmd_buf = [queue commandBuffer];
+
+        ggml_metal_event_encode_wait(ev, cmd_buf);
+
+        [cmd_buf commit];
+
+        [ctx->cmd_bufs_ext addObject:cmd_buf];
+        ctx->cmd_buf_last = cmd_buf;
+
+        [cmd_buf retain];
+    }
+}
+
+ggml_metal_event_t ggml_metal_get_ev_cpy(ggml_metal_t ctx) {
+    return ctx->ev_cpy;
+}
+
 void ggml_metal_set_n_cb(ggml_metal_t ctx, int n_cb) {
    if (ctx->n_cb != n_cb) {
        ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_COMMAND_BUFFERS);
@@ -17,10 +17,12 @@ struct ggml_metal_device_deleter {

 typedef std::unique_ptr<ggml_metal_device, ggml_metal_device_deleter> ggml_metal_device_ptr;

-ggml_metal_device_t ggml_metal_device_get(void) {
-    static ggml_metal_device_ptr ctx { ggml_metal_device_init() };
+ggml_metal_device_t ggml_metal_device_get(int device) {
+    static std::vector<ggml_metal_device_ptr> devs;

-    return ctx.get();
+    devs.emplace_back(ggml_metal_device_init(device));
+
+    return devs.back().get();
 }

 struct ggml_metal_pipelines {
@@ -174,6 +176,26 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows(ggml_me
    return res;
 }

+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag(ggml_metal_library_t lib, const ggml_tensor * op) {
+    char base[256];
+    char name[256];
+
+    const int n = op->src[0]->ne[0];
+
+    snprintf(base, 256, "kernel_diag_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s_n=%d", base, n);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    res.nsg  = 1;
+    res.smem = 0;
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat(ggml_metal_library_t lib, ggml_type tsrc) {
    char base[256];
    char name[256];
@@ -190,61 +212,69 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat(ggml_meta
 }

 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary(ggml_metal_library_t lib, const ggml_tensor * op) {
-    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
-
    char base[256];
    char name[256];

-    const int64_t n = ggml_nelements(op);
+    int op_num = -1;

-    const char * op_str = "undefined";
    switch (op->op) {
-        case GGML_OP_SCALE:      op_str = "scale";      break;
-        case GGML_OP_FILL:       op_str = "fill";       break;
-        case GGML_OP_CLAMP:      op_str = "clamp";      break;
-        case GGML_OP_SQR:        op_str = "sqr";        break;
-        case GGML_OP_SQRT:       op_str = "sqrt";       break;
-        case GGML_OP_SIN:        op_str = "sin";        break;
-        case GGML_OP_COS:        op_str = "cos";        break;
-        case GGML_OP_LOG:        op_str = "log";        break;
-        case GGML_OP_LEAKY_RELU: op_str = "leaky_relu"; break;
+        case GGML_OP_SCALE:      op_num = OP_UNARY_NUM_SCALE;      break;
+        case GGML_OP_FILL:       op_num = OP_UNARY_NUM_FILL;       break;
+        case GGML_OP_CLAMP:      op_num = OP_UNARY_NUM_CLAMP;      break;
+        case GGML_OP_SQR:        op_num = OP_UNARY_NUM_SQR;        break;
+        case GGML_OP_SQRT:       op_num = OP_UNARY_NUM_SQRT;       break;
+        case GGML_OP_SIN:        op_num = OP_UNARY_NUM_SIN;        break;
+        case GGML_OP_COS:        op_num = OP_UNARY_NUM_COS;        break;
+        case GGML_OP_LOG:        op_num = OP_UNARY_NUM_LOG;        break;
+        case GGML_OP_LEAKY_RELU: op_num = OP_UNARY_NUM_LEAKY_RELU; break;
        case GGML_OP_UNARY:
            switch (ggml_get_unary_op(op)) {
-                case GGML_UNARY_OP_TANH:        op_str = "tanh";        break;
-                case GGML_UNARY_OP_RELU:        op_str = "relu";        break;
-                case GGML_UNARY_OP_SIGMOID:     op_str = "sigmoid";     break;
-                case GGML_UNARY_OP_GELU:        op_str = "gelu";        break;
-                case GGML_UNARY_OP_GELU_ERF:    op_str = "gelu_erf";    break;
-                case GGML_UNARY_OP_GELU_QUICK:  op_str = "gelu_quick";  break;
-                case GGML_UNARY_OP_SILU:        op_str = "silu";        break;
-                case GGML_UNARY_OP_ELU:         op_str = "elu";         break;
-                case GGML_UNARY_OP_NEG:         op_str = "neg";         break;
-                case GGML_UNARY_OP_ABS:         op_str = "abs";         break;
-                case GGML_UNARY_OP_SGN:         op_str = "sgn";         break;
-                case GGML_UNARY_OP_STEP:        op_str = "step";        break;
-                case GGML_UNARY_OP_HARDSWISH:   op_str = "hardswish";   break;
-                case GGML_UNARY_OP_HARDSIGMOID: op_str = "hardsigmoid"; break;
-                case GGML_UNARY_OP_EXP:         op_str = "exp";         break;
-                case GGML_UNARY_OP_SOFTPLUS:    op_str = "softplus";    break;
-                case GGML_UNARY_OP_EXPM1:       op_str = "expm1";       break;
+                case GGML_UNARY_OP_TANH:        op_num = OP_UNARY_NUM_TANH;        break;
+                case GGML_UNARY_OP_RELU:        op_num = OP_UNARY_NUM_RELU;        break;
+                case GGML_UNARY_OP_SIGMOID:     op_num = OP_UNARY_NUM_SIGMOID;     break;
+                case GGML_UNARY_OP_GELU:        op_num = OP_UNARY_NUM_GELU;        break;
+                case GGML_UNARY_OP_GELU_ERF:    op_num = OP_UNARY_NUM_GELU_ERF;    break;
+                case GGML_UNARY_OP_GELU_QUICK:  op_num = OP_UNARY_NUM_GELU_QUICK;  break;
+                case GGML_UNARY_OP_SILU:        op_num = OP_UNARY_NUM_SILU;        break;
+                case GGML_UNARY_OP_ELU:         op_num = OP_UNARY_NUM_ELU;         break;
+                case GGML_UNARY_OP_NEG:         op_num = OP_UNARY_NUM_NEG;         break;
+                case GGML_UNARY_OP_ABS:         op_num = OP_UNARY_NUM_ABS;         break;
+                case GGML_UNARY_OP_SGN:         op_num = OP_UNARY_NUM_SGN;         break;
+                case GGML_UNARY_OP_STEP:        op_num = OP_UNARY_NUM_STEP;        break;
+                case GGML_UNARY_OP_HARDSWISH:   op_num = OP_UNARY_NUM_HARDSWISH;   break;
+                case GGML_UNARY_OP_HARDSIGMOID: op_num = OP_UNARY_NUM_HARDSIGMOID; break;
+                case GGML_UNARY_OP_EXP:         op_num = OP_UNARY_NUM_EXP;         break;
+                case GGML_UNARY_OP_SOFTPLUS:    op_num = OP_UNARY_NUM_SOFTPLUS;    break;
+                case GGML_UNARY_OP_EXPM1:       op_num = OP_UNARY_NUM_EXPM1;       break;
                default: GGML_ABORT("fatal error");
            } break;
        default: GGML_ABORT("fatal error");
    };

-    const char * suffix = "";
-    if (n % 4 == 0) {
-        suffix = "_4";
-    }
+    const char * t0_str = ggml_type_name(op->src[0]->type);
+    const char * t_str  = ggml_type_name(op->type);

-    snprintf(base, 256, "kernel_%s_%s%s", op_str, ggml_type_name(op->src[0]->type), suffix);
-    snprintf(name, 256, "%s", base);
+    const bool is_c4 = op->src[0]->ne[0] % 4 == 0;
+    const bool is_cnt = ggml_is_contiguous(op->src[0]) && ggml_nelements(op) < 32768;
+
+    snprintf(base, 256, "kernel_unary_%s_%s%s", t0_str, t_str, is_c4 ? "_4" : "");
+    snprintf(name, 256, "%s_op=%d_cnt=%d", base, op_num, is_cnt);

    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
    if (!res.pipeline) {
-        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, op_num, FC_UNARY + 0);
+        ggml_metal_cv_set_bool (cv, is_cnt, FC_UNARY + 1);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
    }

+    res.c4  = is_c4;
+    res.cnt = is_cnt;
+
    return res;
 }

@@ -532,6 +562,36 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv(ggml_metal_
    return res;
 }

+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri(ggml_metal_library_t lib, const ggml_tensor * op) {
+    char base[256];
+    char name[256];
+
+    const int nsg = 8;
+    const int n   = op->src[1]->ne[1];
+    const int k   = op->src[1]->ne[0];
+
+    snprintf(base, 256, "kernel_solve_tri_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s_nsg=%d_n=%d_k=%d", base, nsg, n, k);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, nsg, FC_SOLVE_TRI + 0);
+        ggml_metal_cv_set_int16(cv, n,   FC_SOLVE_TRI + 1);
+        ggml_metal_cv_set_int16(cv, k,   FC_SOLVE_TRI + 2);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
+    }
+
+    res.nsg  = nsg;
+    res.smem = GGML_PAD(GGML_PAD(n, 32)*nsg*sizeof(float), 16);
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext(ggml_metal_library_t lib, ggml_type tsrc0, ggml_type tsrc1, int nsg, int nxpsg, int r1ptg) {
    char base[256];
    char name[256];
@@ -1340,34 +1400,78 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
    GGML_UNUSED(op);
 }

-ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(
-        ggml_metal_library_t lib,
-        ggml_op op,
-        int32_t n_fuse,
-        bool row) {
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(ggml_metal_library_t lib, const ggml_tensor * op, int32_t n_fuse) {
    char base[256];
    char name[256];

-    const char * op_str = "undefined";
-    switch (op) {
-        case GGML_OP_ADD:   op_str = "add";   break;
-        case GGML_OP_SUB:   op_str = "sub";   break;
-        case GGML_OP_MUL:   op_str = "mul";   break;
-        case GGML_OP_DIV:   op_str = "div";   break;
+    int op_num = -1;
+
+    switch (op->op) {
+        case GGML_OP_ADD: op_num = 0; break;
+        case GGML_OP_SUB: op_num = 1; break;
+        case GGML_OP_MUL: op_num = 2; break;
+        case GGML_OP_DIV: op_num = 3; break;
        default: GGML_ABORT("fatal error");
    };

-    if (row) {
-        snprintf(base, 256, "kernel_%s_row_c4_fuse_%d", op_str, n_fuse);
-    } else {
-        snprintf(base, 256, "kernel_%s_fuse_%d", op_str, n_fuse);
-    }
+    const char * t0_str = ggml_type_name(op->src[0]->type);
+    const char * t1_str = ggml_type_name(op->src[1]->type);
+    const char * t_str  = ggml_type_name(op->type);

-    snprintf(name, 256, "%s", base);
+    const bool is_c4 = (op->src[0]->ne[0] % 4 == 0) && (op->src[1]->ne[0] % 4 == 0);
+
+    const bool is_rb = ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]) && (ggml_nrows(op->src[1]) == 1) && ggml_nelements(op) < 65536;
+
+    snprintf(base, 256, "kernel_bin_fuse_%s_%s_%s%s", t0_str, t1_str, t_str, is_c4 ? "_4" : "");
+    snprintf(name, 256, "%s_op=%d_nf=%d_rb=%d", base, op_num, n_fuse, is_rb);

    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
    if (!res.pipeline) {
-        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, op_num, FC_BIN + 0);
+        ggml_metal_cv_set_int16(cv, n_fuse, FC_BIN + 1);
+        ggml_metal_cv_set_bool (cv, is_rb,  FC_BIN + 2);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
+    }
+
+    res.c4  = is_c4;
+    res.cnt = is_rb;
+
+    return res;
+}
+
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin_one(ggml_metal_library_t lib, ggml_op op) {
+    char base[256];
+    char name[256];
+
+    int op_num = -1;
+
+    switch (op) {
+        case GGML_OP_ADD: op_num = 0; break;
+        case GGML_OP_SUB: op_num = 1; break;
+        case GGML_OP_MUL: op_num = 2; break;
+        case GGML_OP_DIV: op_num = 3; break;
+        default: GGML_ABORT("fatal error");
+    };
+
+    snprintf(base, 256, "kernel_bin_fuse_%s_%s_%s", "f32", "f32", "f32");
+    snprintf(name, 256, "%s_op=%d_nf=%d", base, op_num, 1);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, op_num, FC_BIN + 0);
+        ggml_metal_cv_set_int16(cv, 1,      FC_BIN + 1);
+        ggml_metal_cv_set_bool (cv, false,  FC_BIN + 2);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
    }

    return res;
@@ -1376,13 +1480,15 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm(ggml_metal_library_t lib, const ggml_tensor * op) {
    assert(op->op == GGML_OP_L2_NORM);

-    GGML_ASSERT(op->src[0]->ne[0] % 4 == 0);
-    GGML_ASSERT(ggml_is_contiguous_1(op->src[0]));
-
    char base[256];
    char name[256];

-    snprintf(base, 256, "kernel_l2_norm_f32");
+    const bool is_c4 = op->src[0]->ne[0] % 4 == 0;
+
+    const char * t0_str = ggml_type_name(op->src[0]->type);
+    const char * t_str  = ggml_type_name(op->type);
+
+    snprintf(base, 256, "kernel_l2_norm_%s_%s%s", t0_str, t_str, is_c4 ? "_4" : "");
    snprintf(name, 256, "%s", base);

    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
@@ -1390,6 +1496,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm(ggml_met
        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
    }

+    res.c4   = is_c4;
    res.smem = 32*sizeof(float);

    return res;
@@ -53,6 +53,9 @@ struct ggml_metal_pipeline_with_params {
    int nr1;

    size_t smem;
+
+    bool c4;
+    bool cnt;
 };

 int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_with_params pipeline);
@@ -108,6 +111,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_1d
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op, enum ggml_op_pool op_pool);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_get_rows          (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows          (ggml_metal_library_t lib, enum ggml_type tidx, enum ggml_type tdst);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag              (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary             (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -121,6 +125,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv              (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext        (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1, int nsg, int nxpsg, int r1ptg);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv            (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -132,7 +137,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort_merge     (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k             (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
-struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse );
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin_one           (ggml_metal_library_t lib, enum ggml_op op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_norm              (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse);
@@ -205,7 +211,9 @@ void ggml_metal_rsets_free(ggml_metal_rsets_t rsets);
 //

 struct ggml_metal_device_props {
+    int device;
    char name[128];
+    char desc[128];

    size_t max_buffer_size;
    size_t max_working_set_size;
@@ -224,11 +232,15 @@ struct ggml_metal_device_props {
    int op_offload_min_batch_size;
 };

-ggml_metal_device_t ggml_metal_device_init(void);
+typedef struct ggml_metal_event * ggml_metal_event_t;
+
+void ggml_metal_event_encode_signal(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf);
+void ggml_metal_event_encode_wait  (ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf);
+
+ggml_metal_device_t ggml_metal_device_init(int device);
 void ggml_metal_device_free(ggml_metal_device_t dev);

-// return a singleton that is automatically destroyed when the program exits
-ggml_metal_device_t ggml_metal_device_get(void);
+ggml_metal_device_t ggml_metal_device_get(int device);

 void * ggml_metal_device_get_obj  (ggml_metal_device_t dev); // id<MTLDevice>
 void * ggml_metal_device_get_queue(ggml_metal_device_t dev); // id<MTLCommandQueue>
@@ -240,6 +252,10 @@ void ggml_metal_device_rsets_rm (ggml_metal_device_t dev, ggml_metal_rset_t rset

 void ggml_metal_device_rsets_keep_alive(ggml_metal_device_t dev);

+ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev);
+void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev);
+void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_event_t ev);
+
 void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total);
 bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_tensor * op);

@@ -24,9 +24,6 @@
 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;
-
 #if !GGML_METAL_EMBED_LIBRARY
 // Here to assist with NSBundle Path Hack
@interface GGMLMetalClass : NSObject
@@ -349,10 +346,12 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline(ggml_meta

    struct ggml_metal_pipeline_with_params res = {
        /*.pipeline =*/ nil,
+        /*.nsg      =*/ 0,
        /*.nr0      =*/ 0,
        /*.nr1      =*/ 0,
-        /*.nsg      =*/ 0,
        /*.smem     =*/ 0,
+        /*.c4       =*/ false,
+        /*.cnt      =*/ false,
    };

    res.pipeline = ggml_metal_pipelines_get(lib->pipelines, name);
@@ -365,10 +364,12 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline(ggml_meta
 struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_metal_library_t lib, const char * base, const char * name, ggml_metal_cv_t cv) {
    struct ggml_metal_pipeline_with_params res = {
        /*.pipeline =*/ nil,
+        /*.nsg      =*/ 0,
        /*.nr0      =*/ 0,
        /*.nr1      =*/ 0,
-        /*.nsg      =*/ 0,
        /*.smem     =*/ 0,
+        /*.c4       =*/ false,
+        /*.cnt      =*/ false,
    };

    [lib->lock lock];
@@ -523,6 +524,9 @@ struct ggml_metal_device {
    ggml_metal_library_t library;

    struct ggml_metal_device_props props;
+
+    // virtual address for GPU memory allocations
+    atomic_uintptr_t addr_virt;
 };

 //
@@ -618,7 +622,7 @@ void ggml_metal_rsets_free(ggml_metal_rsets_t rsets) {
    free(rsets);
 }

-ggml_metal_device_t ggml_metal_device_init(void) {
+ggml_metal_device_t ggml_metal_device_init(int device) {
    ggml_metal_device_t dev = calloc(1, sizeof(struct ggml_metal_device));

    assert(dev != NULL);
@@ -632,6 +636,9 @@ ggml_metal_device_t ggml_metal_device_init(void) {
                GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
            }

+            dev->addr_virt = 0x000000400ULL;
+
+            dev->props.device = device;
            dev->props.has_simdgroup_reduction  = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7];
            dev->props.has_simdgroup_reduction |= [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];

@@ -792,7 +799,8 @@ ggml_metal_device_t ggml_metal_device_init(void) {
                dev->props.max_working_set_size   = dev->mtl_device.maxBufferLength;
            }

-            strncpy(dev->props.name, [[dev->mtl_device name] UTF8String], sizeof(dev->props.name) - 1);
+            snprintf(dev->props.name, sizeof(dev->props.name), "%s%d", "MTL", device);
+            snprintf(dev->props.desc, sizeof(dev->props.desc), "%s", [[dev->mtl_device name] UTF8String]);

            dev->library = ggml_metal_library_init(dev);
            if (!dev->library) {
@@ -922,6 +930,59 @@ void ggml_metal_device_rsets_keep_alive(ggml_metal_device_t dev) {
    atomic_store_explicit(&dev->rsets->d_loop, 2*dev->rsets->keep_alive_s, memory_order_relaxed);
 }

+struct ggml_metal_event {
+    void * obj; // id<MTLEvent>
+
+    atomic_int value;
+};
+
+void ggml_metal_event_encode_signal(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf_raw) {
+    id<MTLEvent> event = (id<MTLEvent>)ev->obj;
+
+    id<MTLCommandBuffer> cmd_buf = (id<MTLCommandBuffer>) cmd_buf_raw;
+
+    [cmd_buf encodeSignalEvent:event value:atomic_fetch_add_explicit(&ev->value, 1, memory_order_relaxed) + 1];
+}
+
+void ggml_metal_event_encode_wait(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf_raw) {
+    id<MTLEvent> event = (id<MTLEvent>)ev->obj;
+
+    id<MTLCommandBuffer> cmd_buf = (id<MTLCommandBuffer>) cmd_buf_raw;
+
+    [cmd_buf encodeWaitForEvent:event value:atomic_load_explicit(&ev->value, memory_order_relaxed)];
+}
+
+ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev) {
+    id<MTLEvent> event = [dev->mtl_device newEvent];
+
+    ggml_metal_event_t ev = calloc(1, sizeof(struct ggml_metal_event));
+
+    ev->obj = (__bridge void *)event;
+    ev->value = 0;
+
+    return ev;
+}
+
+void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev) {
+    id<MTLEvent> event = ev->obj;
+    [event release];
+
+    free(ev);
+
+    GGML_UNUSED(dev);
+}
+
+void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_event_t ev) {
+    @autoreleasepool {
+        id<MTLEvent> event = ev->obj;
+
+        id<MTLCommandBuffer> cmd_buf = [dev->mtl_queue commandBuffer];
+        [cmd_buf encodeWaitForEvent:event value:atomic_load_explicit(&ev->value, memory_order_relaxed)];
+        [cmd_buf commit];
+        [cmd_buf waitUntilCompleted];
+    }
+}
+
 void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total) {
    if (@available(macOS 10.12, iOS 16.0, *)) {
        *total = dev->mtl_device.recommendedMaxWorkingSetSize;
@@ -950,6 +1011,15 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
    }

    switch (op->op) {
+        case GGML_OP_SCALE:
+        case GGML_OP_FILL:
+        case GGML_OP_CLAMP:
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_SIN:
+        case GGML_OP_COS:
+        case GGML_OP_LOG:
+            return ggml_is_contiguous_rows(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
        case GGML_OP_UNARY:
            switch (ggml_get_unary_op(op)) {
                case GGML_UNARY_OP_TANH:
@@ -969,7 +1039,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                case GGML_UNARY_OP_EXP:
                case GGML_UNARY_OP_SOFTPLUS:
                case GGML_UNARY_OP_EXPM1:
-                    return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+                    return ggml_is_contiguous_rows(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
                default:
                    return false;
            }
@@ -997,11 +1067,9 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
        case GGML_OP_MUL:
        case GGML_OP_DIV:
        case GGML_OP_ADD_ID:
-            return op->src[0]->type == GGML_TYPE_F32;
+            return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && op->src[0]->type == GGML_TYPE_F32;
        case GGML_OP_ACC:
        case GGML_OP_REPEAT:
-        case GGML_OP_SCALE:
-        case GGML_OP_FILL:
        case GGML_OP_CONV_TRANSPOSE_1D:
            return true;
        case GGML_OP_CONV_TRANSPOSE_2D:
@@ -1009,14 +1077,6 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                (op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32) &&
                op->src[1]->type == GGML_TYPE_F32 &&
                op->type == GGML_TYPE_F32;
-        case GGML_OP_CLAMP:
-            return op->src[0]->type == GGML_TYPE_F32;
-        case GGML_OP_SQR:
-        case GGML_OP_SQRT:
-        case GGML_OP_SIN:
-        case GGML_OP_COS:
-        case GGML_OP_LOG:
-            return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
        case GGML_OP_SUM:
            return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
        case GGML_OP_TRI:
@@ -1026,9 +1086,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
        case GGML_OP_MEAN:
        case GGML_OP_SOFT_MAX:
        case GGML_OP_GROUP_NORM:
-            return has_simdgroup_reduction && ggml_is_contiguous_rows(op->src[0]);
        case GGML_OP_L2_NORM:
-            return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(op->src[0]));
+            return has_simdgroup_reduction && ggml_is_contiguous_rows(op->src[0]);
        case GGML_OP_COUNT_EQUAL:
            return has_simdgroup_reduction &&
                op->src[0]->type == GGML_TYPE_I32 &&
@@ -1096,6 +1155,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
        case GGML_OP_RWKV_WKV6:
        case GGML_OP_RWKV_WKV7:
            return true;
+        case GGML_OP_SOLVE_TRI:
        case GGML_OP_MUL_MAT:
        case GGML_OP_MUL_MAT_ID:
            return has_simdgroup_reduction;
@@ -1177,6 +1237,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                        return false;
                };
            }
+        case GGML_OP_DIAG:
+            return true;
        case GGML_OP_OPT_STEP_ADAMW:
        case GGML_OP_OPT_STEP_SGD:
            return has_simdgroup_reduction;
@@ -1344,8 +1406,8 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
        res->all_data = ggml_metal_host_malloc(size_aligned);
        res->is_shared = true;
    } else {
-        // use virtual address from g_addr_device counter
-        res->all_data = (void *) atomic_fetch_add_explicit(&g_addr_device, size_aligned, memory_order_relaxed);
+        // use virtual address
+        res->all_data = (void *) atomic_fetch_add_explicit(&dev->addr_virt, size_aligned, memory_order_relaxed);
        res->is_shared = false;
    }
    res->all_size = size_aligned;
@@ -78,15 +78,47 @@
 #define FC_MUL_MM                      700
 #define FC_ROPE                        800
 #define FC_SSM_CONV                    900
-#define FC_COUNT_EQUAL                 1000
+#define FC_SOLVE_TRI                   1000
+#define FC_COUNT_EQUAL                 1100
+#define FC_UNARY                       1200
+#define FC_BIN                         1300

 // op-specific constants
-#define OP_FLASH_ATTN_EXT_NQPTG 8
+#define OP_FLASH_ATTN_EXT_NQPSG 8
 #define OP_FLASH_ATTN_EXT_NCPSG 64

-#define OP_FLASH_ATTN_EXT_VEC_NQPTG 1
+#define OP_FLASH_ATTN_EXT_VEC_NQPSG 1
 #define OP_FLASH_ATTN_EXT_VEC_NCPSG 32

+#define OP_UNARY_NUM_SCALE      10
+#define OP_UNARY_NUM_FILL       11
+#define OP_UNARY_NUM_CLAMP      12
+#define OP_UNARY_NUM_SQR        13
+#define OP_UNARY_NUM_SQRT       14
+#define OP_UNARY_NUM_SIN        15
+#define OP_UNARY_NUM_COS        16
+#define OP_UNARY_NUM_LOG        17
+#define OP_UNARY_NUM_LEAKY_RELU 18
+
+#define OP_UNARY_NUM_TANH        100
+#define OP_UNARY_NUM_RELU        101
+#define OP_UNARY_NUM_SIGMOID     102
+#define OP_UNARY_NUM_GELU        103
+#define OP_UNARY_NUM_GELU_ERF    104
+#define OP_UNARY_NUM_GELU_QUICK  105
+#define OP_UNARY_NUM_SILU        106
+#define OP_UNARY_NUM_ELU         107
+#define OP_UNARY_NUM_NEG         108
+#define OP_UNARY_NUM_ABS         109
+#define OP_UNARY_NUM_SGN         110
+#define OP_UNARY_NUM_STEP        111
+#define OP_UNARY_NUM_HARDSWISH   112
+#define OP_UNARY_NUM_HARDSIGMOID 113
+#define OP_UNARY_NUM_EXP         114
+#define OP_UNARY_NUM_SOFTPLUS    115
+#define OP_UNARY_NUM_EXPM1       116
+
+
 // kernel argument structs
 //
 // - element counters (e.g. ne00) typically use int32_t to reduce register usage
@@ -122,6 +154,31 @@ typedef struct {
    int32_t  dim;
 } ggml_metal_kargs_concat;

+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    float    slope;
+    float    scale;
+    float    bias;
+    float    val;
+    float    min;
+    float    max;
+} ggml_metal_kargs_unary;
+
 typedef struct {
    int32_t  ne00;
    int32_t  ne01;
@@ -179,20 +236,6 @@ typedef struct {
    uint64_t nb3;
 } ggml_metal_kargs_repeat;

-typedef struct {
-    float scale;
-    float bias;
-} ggml_metal_kargs_scale;
-
-typedef struct {
-    float val;
-} ggml_metal_kargs_fill;
-
-typedef struct {
-    float min;
-    float max;
-} ggml_metal_kargs_clamp;
-
 typedef struct {
    int64_t  nk0;
    int64_t  ne00;
@@ -496,8 +539,21 @@ typedef struct {

 typedef struct {
    int32_t  ne00;
-    int32_t  ne00_4;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
    float    eps;
 } ggml_metal_kargs_l2_norm;

@@ -733,6 +789,33 @@ typedef struct {
    uint64_t nb0;
 } ggml_metal_kargs_ssm_scan;

+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_solve_tri;
+
 typedef struct {
    int32_t  ne00t;
    int32_t  ne00;
@@ -764,6 +847,25 @@ typedef struct {
    uint64_t nb3;
 } ggml_metal_kargs_set_rows;

+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_diag;
+
 typedef struct {
    int64_t  ne00;
    int64_t  ne01;
@@ -833,10 +935,6 @@ typedef struct {
    int      max_period;
 } ggml_metal_kargs_timestep_embedding;

-typedef struct {
-    float    slope;
-} ggml_metal_kargs_leaky_relu;
-
 typedef struct {
    int32_t  ne00;
    int32_t  ne01;
@@ -287,17 +287,9 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
                n_fuse = ggml_metal_op_acc(ctx, idx);
            } break;
        case GGML_OP_SCALE:
-            {
-                n_fuse = ggml_metal_op_scale(ctx, idx);
-            } break;
        case GGML_OP_FILL:
-            {
-                n_fuse = ggml_metal_op_fill(ctx, idx);
-            } break;
        case GGML_OP_CLAMP:
-            {
-                n_fuse = ggml_metal_op_clamp(ctx, idx);
-            } break;
+        case GGML_OP_LEAKY_RELU:
        case GGML_OP_SQR:
        case GGML_OP_SQRT:
        case GGML_OP_SIN:
@@ -341,6 +333,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
            {
                n_fuse = ggml_metal_op_rwkv(ctx, idx);
            } break;
+        case GGML_OP_SOLVE_TRI:
+            {
+                n_fuse = ggml_metal_op_solve_tri(ctx, idx);
+            } break;
        case GGML_OP_MUL_MAT:
            {
                n_fuse = ggml_metal_op_mul_mat(ctx, idx);
@@ -357,6 +353,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
            {
                n_fuse = ggml_metal_op_set_rows(ctx, idx);
            } break;
+        case GGML_OP_DIAG:
+            {
+                n_fuse = ggml_metal_op_diag(ctx, idx);
+            } break;
        case GGML_OP_L2_NORM:
            {
                n_fuse = ggml_metal_op_l2_norm(ctx, idx);
@@ -418,10 +418,6 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
            {
                n_fuse = ggml_metal_op_top_k(ctx, idx);
            } break;
-        case GGML_OP_LEAKY_RELU:
-            {
-                n_fuse = ggml_metal_op_leaky_relu(ctx, idx);
-            } break;
        case GGML_OP_TRI:
            {
                n_fuse = ggml_metal_op_tri(ctx, idx);
@@ -699,7 +695,7 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
        /*.o1   =*/ { 0 },
    };

-    auto pipeline = ggml_metal_library_get_pipeline_bin(lib, GGML_OP_ADD, 1, false);
+    auto pipeline = ggml_metal_library_get_pipeline_bin_one(lib, GGML_OP_ADD);

    ggml_metal_encoder_set_pipeline(enc, pipeline);
    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -714,119 +710,6 @@ int ggml_metal_op_acc(ggml_metal_op_t ctx, int idx) {
    return 1;
 }

-int ggml_metal_op_scale(ggml_metal_op_t ctx, int idx) {
-    ggml_tensor * op = ctx->node(idx);
-
-    ggml_metal_library_t lib = ctx->lib;
-    ggml_metal_encoder_t enc = ctx->enc;
-
-    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
-    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
-    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
-
-    float scale;
-    float bias;
-    memcpy(&scale, ((const int32_t *) op->op_params) + 0, sizeof(float));
-    memcpy(&bias,  ((const int32_t *) op->op_params) + 1, sizeof(float));
-
-    ggml_metal_kargs_scale args = {
-        /*.scale =*/ scale,
-        /*.bias  =*/ bias,
-    };
-
-    int64_t n = ggml_nelements(op);
-
-    if (n % 4 == 0) {
-        n /= 4;
-    }
-
-    auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
-
-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
-
-    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
-
-    return 1;
-}
-
-int ggml_metal_op_fill(ggml_metal_op_t ctx, int idx) {
-    ggml_tensor * op = ctx->node(idx);
-
-    ggml_metal_library_t lib = ctx->lib;
-    ggml_metal_encoder_t enc = ctx->enc;
-
-    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
-    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
-    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
-
-    const float val = ggml_get_op_params_f32(op, 0);
-
-    ggml_metal_kargs_fill args = {
-        /*.val =*/ val
-    };
-
-    int64_t n = ggml_nelements(op);
-
-    if (n % 4 == 0) {
-        n /= 4;
-    }
-
-    auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
-
-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
-
-    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
-
-    return 1;
-}
-
-int ggml_metal_op_clamp(ggml_metal_op_t ctx, int idx) {
-    ggml_tensor * op = ctx->node(idx);
-
-    ggml_metal_library_t lib = ctx->lib;
-    ggml_metal_encoder_t enc = ctx->enc;
-
-    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
-    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
-    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
-
-    float min;
-    float max;
-    memcpy(&min, ((const int32_t *) op->op_params) + 0, sizeof(float));
-    memcpy(&max, ((const int32_t *) op->op_params) + 1, sizeof(float));
-
-    ggml_metal_kargs_clamp args = {
-        /*.min =*/ min,
-        /*.max =*/ max,
-    };
-
-    int64_t n = ggml_nelements(op);
-
-    if (n % 4 == 0) {
-        n /= 4;
-    }
-
-    auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
-
-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
-
-    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
-
-    return 1;
-}
-
 int ggml_metal_op_unary(ggml_metal_op_t ctx, int idx) {
    ggml_tensor * op = ctx->node(idx);

@@ -838,19 +721,79 @@ int ggml_metal_op_unary(ggml_metal_op_t ctx, int idx) {
    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);

-    int64_t n = ggml_nelements(op);
+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));

-    if (n % 4 == 0) {
-        n /= 4;
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_kargs_unary args = {
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.nb0   =*/ nb0,
+        /*.nb1   =*/ nb1,
+        /*.nb2   =*/ nb2,
+        /*.nb3   =*/ nb3,
+        /*.slope =*/ 0.0,
+        /*.scale =*/ 0.0,
+        /*.bias  =*/ 0.0,
+        /*.val   =*/ 0.0,
+        /*.min   =*/ 0.0,
+        /*.max   =*/ 0.0,
+    };
+
+    if (op->op == GGML_OP_LEAKY_RELU) {
+        args.slope = ggml_get_op_params_f32(op, 0);
+    }
+
+    if (op->op == GGML_OP_SCALE) {
+        args.scale = ggml_get_op_params_f32(op, 0);
+        args.bias  = ggml_get_op_params_f32(op, 1);
+    }
+
+    if (op->op == GGML_OP_FILL) {
+        args.val = ggml_get_op_params_f32(op, 0);
+    }
+
+    if (op->op == GGML_OP_CLAMP) {
+        args.min = ggml_get_op_params_f32(op, 0);
+        args.max = ggml_get_op_params_f32(op, 1);
    }

    auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);

-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         1);
+    if (pipeline.c4) {
+        args.ne00 = ne00/4;
+        args.ne0  = ne0/4;
+    }

-    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);
+
+    if (pipeline.cnt) {
+        const int n = pipeline.c4 ? ggml_nelements(op)/4 : ggml_nelements(op);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+    } else {
+        const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+
+        const int nth = MIN(args.ne00, nth_max);
+
+        const int nk0 = (args.ne00 + nth - 1)/nth;
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, nk0*ne01, ne02, ne03, nth, 1, 1);
+    }

    return 1;
 }
@@ -1255,6 +1198,48 @@ int ggml_metal_op_set_rows(ggml_metal_op_t ctx, int idx) {
    return 1;
 }

+int ggml_metal_op_diag(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS(int32_t,  ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS(int32_t,  ne, op, ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
+
+    ggml_metal_kargs_diag args = {
+        /*.ne00 =*/ne00,
+        /*.ne01 =*/ne01,
+        /*.ne02 =*/ne02,
+        /*.ne03 =*/ne03,
+        /*.nb00 =*/nb00,
+        /*.nb01 =*/nb01,
+        /*.nb02 =*/nb02,
+        /*.nb03 =*/nb03,
+        /*.ne0  =*/ne0,
+        /*.ne1  =*/ne1,
+        /*.ne2  =*/ne2,
+        /*.ne3  =*/ne3,
+        /*.nb0  =*/nb0,
+        /*.nb1  =*/nb1,
+        /*.nb2  =*/nb2,
+        /*.nb3  =*/nb3,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_diag(lib, op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         2);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, 32, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
    ggml_tensor * op = ctx->node(idx);

@@ -1557,6 +1542,63 @@ int ggml_metal_op_rwkv(ggml_metal_op_t ctx, int idx) {
    return 1;
 }

+int ggml_metal_op_solve_tri(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    ggml_metal_kargs_solve_tri args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.ne10 =*/ ne10,
+        /*.ne11 =*/ ne11,
+        /*.ne12 =*/ ne12,
+        /*.ne13 =*/ ne13,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.ne0  =*/ ne0,
+        /*.ne1  =*/ ne1,
+        /*.ne2  =*/ ne2,
+        /*.ne3  =*/ ne3,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_solve_tri(lib, op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
+
+    const int nsg = pipeline.nsg;
+
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, pipeline.smem, 0);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, (ne10 + nsg - 1)/nsg, ne02, ne03, 32, nsg, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_cpy(ggml_metal_op_t ctx, int idx) {
    ggml_tensor * op = ctx->node(idx);

@@ -2295,7 +2337,7 @@ size_t ggml_metal_op_flash_attn_ext_extra_blk(const ggml_tensor * op) {
    //    return res;
    //}

-    const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPTG : OP_FLASH_ATTN_EXT_NQPTG;
+    const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPSG : OP_FLASH_ATTN_EXT_NQPSG;
    const int ncpsg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NCPSG : OP_FLASH_ATTN_EXT_NCPSG;

    const int64_t ne1 = (ne01 + nqptg - 1)/nqptg;
@@ -2411,7 +2453,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {

    if (!ggml_metal_op_flash_attn_ext_use_vec(op)) {
        // half8x8 kernel
-        const int nqptg = OP_FLASH_ATTN_EXT_NQPTG; // queries per threadgroup
+        const int nqptg = OP_FLASH_ATTN_EXT_NQPSG; // queries per threadgroup
        const int ncpsg = OP_FLASH_ATTN_EXT_NCPSG; // cache values per simdgroup

        GGML_ASSERT(nqptg <= 32);
@@ -2578,9 +2620,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
    } else {
        // half4x4 kernel
-        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPTG; // queries per threadgroup
+        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPSG; // queries per threadgroup
        const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
-        const int nkpsg = 1*ncpsg;
+        const int nhptg = 1;                           // heads per threadgroup

        GGML_ASSERT(nqptg <= 32);
        GGML_ASSERT(nqptg  % 1  == 0);
@@ -2632,6 +2674,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            ggml_metal_op_concurrency_reset(ctx);
        }

+        // note: for simplicity assume the K is larger or equal than V
+        GGML_ASSERT(ne10 >= ne20);
+
        // ne00 + 2*ncpsg*(nsg)
        // for each query, we load it as f16 in shared memory (ne00)
        // and store the soft_max values and the mask
@@ -2639,28 +2684,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
        // ne20*(nsg)
        // each simdgroup has a full f32 head vector in shared mem to accumulate results
        //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*GGML_PAD(ne20, 128)*(nsg))*(sizeof(float)/2), 16))
-
-        int64_t nsgmax = 2;
-        while (true) {
-            const size_t smem = FATTN_SMEM(nsgmax);
-            // avoid using more than half of the threadgroup memory - can cause slow downs especially for large head sizes
-            if (smem > props_dev->max_theadgroup_memory_size/2) {
-                break;
-            }
-            nsgmax *= 2;
-        }
-        nsgmax /= 2;
-
-        // simdgroups per threadgroup (a.k.a. warps)
-        //const int64_t nsgt = MAX(2, MIN(nsgmax, MIN((ne11 + nkpsg - 1)/(nkpsg), (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)));
-        const int64_t nsgt = MAX(2, MIN(nsgmax, MIN((ne11 + nkpsg - 1)/(nkpsg), (int64_t) 1024/32)));
+#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg))*(sizeof(float)/2), 16))

        int64_t nsg = 1;
-        while (nsg <= nsgt) {
-            nsg *= 2;
-        }
-        nsg /= 2;

        // workgroups
        // each workgroup handles nsg*nkpsg cache values
@@ -2673,7 +2699,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
        } else {
            nwg = 32;
            nsg = 1;
-            while (2*nwg*nsg*nkpsg < ne11 && nsg < 4) {
+            while (2*nwg*nsg*ncpsg < ne11 && nsg < 4) {
                nsg *= 2;
            }
        }
@@ -2739,7 +2765,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {

            ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);

-            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
+            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, (ne02 + nhptg - 1)/nhptg, ne03*nwg, 32, nsg, 1);
        } else {
            // sanity checks
            assert(ggml_metal_op_flash_attn_ext_extra_tmp(op) != 0);
@@ -2752,7 +2778,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            ggml_metal_encoder_set_buffer(enc, bid_tmp, 7);

            ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
-            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
+            ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, (ne02 + nhptg - 1)/nhptg, ne03*nwg, 32, nsg, 1);

            // sync the 2 kernels
            ggml_metal_op_concurrency_reset(ctx);
@@ -2804,8 +2830,6 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
    GGML_ASSERT(ggml_is_contiguous_rows(op->src[1]));

-    bool bcast_row = false;
-
    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
    ggml_metal_buffer_id bid_src1 = ggml_metal_get_buffer_id(op->src[1]);
    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
@@ -2899,18 +2923,7 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {

    struct ggml_metal_pipeline_with_params pipeline;

-    if (ggml_nelements(op->src[1]) == ne10 && ggml_is_contiguous(op->src[1]) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-        GGML_ASSERT(ggml_is_contiguous(op->src[0]));
-
-        // src1 is a row
-        GGML_ASSERT(ne11 == 1);
-
-        pipeline = ggml_metal_library_get_pipeline_bin(lib, op->op, n_fuse, true);
-
-        bcast_row = true;
-    } else {
-        pipeline = ggml_metal_library_get_pipeline_bin(lib, op->op, n_fuse, false);
-    }
+    pipeline = ggml_metal_library_get_pipeline_bin(lib, op, n_fuse);

    if (n_fuse > 1) {
        bid_dst = ggml_metal_get_buffer_id(ctx->node(idx + n_fuse - 1));
@@ -2924,20 +2937,28 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
        }
    }

+    if (pipeline.c4) {
+        args.ne00 = ne00/4;
+        args.ne10 = ne10/4;
+        args.ne0  = ne0/4;
+    }
+
    ggml_metal_encoder_set_pipeline(enc, pipeline);
    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
    ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
    ggml_metal_encoder_set_buffer  (enc, bid_dst,  3);

-    if (bcast_row) {
-        const int64_t n = ggml_nelements(op)/4;
+    if (pipeline.cnt) {
+        const int n = pipeline.c4 ? ggml_nelements(op)/4 : ggml_nelements(op);

        ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
    } else {
-        int nth = 32;
+        const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));

-        while (16*nth < ne0 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+        int nth = 1;
+
+        while (2*nth < args.ne0 && nth < nth_max) {
            nth *= 2;
        }

@@ -2958,39 +2979,59 @@ int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);

+    GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
+
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_dst  = ggml_metal_get_buffer_id(op);
+
    float eps;
    memcpy(&eps, op->op_params, sizeof(float));

-    int nth = 32; // SIMD width
-
    ggml_metal_kargs_l2_norm args = {
-        /*.ne00   =*/ ne00,
-        /*.ne00_4 =*/ ne00/4,
-        /*.nb01   =*/ nb01,
-        /*.eps    =*/ eps,
+        /*.ne00  =*/ ne00,
+        /*.ne01  =*/ ne01,
+        /*.ne02  =*/ ne02,
+        /*.ne03  =*/ ne03,
+        /*.nb00  =*/ nb00,
+        /*.nb01  =*/ nb01,
+        /*.nb02  =*/ nb02,
+        /*.nb03  =*/ nb03,
+        /*.ne0   =*/ ne0,
+        /*.ne1   =*/ ne1,
+        /*.ne2   =*/ ne2,
+        /*.ne3   =*/ ne3,
+        /*.nb0   =*/ nb0,
+        /*.nb1   =*/ nb1,
+        /*.nb2   =*/ nb2,
+        /*.nb3   =*/ nb3,
+        /*.eps   =*/ eps,
    };

    auto pipeline = ggml_metal_library_get_pipeline_l2_norm(lib, op);

-    while (nth < ne00/4 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
+    if (pipeline.c4) {
+        args.ne00 = ne00/4;
+        args.ne0  = ne0/4;
+    }
+
+    int nth = 32; // SIMD width
+
+    while (nth < ne00 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
        nth *= 2;
    }

    nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
-    nth = std::min(nth, ne00/4);

    const size_t smem = pipeline.smem;

-    const int64_t nrows = ggml_nrows(op->src[0]);
-
    ggml_metal_encoder_set_pipeline(enc, pipeline);
    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+    ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+    ggml_metal_encoder_set_buffer  (enc, bid_dst,  2);

    ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);

-    ggml_metal_encoder_dispatch_threadgroups(enc, nrows, 1, 1, nth, 1, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);

    return 1;
 }
@@ -3998,42 +4039,6 @@ int ggml_metal_op_top_k(ggml_metal_op_t ctx, int idx) {
    return 1;
 }

-int ggml_metal_op_leaky_relu(ggml_metal_op_t ctx, int idx) {
-    ggml_tensor * op = ctx->node(idx);
-
-    ggml_metal_library_t lib = ctx->lib;
-    ggml_metal_encoder_t enc = ctx->enc;
-
-    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
-    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
-    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
-    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
-
-    float slope;
-    memcpy(&slope, op->op_params, sizeof(float));
-
-    ggml_metal_kargs_leaky_relu args = {
-        /*.slope =*/ slope
-    };
-
-    auto pipeline = ggml_metal_library_get_pipeline_unary(lib, op);
-
-    int64_t n = ggml_nelements(op);
-
-    if (n % 4 == 0) {
-        n /= 4;
-    }
-
-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
-
-    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
-
-    return 1;
-}
-
 int ggml_metal_op_tri(ggml_metal_op_t ctx, int idx) {
    ggml_tensor * op = ctx->node(idx);

@@ -46,9 +46,6 @@ size_t ggml_metal_op_flash_attn_ext_extra_tmp(const struct ggml_tensor * op);
 int ggml_metal_op_concat            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_repeat            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_acc               (ggml_metal_op_t ctx, int idx);
-int ggml_metal_op_scale             (ggml_metal_op_t ctx, int idx);
-int ggml_metal_op_fill              (ggml_metal_op_t ctx, int idx);
-int ggml_metal_op_clamp             (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_unary             (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_glu               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_sum               (ggml_metal_op_t ctx, int idx);
@@ -56,10 +53,12 @@ int ggml_metal_op_sum_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_cumsum            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_get_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_set_rows          (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_diag              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_soft_max          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_conv          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_scan          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_rwkv              (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_solve_tri         (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_cpy               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_1d           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pool_2d           (ggml_metal_op_t ctx, int idx);
@@ -84,7 +83,6 @@ int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argmax            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argsort           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_top_k             (ggml_metal_op_t ctx, int idx);
-int ggml_metal_op_leaky_relu        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_tri               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_opt_step_adamw    (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_opt_step_sgd      (ggml_metal_op_t ctx, int idx);
@@ -7,11 +7,15 @@
 #include "ggml-metal-context.h"
 #include "ggml-metal-ops.h"

-// globals
+#include <mutex>
+#include <string>

-// initialized in ggml_backend_metal_reg
-static ggml_backend_reg    g_ggml_metal_reg;
-static ggml_backend_device g_ggml_metal_device;
+#define GGML_METAL_NAME "MTL"
+#define GGML_METAL_MAX_DEVICES 16
+
+// number of Metal devices
+// note: can be overriden with GGML_METAL_DEVICES env to simulate virtual devices
+static int g_devices = 1;

 ////////////////////////////////////////////////////////////////////////////////
 // backend interface
@@ -165,10 +169,28 @@ static ggml_backend_buffer_i ggml_backend_metal_buffer_private_i = {
    /* .reset           = */ NULL,
 };

+static bool ggml_backend_buffer_is_metal(ggml_backend_buffer_t buffer) {
+    return buffer->iface.free_buffer == ggml_backend_metal_buffer_shared_free_buffer ||
+           buffer->iface.free_buffer == ggml_backend_metal_buffer_private_free_buffer;
+}
+
 //
 // buffer types
 //

+struct ggml_backend_metal_buffer_type {
+    int device;
+    std::string name;
+};
+
+struct ggml_backend_metal_buffer_type_deleter {
+    void operator()(ggml_backend_metal_buffer_type * ctx) const {
+        delete ctx;
+    }
+};
+
+typedef std::unique_ptr<ggml_backend_metal_buffer_type, ggml_backend_metal_buffer_type_deleter> ggml_backend_metal_buffer_type_ptr;
+
 // common method for allocating shread or private Metal buffers
 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size, bool shared) {
    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)buft->device->context;
@@ -218,9 +240,9 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
 // default (shared) buffer type

 static const char * ggml_backend_metal_buffer_type_shared_get_name(ggml_backend_buffer_type_t buft) {
-    return "Metal";
+    ggml_backend_metal_buffer_type * ctx = (ggml_backend_metal_buffer_type *)buft->context;

-    GGML_UNUSED(buft);
+    return ctx->name.c_str();
 }

 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_shared_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -249,29 +271,54 @@ static bool ggml_backend_metal_buffer_type_shared_is_host(ggml_backend_buffer_ty
    GGML_UNUSED(buft);
 }

-static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_shared(void) {
-    static ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
-        /* .iface = */ {
-            /* .get_name         = */ ggml_backend_metal_buffer_type_shared_get_name,
-            /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_shared_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_metal_buffer_type_shared_get_alignment,
-            /* .get_max_size     = */ ggml_backend_metal_buffer_type_shared_get_max_size,
-            /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_shared_get_alloc_size,
-            /* .is_host          = */ ggml_backend_metal_buffer_type_shared_is_host,
-        },
-        /* .device  = */ &g_ggml_metal_device,
-        /* .context = */ NULL,
-    };
+static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_shared(int device) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);

-    return &ggml_backend_buffer_type_metal;
+    static std::vector<ggml_backend_buffer_type> bufts;
+    static std::vector<ggml_backend_metal_buffer_type_ptr> ctxs;
+
+    static bool initialized = false;
+    if (!initialized) {
+        bufts.reserve(g_devices);
+        ctxs.reserve(g_devices);
+
+        for (int i = 0; i < g_devices; ++i) {
+            ggml_backend_metal_buffer_type * raw_ctx =
+                new ggml_backend_metal_buffer_type {
+                    /* .device = */ i,
+                    /* .name   = */ GGML_METAL_NAME + std::to_string(i),
+                };
+            ctxs.emplace_back(raw_ctx);
+
+            ggml_backend_buffer_type buft = {
+                /* .iface = */ {
+                    /* .get_name         = */ ggml_backend_metal_buffer_type_shared_get_name,
+                    /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_shared_alloc_buffer,
+                    /* .get_alignment    = */ ggml_backend_metal_buffer_type_shared_get_alignment,
+                    /* .get_max_size     = */ ggml_backend_metal_buffer_type_shared_get_max_size,
+                    /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_shared_get_alloc_size,
+                    /* .is_host          = */ ggml_backend_metal_buffer_type_shared_is_host,
+                },
+                /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_metal_reg(), i),
+                /* .context = */ raw_ctx,
+            };
+
+            bufts.emplace_back(buft);
+        }
+
+        initialized = true;
+    }
+
+    return &bufts[device];
 }

 // default (private) buffer type

 static const char * ggml_backend_metal_buffer_type_private_get_name(ggml_backend_buffer_type_t buft) {
-    return "Metal_Private";
+    ggml_backend_metal_buffer_type * ctx = (ggml_backend_metal_buffer_type *)buft->context;

-    GGML_UNUSED(buft);
+    return ctx->name.c_str();
 }

 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_private_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -300,29 +347,53 @@ static bool ggml_backend_metal_buffer_type_private_is_host(ggml_backend_buffer_t
    GGML_UNUSED(buft);
 }

-static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_private(void) {
-    static ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
-        /* .iface = */ {
-            /* .get_name         = */ ggml_backend_metal_buffer_type_private_get_name,
-            /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_private_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_metal_buffer_type_private_get_alignment,
-            /* .get_max_size     = */ ggml_backend_metal_buffer_type_private_get_max_size,
-            /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_private_get_alloc_size,
-            /* .is_host          = */ ggml_backend_metal_buffer_type_private_is_host,
-        },
-        /* .device  = */ &g_ggml_metal_device,
-        /* .context = */ NULL,
-    };
+static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_private(int device) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);

-    return &ggml_backend_buffer_type_metal;
+    static std::vector<ggml_backend_buffer_type> bufts;
+    static std::vector<ggml_backend_metal_buffer_type_ptr> ctxs;
+
+    static bool initialized = false;
+    if (!initialized) {
+        bufts.reserve(g_devices);
+        ctxs.reserve(g_devices);
+
+        for (int i = 0; i < g_devices; ++i) {
+            ggml_backend_metal_buffer_type * raw_ctx = new ggml_backend_metal_buffer_type{
+                /* .device = */ i,
+                /* .name   = */ GGML_METAL_NAME + std::to_string(i) + "_Private"
+            };
+            ctxs.emplace_back(raw_ctx);
+
+            ggml_backend_buffer_type buft = {
+                /* .iface = */ {
+                    /* .get_name         = */ ggml_backend_metal_buffer_type_private_get_name,
+                    /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_private_alloc_buffer,
+                    /* .get_alignment    = */ ggml_backend_metal_buffer_type_private_get_alignment,
+                    /* .get_max_size     = */ ggml_backend_metal_buffer_type_private_get_max_size,
+                    /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_private_get_alloc_size,
+                    /* .is_host          = */ ggml_backend_metal_buffer_type_private_is_host,
+                },
+                /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_metal_reg(), i),
+                /* .context = */ raw_ctx,
+            };
+
+            bufts.emplace_back(buft);
+        }
+
+        initialized = true;
+    }
+
+    return &bufts[device];
 }

 // mapped buffer type

 static const char * ggml_backend_metal_buffer_type_mapped_get_name(ggml_backend_buffer_type_t buft) {
-    return "Metal_Mapped";
+    ggml_backend_metal_buffer_type * ctx = (ggml_backend_metal_buffer_type *)buft->context;

-    GGML_UNUSED(buft);
+    return ctx->name.c_str();
 }

 static ggml_backend_buffer_t ggml_backend_metal_buffer_type_mapped_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -352,31 +423,55 @@ static bool ggml_backend_metal_buffer_type_mapped_is_host(ggml_backend_buffer_ty
    GGML_UNUSED(buft);
 }

-static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_mapped(void) {
-    // note: not obvious, but this buffer type still needs to implement .alloc_buffer:
-    //       https://github.com/ggml-org/llama.cpp/pull/15832#discussion_r2333177099
-    static ggml_backend_buffer_type ggml_backend_buffer_type_mapped_metal = {
-        /* .iface = */ {
-            /* .get_name         = */ ggml_backend_metal_buffer_type_mapped_get_name,
-            /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_mapped_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_metal_buffer_type_mapped_get_alignment,
-            /* .get_max_size     = */ ggml_backend_metal_buffer_type_mapped_get_max_size,
-            /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_mapped_get_alloc_size,
-            /* .is_host          = */ ggml_backend_metal_buffer_type_mapped_is_host,
-        },
-        /* .device  = */ &g_ggml_metal_device,
-        /* .context = */ NULL,
-    };
+static ggml_backend_buffer_type_t ggml_backend_metal_buffer_type_mapped(int device) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);

-    return &ggml_backend_buffer_type_mapped_metal;
+    static std::vector<ggml_backend_buffer_type> bufts;
+    static std::vector<ggml_backend_metal_buffer_type_ptr> ctxs;
+
+    static bool initialized = false;
+    if (!initialized) {
+        bufts.reserve(g_devices);
+        ctxs.reserve(g_devices);
+
+        for (int i = 0; i < g_devices; ++i) {
+            ggml_backend_metal_buffer_type * raw_ctx = new ggml_backend_metal_buffer_type{
+                /* .device = */ i,
+                /* .name   = */ GGML_METAL_NAME + std::to_string(i) + "_Mapped"
+            };
+            ctxs.emplace_back(raw_ctx);
+
+            // note: not obvious, but this buffer type still needs to implement .alloc_buffer:
+            //       https://github.com/ggml-org/llama.cpp/pull/15832#discussion_r2333177099
+            ggml_backend_buffer_type buft = {
+                /* .iface = */ {
+                    /* .get_name         = */ ggml_backend_metal_buffer_type_mapped_get_name,
+                    /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_mapped_alloc_buffer,
+                    /* .get_alignment    = */ ggml_backend_metal_buffer_type_mapped_get_alignment,
+                    /* .get_max_size     = */ ggml_backend_metal_buffer_type_mapped_get_max_size,
+                    /* .get_alloc_size   = */ ggml_backend_metal_buffer_type_mapped_get_alloc_size,
+                    /* .is_host          = */ ggml_backend_metal_buffer_type_mapped_is_host,
+                },
+                /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_metal_reg(), i),
+                /* .context = */ raw_ctx,
+            };
+
+            bufts.emplace_back(buft);
+        }
+
+        initialized = true;
+    }
+
+    return &bufts[device];
 }

 // backend

 static const char * ggml_backend_metal_name(ggml_backend_t backend) {
-    return "Metal";
+    ggml_metal_t ctx = (ggml_metal_t)backend->context;

-    GGML_UNUSED(backend);
+    return ggml_metal_get_name(ctx);
 }

 static void ggml_backend_metal_free(ggml_backend_t backend) {
@@ -409,12 +504,24 @@ static void ggml_backend_metal_get_tensor_async(ggml_backend_t backend, const gg
 }

 static bool ggml_backend_metal_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
-    return false;
+    if (!ggml_backend_is_metal(backend_src) || !ggml_backend_is_metal(backend_dst)) {
+        return false;
+    }

-    GGML_UNUSED(backend_src);
-    GGML_UNUSED(backend_dst);
-    GGML_UNUSED(src);
-    GGML_UNUSED(dst);
+    if (!ggml_backend_buffer_is_metal(src->buffer) || !ggml_backend_buffer_is_metal(dst->buffer)) {
+        return false;
+    }
+
+    ggml_metal_t ctx_src = (ggml_metal_t)backend_src->context;
+    ggml_metal_t ctx_dst = (ggml_metal_t)backend_dst->context;
+
+    //ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
+    //ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
+
+    //ggml_metal_buffer_t buf_ctx_src = (ggml_metal_buffer_t)buf_src->context;
+    //ggml_metal_buffer_t buf_ctx_dst = (ggml_metal_buffer_t)buf_dst->context;
+
+    return ggml_metal_cpy_tensor_async(ctx_src, ctx_dst, src, dst);
 }

 static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
@@ -423,6 +530,20 @@ static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend,
    return ggml_metal_graph_compute(ctx, cgraph);
 }

+static void ggml_backend_metal_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_metal_t ctx = (ggml_metal_t)backend->context;
+    ggml_metal_event_t ev = (ggml_metal_event_t)event->context;
+
+    ggml_metal_event_record(ctx, ev);
+}
+
+static void ggml_backend_metal_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_metal_t ctx = (ggml_metal_t)backend->context;
+    ggml_metal_event_t ev = (ggml_metal_event_t)event->context;
+
+    ggml_metal_event_wait(ctx, ev);
+}
+
 static void ggml_backend_metal_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
    ggml_metal_t ctx = (ggml_metal_t)backend->context;

@@ -435,7 +556,6 @@ static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
    ggml_metal_t ctx = (ggml_metal_t)backend->context;

    ggml_metal_set_n_cb(ctx, n_cb);
-
 }

 static ggml_backend_i ggml_backend_metal_i = {
@@ -450,12 +570,8 @@ static ggml_backend_i ggml_backend_metal_i = {
    /* .graph_plan_update       = */ NULL,
    /* .graph_plan_compute      = */ NULL,
    /* .graph_compute           = */ ggml_backend_metal_graph_compute,
-
-    // the events API is needed only for multi-GPU setups, so likely no need to implement it for Metal
-    // in any case, these docs seem relevant if we ever decide to implement it:
-    // https://developer.apple.com/documentation/metal/mtlcommandbuffer#Synchronizing-Passes-with-Events
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
+    /* .event_record            = */ ggml_backend_metal_event_record,
+    /* .event_wait              = */ ggml_backend_metal_event_wait,
    /* .graph_optimize          = */ ggml_backend_metal_graph_optimize,
 };

@@ -519,15 +635,17 @@ void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
 // backend device

 static const char * ggml_backend_metal_device_get_name(ggml_backend_dev_t dev) {
-    return "Metal";
+    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;

-    GGML_UNUSED(dev);
+    const ggml_metal_device_props * props_dev = ggml_metal_device_get_props(ctx_dev);
+
+    return props_dev->name;
 }

 static const char * ggml_backend_metal_device_get_description(ggml_backend_dev_t dev) {
    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;

-    return ggml_metal_device_get_props(ctx_dev)->name;
+    return ggml_metal_device_get_props(ctx_dev)->desc;
 }

 static void ggml_backend_metal_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
@@ -550,14 +668,14 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_bac
    ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);

    props->caps = {
-        /* .async                 = */ true,
-        /* .host_buffer           = */ false,
-        /* .buffer_from_host_ptr  = */ true,
-        /* .events                = */ false,
+        /* .async                = */ true,
+        /* .host_buffer          = */ false,
+        /* .buffer_from_host_ptr = */ true,
+        /* .events               = */ true,
    };
 }

-static ggml_backend_t ggml_backend_metal_device_init(ggml_backend_dev_t dev, const char * params) {
+static ggml_backend_t ggml_backend_metal_device_init_backend(ggml_backend_dev_t dev, const char * params) {
    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;

    ggml_metal_t ctx = ggml_metal_init(ctx_dev);
@@ -587,7 +705,7 @@ static ggml_backend_buffer_type_t ggml_backend_metal_device_get_buffer_type(ggml

    const ggml_metal_device_props * props_dev = ggml_metal_device_get_props(ctx_dev);

-    return props_dev->use_shared_buffers ? ggml_backend_metal_buffer_type_shared() : ggml_backend_metal_buffer_type_private();
+    return props_dev->use_shared_buffers ? ggml_backend_metal_buffer_type_shared(props_dev->device) : ggml_backend_metal_buffer_type_private(props_dev->device);
 }

 static ggml_backend_buffer_t ggml_backend_metal_device_buffer_mapped(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
@@ -595,7 +713,9 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_mapped(ggml_backen

    ggml_metal_buffer_t res = ggml_metal_buffer_map(ctx_dev, ptr, size, max_tensor_size);

-    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type_mapped(), ggml_backend_metal_buffer_shared_i, res, size);
+    const ggml_metal_device_props * props_dev = ggml_metal_device_get_props(ctx_dev);
+
+    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type_mapped(props_dev->device), ggml_backend_metal_buffer_shared_i, res, size);
 }

 static bool ggml_backend_metal_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
@@ -606,9 +726,10 @@ static bool ggml_backend_metal_device_supports_op(ggml_backend_dev_t dev, const

 static bool ggml_backend_metal_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
    return
+        buft->device == dev && (
        buft->iface.get_name == ggml_backend_metal_buffer_type_shared_get_name ||
        buft->iface.get_name == ggml_backend_metal_buffer_type_private_get_name ||
-        buft->iface.get_name == ggml_backend_metal_buffer_type_mapped_get_name;
+        buft->iface.get_name == ggml_backend_metal_buffer_type_mapped_get_name);

    GGML_UNUSED(dev);
 }
@@ -632,45 +753,97 @@ static bool ggml_backend_metal_device_offload_op(ggml_backend_dev_t dev, const g
            get_op_batch_size(op) >= ggml_metal_device_get_props(ctx_dev)->op_offload_min_batch_size;
 }

+static ggml_backend_event_t ggml_backend_metal_device_event_new(ggml_backend_dev_t dev) {
+    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
+
+    ggml_metal_event_t event = ggml_metal_device_event_init(ctx_dev);
+    GGML_ASSERT(event);
+
+    ggml_backend_event_t ev = new ggml_backend_event {
+        /* .device  = */ dev,
+        /* .context = */ event,
+    };
+
+    return ev;
+}
+
+static void ggml_backend_metal_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
+
+    ggml_metal_event_t ev = (ggml_metal_event_t)event->context;
+
+    ggml_metal_device_event_free(ctx_dev, ev);
+
+    delete event;
+}
+
+static void ggml_backend_metal_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
+
+    ggml_metal_event_t evt = (ggml_metal_event_t)event->context;
+
+    ggml_metal_device_event_synchronize(ctx_dev, evt);
+}
+
 static ggml_backend_device_i ggml_backend_metal_device_i = {
    /* .get_name             = */ ggml_backend_metal_device_get_name,
    /* .get_description      = */ ggml_backend_metal_device_get_description,
    /* .get_memory           = */ ggml_backend_metal_device_get_memory,
    /* .get_type             = */ ggml_backend_metal_device_get_type,
    /* .get_props            = */ ggml_backend_metal_device_get_props,
-    /* .init_backend         = */ ggml_backend_metal_device_init,
+    /* .init_backend         = */ ggml_backend_metal_device_init_backend,
    /* .get_buffer_type      = */ ggml_backend_metal_device_get_buffer_type,
    /* .get_host_buffer_type = */ NULL,
    /* .buffer_from_host_ptr = */ ggml_backend_metal_device_buffer_mapped,
    /* .supports_op          = */ ggml_backend_metal_device_supports_op,
    /* .supports_buft        = */ ggml_backend_metal_device_supports_buft,
    /* .offload_op           = */ ggml_backend_metal_device_offload_op,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
+    /* .event_new            = */ ggml_backend_metal_device_event_new,
+    /* .event_free           = */ ggml_backend_metal_device_event_free,
+    /* .event_synchronize    = */ ggml_backend_metal_device_event_synchronize,
 };

 // backend registry

+struct ggml_backend_metal_reg {
+    std::vector<ggml_backend_dev_t> devices;
+};
+
+typedef struct ggml_backend_metal_reg * ggml_backend_metal_reg_t;
+
+static ggml_backend_metal_reg_t ggml_backend_metal_reg_init(void) {
+    ggml_backend_metal_reg_t ctx = new struct ggml_backend_metal_reg;
+
+    return ctx;
+}
+
+static void ggml_backend_metal_reg_free(ggml_backend_metal_reg_t ctx) {
+    delete ctx;
+}
+
+struct ggml_backend_metal_reg_deleter {
+    void operator()(ggml_backend_metal_reg_t ctx) {
+        ggml_backend_metal_reg_free(ctx);
+    }
+};
+
+typedef std::unique_ptr<struct ggml_backend_metal_reg, ggml_backend_metal_reg_deleter> ggml_backend_metal_reg_ptr;
+
 static const char * ggml_backend_metal_reg_get_name(ggml_backend_reg_t reg) {
-    return "Metal";
+    return GGML_METAL_NAME;

    GGML_UNUSED(reg);
 }

 static size_t ggml_backend_metal_reg_device_count(ggml_backend_reg_t reg) {
-    return 1;
-
-    GGML_UNUSED(reg);
+    ggml_backend_metal_reg_t ctx = (ggml_backend_metal_reg_t)reg->context;
+    return ctx->devices.size();
 }

 static ggml_backend_dev_t ggml_backend_metal_reg_device_get(ggml_backend_reg_t reg, size_t index) {
-    GGML_ASSERT(index == 0);
-
-    return &g_ggml_metal_device;
-
-    GGML_UNUSED(reg);
-    GGML_UNUSED(index);
+    ggml_backend_metal_reg_t ctx = (ggml_backend_metal_reg_t)reg->context;
+    GGML_ASSERT(index < ctx->devices.size());
+    return ctx->devices[index];
 }

 static ggml_backend_feature g_ggml_backend_metal_features[] = {
@@ -698,27 +871,67 @@ static void * ggml_backend_metal_get_proc_address(ggml_backend_reg_t reg, const

 static ggml_backend_reg_i ggml_backend_metal_reg_i = {
    /* .get_name         = */ ggml_backend_metal_reg_get_name,
-    /* .device_count     = */ ggml_backend_metal_reg_device_count,
-    /* .device_get       = */ ggml_backend_metal_reg_device_get,
+    /* .get_device_count = */ ggml_backend_metal_reg_device_count,
+    /* .get_device       = */ ggml_backend_metal_reg_device_get,
    /* .get_proc_address = */ ggml_backend_metal_get_proc_address,
 };

-ggml_backend_reg_t ggml_backend_metal_reg(void) {
-    {
-        g_ggml_metal_reg = {
-            /* .api_version = */ GGML_BACKEND_API_VERSION,
-            /* .iface       = */ ggml_backend_metal_reg_i,
-            /* .context     = */ NULL,
-        };
+static ggml_backend_dev_t ggml_backend_metal_device_init(ggml_backend_reg_t reg, int device) {
+    return new ggml_backend_device {
+        /* .iface   = */ ggml_backend_metal_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ ggml_metal_device_get(device),
+    };
+}

-        g_ggml_metal_device = {
-            /* .iface   = */ ggml_backend_metal_device_i,
-            /* .reg     = */ &g_ggml_metal_reg,
-            /* .context = */ ggml_metal_device_get(),
-        };
+static void ggml_backend_metal_device_free(ggml_backend_dev_t dev) {
+    delete dev;
+}
+
+struct ggml_backend_device_deleter {
+    void operator()(ggml_backend_dev_t ctx) {
+        ggml_backend_metal_device_free(ctx);
+    }
+};
+
+typedef std::unique_ptr<ggml_backend_device, ggml_backend_device_deleter> ggml_backend_device_ptr;
+
+ggml_backend_reg_t ggml_backend_metal_reg(void) {
+    static ggml_backend_reg reg;
+    static bool initialized = false;
+
+    {
+        static std::mutex mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+
+        const char * env = getenv("GGML_METAL_DEVICES");
+        if (env) {
+            g_devices = atoi(env);
+        }
+
+        static std::vector<ggml_backend_device_ptr> devs;
+
+        if (!initialized) {
+            static ggml_backend_metal_reg_ptr reg_ctx(ggml_backend_metal_reg_init());
+
+            for (int i = 0; i < g_devices; ++i) {
+                auto * dev = ggml_backend_metal_device_init(&reg, i);
+                devs.emplace_back(dev);
+
+                reg_ctx->devices.push_back(dev);
+            }
+
+            reg = {
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_metal_reg_i,
+                /* .context     = */ reg_ctx.get(),
+            };
+        }
+
+        initialized = true;
    }

-    return &g_ggml_metal_reg;
+    return &reg;
 }

 GGML_BACKEND_DL_IMPL(ggml_backend_metal_reg)
--- a/Show More
+++ b/Show More