metal : null-check buffer alloc to fix OOM crash (#25371)

* metal : null-check ggml_metal_buffer_init result to avoid OOM crash

ggml_backend_metal_buffer_type_alloc_buffer used the result of
ggml_metal_buffer_init without checking for NULL. ggml_metal_buffer_init
returns NULL when the underlying Metal allocation fails (e.g. an
out-of-memory condition), and the following ggml_metal_buffer_is_shared(res)
call dereferences it, turning a recoverable allocation failure into a hard
crash (EXC_BAD_ACCESS). This is easy to hit on memory-constrained devices
such as iOS when a model/context exceeds the available Metal budget.

Log the failure using the existing GGML_LOG_ERROR convention and return
NULL so the allocator surfaces a diagnosable error up the stack instead of
crashing.

* cont : fix log

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.github/actions/ccache-clear/action.yml

@@ -21,68 +21,30 @@ inputs:
 runs:
   using: "composite"
   steps:
+    - name: Install GitHub CLI if missing
+      shell: bash
+      run: |
+        # e.g. in container jobs, where it is not preinstalled
+        if ! command -v gh >/dev/null 2>&1; then
+          echo "GitHub CLI not found, installing..."
+          if ! command -v curl >/dev/null 2>&1; then
+            apt-get update >/dev/null 2>&1 || true
+            apt-get install -y curl >/dev/null 2>&1 || true
+          fi
+          mkdir -p -m 755 /etc/apt/keyrings
+          curl -fsSL https://cli.github.com/packages/githubcli-archive-keyring.gpg | tee /etc/apt/keyrings/githubcli-archive-keyring.gpg >/dev/null
+          chmod go+r /etc/apt/keyrings/githubcli-archive-keyring.gpg
+          echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" > /etc/apt/sources.list.d/github-cli.list
+          apt-get update >/dev/null 2>&1 || true
+          apt-get install -y gh || { echo "Failed to install GitHub CLI (gh)" >&2; exit 1; }
+        fi
+        command -v gh >/dev/null 2>&1 || { echo "GitHub CLI (gh) is required but could not be installed" >&2; exit 1; }
+
     - name: Clear caches
       shell: bash
-      env:
-        CLEAR_KEY: ${{ inputs.key }}
-        CLEAR_OLDER: ${{ inputs.older }}
-        CLEAR_MIN: ${{ inputs.min }}
-        CLEAR_DRY_RUN: ${{ inputs.dry-run }}
       run: |
-        # Convert a duration (e.g. 90m, 1h, 1d, plain seconds) to seconds
-        to_seconds() {
-          local val="$1"
-          [[ "$val" =~ ^[0-9]+$ ]] && { echo "$val"; return 0; }
-          local num="${val%?}" unit="${val: -1}" mult
-          [[ "$num" =~ ^[0-9]+$ ]] || return 1
-          case "$unit" in
-            s) mult=1 ;;
-            m) mult=60 ;;
-            h) mult=3600 ;;
-            d) mult=86400 ;;
-            *) return 1 ;;
-          esac
-          echo $((num * mult))
-        }
-
-        [[ "$CLEAR_MIN" =~ ^[0-9]+$ ]] || { echo "Invalid min value: $CLEAR_MIN" >&2; exit 1; }
-        [[ "$CLEAR_DRY_RUN" =~ ^(true|false)$ ]] || { echo "Invalid dry-run value: $CLEAR_DRY_RUN" >&2; exit 1; }
-
-        CACHES=$(gh cache list --key "ccache-$CLEAR_KEY" --json id,key,createdAt --jq '.[] | [.createdAt, .id, .key] | @tsv' 2>/dev/null | LC_ALL=C sort)
-        if [ -z "$CACHES" ]; then
-          echo "No caches found with key prefix: $CLEAR_KEY"
-          exit 0
-        fi
-
-        TOTAL=$(( $(wc -l <<< "$CACHES") ))
-
-        echo "Found $TOTAL cache(s) with key prefix: $CLEAR_KEY (oldest first):"
-        while IFS=$'\t' read -r CREATED ID KEY; do
-          printf '  %s  %s  %s\n' "$CREATED" "$ID" "$KEY"
-        done <<< "$CACHES"
-
-        CUTOFF=""
-        if [ -n "$CLEAR_OLDER" ]; then
-          OLDER_SECONDS=$(to_seconds "$CLEAR_OLDER") || { echo "Invalid older value: $CLEAR_OLDER (expected e.g. 90m, 1h, 1d)" >&2; exit 1; }
-          CUTOFF=$(( $(date +%s) - OLDER_SECONDS ))
-        fi
-
-        # Caches are sorted oldest first
-        DELETED=0
-        while IFS=$'\t' read -r CREATED ID KEY; do
-          if [ -n "$CUTOFF" ] && [ "$(date -d "$CREATED" +%s)" -ge "$CUTOFF" ]; then
-            echo "Rest are not older than $CLEAR_OLDER, stopping"
-            break
-          fi
-          if [ $((TOTAL - DELETED - 1)) -lt "$CLEAR_MIN" ]; then
-            echo "Keeping at least $CLEAR_MIN cache(s), stopping"
-            break
-          fi
-          if [ "$CLEAR_DRY_RUN" = "true" ]; then
-            echo "Would delete cache: $ID ($KEY)"
-          else
-            echo "Deleting cache: $ID ($KEY)"
-            gh cache delete "$ID"
-          fi
-          DELETED=$((DELETED + 1))
-        done <<< "$CACHES"
+        bash scripts/ccache-clear.sh \
+          --key "${{ inputs.key }}" \
+          --older "${{ inputs.older }}" \
+          --min "${{ inputs.min }}" \
+          ${{ inputs.dry-run == 'true' && '--dry-run' || '' }}

.github/workflows/build-apple.yml

@@ -73,6 +73,16 @@ jobs:
           cd build
           ctest -L main -E "test-llama-archs" --verbose --timeout 900
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: apple-arm64
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   macos-latest-x64:
     runs-on: macos-15-intel
 
@@ -109,6 +119,16 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: apple-x64
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   macos-latest-ios-xcode:
     runs-on: macos-latest
 
@@ -163,14 +183,6 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
-      # TODO: this likely does not do anything - if yes, remove it
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: apple-tvos
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
       - name: Build
         id: cmake_build
         run: |
@@ -196,14 +208,6 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
-      # TODO: this likely does not do anything - if yes, remove it
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: apple-visionos
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
       - name: Build
         id: cmake_build
         run: |
@@ -234,14 +238,6 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
-      # TODO: this likely does not do anything - if yes, remove it
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: apple-swift
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
       - name: Download xcframework artifact
         uses: actions/download-artifact@v7
         with:

.github/workflows/build-cpu.yml

@@ -125,7 +125,7 @@ jobs:
           GH_TOKEN: ${{ github.token }}
         with:
           key: cpu-${{ matrix.os }}
-          older: 1h
+          older: 5m
           min: 1
           dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
 
@@ -215,3 +215,13 @@ jobs:
       #     cd build
       #     $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
       #     & $sde -future -- ctest -L main -C Release --verbose --timeout 900
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: cpu-windows-2025-${{ matrix.build }}
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-cuda-ubuntu.yml

@@ -72,6 +72,16 @@ jobs:
             -DGGML_CUDA_CUB_3DOT2=ON
           cmake --build build
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: cuda-ubuntu-24.04-cuda
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   hip:
     runs-on: ubuntu-22.04
     container: rocm/dev-ubuntu-22.04:6.1.2
@@ -103,6 +113,16 @@ jobs:
             -DGGML_HIP=ON
           cmake --build build --config Release -j $(nproc)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: cuda-ubuntu-22.04-hip
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   musa:
     runs-on: ubuntu-22.04
     container: mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64
@@ -131,3 +151,13 @@ jobs:
           cmake -B build -S . \
             -DGGML_MUSA=ON
           time cmake --build build --config Release -j $(nproc)
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: cuda-ubuntu-22.04-musa
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-opencl.yml

@@ -80,3 +80,13 @@ jobs:
         run: |
           cmake -S . -B build -G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: opencl-windows-2025-x64
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-openvino.yml

@@ -167,3 +167,13 @@ jobs:
 
           cd build
           ctest --test-dir ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" -C Release --verbose --timeout 3000
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: openvino-windows-2022
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-sycl.yml

@@ -96,6 +96,16 @@ jobs:
             -DGGML_SYCL_F16=${{ matrix.fp16 }}
           time cmake --build build --config Release -j $(nproc)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: sycl-ubuntu-24-${{ matrix.build }}
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   windows-latest-sycl:
     runs-on: windows-2022
 
@@ -139,3 +149,13 @@ jobs:
       - name: Build
         id: cmake_build
         run:  examples/sycl/win-build-sycl.bat
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: sycl-windows-latest
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-vulkan.yml

@@ -55,7 +55,7 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: vulkan-ubuntu-24.04-arm-new
+          key: vulkan-ubuntu-24.04-arm
           variant: ccache
           evict-old-files: 1d
           save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
@@ -73,6 +73,16 @@ jobs:
         run: |
           time cmake --build build -j $(nproc)
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: vulkan-ubuntu-24.04-arm
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   ubuntu-llvmpipe:
     runs-on: ubuntu-24.04
 
@@ -128,6 +138,16 @@ jobs:
           # test-backend-ops is too slow on llvmpipe, skip it
           ctest -L main -E test-backend-ops --verbose --timeout 900
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: vulkan-ubuntu-24.04-llvmpipe
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   windows:
     runs-on: windows-2025
 
@@ -180,3 +200,13 @@ jobs:
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: cpu-windows-2025-x64-vulkan
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-wasm.yml

@@ -88,3 +88,13 @@ jobs:
             -DEMDAWNWEBGPU_DIR=emdawnwebgpu_pkg
 
           time cmake --build build-wasm --config Release --target test-backend-ops -j $(nproc)
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: webgpu-ubuntu-24.04-arm-wasm
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-webgpu.yml

@@ -101,6 +101,16 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: webgpu-macos-latest
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   ubuntu:
     runs-on: ubuntu-24.04
 
@@ -153,3 +163,13 @@ jobs:
           # This is using llvmpipe and runs slower than other backends
           # test-backend-ops is too slow on llvmpipe, skip it
           ctest -L main -E test-backend-ops --verbose --timeout 900
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: webgpu-ubuntu-24.04
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/hip-quality-check.yml

@@ -84,3 +84,13 @@ jobs:
           cd build
           make -j $(nproc) 2>&1 | tee metrics.log | grep -v 'Rpass-analysis=kernel-resource-usage\|remark:\|^$'
           python3 ../scripts/hip/gcn-cdna-vgpr-check.py metrics.log
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: hip-quality-check-ubuntu-22.04
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/make-release.yml

@@ -84,11 +84,13 @@ jobs:
 
             New version has been released.
 
+            ## Assets
+
             ${{ steps.desc.outputs.nightly }}
 
-            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
+            ## More info
 
-            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
+            - [Releases and versioning of `ggml-org` projects](https://github.com/ggml-org/ggml/discussions/1579)
 
             ## ${{ steps.desc.outputs.changelog_title }}
 

.github/workflows/server.yml

@@ -128,6 +128,16 @@ jobs:
           export LLAMA_ARG_BACKEND_SAMPLING=1
           SLOW_TESTS=1 ./tests.sh
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: server-ubuntu-24.04-arm
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   windows:
     runs-on: windows-2025
 
@@ -181,3 +191,13 @@ jobs:
           cd tools/server/tests
           export SLOW_TESTS="1"
           ./tests.sh
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: server-windows-2025-x64
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

CMakeLists.txt

@@ -4,7 +4,7 @@ include(CheckIncludeFileCXX)
 
 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 2)
+set(LLAMA_VERSION_MINOR 3)
 set(LLAMA_VERSION_PATCH 0)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")
 

CONTRIBUTING.md

@@ -74,6 +74,7 @@ For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 - If a PR does not warrant a new release, add `[no release]` in the squashed commit to spare CI resources
 - Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
 - Add the ["merge ready"](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+is%3Aopen+draft%3Ano+sort%3Aupdated-desc+label%3A%22merge+ready%22+) label to a PR to indicate when a PR can be fast-merged without waiting for 2 independent reviews. [(more info)](https://github.com/ggml-org/llama.cpp/pull/26178)
+- Wait for CI results before merging
 
 Maintainers reserve the right to decline review or close pull requests for any reason, without any questions, particularly under any of the following conditions:
 - The proposed change is already mentioned in the roadmap or an existing issue, and it has been assigned to someone.

README.md

@@ -13,7 +13,7 @@
 [![Docker](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/docker.yml?label=Docker)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
 [![Winget](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/winget.yml?label=Winget)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)
 
-[manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
+[ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Anikwen%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3Amarty1885%20OR%20author%3A0cc4m%20OR%20author%3ATitaniumtown%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [dev stats](https://github.com/ggml-org/llama.cpp-dev) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
 
 </div>
 

ci/run.sh

@@ -300,6 +300,40 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
+# test_llama_archs_tensor_split
+
+function gg_run_test_llama_archs_tensor_split {
+    cd ${SRC}
+
+    set -e
+
+    if [ ! -z ${GG_BUILD_CUDA} ]; then
+        GGML_CUDA_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+        GGML_CUDA_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+        GGML_CUDA_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+        GGML_CUDA_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+    fi
+
+    if [ ! -z ${GG_BUILD_METAL} ]; then
+        GGML_METAL_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+        GGML_METAL_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+        GGML_METAL_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+        GGML_METAL_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+    fi
+
+    set +e
+}
+
+function gg_sum_test_llama_archs_tensor_split {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs test-llama-archs with 1 to 4 devices\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}.log)"
+    gg_printf '```\n'
+}
+
 # test_scripts
 
 function gg_run_test_scripts {
@@ -751,6 +785,8 @@ ret=0
 test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
+test $ret -eq 0 && gg_run test_llama_archs_tensor_split
+
 if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
     test $ret -eq 0 && gg_run test_backend_ops_cpu
 fi

common/common.cpp

@@ -402,10 +402,11 @@ void common_params_print_info(const common_params & params, bool print_devices)
 #endif
     COM_TRC("%s: build %d (%s) with %s for %s%s\n", __func__, llama_build_number(), llama_commit(), llama_compiler(), llama_build_target(), build_type);
 
-    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, common_log_get_verbosity_thold());
+    const int verbosity = common_log_get_verbosity_thold();
+    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, verbosity);
 
     // device enumeration creates a primary context on CUDA backends, skip it when the caller does not own any device
-    if (print_devices) {
+    if (print_devices && verbosity >= LOG_LEVEL_TRACE) {
         COM_TRC("%s", "device_info:\n");
         for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
             auto * dev = ggml_backend_dev_get(i);

common/json.cpp

@@ -78,19 +78,21 @@ common_json_value::common_json_value(const common_json & val) :
 common_json_value::common_json_value(common_json && val) :
     type(VAL_JSON), val_json(std::make_shared<common_json>(std::move(val))) {}
 
+// the ctors and get<T>() below are explicit specializations, giving strong symbols
+// an explicit instantiation is a weak symbol, dropped by some LTO builds (clang-cl)
 template <typename T>
-common_json_value::common_json_value(const std::set<T> & vals) : type(VAL_JSON) {
+static std::shared_ptr<common_json> set_json(const std::set<T> & vals) {
     common_json out = common_json::array();
 
     for (const auto & val : vals) {
         out.push_back(val);
     }
 
-    val_json = std::make_shared<common_json>(std::move(out));
+    return std::make_shared<common_json>(std::move(out));
 }
 
 // a set value is usable only for the types below
-#define COMMON_JSON_SET(...) template common_json_value::common_json_value(const std::set<__VA_ARGS__> &);
+#define COMMON_JSON_SET(...) template <> common_json_value::common_json_value(const std::set<__VA_ARGS__> & vals) : type(VAL_JSON), val_json(set_json(vals)) {}
 
 COMMON_JSON_SET(int)
 COMMON_JSON_SET(std::string)
@@ -98,56 +100,45 @@ COMMON_JSON_SET(std::string)
 #undef COMMON_JSON_SET
 
 template <typename T>
-common_json_value::common_json_value(const std::map<std::string, T> & vals) : type(VAL_JSON) {
+static std::shared_ptr<common_json> map_json(const T & vals) {
     common_json out = common_json::object();
 
     for (const auto & val : vals) {
         out.set({ val.first, val.second });
     }
 
-    val_json = std::make_shared<common_json>(std::move(out));
+    return std::make_shared<common_json>(std::move(out));
 }
 
 // a map value is usable only for the types below
-#define COMMON_JSON_MAP(...) template common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> &);
+#define COMMON_JSON_MAP(...) template <> common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
 
 COMMON_JSON_MAP(bool)
 COMMON_JSON_MAP(std::string)
 
 #undef COMMON_JSON_MAP
 
-template <typename T>
-common_json_value::common_json_value(const std::unordered_map<std::string, T> & vals) : type(VAL_JSON) {
-    common_json out = common_json::object();
-
-    for (const auto & val : vals) {
-        out.set({ val.first, val.second });
-    }
-
-    val_json = std::make_shared<common_json>(std::move(out));
-}
-
 // an unordered map value is usable only for the types below
-#define COMMON_JSON_UMAP(...) template common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> &);
+#define COMMON_JSON_UMAP(...) template <> common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
 
 COMMON_JSON_UMAP(size_t)
 
 #undef COMMON_JSON_UMAP
 
 template <typename T>
-common_json_value::common_json_value(const std::vector<T> & vals) : type(VAL_JSON) {
+static std::shared_ptr<common_json> vec_json(const std::vector<T> & vals) {
     common_json out = common_json::array();
 
     for (const auto & val : vals) {
         out.push_back(val);
     }
 
-    val_json = std::make_shared<common_json>(std::move(out));
+    return std::make_shared<common_json>(std::move(out));
 }
 
 // a vector value is usable only for the types below
 // note: std::vector<bool> is not here, its proxy reference does not convert
-#define COMMON_JSON_VEC(...) template common_json_value::common_json_value(const std::vector<__VA_ARGS__> &);
+#define COMMON_JSON_VEC(...) template <> common_json_value::common_json_value(const std::vector<__VA_ARGS__> & vals) : type(VAL_JSON), val_json(vec_json(vals)) {}
 
 COMMON_JSON_VEC(int)
 COMMON_JSON_VEC(unsigned char)
@@ -404,10 +395,6 @@ common_json::items_view common_json::items() const {
     return items_view(const_cast<common_json *>(this), size());
 }
 
-template <typename T> T common_json::get() const {
-    return guard([&] { return as_json(this).get<T>(); });
-}
-
 // the backing library cannot build a common_json, so this one is just a copy
 template <> common_json common_json::get<common_json>() const {
     return *this;
@@ -415,7 +402,7 @@ template <> common_json common_json::get<common_json>() const {
 
 // get<T>() is usable only for the types below
 
-#define COMMON_JSON_GET(...) template __VA_ARGS__ common_json::get<__VA_ARGS__>() const;
+#define COMMON_JSON_GET(...) template <> __VA_ARGS__ common_json::get<__VA_ARGS__>() const { return guard([&] { return as_json(this).get<__VA_ARGS__>(); }); }
 
 COMMON_JSON_GET(bool)
 COMMON_JSON_GET(int)
@@ -435,3 +422,12 @@ COMMON_JSON_GET(std::vector<size_t>)
 COMMON_JSON_GET(std::unordered_map<std::string, size_t>)
 
 #undef COMMON_JSON_GET
+
+// must stay below the get<std::string> specialization
+common_json::operator std::string() const {
+    return get<std::string>();
+}
+
+std::string common_json::value(const std::string & key, const char * def) const {
+    return contains(key) ? at(key).get<std::string>() : std::string(def);
+}

common/json.h

@@ -221,16 +221,14 @@ class common_json {
     // implicit get<T>() for plain values, so they can be assigned to their C++ type directly
     // note: kept to this short list on purpose, a wider one makes j["key"] ambiguous
     // note: a numeric one would make "str = json;" ambiguous, a number converts to char too
-    operator std::string() const { return get<std::string>(); }
+    operator std::string() const;
 
     template <typename T>
     T value(const std::string & key, T def) const {
         return contains(key) ? at(key).get<T>() : def;
     }
 
-    std::string value(const std::string & key, const char * def) const {
-        return contains(key) ? at(key).get<std::string>() : std::string(def);
-    }
+    std::string value(const std::string & key, const char * def) const;
 
     // a JSON default needs no get<T>(), it is already the right type
     common_json value(const std::string & key, const common_json & def) const {

conversion/glm.py

@@ -112,12 +112,38 @@ class GlmOCRModel(Glm4Model):
 @ModelBase.example("zai-org/GLM-4.5-Air")
 class Glm4MoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GLM4_MOE
+    supports_mtp_export = True
+    _n_main_layers: int | None = None
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        # GLM4_MOE has num_hidden_layers + 1 actual layers (including NextN layer)
-        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
-        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+        if not self.no_mtp:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
+        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
+        type(self)._n_main_layers = hparams.get(key)
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        assert cls._n_main_layers is not None
+        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
+
+        if is_mtp and cls.no_mtp:
+            return None
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return name, gen
 
     def set_vocab(self):
         return self._set_vocab_glm()
@@ -153,10 +179,22 @@ class Glm4MoeModel(TextModel):
         if (norm_topk_prob := self.hparams.get("norm_topk_prob")) is not None:
             self.gguf_writer.add_expert_weights_norm(norm_topk_prob)
 
-        # NextN/MTP prediction layers
-        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        if not self.no_mtp and (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
             self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
 
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
     _experts: list[dict[str, Tensor]] | None = None
 
     # note: unlike GLM4V non-MoE, we don't need to permute Q/K here since GLM4V_MOE uses Neox ordering already
@@ -348,6 +386,7 @@ class GlmMoeDsaModel(DeepseekV2Model):
 @ModelBase.example("upstage/Solar-Open-100B")
 class SolarOpenModel(Glm4MoeModel):
     model_arch = gguf.MODEL_ARCH.GLM4_MOE
+    supports_mtp_export = False
 
     def set_vocab(self):
         from transformers import AutoTokenizer

docs/autoparser.md

@@ -443,21 +443,21 @@ Each returned parser is wrapped by `wrap_for_generation_prompt()`, which prepend
 |                                           | `wrap_for_generation_prompt()`, string helpers                                  |
 | `common/chat-peg-parser.h/cpp`            | `common_chat_peg_builder`, `common_chat_peg_mapper`, and helpers                |
 | `common/chat.cpp`                         | Entry point: `common_chat_templates_apply_jinja()`                              |
-| `tools/parser/debug-template-parser.cpp`  | Debug tool for template analysis                                                |
-| `tools/parser/template-analysis.cpp`      | Template analysis tool                                                          |
+| `tests/test-chat-auto-parser.cpp`         | Auto-parser unit tests; also a debug tool when given a template path            |
+| `tests/test-chat-analysis.cpp`            | Template differential analysis debug tool                                       |
 
 ## Testing & Debugging
 
 ### Debug Tools
 
-**Template Debugger**: `tools/parser/debug-template-parser.cpp`
+**Template Debugger**: `tests/test-chat-auto-parser.cpp`
 
-- Usage: `./bin/llama-debug-template-parser path/to/template.jinja`
+- Usage: `./bin/test-chat-auto-parser path/to/template.jinja` (without a path, it runs the automated tests)
 - Shows detected format, markers, generated parser, and GBNF grammar
 
-**Template Analysis**: `tools/parser/template-analysis.cpp`
+**Template Analysis**: `tests/test-chat-analysis.cpp`
 
-- Usage: `./bin/llama-template-analysis path/to/template.jinja`
+- Usage: `./bin/test-chat-analysis --template-file path/to/template.jinja` (without arguments, it runs on all templates from the test suite)
 
 **Debug Logging**: Enable with `LLAMA_ARG_LOG_VERBOSITY=2`
 
@@ -519,7 +519,7 @@ The following templates have active tests in `tests/test-chat.cpp`:
 
 To support a new template format:
 
-1. **If it follows standard patterns** — The auto-parser should detect it automatically. Run `llama-debug-template-parser` to verify markers are correctly extracted.
+1. **If it follows standard patterns** — The auto-parser should detect it automatically. Run `test-chat-auto-parser <template_path>` to verify markers are correctly extracted.
 2. **If differential analysis extracts incorrect markers** — Add a workaround lambda to the `workarounds` vector in `common/chat-diff-analyzer.cpp`. Inspect the template source for a unique identifying substring.
 3. **If it needs fundamentally different handling** — Add a dedicated handler function in `chat.cpp` before the auto-parser block (as done for GPT-OSS, Functionary v3.2, and Ministral).
 

ggml/CMakeLists.txt

@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 21)
+set(GGML_VERSION_MINOR 22)
 set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
@@ -342,9 +342,6 @@ set(GGML_PUBLIC_HEADERS
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
-#if (GGML_METAL)
-#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
-#endif()
 install(TARGETS ggml LIBRARY PUBLIC_HEADER)
 install(TARGETS ggml-base LIBRARY)
 

ggml/include/ggml.h

@@ -1724,6 +1724,19 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_past);
 
+    GGML_API struct ggml_tensor * ggml_clamp(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_clamp_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
+
     GGML_API struct ggml_tensor * ggml_soft_max(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1990,14 +2003,6 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_offs);
 
-    // clamp
-    // in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_clamp(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            float                 min,
-            float                 max);
-
     // im2col
     // converts data into a format that effectively results in a convolution when combined with matrix multiplication
     GGML_API struct ggml_tensor * ggml_im2col(

ggml/src/ggml-alloc.c

@@ -40,6 +40,7 @@ bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_SILU_BACK:
         case GGML_OP_RMS_NORM:
         case GGML_OP_RMS_NORM_BACK:
+        case GGML_OP_CLAMP:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
             return true;

ggml/src/ggml-backend-meta.cpp

@@ -592,7 +592,18 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             GGML_ASSERT(split_states_equal(src_ss[0], src_ss[1]));
             return {assume_sync ? GGML_BACKEND_SPLIT_AXIS_MIRRORED : GGML_BACKEND_SPLIT_AXIS_PARTIAL, {0}, {1}, 1};
         }
-        GGML_ABORT("fatal error");
+        if (src_ss[0].axis == src_ss[1].axis && src_ss[0].axis >= GGML_BACKEND_SPLIT_AXIS_2 &&
+                src_ss[0].axis < GGML_MAX_DIMS) {
+            GGML_ASSERT(split_states_equal(src_ss[0], src_ss[1]));
+            return src_ss[0];
+        }
+        // batched matmul with the batches split across devices and a replicated activation
+        if (src_ss[0].axis >= GGML_BACKEND_SPLIT_AXIS_2 && src_ss[0].axis < GGML_MAX_DIMS &&
+                src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+            return src_ss[0];
+        }
+        GGML_ABORT("unsupported mul_mat split states: node=%s src0=%s axis=%d src1=%s axis=%d",
+            tensor->name, tensor->src[0]->name, (int) src_ss[0].axis, tensor->src[1]->name, (int) src_ss[1].axis);
         //return {GGML_BACKEND_SPLIT_AXIS_UNKNOWN, {0}, {1}, 1};
     };
 
@@ -602,27 +613,40 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             case GGML_BACKEND_SPLIT_AXIS_1:
             case GGML_BACKEND_SPLIT_AXIS_2:
             case GGML_BACKEND_SPLIT_AXIS_3: {
-                GGML_ASSERT(src_ss[0].n_segments == 1);
-                if (src_ss[0].axis == ggml_n_dims(tensor->src[0]) - 1 && src_ss[0].nr[0] == 1) {
-                    return {ggml_backend_meta_split_axis(ggml_n_dims(tensor) - 1), {0}, {1}, 1};
-                }
-                int64_t base_ne_in = tensor->src[0]->ne[0];
-                for (int dim = 1; dim <= src_ss[0].axis; dim++) {
+                int64_t base_ne_in = 1;
+                for (int dim = 0; dim <= src_ss[0].axis; dim++) {
                     base_ne_in *= tensor->src[0]->ne[dim];
                 }
-                base_ne_in /= src_ss[0].nr[0];
+                if (src_ss[0].n_segments == 1) {
+                    base_ne_in /= src_ss[0].nr[0];
+                    if (src_ss[0].axis == ggml_n_dims(tensor->src[0]) - 1 && src_ss[0].nr[0] == 1) {
+                        return {ggml_backend_meta_split_axis(ggml_n_dims(tensor) - 1), {0}, {1}, 1};
+                    }
+                    if (src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_0 && tensor->ne[0] == tensor->src[0]->ne[0] &&
+                            tensor->ne[1] == 1 && src_ss[0].nr[0] == 1) {
+                        bool complete_rows = true;
+                        for (size_t j = 0; j < n_bufs; j++) {
+                            const int64_t ne = src_ss[0].ne[j];
+                            complete_rows = complete_rows && (ne == 0 || ne == tensor->src[0]->ne[0]);
+                        }
+                        if (complete_rows) {
+                            // Move a complete dim-0 split to the following singleton dimension.
+                            return {GGML_BACKEND_SPLIT_AXIS_1, {0}, {1}, 1};
+                        }
+                    }
+                }
+                // Reshape outputs use one segment; split-state propagation merges source segments.
                 int64_t base_ne_out = 1;
                 for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
-                    const int64_t base_ne_out_next = base_ne_out *= tensor->ne[dim];
-                    if (base_ne_out_next % base_ne_in == 0) {
-                        return {ggml_backend_meta_split_axis(dim), {0}, {uint32_t(base_ne_out_next/base_ne_in)}, 1};
+                    base_ne_out *= tensor->ne[dim];
+                    if (base_ne_out % base_ne_in == 0) {
+                        return {ggml_backend_meta_split_axis(dim), {0}, {uint32_t(base_ne_out/base_ne_in)}, 1};
                     }
-                    if (base_ne_out_next > base_ne_in) {
+                    if (base_ne_out > base_ne_in) {
                         GGML_ASSERT(src_ss[0].n_segments == 1);
                         GGML_ASSERT(src_ss[0].nr[0]      == 1);
                         return {ggml_backend_meta_split_axis(dim), {0}, {1}, 1};
                     }
-                    base_ne_out = base_ne_out_next;
                 }
                 GGML_ABORT("shape mismatch for %s", ggml_op_name(tensor->op));
             }
@@ -747,14 +771,33 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
     };
 
     auto handle_flash_attn_ext = [&](const std::vector<ggml_backend_meta_split_state> & src_ss) -> ggml_backend_meta_split_state {
-        GGML_ASSERT(                             src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_2);
-        GGML_ASSERT(                             src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_2);
-        GGML_ASSERT(                             src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_2);
-        GGML_ASSERT(tensor->src[4] == nullptr || src_ss[3].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+        GGML_ASSERT(tensor->src[3] == nullptr || src_ss[3].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+
+        if (src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+            GGML_ASSERT(src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+            GGML_ASSERT(src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+            GGML_ASSERT(tensor->src[4] == nullptr || src_ss[4].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+            return {GGML_BACKEND_SPLIT_AXIS_MIRRORED, {0}, {1}, 1};
+        }
+
+        GGML_ASSERT(src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_2);
+        const bool kv_split = src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_2 &&
+                src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_2;
+        const bool kv_mirrored = src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED &&
+                src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED;
+        GGML_ASSERT(kv_split || kv_mirrored);
         GGML_ASSERT(tensor->src[4] == nullptr || src_ss[4].axis == GGML_BACKEND_SPLIT_AXIS_0);
         return {GGML_BACKEND_SPLIT_AXIS_1, {0}, {1}, 1};
     };
 
+    auto handle_lightning_indexer = [&](
+            const std::vector<ggml_backend_meta_split_state> & src_ss) -> ggml_backend_meta_split_state {
+        for (size_t i = 0; i < 4; i++) {
+            GGML_ASSERT(src_ss[i].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+        }
+        return {GGML_BACKEND_SPLIT_AXIS_MIRRORED, {0}, {1}, 1};
+    };
+
     auto handle_ssm_conv = [&](const std::vector<ggml_backend_meta_split_state> & src_ss) -> ggml_backend_meta_split_state {
         if (src_ss[0].axis == src_ss[1].axis) {
             if (src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_0) {
@@ -792,7 +835,7 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             ggml_backend_dev_t dev = ggml_backend_buft_get_device(ggml_backend_buffer_get_type(tensor->buffer));
             const ggml_backend_meta_device_context * dev_ctx = (const ggml_backend_meta_device_context *) dev->context;
             ggml_backend_meta_split_state ret = dev_ctx->get_split_state(tensor, dev_ctx->get_split_state_ud);
-            if (ret.axis >= 0 && ret.axis <= GGML_MAX_DIMS) {
+            if (ret.axis >= 0 && ret.axis < GGML_MAX_DIMS) {
                 const int64_t granularity = ret.axis == GGML_BACKEND_SPLIT_AXIS_0 ? ggml_blck_size(tensor->type) : 1;
                 int64_t ne_sum = 0;
                 for (size_t s = 0; s < ret.n_segments; s++) {
@@ -802,6 +845,9 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
                     }
                 }
                 GGML_ASSERT(ne_sum == tensor->ne[ret.axis]);
+            } else if (ret.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
+                GGML_ASSERT(ret.n_segments == 1);
+                GGML_ASSERT(ret.nr[0] == 1);
             }
             return ret;
         }
@@ -922,7 +968,7 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
                 split_state = handle_rope(src_ss);
             } break;
             case GGML_OP_ROPE_BACK: {
-                split_state = handle_generic(src_ss, /*scalar_only =*/ true);
+                split_state = handle_rope(src_ss);
             } break;
             case GGML_OP_CLAMP: {
                 split_state = handle_generic(src_ss, /*scalar_only =*/ false);
@@ -986,6 +1032,9 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             case GGML_OP_GATED_DELTA_NET: {
                 split_state = handle_gated_delta_net(src_ss);
             } break;
+            case GGML_OP_LIGHTNING_INDEXER: {
+                split_state = handle_lightning_indexer(src_ss);
+            } break;
             case GGML_OP_DSV4_HC_COMB:
             case GGML_OP_DSV4_HC_PRE:
             case GGML_OP_DSV4_HC_POST: {
@@ -1070,13 +1119,14 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
         if (buf_ctx->debug > 0) {
             std::string srcs_info;
             for (size_t i = 0; i < GGML_MAX_SRC; i++) {
-                if (tensor->src[i] == nullptr) {
+                if (tensor->src[i] == nullptr || tensor->src[i] == tensor) {
                     continue;
                 }
                 if (!srcs_info.empty()) {
                     srcs_info += ", ";
                 }
-                const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor->src[0], true);
+                const ggml_backend_meta_split_state split_state =
+                        ggml_backend_meta_get_split_state(tensor->src[i], true);
                 GGML_ASSERT(split_state.n_segments == 1);
                 const char * axis_name = ggml_backend_meta_split_axis_name(split_state.axis);
                 std::string ne_info;
@@ -1255,6 +1305,108 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
     return ggml_backend_meta_buffer_init_tensor_impl(buf_ctx->get_simple_tensor_container(tensor), tensor);
 }
 
+static void ggml_backend_meta_buffer_memset_tensor(
+        ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    const size_t n_bufs = ggml_backend_meta_buffer_n_bufs(buffer);
+    const ggml_backend_meta_split_state split_state =
+            ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
+    GGML_ASSERT(ggml_is_contiguous(tensor) || split_state.axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+
+    if (split_state.n_segments != 1 || split_state.nr[0] != 1) {
+        GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
+        GGML_ASSERT(split_state.nr[0] != 0);
+        GGML_ASSERT(tensor->ne[3] == 1);
+
+        std::vector<size_t> simple_offsets(n_bufs, 0);
+        if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_0) {
+            GGML_ASSERT(tensor->ne[2] == 1);
+
+            const size_t row_stride = tensor->nb[1];
+            GGML_ASSERT(offset % row_stride == 0);
+            GGML_ASSERT(size   % row_stride == 0);
+            const int64_t row_start = offset / row_stride;
+            const int64_t row_count = size   / row_stride;
+            GGML_ASSERT(row_start + row_count <= tensor->ne[1]);
+
+            const int64_t blck_size = ggml_blck_size(tensor->type);
+            for (size_t s = 0; s < split_state.n_segments; s++) {
+                for (size_t r = 0; r < split_state.nr[s]; r++) {
+                    for (size_t j = 0; j < n_bufs; j++) {
+                        ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                        GGML_ASSERT(split_state.ne[s*n_bufs + j] % blck_size == 0);
+                        const size_t nbytes = split_state.ne[s*n_bufs + j]/blck_size * tensor->nb[0];
+                        for (int64_t row = 0; row < row_count; row++) {
+                            ggml_backend_tensor_memset(simple_tensor, value,
+                                    simple_offsets[j] + (row_start + row)*simple_tensor->nb[1], nbytes);
+                        }
+                        simple_offsets[j] += nbytes;
+                    }
+                }
+            }
+            return;
+        }
+
+        GGML_ASSERT(split_state.axis == GGML_BACKEND_SPLIT_AXIS_1);
+
+        const size_t row_stride = tensor->nb[2];
+        GGML_ASSERT(offset % row_stride == 0);
+        GGML_ASSERT(size   % row_stride == 0);
+        const int64_t row_start = offset / row_stride;
+        const int64_t row_count = size   / row_stride;
+        GGML_ASSERT(row_start + row_count <= tensor->ne[2]);
+
+        for (size_t s = 0; s < split_state.n_segments; s++) {
+            for (size_t r = 0; r < split_state.nr[s]; r++) {
+                for (size_t j = 0; j < n_bufs; j++) {
+                    ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                    const size_t nbytes = split_state.ne[s*n_bufs + j] * tensor->nb[1];
+                    for (int64_t row = 0; row < row_count; row++) {
+                        ggml_backend_tensor_memset(simple_tensor, value,
+                                simple_offsets[j] + (row_start + row)*simple_tensor->nb[2], nbytes);
+                    }
+                    simple_offsets[j] += nbytes;
+                }
+            }
+        }
+        return;
+    }
+
+    switch (split_state.axis) {
+        case GGML_BACKEND_SPLIT_AXIS_0:
+        case GGML_BACKEND_SPLIT_AXIS_1:
+        case GGML_BACKEND_SPLIT_AXIS_2: {
+            const size_t chunk_size_full = tensor->nb[split_state.axis + 1];
+            GGML_ASSERT(offset % chunk_size_full == 0);
+            GGML_ASSERT(size   % chunk_size_full == 0);
+            const int64_t i_start =  offset        / chunk_size_full;
+            const int64_t i_stop  = (offset + size) / chunk_size_full;
+            for (size_t j = 0; j < n_bufs; j++) {
+                ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                const size_t chunk_size = simple_tensor->nb[split_state.axis + 1];
+                if (chunk_size == 0) {
+                    continue;
+                }
+                for (int64_t i = i_start; i < i_stop; i++) {
+                    ggml_backend_tensor_memset(simple_tensor, value, i*chunk_size, chunk_size);
+                }
+            }
+        } break;
+        case GGML_BACKEND_SPLIT_AXIS_PARTIAL: {
+            GGML_ASSERT(value == 0);
+            [[fallthrough]];
+        }
+        case GGML_BACKEND_SPLIT_AXIS_MIRRORED: {
+            for (size_t j = 0; j < n_bufs; j++) {
+                ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                ggml_backend_tensor_memset(simple_tensor, value, offset, size);
+            }
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        }
+    }
+}
+
 static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     const size_t n_bufs = ggml_backend_meta_buffer_n_bufs(buffer);
     const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
@@ -1352,15 +1504,29 @@ static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, gg
         } break;
         case GGML_BACKEND_SPLIT_AXIS_PARTIAL: {
             GGML_ASSERT(tensor->type == GGML_TYPE_F32);
-            const int64_t ne = ggml_nelements(tensor);
-            std::vector<float> tmp;
-            tmp.reserve(ne);
-            for (int64_t i = 0; i < ne; i++) {
-                tmp.push_back(((const float *) data)[i] / n_bufs);
+            GGML_ASSERT(offset % sizeof(float) == 0);
+            GGML_ASSERT(size   % sizeof(float) == 0);
+            const size_t n_values = size / sizeof(float);
+            size_t n_contributors = 0;
+            for (size_t j = 0; j < n_bufs; j++) {
+                n_contributors += split_state.ne[j] != 0;
+            }
+            const bool has_contributor_mask = n_contributors != 0;
+            if (!has_contributor_mask) {
+                n_contributors = n_bufs;
+            }
+            std::vector<float> tmp(n_values);
+            for (size_t i = 0; i < n_values; i++) {
+                tmp[i] = ((const float *) data)[i] / n_contributors;
+            }
+            std::vector<float> zero;
+            if (has_contributor_mask) {
+                zero.resize(n_values, 0.0f);
             }
             for (size_t j = 0; j < n_bufs; j++) {
                 ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
-                ggml_backend_tensor_set(simple_tensor, tmp.data(), offset, size);
+                const float * partial = has_contributor_mask && split_state.ne[j] == 0 ? zero.data() : tmp.data();
+                ggml_backend_tensor_set(simple_tensor, partial, offset, size);
             }
         } break;
         default: {
@@ -1488,7 +1654,7 @@ static const ggml_backend_buffer_i ggml_backend_meta_buffer_iface = {
     /* .free_buffer     = */ ggml_backend_meta_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_meta_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_meta_buffer_init_tensor,
-    /* .memset_tensor   = */ nullptr, // TODO implement
+    /* .memset_tensor   = */ ggml_backend_meta_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_meta_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_meta_buffer_get_tensor,
     /* .set_tensor_2d   = */ nullptr,
@@ -1841,7 +2007,7 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
 
         {
             // For MoE models it may make sense to delay the AllReduce in order to reduce I/O:
-            auto get_i_delayed = [&](const int i) -> int {
+            auto get_i_delayed_branch = [&](const int i) -> int {
                 int id = i; // i_delayed
                 int idr = i; // i_delayed return, last safe return value
 
@@ -1941,6 +2107,62 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
                 return idr;
             };
 
+            // AllReduce(a) + AllReduce(b) == AllReduce(a + b) for independent partial branches.
+            auto get_i_delayed = [&](const int i) -> int {
+                const int i_delayed = get_i_delayed_branch(i);
+                ggml_tensor * node = cgraph->nodes[i_delayed];
+
+                if (ggml_node_get_use_count(cgraph, i_delayed) != 1) {
+                    return i_delayed;
+                }
+
+                for (int id = i_delayed + 1; id < cgraph->n_nodes; id++) {
+                    ggml_tensor * next = cgraph->nodes[id];
+                    if (next->view_src == node) {
+                        return i_delayed;
+                    }
+                    for (int s = 0; s < GGML_MAX_SRC; s++) {
+                        if (next->src[s] == node) {
+                            return i_delayed;
+                        }
+                    }
+
+                    if (next->view_src != nullptr && next->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(next->view_src->buffer)) {
+                        continue;
+                    }
+                    if (ggml_backend_meta_get_split_state(next, false).axis != GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
+                        continue;
+                    }
+
+                    const int i_other = id;
+                    const int i_other_delayed = get_i_delayed_branch(i_other);
+                    ggml_tensor * other = cgraph->nodes[i_other_delayed];
+                    if (ggml_node_get_use_count(cgraph, i_other_delayed) != 1 || i_other_delayed + 1 >= cgraph->n_nodes) {
+                        return i_delayed;
+                    }
+
+                    ggml_tensor * sum = cgraph->nodes[i_other_delayed + 1];
+                    if (sum->op != GGML_OP_ADD ||
+                            !ggml_are_same_shape(node, other) || node->type != other->type || sum->type != node->type ||
+                            !((sum->src[0] == node && sum->src[1] == other) ||
+                              (sum->src[0] == other && sum->src[1] == node)) ||
+                            ggml_backend_meta_get_split_state(sum, false).axis != GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        return i_delayed;
+                    }
+
+                    for (size_t j = 0; j < n_backends; j++) {
+                        auto & bcj = backend_ctx->backend_configs[j];
+                        const bool compute       = bcj.nodes[i]->flags       & GGML_TENSOR_FLAG_COMPUTE;
+                        const bool compute_other = bcj.nodes[i_other]->flags & GGML_TENSOR_FLAG_COMPUTE;
+                        if (compute != compute_other) {
+                            return i_delayed;
+                        }
+                    }
+                    return i_other_delayed + 1;
+                }
+                return i_delayed;
+            };
+
             int i_start = 0;
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];

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

@@ -38,6 +38,7 @@
 #include "ggml-cuda/out-prod.cuh"
 #include "ggml-cuda/pad.cuh"
 #include "ggml-cuda/pool2d.cuh"
+#include "ggml-cuda/pool1d.cuh"
 #include "ggml-cuda/quantize.cuh"
 #include "ggml-cuda/rope.cuh"
 #include "ggml-cuda/roll.cuh"
@@ -2326,6 +2327,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_POOL_2D:
             ggml_cuda_op_pool2d(ctx, dst);
             break;
+        case GGML_OP_POOL_1D:
+            ggml_cuda_op_pool1d(ctx, dst);
+            break;
         case GGML_OP_SUM:
             ggml_cuda_op_sum(ctx, dst);
             break;
@@ -4607,8 +4611,8 @@ static std::string ggml_cuda_device_description(int device) {
     const ggml_cuda_device_info & info = ggml_cuda_info();
     std::string description = prop.name;
     if (info.device_count > info.physical_device_count) {
-        description += " (physical device " + std::to_string(info.devices[device].physical_device) +
-                       ", virtual device " + std::to_string(info.devices[device].virtual_index) + ")";
+        description += " (dev p" + std::to_string(info.devices[device].physical_device) +
+                       "/v" + std::to_string(info.devices[device].virtual_index) + ")";
     }
     return description;
 }
@@ -5245,6 +5249,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_2D_DW:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_TRANSPOSE_2D:
+        case GGML_OP_POOL_1D:
         case GGML_OP_POOL_2D:
             return true;
         case GGML_OP_ACC:

ggml/src/ggml-cuda/pool1d.cu

@@ -0,0 +1,85 @@
+#include "pool1d.cuh"
+
+static __global__ void pool1d_nchw_kernel(
+        const int iw, const int ow,
+        const int kw, const int sw, const int pw,
+        const int parallel_elements,
+        const float * src, float * dst, const enum ggml_op_pool op) {
+    const int idx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (idx >= parallel_elements) {
+        return;
+    }
+
+    const int nc = idx / ow;
+    const int cur_ow = idx % ow;
+
+    const float * i_ptr = src + nc * iw;
+    float * o_ptr = dst + nc * ow;
+
+    const int start = cur_ow * sw - pw;
+    const int b = max(0, start);
+    const int e = min(iw, start + kw);
+
+    float res;
+    switch (op) {
+        case GGML_OP_POOL_AVG: res = 0.0f;     break;
+        case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+        default: return;
+    }
+
+    int count = 0;
+    for (int i = b; i < e; i++) {
+#if __CUDA_ARCH__ >= 350
+        float cur = __ldg(i_ptr + i);
+#else
+        float cur = i_ptr[i];
+#endif
+        switch (op) {
+            case GGML_OP_POOL_AVG: res += cur;                break;
+            case GGML_OP_POOL_MAX: res = max(res, cur);       break;
+            default: break;
+        }
+        count++;
+    }
+
+    if (op == GGML_OP_POOL_AVG) {
+        res = (count > 0) ? (res / count) : 0.0f;
+    }
+
+    o_ptr[cur_ow] = res;
+}
+
+static void pool1d_nchw_kernel_f32_f32_cuda(
+        const int iw, const int ow,
+        const int kw, const int sw, const int pw,
+        const int parallel_elements,
+        const float * src, float * dst, const enum ggml_op_pool op,
+        cudaStream_t stream) {
+    const int num_blocks = (parallel_elements + CUDA_POOL1D_BLOCK_SIZE - 1) / CUDA_POOL1D_BLOCK_SIZE;
+    dim3 block_nums(num_blocks);
+    pool1d_nchw_kernel<<<block_nums, CUDA_POOL1D_BLOCK_SIZE, 0, stream>>>(iw, ow, kw, sw, pw, parallel_elements, src, dst, op);
+}
+
+void ggml_cuda_op_pool1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
+    const int k0 = opts[1];
+    const int s0 = opts[2];
+    const int p0 = opts[3];
+
+    const int64_t IW = src0->ne[0];
+    const int64_t OW = dst->ne[0];
+    const int64_t nr = ggml_nrows(src0);
+
+    const int parallel_elements = (int)(nr * OW);
+
+    pool1d_nchw_kernel_f32_f32_cuda(IW, OW, k0, s0, p0, parallel_elements, src0_d, dst_d, op, stream);
+}

ggml/src/ggml-cuda/pool1d.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_POOL1D_BLOCK_SIZE 256
+
+void ggml_cuda_op_pool1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-metal/CMakeLists.txt

@@ -11,6 +11,7 @@ ggml_add_backend_library(ggml-metal
                          ggml-metal-common.cpp
                          ggml-metal-context.m
                          ggml-metal-ops.cpp
+                         ggml-metal-tuning.cpp
                         )
 
 target_link_libraries(ggml-metal PRIVATE
@@ -24,62 +25,119 @@ if (GGML_METAL_NDEBUG)
 endif()
 
 set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
+set(METALLIB_KERNELS_COMMON     "${CMAKE_CURRENT_SOURCE_DIR}/kernels/common.h")
+set(METALLIB_KERNELS_DEQUANTIZE "${CMAKE_CURRENT_SOURCE_DIR}/kernels/dequantize.h")
+set(METALLIB_KERNELS_QUANTIZE   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantize.h")
+
+set(METALLIB_KERNEL_SOURCES
+    kernels/fa.metal
+    kernels/mul_mv.metal
+    kernels/mul_mm.metal
+    kernels/quantize.metal
+    kernels/softmax.metal
+    kernels/norm.metal
+    kernels/unary.metal
+    kernels/binbcast.metal
+    kernels/reduce.metal
+    kernels/tri.metal
+    kernels/ssm.metal
+    kernels/wkv.metal
+    kernels/gated_delta_net.metal
+    kernels/solve_tri.metal
+    kernels/rope.metal
+    kernels/conv.metal
+    kernels/upscale.metal
+    kernels/argsort.metal
+    kernels/pool.metal
+    kernels/misc.metal
+)
+
 if (GGML_METAL_EMBED_LIBRARY)
     enable_language(ASM)
 
     add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
 
-    set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
+    set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
 
     file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
 
-    # merge ggml-common.h and ggml-metal.metal into a single file
-    set(METALLIB_EMBED_ASM        "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-    set(METALLIB_SOURCE_EMBED     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
+    set(METALLIB_EMBED_ASM_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(kind ${src} NAME_WE)
+        # symbol names must be valid C identifiers ('-' is not allowed)
+        string(REPLACE "-" "_" kind_sym ${kind})
 
-    add_custom_command(
-        OUTPUT "${METALLIB_EMBED_ASM}"
-        COMMAND echo "Embedding Metal library"
-        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
-        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
-        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
-        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
-        COMMENT "Generate assembly for embedded Metal library"
-        VERBATIM
-    )
+        set(SRC   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${kind}.metal")
+        set(EMBED "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.metal")
+        set(ASM   "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.s")
 
-    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
+        # only prepend headers that this source actually includes
+        set(HEADERS_FOR_SRC ${METALLIB_KERNELS_COMMON})
+        file(STRINGS ${SRC} _has_dequantize REGEX "#include \"dequantize\\.h\"")
+        file(STRINGS ${SRC} _has_quantize   REGEX "#include \"quantize\\.h\"")
+        if(_has_dequantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_DEQUANTIZE})
+        endif()
+        if(_has_quantize)
+            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_QUANTIZE})
+        endif()
+
+        add_custom_command(
+            OUTPUT "${ASM}"
+            # Step 1: concatenate shared headers + this kernel source
+            COMMAND cat ${HEADERS_FOR_SRC} ${SRC} > "${EMBED}.tmp1"
+            # Step 2: remove internal #include and #pragma once
+            COMMAND sed -e "/\#include \"common.h\"/d" -e "/\#include \"dequantize.h\"/d" -e "/\#include \"quantize.h\"/d" -e "/\#pragma once/d" < "${EMBED}.tmp1" > "${EMBED}.tmp2"
+            # Step 3: inline ggml-common.h (replacing __embed_ggml-common.h__ sentinel)
+            COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}" -e "/__embed_ggml-common.h__/d" < "${EMBED}.tmp2" > "${EMBED}.tmp3"
+            # Step 4: inline ggml-metal-impl.h
+            COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${EMBED}.tmp3" > "${EMBED}"
+            # Step 5: emit an asm chunk with kind-specific start/end symbols
+            #   note: '-' is illegal in C symbols, so we use kind_sym; the macOS
+            #   section name is limited to 16 chars so we keep it shared
+            #   across kinds (__ggml_metallib) and only vary the global symbols.
+            COMMAND echo ".section __DATA,__ggml_metallib"                       >  "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_start"               >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_start:"                     >> "${ASM}"
+            COMMAND echo .incbin "\"${EMBED}\""                                  >> "${ASM}"
+            COMMAND echo ".globl _ggml_metallib_${kind_sym}_end"                 >> "${ASM}"
+            COMMAND echo "_ggml_metallib_${kind_sym}_end:"                       >> "${ASM}"
+            DEPENDS ../ggml-common.h ggml-metal-impl.h
+                    kernels/common.h kernels/dequantize.h kernels/quantize.h
+                    kernels/${kind}.metal
+            COMMENT "Generate embedded Metal library for ${kind}"
+            VERBATIM
+        )
+
+        list(APPEND METALLIB_EMBED_ASM_FILES "${ASM}")
+    endforeach()
+
+    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM_FILES})
 else()
-    # copy metal files to bin directory
+    # copy header files to bin directory
     configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
-    configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
     configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
 
+    file(MAKE_DIRECTORY "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels")
+    configure_file(kernels/common.h     ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/common.h     COPYONLY)
+    configure_file(kernels/dequantize.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/dequantize.h COPYONLY)
+    configure_file(kernels/quantize.h   ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/quantize.h   COPYONLY)
+
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        configure_file(${src} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} COPYONLY)
+    endforeach()
+
     if (GGML_METAL_SHADER_DEBUG)
-        # custom command to do the following:
-        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: disabling fast math is needed in order to pass tests/test-backend-ops
         # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
         # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
         #       ref: https://github.com/ggml-org/whisper.cpp/issues/1720
         # note: adding -g causes segmentation fault during compile
-        #set(XC_FLAGS -fno-fast-math -fno-inline -g)
         set(XC_FLAGS -fno-fast-math -fno-inline)
     else()
         set(XC_FLAGS -O3)
     endif()
 
-    # Append macOS metal versioning flags
     if (GGML_METAL_MACOSX_VERSION_MIN)
         message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
         list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
@@ -90,35 +148,46 @@ else()
         list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
     endif()
 
+    # Compile each kernel source to .air, then link into default.metallib
+    set(AIR_FILES "")
+    foreach(src ${METALLIB_KERNEL_SOURCES})
+        get_filename_component(name ${src} NAME_WE)
+        set(AIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${name}.air")
+        list(APPEND AIR_FILES ${AIR})
+        add_custom_command(
+            OUTPUT ${AIR}
+            COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} -o ${AIR}
+            DEPENDS ${src} kernels/common.h kernels/dequantize.h kernels/quantize.h ${METALLIB_COMMON} ggml-metal-impl.h
+            COMMENT "Compiling ${src}"
+            VERBATIM
+        )
+    endforeach()
+
     add_custom_command(
         OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
-                xcrun -sdk macosx metallib        - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
         COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-        DEPENDS ggml-metal.metal ${METALLIB_COMMON}
-        COMMENT "Compiling Metal kernels"
-        )
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h
+        COMMAND rm -rf ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels
+        DEPENDS ${AIR_FILES}
+        COMMENT "Linking Metal kernels into default.metallib"
+    )
 
-    # FIXME: only add to the ggml-metal target?
     add_custom_target(
         ggml-metal-lib ALL
         DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
+    )
 endif() # GGML_METAL_EMBED_LIBRARY
 
 if (NOT GGML_METAL_EMBED_LIBRARY)
     install(
-        FILES src/ggml-metal/ggml-metal.metal
-        PERMISSIONS
-            OWNER_READ
-            OWNER_WRITE
-            GROUP_READ
-            WORLD_READ
-        DESTINATION ${CMAKE_INSTALL_BINDIR})
+        DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/kernels/
+        DESTINATION ${CMAKE_INSTALL_BINDIR}/kernels
+        FILES_MATCHING PATTERN "*.metal" PATTERN "*.h"
+    )
 
-        install(
-            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DESTINATION ${CMAKE_INSTALL_BINDIR}
-        )
+    install(
+        FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        DESTINATION ${CMAKE_INSTALL_BINDIR}
+    )
 endif()

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

@@ -1,6 +1,7 @@
 #include "ggml-metal-device.h"
 
 #include "ggml-metal-impl.h"
+#include "ggml-metal-tuning.h"
 
 #include "ggml-impl.h"
 
@@ -17,10 +18,10 @@ 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(int device) {
+ggml_metal_device_t ggml_metal_device_get(int device, int n_devices) {
     static std::vector<ggml_metal_device_ptr> devs;
 
-    devs.emplace_back(ggml_metal_device_init(device));
+    devs.emplace_back(ggml_metal_device_init(device, n_devices));
 
     return devs.back().get();
 }
@@ -1544,6 +1545,8 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
         bool    has_bias,
         bool    has_scap,
         bool    has_kvpad,
+        int32_t nqpsg,
+        int32_t ne,
         int32_t nsg,
         int32_t nwg,
         bool    use_kv_f16,
@@ -1559,11 +1562,17 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
 
     const char * type = use_kv_f16 ? "f16" : ggml_type_name(op->src[1]->type);
 
-    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
+    char qne_suffix[16] = {0};
+    if (!(nqpsg == 1 && ne == ggml_metal_tuning::fa_vec_baseline_ne(dk, dv))) {
+        snprintf(qne_suffix, sizeof(qne_suffix), "_q%d_ne%d", nqpsg, ne);
+    }
+
+    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d%s",
             "flash_attn_ext_vec",
             type,
             dk,
-            dv);
+            dv,
+            qne_suffix);
 
     snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_scap=%d_kvpad=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
             base,

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

@@ -207,6 +207,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
         bool    has_bias,
         bool    has_scap,
         bool    has_kvpad,
+        int32_t nqpsg,
+        int32_t ne,
         int32_t nsg,
         int32_t nwg,
         bool    use_kv_f16,
@@ -257,8 +259,12 @@ enum ggml_metal_device_id {
     GGML_METAL_DEVICE_M5_ULTRA,
 };
 
+const char * ggml_metal_device_id_token(enum ggml_metal_device_id id);
+
 struct ggml_metal_device_props {
     int device;
+    int device_phys;
+    int device_virt;
     char name[128];
     char desc[128];
 
@@ -277,6 +283,7 @@ struct ggml_metal_device_props {
     bool supports_gpu_family_apple7;
 
     enum ggml_metal_device_id device_id;
+    int gpu_family;
 
     int op_offload_min_batch_size;
 };
@@ -286,10 +293,10 @@ 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);
+ggml_metal_device_t ggml_metal_device_init(int device, int n_devices);
 void ggml_metal_device_free(ggml_metal_device_t dev);
 
-ggml_metal_device_t ggml_metal_device_get(int device);
+ggml_metal_device_t ggml_metal_device_get(int device, int n_devices);
 
 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>

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

@@ -95,8 +95,63 @@ int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_wi
     return pipeline.pipeline->obj.maxTotalThreadsPerThreadgroup;
 }
 
+//
+// MTLLibrary collection (one library per op-source, compiled separately)
+//
+
+// Single source of truth for the per-kind metal libraries. The order here
+// defines the enum values and every per-kind table below, so adding a library
+// is a one-line change here (plus adding its source to CMakeLists.txt).
+//   X(suffix, name): name is both the kernels/<name>.metal basename and the
+//   ggml_metallib_<name>_{start,end} embed-symbol stem.
+#define GGML_METAL_LIBS \
+    X(FA,              fa)             \
+    X(MUL_MV,          mul_mv)         \
+    X(MUL_MM,          mul_mm)         \
+    X(QUANTIZE,        quantize)       \
+    X(SOFTMAX,         softmax)        \
+    X(NORM,            norm)           \
+    X(UNARY,           unary)          \
+    X(BINBCAST,        binbcast)       \
+    X(REDUCE,          reduce)         \
+    X(TRI,             tri)            \
+    X(SSM,             ssm)            \
+    X(WKV,             wkv)            \
+    X(GATED_DELTA_NET, gated_delta_net)\
+    X(SOLVE_TRI,       solve_tri)      \
+    X(ROPE,            rope)           \
+    X(CONV,            conv)           \
+    X(UPSCALE,         upscale)        \
+    X(ARGSORT,         argsort)        \
+    X(POOL,            pool)           \
+    X(MISC,            misc)
+
+enum ggml_metal_lib_kind {
+#define X(e, s) GGML_METAL_LIB_##e,
+    GGML_METAL_LIBS
+#undef X
+    GGML_METAL_LIB_COUNT,
+};
+
+static const char * const k_lib_names[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = #s,
+    GGML_METAL_LIBS
+#undef X
+};
+
 struct ggml_metal_library {
-    id<MTLLibrary> obj;
+    // Per-kind compiled libraries. When single_library is true, the whole library
+    // (e.g. a pre-compiled default.metallib or a from-source build) lives at
+    // objs[0] and the remaining slots are nil.
+    id<MTLLibrary> objs[GGML_METAL_LIB_COUNT];
+    bool single_library; // true: combined library at objs[0]; false: per-kind libs in objs[*]
+
+    // Routing table: kernel function name -> objs[] index, populated from each
+    // compiled library's -[MTLLibrary functionNames]. The actual compiled
+    // libraries are the single source of truth for which library owns a kernel,
+    // so adding kernels later requires no manual routing maintenance.
+    // nil in single_library mode (everything resolves to objs[0]).
+    NSMutableDictionary<NSString *, NSNumber *> * fn_to_lib;
 
     ggml_metal_device_t dev;
     ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
@@ -104,160 +159,376 @@ struct ggml_metal_library {
     NSLock * lock;
 };
 
-ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
-    id<MTLLibrary> library = nil;
-    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+// Build the fn_to_lib routing table by querying each compiled library's public
+// function names. Call once after all per-kind libraries have been compiled.
+static void ggml_metal_library_build_index(ggml_metal_library_t lib) {
+    @autoreleasepool {
+        NSMutableDictionary<NSString *, NSNumber *> * index = [[NSMutableDictionary alloc] init];
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            for (NSString * fname in [lib->objs[kind] functionNames]) {
+                index[fname] = @(kind);
+            }
+        }
+        lib->fn_to_lib = index;
+    }
+}
 
-    // load library
-    //
-    // - first check if the library is embedded
-    // - then check if the library is in the bundle
-    // - if not found, load the source and compile it
-    // - if that fails, return NULL
-    //
-    // TODO: move to a function
-    {
-        const int64_t t_start = ggml_time_us();
+// Parse a `#include "name"` line. Returns the quoted name in *include_name on
+// success. Whitespace-tolerant; ignores `#include <...>` (system headers).
+static bool ggml_metal_library_parse_quoted_include(NSString * line, NSString ** include_name) {
+    NSScanner * scanner = [NSScanner scannerWithString:line];
+    scanner.charactersToBeSkipped = [NSCharacterSet whitespaceCharacterSet];
 
-        NSError * error = nil;
-        NSString * src = nil;
+    if (![scanner scanString:@"#" intoString:NULL] ||
+        ![scanner scanString:@"include" intoString:NULL] ||
+        ![scanner scanString:@"\"" intoString:NULL]) {
+        return false;
+    }
 
-#if GGML_METAL_EMBED_LIBRARY
-        GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+    NSString * name = nil;
+    if (![scanner scanUpToString:@"\"" intoString:&name]) {
+        return false;
+    }
 
-        extern const char ggml_metallib_start[];
-        extern const char ggml_metallib_end[];
+    if (include_name) {
+        *include_name = name;
+    }
+    return true;
+}
 
-        src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
+// Recursively inline `#include "name"` directives. System includes (<...>),
+// `#if/#else/#endif`, and other preprocessor lines are passed through to the
+// Metal compiler unchanged. `#pragma once` is dropped since `seen` already
+// guards against double-inclusion.
+static bool ggml_metal_library_flatten_file(NSMutableString * dst, NSString * path,
+                                            NSArray<NSString *> * search_paths,
+                                            NSMutableSet<NSString *> * seen, NSError ** error) {
+    NSString * key = [path stringByStandardizingPath];
+    if ([seen containsObject:key]) {
+        return true;
+    }
+    [seen addObject:key];
 
-#ifdef SWIFT_PACKAGE
-        NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
-#else
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-#endif
+    NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:error];
+    if (!src) {
+        return false;
+    }
 
-        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (path_lib == nil) {
-            // Try to find the resource in the directory where the current binary located.
-            NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
-            NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+    NSFileManager * fm = [NSFileManager defaultManager];
+    for (NSString * line in [src componentsSeparatedByString:@"\n"]) {
+        NSString * trimmed = [line stringByTrimmingCharactersInSet:[NSCharacterSet whitespaceCharacterSet]];
+        if ([trimmed isEqualToString:@"#pragma once"]) {
+            continue;
+        }
 
-            NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
-            if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
-
-                NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
-                if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
-                    // Optionally, if this is a symlink, try to resolve it.
-                    path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
-                    if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
-                        // It is a relative path, adding the binary directory as directory prefix.
-                        path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
-                    }
-                    if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                        // Link to the resource could not be resolved.
-                        path_lib_default = nil;
-                    } else {
-                        GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
-                    }
+        NSString * include_name = nil;
+        if (ggml_metal_library_parse_quoted_include(line, &include_name)) {
+            NSString * resolved = nil;
+            for (NSString * dir in search_paths) {
+                NSString * candidate = [dir stringByAppendingPathComponent:include_name];
+                if ([fm isReadableFileAtPath:candidate]) {
+                    resolved = candidate;
+                    break;
                 }
-            } else {
-                // The resource couldn't be found in the binary's directory.
-                path_lib_default = nil;
             }
-
-            path_lib = path_lib_default;
+            if (!resolved) {
+                if (error) {
+                    NSString * msg = [NSString stringWithFormat:@"could not resolve include \"%@\" from '%@'", include_name, path];
+                    *error = [NSError errorWithDomain:@"ggml-metal-source-flatten" code:1
+                                             userInfo:@{NSLocalizedDescriptionKey: msg}];
+                }
+                return false;
+            }
+            if (!ggml_metal_library_flatten_file(dst, resolved, search_paths, seen, error)) {
+                return false;
+            }
+            continue;
         }
 
-        if (path_lib != nil) {
-            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+        [dst appendString:line];
+        [dst appendString:@"\n"];
+    }
 
-            library = [device newLibraryWithURL:libURL error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        } else {
-            GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+    return true;
+}
 
-            NSString * path_source;
-            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+static NSString * ggml_metal_library_flatten_source(NSString * path_source, NSError ** error) {
+    // Search paths cover both runtime layout (build/bin/kernels + build/bin)
+    // and source-tree layout (ggml/src/ggml-metal/kernels + ggml/src/ggml-metal + ggml/src).
+    NSString * path_kernels = [path_source stringByDeletingLastPathComponent];
+    NSString * path_base    = [path_kernels stringByDeletingLastPathComponent];
+    NSArray<NSString *> * search_paths = @[
+        path_kernels,
+        path_base,
+        [path_base stringByDeletingLastPathComponent],
+    ];
 
-            GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
+    NSMutableString * src = [[NSMutableString alloc] init];
+    NSMutableSet<NSString *> * seen = [NSMutableSet set];
 
-            if (path_resource) {
-                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
-            } else {
-                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+    if (!ggml_metal_library_flatten_file(src, path_source, search_paths, seen, error)) {
+        [src release];
+        return nil;
+    }
+    return src;
+}
+
+// Compile all per-kind libraries in parallel. `source_for_kind` returns the MSL
+// source for a kind (the helper takes ownership and releases it), or nil with
+// *err set on failure. On success the objs[] slots are populated and the routing
+// index is built; on any failure every error is logged and false is returned
+// (the caller is responsible for freeing `res`).
+static bool ggml_metal_library_compile_all(
+        ggml_metal_library_t res,
+        id<MTLDevice> device,
+        NSDictionary * prep,
+        NSString * (^source_for_kind)(int kind, NSError ** err),
+        const char * origin) {
+    const int64_t t_start = ggml_time_us();
+
+    int64_t  * t_per_lib   = calloc(GGML_METAL_LIB_COUNT, sizeof(int64_t));
+    NSError ** err_per_lib = calloc(GGML_METAL_LIB_COUNT, sizeof(NSError *));
+    __block atomic_bool any_failure = false;
+
+    dispatch_group_t group = dispatch_group_create();
+    dispatch_queue_t queue = dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0);
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        dispatch_group_async(group, queue, ^{
+
+            const int64_t t0 = ggml_time_us();
+
+            NSError * error = nil;
+
+            NSString * src = source_for_kind(kind, &error);
+            if (!src) {
+                err_per_lib[kind] = [error retain];
+                atomic_store(&any_failure, true);
+                return;
             }
 
-            if (path_source == nil) {
-                GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                path_source = @"ggml-metal.metal";
-            }
+            id<MTLLibrary> lib = nil;
 
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
-
-            src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        }
-#endif
-
-        if (!library) {
             @autoreleasepool {
-                // dictionary of preprocessor macros
-                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
-
-                if (ggml_metal_device_get_props(dev)->has_bfloat) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
-                }
-
-                if (ggml_metal_device_get_props(dev)->has_tensor) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
-                }
-
-#if GGML_METAL_EMBED_LIBRARY
-                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
-#endif
-
                 MTLCompileOptions * options = [MTLCompileOptions new];
                 options.preprocessorMacros = prep;
 
-                //[options setFastMathEnabled:false];
+                lib = [device newLibraryWithSource:src options:options error:&error];
 
-                library = [device newLibraryWithSource:src options:options error:&error];
-                if (error) {
-                    GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                    return nil;
-                }
-
-#if !__has_feature(objc_arc)
                 [options release];
-#endif
+
+                // retain the error before the autorelease pool drains it
+                if (!lib) {
+                    err_per_lib[kind] = [error retain];
+                }
+            }
+
+            [src release];
+
+            t_per_lib[kind] = ggml_time_us() - t0;
+
+            if (!lib) {
+                atomic_store(&any_failure, true);
+                return;
+            }
+
+            res->objs[kind] = lib;
+        });
+    }
+    dispatch_group_wait(group, DISPATCH_TIME_FOREVER);
+    dispatch_release(group);
+
+    const bool ok = !atomic_load(&any_failure);
+
+    if (ok) {
+        const int64_t t_total = ggml_time_us() - t_start;
+        int64_t t_max = 0;
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            GGML_LOG_DEBUG("%s: compiled '%s' library in %.3f sec\n",
+                           __func__, k_lib_names[kind], t_per_lib[kind] / 1e6);
+            if (t_per_lib[kind] > t_max) t_max = t_per_lib[kind];
+        }
+        GGML_LOG_INFO("%s: loaded %d libraries from %s in %.3f sec (max single = %.3f sec)\n",
+                      __func__, GGML_METAL_LIB_COUNT, origin, t_total / 1e6, t_max / 1e6);
+
+        ggml_metal_library_build_index(res);
+    } else {
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            if (err_per_lib[kind]) {
+                GGML_LOG_ERROR("%s: failed to build '%s' library: %s\n", __func__,
+                               k_lib_names[kind], [[err_per_lib[kind] description] UTF8String]);
+                [err_per_lib[kind] release];
             }
         }
-
-#if GGML_METAL_EMBED_LIBRARY
-        [src release];
-#endif // GGML_METAL_EMBED_LIBRARY
-
-        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
     }
 
-    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
+    free(err_per_lib);
+    free(t_per_lib);
 
-    res->obj       = library;
+    return ok;
+}
+
+ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+
+    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
     res->dev       = dev;
     res->pipelines = ggml_metal_pipelines_init();
     res->lock      = [NSLock new];
 
+    // shared MTLCompileOptions preprocessor macros (matches the build-time defines)
+    NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+    if (ggml_metal_device_get_props(dev)->has_bfloat) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
+    }
+    if (ggml_metal_device_get_props(dev)->has_tensor) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
+    }
+#if GGML_METAL_EMBED_LIBRARY
+    [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
+#endif
+
+#if GGML_METAL_EMBED_LIBRARY
+    GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+
+    // start/end symbols emitted by CMake (see CMakeLists.txt), one pair per kind
+#define X(e, s) extern const char ggml_metallib_##s##_start[]; extern const char ggml_metallib_##s##_end[];
+    GGML_METAL_LIBS
+#undef X
+
+    static const char * const lib_start[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_start,
+    GGML_METAL_LIBS
+#undef X
+    };
+    static const char * const lib_end[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_end,
+    GGML_METAL_LIBS
+#undef X
+    };
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            (void) err;
+            return [[NSString alloc] initWithBytes:lib_start[kind]
+                                            length:(lib_end[kind] - lib_start[kind])
+                                          encoding:NSUTF8StringEncoding];
+        }, "embedded data");
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
     return res;
+#else
+#ifdef SWIFT_PACKAGE
+    NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
+#else
+    NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
+
+    const int64_t t_start = ggml_time_us();
+
+    NSError * error = nil;
+    NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+    if (path_lib == nil) {
+        // Try to find the resource in the directory where the current binary located.
+        NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
+        NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+
+        NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
+        if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+            GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
+
+            NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
+            if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
+                // Optionally, if this is a symlink, try to resolve it.
+                path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
+                if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
+                    // It is a relative path, adding the binary directory as directory prefix.
+                    path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
+                }
+                if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+                    // Link to the resource could not be resolved.
+                    path_lib_default = nil;
+                } else {
+                    GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
+                }
+            }
+        } else {
+            // The resource couldn't be found in the binary's directory.
+            path_lib_default = nil;
+        }
+
+        path_lib = path_lib_default;
+    }
+
+    if (path_lib != nil) {
+        // pre-compiled library found: a single combined default.metallib
+        NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+        GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
+        res->objs[0]        = [device newLibraryWithURL:libURL error:&error];
+        res->single_library = true;
+        if (!res->objs[0]) {
+            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            ggml_metal_library_free(res);
+            return NULL;
+        }
+
+        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
+        return res;
+    }
+
+    // no pre-compiled metallib: fall back to compiling each kernel source separately
+    GGML_LOG_INFO("%s: default.metallib not found, loading kernel sources\n", __func__);
+
+    NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+    if (path_resource) {
+        GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, [path_resource UTF8String]);
+    }
+
+    // resolve each kind's source path up front (file lookup/logging stays on the calling thread)
+    NSString ** path_per_kind = calloc(GGML_METAL_LIB_COUNT, sizeof(NSString *));
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        NSString * rel = [NSString stringWithFormat:@"kernels/%s.metal", k_lib_names[kind]];
+
+        NSString * path_source = nil;
+        if (path_resource) {
+            path_source = [path_resource stringByAppendingPathComponent:rel];
+        } else {
+            NSString * stem = [NSString stringWithFormat:@"kernels/%s", k_lib_names[kind]];
+            path_source = [bundle pathForResource:stem ofType:@"metal"];
+        }
+
+        if (path_source == nil || ![[NSFileManager defaultManager] isReadableFileAtPath:path_source]) {
+            GGML_LOG_WARN("%s: could not locate %s in bundle, falling back to cwd\n", __func__, [rel UTF8String]);
+            path_source = rel;
+        }
+
+        GGML_LOG_DEBUG("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+        path_per_kind[kind] = [path_source retain];
+    }
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            return ggml_metal_library_flatten_source(path_per_kind[kind], err);
+        }, "source");
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        [path_per_kind[kind] release];
+    }
+    free(path_per_kind);
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
+    return res;
+#endif
 }
 
 ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose) {
@@ -319,10 +590,11 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
         return NULL;
     }
 
-    res->obj       = library;
-    res->dev       = dev;
-    res->pipelines = ggml_metal_pipelines_init();
-    res->lock      = [NSLock new];
+    res->objs[0]        = library;
+    res->single_library = true;
+    res->dev            = dev;
+    res->pipelines      = ggml_metal_pipelines_init();
+    res->lock           = [NSLock new];
 
     return res;
 }
@@ -332,8 +604,14 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
         return;
     }
 
-    if (lib->obj) {
-        [lib->obj release];
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        if (lib->objs[kind]) {
+            [lib->objs[kind] release];
+        }
+    }
+
+    if (lib->fn_to_lib) {
+        [lib->fn_to_lib release];
     }
 
     ggml_metal_pipelines_free(lib->pipelines);
@@ -394,11 +672,28 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_
 
         GGML_LOG_DEBUG("%s: compiling pipeline: base = '%s', name = '%s'\n", __func__, base, name);
 
+        // route to the library that actually defines this kernel; fn_to_lib is
+        // built from -[MTLLibrary functionNames] so it's always in sync
+        int lib_idx = 0;
+        if (!lib->single_library) {
+            NSNumber * idx = lib->fn_to_lib[base_func];
+            if (!idx) {
+                [lib->lock unlock];
+
+                GGML_LOG_ERROR("%s: kernel not found in any metal library: base = '%s', name = '%s'\n", __func__, base, name);
+
+                return res;
+            }
+            lib_idx = [idx intValue];
+        }
+
+        id<MTLLibrary> mtl_lib = lib->objs[lib_idx];
+
         id<MTLFunction> mtl_function;
         if (!cv) {
-            mtl_function = [lib->obj newFunctionWithName:base_func];
+            mtl_function = [mtl_lib newFunctionWithName:base_func];
         } else {
-            mtl_function = [lib->obj newFunctionWithName:base_func constantValues:cv->obj error:&error];
+            mtl_function = [mtl_lib newFunctionWithName:base_func constantValues:cv->obj error:&error];
         }
         if (!mtl_function) {
             [lib->lock unlock];
@@ -667,6 +962,34 @@ void ggml_metal_rsets_free(ggml_metal_rsets_t rsets) {
     free(rsets);
 }
 
+static const struct {
+    const char *              name;
+    const char *              token;
+    enum ggml_metal_device_id id;
+} k_metal_devices[] = {
+#define DEV(name, id) { name, #id, id }
+    DEV("M1",       GGML_METAL_DEVICE_M1),
+    DEV("M1 Pro",   GGML_METAL_DEVICE_M1_PRO),
+    DEV("M1 Max",   GGML_METAL_DEVICE_M1_MAX),
+    DEV("M1 Ultra", GGML_METAL_DEVICE_M1_ULTRA),
+    DEV("M2",       GGML_METAL_DEVICE_M2),
+    DEV("M2 Pro",   GGML_METAL_DEVICE_M2_PRO),
+    DEV("M2 Max",   GGML_METAL_DEVICE_M2_MAX),
+    DEV("M2 Ultra", GGML_METAL_DEVICE_M2_ULTRA),
+    DEV("M3",       GGML_METAL_DEVICE_M3),
+    DEV("M3 Pro",   GGML_METAL_DEVICE_M3_PRO),
+    DEV("M3 Max",   GGML_METAL_DEVICE_M3_MAX),
+    DEV("M3 Ultra", GGML_METAL_DEVICE_M3_ULTRA),
+    DEV("M4",       GGML_METAL_DEVICE_M4),
+    DEV("M4 Pro",   GGML_METAL_DEVICE_M4_PRO),
+    DEV("M4 Max",   GGML_METAL_DEVICE_M4_MAX),
+    DEV("M5",       GGML_METAL_DEVICE_M5),
+    DEV("M5 Pro",   GGML_METAL_DEVICE_M5_PRO),
+    DEV("M5 Max",   GGML_METAL_DEVICE_M5_MAX),
+    DEV("M5 Ultra", GGML_METAL_DEVICE_M5_ULTRA),
+#undef DEV
+};
+
 static enum ggml_metal_device_id ggml_metal_device_id_parse(const char * name) {
     if (!name) {
         return GGML_METAL_DEVICE_GENERIC;
@@ -678,40 +1001,24 @@ static enum ggml_metal_device_id ggml_metal_device_id_parse(const char * name) {
     }
     const char * suffix = name + sizeof(prefix) - 1;
 
-    static const struct {
-        const char * name;
-        enum ggml_metal_device_id id;
-    } table[] = {
-        {"M1",       GGML_METAL_DEVICE_M1},
-        {"M1 Pro",   GGML_METAL_DEVICE_M1_PRO},
-        {"M1 Max",   GGML_METAL_DEVICE_M1_MAX},
-        {"M1 Ultra", GGML_METAL_DEVICE_M1_ULTRA},
-        {"M2",       GGML_METAL_DEVICE_M2},
-        {"M2 Pro",   GGML_METAL_DEVICE_M2_PRO},
-        {"M2 Max",   GGML_METAL_DEVICE_M2_MAX},
-        {"M2 Ultra", GGML_METAL_DEVICE_M2_ULTRA},
-        {"M3",       GGML_METAL_DEVICE_M3},
-        {"M3 Pro",   GGML_METAL_DEVICE_M3_PRO},
-        {"M3 Max",   GGML_METAL_DEVICE_M3_MAX},
-        {"M3 Ultra", GGML_METAL_DEVICE_M3_ULTRA},
-        {"M4",       GGML_METAL_DEVICE_M4},
-        {"M4 Pro",   GGML_METAL_DEVICE_M4_PRO},
-        {"M4 Max",   GGML_METAL_DEVICE_M4_MAX},
-        {"M5",       GGML_METAL_DEVICE_M5},
-        {"M5 Pro",   GGML_METAL_DEVICE_M5_PRO},
-        {"M5 Max",   GGML_METAL_DEVICE_M5_MAX},
-        {"M5 Ultra", GGML_METAL_DEVICE_M5_ULTRA},
-    };
-
-    for (size_t i = 0; i < sizeof(table)/sizeof(table[0]); ++i) {
-        if (strcmp(suffix, table[i].name) == 0) {
-            return table[i].id;
+    for (size_t i = 0; i < sizeof(k_metal_devices)/sizeof(k_metal_devices[0]); ++i) {
+        if (strcmp(suffix, k_metal_devices[i].name) == 0) {
+            return k_metal_devices[i].id;
         }
     }
     return GGML_METAL_DEVICE_GENERIC;
 }
 
-ggml_metal_device_t ggml_metal_device_init(int device) {
+const char * ggml_metal_device_id_token(enum ggml_metal_device_id id) {
+    for (size_t i = 0; i < sizeof(k_metal_devices)/sizeof(k_metal_devices[0]); ++i) {
+        if (k_metal_devices[i].id == id) {
+            return k_metal_devices[i].token;
+        }
+    }
+    return "GGML_METAL_DEVICE_GENERIC";
+}
+
+ggml_metal_device_t ggml_metal_device_init(int device, int n_devices) {
     ggml_metal_device_t dev = calloc(1, sizeof(struct ggml_metal_device));
 
     assert(dev != NULL);
@@ -728,6 +1035,12 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
             dev->addr_virt = 0x000000400ULL;
 
             dev->props.device = device;
+
+            // the Metal backend uses the system default device as the single physical device;
+            // additional (virtual) devices are emulated on top of it via GGML_METAL_DEVICES
+            dev->props.device_phys = 0;
+            dev->props.device_virt = device;
+
             dev->props.has_simdgroup_reduction  = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7];
             dev->props.has_simdgroup_reduction |= [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
 
@@ -891,7 +1204,13 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
             }
 
             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]);
+            const char * gpu_name = [[dev->mtl_device name] UTF8String];
+            if (n_devices > 1) {
+                snprintf(dev->props.desc, sizeof(dev->props.desc), "%s (dev p%d/v%d)",
+                         gpu_name, dev->props.device_phys, dev->props.device_virt);
+            } else {
+                snprintf(dev->props.desc, sizeof(dev->props.desc), "%s", gpu_name);
+            }
 
             dev->library = ggml_metal_library_init(dev);
             if (!dev->library) {
@@ -913,7 +1232,8 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
             {
                 for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
                     if ([dev->mtl_device supportsFamily:i]) {
-                        GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d  (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
+                        dev->props.gpu_family = i - (int) MTLGPUFamilyApple1 + 1;
+                        GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d  (%d)\n", __func__, dev->props.gpu_family, i);
                         break;
                     }
                 }

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

@@ -7,6 +7,7 @@
 #include "ggml-metal-impl.h"
 #include "ggml-metal-common.h"
 #include "ggml-metal-device.h"
+#include "ggml-metal-tuning.h"
 
 #include <cassert>
 #include <algorithm>
@@ -3346,12 +3347,18 @@ 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_NQPSG; // queries per threadgroup
+        auto cfg = ggml_metal_tuning::fa_vec_pick(
+                props_dev->device_id,
+                props_dev->gpu_family,
+                (int) op->src[1]->type,
+                (int) ne00, (int) ne20,   // dk, dv (ne00 == dk for FA)
+                ne11, ne01);
+        int nqptg = cfg.Q;                             // queries per threadgroup
         const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
         const int nhptg = 1;                           // heads per threadgroup
 
         GGML_ASSERT(nqptg <= 32);
-        GGML_ASSERT(nqptg  % 1  == 0);
+        GGML_ASSERT(nqptg == 1 || nqptg == 2 || nqptg == 4);  // only instantiated Q values
         GGML_ASSERT(ncpsg  % 32 == 0);
 
         bool need_sync = false;
@@ -3410,7 +3417,7 @@ 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(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg)*nqptg)*(sizeof(float)/2), 16))
 
         int64_t nsg = 1;
 
@@ -3430,6 +3437,12 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             }
         }
 
+        // fall back to baseline (Q=1) if the tuned config exceeds threadgroup memory
+        if ((size_t) FATTN_SMEM(nsg) > props_dev->max_theadgroup_memory_size) {
+            cfg   = ggml_metal_tuning::fa_vec_baseline_cfg((int) ne00, (int) ne20);
+            nqptg = cfg.Q;  // = 1
+        }
+
         const int32_t ns10 = nb11_attn/nb10_attn;
         const int32_t ns20 = nb21_attn/nb20_attn;
 
@@ -3468,7 +3481,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.logit_softcap =*/ logit_softcap,
         };
 
-        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg, use_kv_f16, ns10, ns20);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nqptg, cfg.NE, nsg, nwg, use_kv_f16, ns10, ns20);
 
         GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 

ggml/src/ggml-metal/ggml-metal-tuning.h

@@ -0,0 +1,77 @@
+#pragma once
+
+#include "ggml-metal-device.h"  // enum ggml_metal_device_id
+#include "ggml.h"
+
+#include <cstdint>
+#include <vector>
+
+namespace ggml_metal_tuning {
+
+// FA vec selection buckets. ne01 (query rows) splits decode (==1) from batch (>=2), the
+// batch side refined into {2,3,4,5}: Q>1 reuses one K/V load across rows, so it only pays
+// off once ne01 aligns with Q. ne11 (KV length) is bucketed too, as the Q>1 crossover is
+// head-size dependent (small dk crosses late, large dk wins even at short KV).
+constexpr int FA_VEC_NE11_BUCKETS[] = { 1024, 4096, 16384 };
+constexpr int FA_VEC_NE01_BUCKETS[] = { 2, 3, 4, 5 };
+
+int fa_vec_ne11_bucket(int64_t ne11);
+int fa_vec_ne01_bucket(int64_t ne01);
+
+// NE baked into each (dk,dv) baseline instantiation in kernels/fa.metal.
+// Hand-maintained mirror; keep in sync with those instantiations.
+// The Metal test slice covers every legal config for dk=128 and dk=576.
+int fa_vec_baseline_ne(int dk, int dv);
+
+// Tuned table has two row kinds. Exact rows key a (ne11_b, ne01_b) bucket. Default rows
+// collapse ne11 over one ne01 domain: ne11_b == FA_VEC_NE11_DEFAULT and ne01_b holds the
+// domain. fa_vec_pick tries exact bucket -> domain default -> baseline; short KV
+// (ne11 < FA_VEC_NE11_BUCKETS[0]) always uses baseline.
+constexpr int8_t FA_VEC_NE11_DEFAULT  = -1;
+constexpr int8_t FA_VEC_DOMAIN_DECODE = 0;  // ne01 == 1
+constexpr int8_t FA_VEC_DOMAIN_BATCH  = 1;  // ne01 >= 2
+
+struct fa_vec_key_t {
+    int8_t  device_id;
+    int8_t  dtype;
+    int16_t dk;
+    int16_t dv;
+    int8_t  ne11_b;
+    int8_t  ne01_b;
+};
+
+static_assert(sizeof(fa_vec_key_t) == 8, "fa_vec_key_t must be tightly packed for memcmp");
+
+struct fa_vec_cfg_t {
+    int8_t Q;
+    int8_t NE;
+};
+
+struct fa_vec_entry_t {
+    fa_vec_key_t key;
+    fa_vec_cfg_t cfg;
+};
+
+// legal NE values for a (dk,dv): NL = 32/NE, require (dk/4)%NL==0 && (dv/4)%NL==0.
+// single source shared by the offline tuner and test-backend-ops.
+inline std::vector<int> fa_vec_legal_ne(int dk, int dv) {
+    std::vector<int> r;
+    for (int ne : { 1, 2, 4 }) {
+        const int nl = 32 / ne;
+        if ((dk / 4) % nl == 0 && (dv / 4) % nl == 0) {
+            r.push_back(ne);
+        }
+    }
+    return r;
+}
+
+// test/tune-only override; when set, fa_vec_pick returns it directly.
+void         fa_vec_set_override(fa_vec_cfg_t cfg);
+void         fa_vec_clear_override();
+fa_vec_cfg_t fa_vec_baseline_cfg(int dk, int dv);
+
+// device_id selects a per-SKU row; on a miss, gpu_family (0 if unknown) maps to a representative
+// SKU and the table is retried. No match -> baseline.
+fa_vec_cfg_t fa_vec_pick(enum ggml_metal_device_id device_id, int gpu_family, int dtype, int dk, int dv, int64_t ne11, int64_t ne01);
+
+}  // namespace ggml_metal_tuning

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

@@ -6,6 +6,7 @@
 #include "ggml-metal-device.h"
 #include "ggml-metal-context.h"
 #include "ggml-metal-ops.h"
+#include "ggml-metal-tuning.h"
 
 #include <mutex>
 #include <string>
@@ -203,6 +204,11 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
     ggml_metal_device_t ctx_dev = (ggml_metal_device_t)buft->device->context;
     ggml_metal_buffer_t res = ggml_metal_buffer_init(ctx_dev, size, shared);
 
+    if (res == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate Metal buffer of %zu bytes (out of memory)\n", __func__, size);
+        return NULL;
+    }
+
     ggml_backend_buffer_i buf_i = ggml_metal_buffer_is_shared(res)
         ? ggml_backend_metal_buffer_shared_i
         : ggml_backend_metal_buffer_private_i;
@@ -870,10 +876,55 @@ static ggml_backend_feature * ggml_backend_metal_get_features(ggml_backend_reg_t
     GGML_UNUSED(reg);
 }
 
+// test/tune-only override for the FA vec (Q, NE) selection, reached via proc_address.
+static void ggml_backend_metal_tuning_set_fa_vec_override(int Q, int NE) {
+    ggml_metal_tuning::fa_vec_set_override({ (int8_t) Q, (int8_t) NE });
+}
+
+static void ggml_backend_metal_tuning_clear_fa_vec_override(void) {
+    ggml_metal_tuning::fa_vec_clear_override();
+}
+
+static int ggml_backend_metal_tuning_fa_vec_ne11_bucket(int64_t ne11) {
+    return ggml_metal_tuning::fa_vec_ne11_bucket(ne11);
+}
+
+static int ggml_backend_metal_tuning_fa_vec_ne01_bucket(int64_t ne01) {
+    return ggml_metal_tuning::fa_vec_ne01_bucket(ne01);
+}
+
+static int ggml_backend_metal_tuning_fa_vec_baseline_ne(int dk, int dv) {
+    return ggml_metal_tuning::fa_vec_baseline_ne(dk, dv);
+}
+
+static const char * ggml_backend_metal_tuning_device_token(ggml_backend_dev_t dev) {
+    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
+
+    return ggml_metal_device_id_token(ggml_metal_device_get_props(ctx_dev)->device_id);
+}
+
 static void * ggml_backend_metal_get_proc_address(ggml_backend_reg_t reg, const char * name) {
     if (strcmp(name, "ggml_backend_get_features") == 0) {
         return (void *)ggml_backend_metal_get_features;
     }
+    if (strcmp(name, "ggml_backend_metal_tuning_set_fa_vec_override") == 0) {
+        return (void *)ggml_backend_metal_tuning_set_fa_vec_override;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_clear_fa_vec_override") == 0) {
+        return (void *)ggml_backend_metal_tuning_clear_fa_vec_override;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_ne11_bucket") == 0) {
+        return (void *)ggml_backend_metal_tuning_fa_vec_ne11_bucket;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_ne01_bucket") == 0) {
+        return (void *)ggml_backend_metal_tuning_fa_vec_ne01_bucket;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_baseline_ne") == 0) {
+        return (void *)ggml_backend_metal_tuning_fa_vec_baseline_ne;
+    }
+    if (strcmp(name, "ggml_backend_metal_tuning_device_token") == 0) {
+        return (void *)ggml_backend_metal_tuning_device_token;
+    }
 
     return NULL;
 
@@ -891,7 +942,7 @@ static ggml_backend_dev_t ggml_backend_metal_device_init(ggml_backend_reg_t reg,
     return new ggml_backend_device {
         /* .iface   = */ ggml_backend_metal_device_i,
         /* .reg     = */ reg,
-        /* .context = */ ggml_metal_device_get(device),
+        /* .context = */ ggml_metal_device_get(device, g_devices),
     };
 }
 

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

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

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

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

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

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

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

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

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

@@ -0,0 +1,735 @@
+#pragma once
+
+#include "common.h"
+
+#define GGML_COMMON_DECL_METAL
+#define GGML_COMMON_IMPL_METAL
+#if defined(GGML_METAL_EMBED_LIBRARY)
+__embed_ggml-common.h__
+#else
+#include "ggml-common.h"
+#endif
+
+#define QK_NL 16 // shared by mul_mm and get_rows_q instantiations
+
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f32_t4(device const float4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*src);
+}
+
+template <typename type4x4>
+void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+
+#if defined(GGML_METAL_HAS_BF16)
+template <typename type4x4>
+void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_bf16_t4(device const bfloat4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+#endif
+
+template <typename type4x4>
+void dequantize_q1_0(device const block_q1_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+    const float neg_d = -d;
+
+    const int byte_offset = il * 2;  // il*16 bits = il*2 bytes
+    const uint8_t b0 = qs[byte_offset];
+    const uint8_t b1 = qs[byte_offset + 1];
+
+    float4x4 reg_f;
+
+    reg_f[0][0] = select(neg_d, d, bool(b0 & 0x01));
+    reg_f[0][1] = select(neg_d, d, bool(b0 & 0x02));
+    reg_f[0][2] = select(neg_d, d, bool(b0 & 0x04));
+    reg_f[0][3] = select(neg_d, d, bool(b0 & 0x08));
+    reg_f[1][0] = select(neg_d, d, bool(b0 & 0x10));
+    reg_f[1][1] = select(neg_d, d, bool(b0 & 0x20));
+    reg_f[1][2] = select(neg_d, d, bool(b0 & 0x40));
+    reg_f[1][3] = select(neg_d, d, bool(b0 & 0x80));
+
+    reg_f[2][0] = select(neg_d, d, bool(b1 & 0x01));
+    reg_f[2][1] = select(neg_d, d, bool(b1 & 0x02));
+    reg_f[2][2] = select(neg_d, d, bool(b1 & 0x04));
+    reg_f[2][3] = select(neg_d, d, bool(b1 & 0x08));
+    reg_f[3][0] = select(neg_d, d, bool(b1 & 0x10));
+    reg_f[3][1] = select(neg_d, d, bool(b1 & 0x20));
+    reg_f[3][2] = select(neg_d, d, bool(b1 & 0x40));
+    reg_f[3][3] = select(neg_d, d, bool(b1 & 0x80));
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q1_0_t4(device const block_q1_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const float neg_d = -d;
+    const int base = il * 4;
+    const uint8_t byte = xb->qs[base / 8];
+    const int s = base % 8;
+
+    float4 reg_f;
+    reg_f[0] = select(neg_d, d, bool((byte >> (s    )) & 1));
+    reg_f[1] = select(neg_d, d, bool((byte >> (s + 1)) & 1));
+    reg_f[2] = select(neg_d, d, bool((byte >> (s + 2)) & 1));
+    reg_f[3] = select(neg_d, d, bool((byte >> (s + 3)) & 1));
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q2_0(device const block_q2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    const int byte_offset = il * 4;  // il*16 elements = il*4 bytes (4 elements per byte)
+    float4x4 reg_f;
+
+    for (int i = 0; i < 4; i++) {
+        const uint8_t b = qs[byte_offset + i];
+        reg_f[i][0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+        reg_f[i][1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+        reg_f[i][2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+        reg_f[i][3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q2_0_t4(device const block_q2_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const uint8_t b = xb->qs[il];
+
+    float4 reg_f;
+    reg_f[0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+    reg_f[1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+    reg_f[2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+    reg_f[3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = d1 * (qs[i] & mask0) + md;
+        reg_f[i/2][2*(i%2) + 1] = d2 * (qs[i] & mask1) + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + md;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + md;
+    }
+}
+
+
+
+template <typename type4x4>
+void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = ((qs[i] & mask0) * d1) + m;
+        reg_f[i/2][2*(i%2) + 1] = ((qs[i] & mask1) * d2) + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_1_t4(device const block_q4_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + m;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + md;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_0_t4(device const block_q5_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + md;
+        reg[2*ii + 1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + m;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + m;
+        reg[2*ii + 1] = d * x1 + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const packed_char4 * qs = (device const packed_char4 *) xb->qs;
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 4; ++i) {
+        reg_f[i] = float4(qs[4*il + i]) * d;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
+    device const packed_char4 * qs = (device const packed_char4 *) xb->qs;
+    const float d = xb->d;
+
+    reg = (type4) (float4(qs[il]) * d);
+}
+
+template <typename type4x4>
+void dequantize_mxfp4(device const block_mxfp4 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const uint8_t shr = il >= 1 ? 4 : 0;
+
+    for (int i = 0; i < 4; ++i) {
+        reg[i][0] = d * kvalues_mxfp4_f[(q2[4*i + 0] >> shr) & 0x0F];
+        reg[i][1] = d * kvalues_mxfp4_f[(q2[4*i + 1] >> shr) & 0x0F];
+        reg[i][2] = d * kvalues_mxfp4_f[(q2[4*i + 2] >> shr) & 0x0F];
+        reg[i][3] = d * kvalues_mxfp4_f[(q2[4*i + 3] >> shr) & 0x0F];
+    }
+}
+
+template <typename type4>
+void dequantize_mxfp4_t4(device const block_mxfp4 * xb, short il, thread type4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const short il4 = il%4;
+
+    const uint8_t shr = il >= 4 ? 4 : 0;
+
+    reg[0] = d * kvalues_mxfp4_f[(q2[4*il4 + 0] >> shr) & 0x0F];
+    reg[1] = d * kvalues_mxfp4_f[(q2[4*il4 + 1] >> shr) & 0x0F];
+    reg[2] = d * kvalues_mxfp4_f[(q2[4*il4 + 2] >> shr) & 0x0F];
+    reg[3] = d * kvalues_mxfp4_f[(q2[4*il4 + 3] >> shr) & 0x0F];
+}
+
+template <typename type4x4>
+void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
+    const float d = xb->d;
+    const float min = xb->dmin;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    float dl, ml;
+    uint8_t sc = xb->scales[il];
+
+    q = q + 32*(il/8) + 16*(il&1);
+    il = (il/2)%4;
+
+    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    device const uint8_t * h = (device const uint8_t *)xb->hmask;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    q = q + 32 * (il/8) + 16 * (il&1);
+    h = h + 16 * (il&1);
+    uint8_t m = 1 << (il/2);
+    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
+                                 ((il/4)>0 ? 12  : 3);
+    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
+    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
+
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
+
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
+    }
+}
+
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
+template <typename type4x4>
+void dequantize_q4_K(device const block_q4_K * xb, short il, thread type4x4 & reg) {
+    device const uchar * q = xb->qs;
+
+    short is = (il/4) * 2;
+    q = q + (il/4) * 32 + 16 * (il&1);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d   = il < 2 ? xb->d : xb->d / 16.h;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask = il < 2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q  = xb->qs;
+    device const uint8_t * qh = xb->qh;
+
+    short is = (il/4) * 2;
+    q  = q + 32 * (il/4) + 16 * (il&1);
+    qh = qh + 16 * (il&1);
+    uint8_t ul = 1 << (il/2);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint16_t * ql = (device const uint16_t *)xb->ql;
+    device const uint16_t * qh = (device const uint16_t *)xb->qh;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    ql = ql + 32*(il/8) + 16*((il/2)&1) + 8*(il&1);
+    qh = qh + 16*(il/8) + 8*(il&1);
+    float sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint32_t kmask1 = il>1 ? (il>2 ? 0xC0C0C0C0 : 0x30303030) : (il>0 ? 0x0C0C0C0C : 0x03030303);
+    const uint32_t kmask2 = il>1 ? 0xF0F0F0F0                       : 0x0F0F0F0F;
+    const float ml = d_all * sc * 32.f;
+    const float dl0 = d_all * sc;
+    const float dl1 = dl0 / 256.f;
+    const float dl2 = dl0 / (256.f * 256.f);
+    const float dl3 = dl0 / (256.f * 256.f * 256.f);
+    const uint8_t shr_h = il>2 ? 2 : 0;
+    const uint8_t shl_h = il>1 ? 0 : (il>0 ? 2 : 4);
+    const uint8_t shr_l = il>1 ? 4 : 0;
+    for (int i = 0; i < 4; ++i) {
+        const uint32_t  low = (ql[2*i] | (uint32_t)(ql[2*i+1] << 16)) & kmask2;
+        const uint32_t high = (qh[2*i] | (uint32_t)(qh[2*i+1] << 16)) & kmask1;
+        const uint32_t q = ((high << shl_h) >> shr_h) | (low >> shr_l);
+        reg[i][0] = dl0 *  ((half)(q & 0xFF))       - ml;
+        reg[i][1] = dl1 * ((float)(q & 0xFF00))     - ml;
+        reg[i][2] = dl2 * ((float)(q & 0xFF0000))   - ml;
+        reg[i][3] = dl3 * ((float)(q & 0xFF000000)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * signs = qs + QK_K/8;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
+        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    const float d = xb->d;
+    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint16_t * qh = xb->qh;
+    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
+    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
+    const uint16_t h = qh[ib32] >> 6*il;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    device const uint16_t * sc = (device const uint16_t *)xb->scales;
+
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = scale.f16;
+
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * qh = xb->qh + 2*ib32 + il;
+
+    const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
+    const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml1;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml2;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4>
+void dequantize_iq4_nl_t4(device const block_iq4_nl * xb, short il, thread type4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    aux32 = ((q4[2*(il%4)] | (q4[2*(il%4)+1] << 16)) >> 4*(il/4)) & 0x0f0f0f0f;
+    reg[0] = d * kvalues_iq4nl_f[q8[0]];
+    reg[1] = d * kvalues_iq4nl_f[q8[1]];
+    reg[2] = d * kvalues_iq4nl_f[q8[2]];
+    reg[3] = d * kvalues_iq4nl_f[q8[3]];
+}
+
+template <typename type4x4>
+void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
+    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)xb->d * (ls - 32);
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4x4>
+void dequantize_tq2_0(device const block_tq2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    // 2 bits per element, 4 elements per byte, 128 elements per 32-byte group
+    const short base = il * 16;
+    for (int k = 0; k < 16; k++) {
+        const int i = base + k;
+        const int byte = ((i >> 7) & 1) * 32 + (i & 31);
+        const int l = (i >> 5) & 3;
+        reg_f[k/4][k%4] = d * (float)(((qs[byte] >> (2*l)) & 3) - 1);
+    }
+
+    reg = (type4x4) reg_f;
+}

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

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

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

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

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

@@ -0,0 +1,853 @@
+#include "common.h"
+#include "dequantize.h"
+
+constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
+constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
+constant short FC_mul_mm_ne12  [[function_constant(FC_MUL_MM + 2)]];
+constant short FC_mul_mm_ne13  [[function_constant(FC_MUL_MM + 3)]];
+constant short FC_mul_mm_r2    [[function_constant(FC_MUL_MM + 4)]];
+constant short FC_mul_mm_r3    [[function_constant(FC_MUL_MM + 5)]];
+
+// each block_q contains 16*nl weights
+#ifdef GGML_METAL_HAS_TENSOR
+template<
+    typename SA, typename SA_4x4, typename SA_8x8,
+    typename SB, typename SB_2x4, typename SB_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread SA_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * srcA,
+        device const char * srcB,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig [[threadgroup_position_in_grid]],
+        ushort tiitg [[thread_index_in_threadgroup]],
+        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
+    (void) sgitg;
+
+    // Matrix dimensions: A(M,K) x B(K,N) -> C(M,N)
+    const int K = args.ne00;
+    const int M = args.ne0;
+    const int N = args.ne1;
+
+    // Batch dimension handling
+    const int im = tgpig.z;
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    // Batch offsets for srcA and srcB
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+
+    // Tile dimensions
+    constexpr int NRB = SZ_SIMDGROUP * N_MM_BLOCK_X * N_MM_SIMD_GROUP_X;
+    constexpr int NRA = SZ_SIMDGROUP * N_MM_BLOCK_Y * N_MM_SIMD_GROUP_Y;
+
+    // Tile offsets in output matrix
+    const int ra = tgpig.y * NRA;
+    const int rb = tgpig.x * NRB;
+
+    // Threadgroup memory for dequantized A tile only
+    threadgroup SA * sa = (threadgroup SA *)(shmem);
+
+    // Work-item count for A loading
+    constexpr int A_WORK_ITEMS = NRA * N_MM_NK;
+    constexpr int NUM_THREADS = N_SIMDWIDTH * N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y;
+
+    // tA wraps threadgroup memory
+    auto tA = tensor(sa, dextents<int32_t, 2>(N_MM_NK_TOTAL, NRA));
+
+    // tB wraps device memory directly
+    device T1 * ptrB = (device T1 *)(srcB + args.nb12*i12 + args.nb13*i13);
+    const int strideB = args.nb11 / sizeof(T1);
+    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));
+
+    // Configure matmul operation
+    // note: K is dynamic_extent (clamped to the valid range in PHASE 2), since a static
+    //       N_MM_NK_TOTAL K tile would read src1 out of bounds when K % N_MM_NK_TOTAL != 0
+    // ref: https://github.com/ggml-org/llama.cpp/pull/27064
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(
+            NRB, NRA, static_cast<int>(dynamic_extent), false, true, true,
+            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tB), decltype(tA), float>();
+
+    // Accumulate partial results over K dimension
+    for (int loop_k = 0; loop_k < K; loop_k += N_MM_NK_TOTAL) {
+        // === PHASE 1: Dequantization of A into threadgroup memory ===
+        for (int work = tiitg; work < A_WORK_ITEMS; work += NUM_THREADS) {
+            const int row = work / N_MM_NK;
+            const int k_chunk = work % N_MM_NK;
+            const int k_pos = loop_k + k_chunk * 16;
+            const short k_base = k_chunk * 16;
+
+            // Bounds check: skip device read if row is out of matrix bounds
+            if (ra + row < M) {
+                if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+                    // Element-wise reads when K is not aligned (nb01 not aligned for half4x4/float4x4).
+                    // MSL spec Table 2.5: half4x4 requires 8-byte alignment. When K is odd,
+                    // nb01 = K*2 is not 8-byte aligned, so odd-row pointers are misaligned.
+                    // Mirrors the legacy kernel's existing guard.
+                    device const T0 * row_ptr = (device const T0 *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? (SA) row_ptr[k_pos + i] : (SA)0;
+                    }
+                } else {
+                    const int block_idx = k_pos / (16 * nl);
+                    const short il = (k_pos / 16) % nl;
+
+                    device const block_q * row_ptr = (device const block_q *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    SA_4x4 temp_a;
+                    dequantize_func(row_ptr + block_idx, il, temp_a);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        // Zero-pad A for K positions beyond valid range (handles partial K iterations)
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? temp_a[i/4][i%4] : (SA)0;
+                    }
+                }
+            } else {
+                // Zero-pad rows beyond matrix bounds
+                FOR_UNROLL (short i = 0; i < 16; i++) {
+                    sa[row * N_MM_NK_TOTAL + (k_base + i)] = (SA)0;
+                }
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // === PHASE 2: Tensor matmul ===
+        // Clamp the K extent of both operand tensors to the remaining valid K range so
+        // the dynamic-K op never reads past the K extent of src1 (or the staged A tile).
+        const int kExt = min(N_MM_NK_TOTAL, K - loop_k);
+
+        auto tAv = tensor(sa, dextents<int32_t, 2>(kExt, NRA), array<int, 2>({1, N_MM_NK_TOTAL}));
+        auto tBv = tensor(ptrB + loop_k + rb * strideB, dextents<int32_t, 2>(kExt, N - rb), array<int, 2>({1, strideB}));
+
+        mm.run(tBv, tAv, cT);
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // Store result tile to output matrix (with batch offset)
+    // cT.store handles bounds checking via tD's extents (M, N)
+    device float * dstBatch = (device float *)dst + im * N * M;
+
+    auto tD = tensor(dstBatch, dextents<int32_t, 2>(M, N), array<int, 2>({1, M}));
+    cT.store(tD.slice(ra, rb));
+}
+
+#else
+
+template<
+    typename S0, typename S0_4x4, typename S0_8x8,
+    typename S1, typename S1_2x4, typename S1_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z;
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (args.ne1 - r1 < NR1) ? (args.ne1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int i12 = im % FC_mul_mm_ne12;
+    const int i13 = im / FC_mul_mm_ne12;
+
+    const uint64_t offset0 = (i12/FC_mul_mm_r2)*args.nb02 + (i13/FC_mul_mm_r3)*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1 + lr1)
+        + args.nb10*iy);
+
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+    }
+
+    if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
+        // if no bounds checks on the output are needed, we can directly write to device memory
+        device float * C = (device float *) dst +
+            (r0 + 32*(sgitg &  1)) + \
+            (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
+        }
+    } else {
+        // block is smaller than 64x32, we should avoid writing data outside of the matrix
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+        for (short i = 0; i < 8; i++) {
+            simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (sgitg == 0) {
+            for (int j = tiitg; j < nr1; j += NR1) {
+                device float  * D  = (device float  *) dst + r0 + (r1 + j)*args.ne0 + im*args.ne1*args.ne0;
+                device float4 * D4 = (device float4 *) D;
+
+                threadgroup float  * C  = temp_str + (j*NR0);
+                threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+                int i = 0;
+                for (; i < nr0/4; i++) {
+                    *(D4 + i) = *(C4 + i);
+                }
+
+                i *= 4;
+                for (; i < nr0; i++) {
+                    *(D + i) = *(C + i);
+                }
+            }
+        }
+    }
+}
+
+#endif // GGML_METAL_HAS_TENSOR
+
+template<short ne20> // n_expert_used
+kernel void kernel_mul_mm_id_map0(
+        constant ggml_metal_kargs_mul_mm_id_map0 & args,
+        device  const char * src2,
+        device        char * htpe,
+        device        char * hids,
+        threadgroup   char * shmem [[threadgroup(0)]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    const short ide = tpitg; // expert id
+
+    uint32_t n_all = 0;
+
+    device int32_t * ids_i32 = (device int32_t *) hids + ide*args.ne21;
+
+    for (int i21 = 0; i21 < args.ne21; i21 += ntg) { // n_tokens
+        if (i21 + tpitg < args.ne21) {
+            device const int32_t * src2_i32 = (device const int32_t *) (src2 + (i21 + tpitg)*args.nb21);
+
+            threadgroup uint16_t * sids = (threadgroup uint16_t *) shmem + tpitg*ne20;
+
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sids[i20] = src2_i32[i20];
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (short t = 0; t < ntg; t++) {
+            if (i21 + t >= args.ne21) {
+                break;
+            }
+
+            threadgroup const uint16_t * sids = (threadgroup const uint16_t *) shmem + t*ne20;
+
+            short sel = 0;
+            #pragma unroll(ne20)
+            for (short i20 = 0; i20 < ne20; i20++) {
+                sel += (sids[i20] == ide)*(i20 + 1);
+            }
+
+            ids_i32[n_all] = (i21 + t)*ne20 + sel - 1;
+
+            n_all += sel > 0;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    device uint32_t * tpe_u32 = (device uint32_t *) (htpe);
+    tpe_u32[ide] = n_all;
+}
+
+typedef decltype(kernel_mul_mm_id_map0<1>) kernel_mul_mm_id_map0_t;
+
+template [[host_name("kernel_mul_mm_id_map0_ne20_1" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<1>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_2" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<2>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_4" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<4>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_5" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<5>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_6" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<6>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_22")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<22>;
+
+template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm_id(
+        constant ggml_metal_kargs_mul_mm_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * htpe,
+        device const char * hids,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    threadgroup S0 * sa = (threadgroup S0 *)(shmem);
+    threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    threadgroup float * sc = (threadgroup float *)(shmem);
+#endif
+
+    constexpr int NR0 = 64;
+    constexpr int NR1 = 32;
+
+    constexpr int NK  = 32;
+    constexpr int NL0 = NK/16;
+    constexpr int NL1 = NK/8;
+
+    const int im = tgpig.z; // expert
+    const int r0 = tgpig.y*NR0;
+    const int r1 = tgpig.x*NR1;
+
+    device const uint32_t * tpe_u32 = (device const uint32_t *) (htpe);
+    device const int32_t  * ids_i32 = (device const int32_t  *) (hids);
+
+    const int32_t neh1 = tpe_u32[im];
+
+    if (r1 >= neh1) {
+        return;
+    }
+
+    // if this block is of 64x32 shape or smaller
+    const short nr0 = (args.ne0 - r0 < NR0) ? (args.ne0 - r0) : NR0;
+    const short nr1 = (    neh1 - r1 < NR1) ? (    neh1 - r1) : NR1;
+
+    // a thread shouldn't load data outside of the matrix
+    const short lr0 = ((short)tiitg/NL0) < nr0 ? ((short)tiitg/NL0) : nr0 - 1; // 0 .. 63
+    const short lr1 = ((short)tiitg/NL1) < nr1 ? ((short)tiitg/NL1) : nr1 - 1; // 0 .. 31
+
+    const short il0 = (tiitg % NL0);
+
+    short il = il0;
+
+    const int id = ids_i32[im*args.ne21 + r1 + lr1];
+
+    const short i11 = (id % args.ne20) % args.ne11;
+    const short i12 = (id / args.ne20);
+    const short i13 = 0;
+
+    const uint64_t offset0 = im*args.nb02 + i13*args.nb03;
+    const short    offset1 = il0/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + args.nb01*(r0 + lr0) + offset0) + offset1;
+
+    const short iy = 8*(tiitg % NL1);
+
+    device const T1 * y = (device const T1 *)(src1
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*i11
+        + args.nb10*iy);
+
+#ifndef GGML_METAL_HAS_TENSOR
+    S0_8x8 ma[4];
+    S1_8x8 mb[2];
+
+    simdgroup_float8x8 mc[8];
+
+    for (short i = 0; i < 8; i++){
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+#else
+    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
+    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
+
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<4>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
+#endif
+
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
+#ifndef GGML_METAL_HAS_TENSOR
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? (S0) *((device T0 *) x + i) : (S0) 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+              //const short lx = i%8;
+              //const short ly = (tiitg/NL0)%8;
+                const short lx = (tiitg/NL0)%8;
+                const short ly = i%8;
+
+                const short ib = 8*sx + sy;
+
+                // NOTE: this is massively slower.. WTF?
+                //sa[64*ib + 8*ly + lx] = temp_a[i/4][i%4];
+
+                *(sa + 64*ib + 8*ly + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+              //const short lx = (tiitg/NL1)%8;
+              //const short ly = i;
+
+                const short ib = 4*sx + sy;
+
+                *(sb + 64*ib + 8*ly + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+          //const short dx = sx;
+          //const short dy = sy;
+
+            const short ly = (tiitg/NL1)%8;
+
+            const short ib = 4*sx + sy;
+
+            *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#else
+        // load data and store to threadgroup memory
+        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            // no need for dequantization
+            for (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+            }
+        } else {
+            S0_4x4 temp_a;
+            dequantize_func(x, il, temp_a);
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            FOR_UNROLL (short i = 0; i < 16; i++) {
+                const short sx = 2*il0 + i/8;
+                const short sy = (tiitg/NL0)/8;
+
+                const short lx = i%8;
+                const short ly = (tiitg/NL0)%8;
+                //const short lx = (tiitg/NL0)%8;
+                //const short ly = i%8;
+
+                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
+            }
+        }
+
+        if (FC_mul_mm_bc_inp) {
+            for (short i = 0; i < 8; ++i) {
+                const short sx = (tiitg%NL1);
+                const short sy = (tiitg/NL1)/8;
+
+                const short lx = i;
+                const short ly = (tiitg/NL1)%8;
+                //const short lx = (tiitg/NL1)%8;
+                //const short ly = i;
+
+                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
+            }
+        } else {
+            const short sx = (tiitg%NL1);
+            const short sy = (tiitg/NL1)/8;
+
+            //const short lx = i;
+            const short ly = (tiitg/NL1)%8;
+            //const short lx = (tiitg/NL1)%8;
+            //const short ly = i;
+
+            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
+        }
+#endif
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
+
+        y += NK;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifndef GGML_METAL_HAS_TENSOR
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
+        threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
+
+        FOR_UNROLL (short ik = 0; ik < NK/8; ik++) {
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + 64*i, 8, 0, false);
+            }
+
+            simdgroup_barrier(mem_flags::mem_none);
+
+            FOR_UNROLL (short i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
+            }
+
+            lsma += 8*64;
+            lsmb += 4*64;
+        }
+#else
+        auto sA = tA.slice(0, 0);
+        auto sB = tB.slice(0, 0);
+
+        mm.run(sB, sA, cT);
+#endif
+    }
+
+    // block is smaller than 64x32, we should avoid writing data outside of the matrix
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+#ifdef GGML_METAL_HAS_TENSOR
+    auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
+    cT.store(tC);
+#else
+    threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
+
+    for (short i = 0; i < 8; i++) {
+        simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
+    }
+#endif
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (short j = sgitg; j < nr1; j += 4) {
+        const int id = ids_i32[im*args.ne21 + r1 + j];
+
+        const short ide = id % args.ne20;
+        const short idt = id / args.ne20;
+
+        device float  * D  = (device float  *) dst + r0 + ide*args.ne0 + idt*args.ne1*args.ne0;
+        device float4 * D4 = (device float4 *) D;
+
+        threadgroup float  * C  = (threadgroup float  *) shmem + j*NR0;
+        threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+        int i = tiisg;
+        for (; i < nr0/4; i += 32) {
+            *(D4 + i) = *(C4 + i);
+        }
+
+        i = (4*(nr0/4)) + tiisg;
+        for (; i < nr0; i += 32) {
+            *(D + i) = *(C + i);
+        }
+    }
+}
+
+//
+// matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_t;
+
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_bf16_f32")]]    kernel mul_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_q1_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_tq2_0_f32")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_f32_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_f16_f16")]]     kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_q1_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_1_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q8_0_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_mxfp4_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q2_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q3_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q4_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q5_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_q6_K_f16")]]    kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f16")]] kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq3_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq2_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_s_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq1_m_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_nl_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_iq4_xs_f16")]]  kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_tq2_0_f16")]]   kernel mul_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  half, half2x4>;
+
+//
+// indirect matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm_id<half, half4x4, simdgroup_half8x8, half, half2x4, simdgroup_half8x8, float4x4, 1, dequantize_f32, float, float4x4, float, float2x4>) mul_mm_id;
+
+template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   float, float2x4>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_mul_mm_id_bf16_f32")]]    kernel mul_mm_id kernel_mul_mm_id<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat, bfloat2x4, simdgroup_bfloat8x8, bfloat4x4,     1,     dequantize_bf16,    bfloat, bfloat4x4, float, float2x4>;
+#endif
+template [[host_name("kernel_mul_mm_id_q1_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  float, float2x4>;
+template [[host_name("kernel_mul_mm_id_tq2_0_f32")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  float, float2x4>;
+
+template [[host_name("kernel_mul_mm_id_f32_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32,     float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_f16_f16")]]     kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16,     half,   half4x4,   half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q1_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q1_0,    8,     dequantize_q1_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_0,    4,     dequantize_q2_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_1_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q8_0_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_mxfp4_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_mxfp4,   2,     dequantize_mxfp4,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q2_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q3_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q4_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q5_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_q6_K_f16")]]    kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K,    float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f16")]] kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs, float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m,   float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f16")]]  kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs,  float,  float4x4,  half, half2x4>;
+template [[host_name("kernel_mul_mm_id_tq2_0_f16")]]   kernel mul_mm_id kernel_mul_mm_id<half,   half4x4,   simdgroup_half8x8,   half,   half2x4,   simdgroup_half8x8,   block_tq2_0,   QK_NL, dequantize_tq2_0,   float,  float4x4,  half, half2x4>;

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

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

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

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

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

@@ -0,0 +1,262 @@
+#pragma once
+
+#include "common.h"
+
+void quantize_q1_0(device const float * src, device block_q1_0 & dst) {
+    float sum_abs = 0.0f;
+    for (int j = 0; j < QK1_0; j++) {
+        sum_abs += fabs(src[j]);
+    }
+    dst.d = sum_abs / QK1_0;
+
+    for (int j = 0; j < QK1_0 / 8; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK1_0; j++) {
+        if (src[j] >= 0.0f) {
+            dst.qs[j / 8] |= (1 << (j % 8));
+        }
+    }
+}
+
+void quantize_q2_0(device const float * src, device block_q2_0 & dst) {
+    float amax = 0.0f;
+    for (int j = 0; j < QK2_0; j++) {
+        float a = fabs(src[j]);
+        if (a > amax) amax = a;
+    }
+    const float d = amax;
+    dst.d = d;
+
+    const float id = d > 0.0f ? 1.0f / d : 0.0f;
+
+    for (int j = 0; j < QK2_0 / 4; j++) {
+        dst.qs[j] = 0;
+    }
+    for (int j = 0; j < QK2_0; j++) {
+        int q = (int)round(src[j] * id) + 1;
+        q = max(0, min(3, q));
+        dst.qs[j / 4] |= (q << (2 * (j % 4)));
+    }
+}
+
+void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float min = FLT_MAX;
+    float max = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; j++) {
+        const float v = src[j];
+        if (min > v) min = v;
+        if (max < v) max = v;
+    }
+
+    const float d = (max - min) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK5_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float max = src[0];
+    float min = src[0];
+
+    for (int j = 1; j < QK5_1; j++) {
+        const float v = src[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = src[j]*id;
+
+        dst.qs[j] = round(x0);
+    }
+}
+
+void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_NL; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / kvalues_iq4nl_f[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = src[0        + j]*id;
+        const float x1 = src[QK4_NL/2 + j]*id;
+
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+        dst.qs[j] = xi0 | (xi1 << 4);
+
+        const float v0 = kvalues_iq4nl_f[xi0];
+        const float v1 = kvalues_iq4nl_f[xi1];
+        const float w0 = src[0        + j]*src[0        + j];
+        const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+        sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+
+    }
+
+    dst.d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
+
+void quantize_tq2_0(device const float * src, device block_tq2_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK_K; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = (half) d;
+
+    for (int j = 0; j < QK_K/4; j += 32) {
+        for (int m = 0; m < 32; ++m) {
+            uint8_t q = 0;
+            for (int n = 0; n < 4; ++n) {
+                // -1, 0, 1 -> 0, 1, 2
+                int xi = (int)round(src[m + n*32] * id) + 1;
+                q += (uint8_t)((xi & 3) << (2*n));
+            }
+            dst.qs[j + m] = q;
+        }
+        src += 4*32;
+    }
+}

ggml/src/ggml-metal/kernels/quantize.metal

@@ -0,0 +1,435 @@
+#include "common.h"
+#include "dequantize.h"
+#include "quantize.h"
+
+template<typename T0, typename T1>
+kernel void kernel_cpy_t_t(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.ne00;) {
+        device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        dst_data[i00] = (T1) src[0];
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_t_t<float, float>) kernel_cpy_t;
+
+template [[host_name("kernel_cpy_f32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   float>;
+template [[host_name("kernel_cpy_f32_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   half>;
+template [[host_name("kernel_cpy_f32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<float,   int32_t>;
+template [[host_name("kernel_cpy_i32_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, float>;
+template [[host_name("kernel_cpy_i32_i32")]]   kernel kernel_cpy_t kernel_cpy_t_t<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_f32_bf16")]]  kernel kernel_cpy_t kernel_cpy_t_t<float,   bfloat>;
+#endif
+template [[host_name("kernel_cpy_f16_f32")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    float>;
+template [[host_name("kernel_cpy_f16_f16")]]   kernel kernel_cpy_t kernel_cpy_t_t<half,    half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_cpy_bf16_f32")]]  kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  float>;
+template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy_t_t<bfloat,  bfloat>;
+#endif
+
+template<short QK,
+         typename block_q,
+         void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_cpy_f32_q(
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK;
+
+    device block_q * dst_data = (device block_q *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        device const float * src = (device const float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + (i00*QK)*args.nb00);
+
+        quantize_func(src, dst_data[i00]);
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_f32_q<QK8_0,  block_q8_0,  quantize_q8_0>)  cpy_f_q_t;
+
+template [[host_name("kernel_cpy_f32_q8_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK8_0,  block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_cpy_f32_q1_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK1_0,  block_q1_0,   quantize_q1_0>;
+template [[host_name("kernel_cpy_f32_q2_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK2_0,  block_q2_0,   quantize_q2_0>;
+template [[host_name("kernel_cpy_f32_q4_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_0,  block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_cpy_f32_q4_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK4_1,  block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_cpy_f32_q5_0")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_0,  block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_cpy_f32_q5_1")]]   kernel cpy_f_q_t kernel_cpy_f32_q<QK5_1,  block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_cpy_f32_iq4_nl")]] kernel cpy_f_q_t kernel_cpy_f32_q<QK4_NL, block_iq4_nl, quantize_iq4_nl>;
+template [[host_name("kernel_cpy_f32_tq2_0")]]  kernel cpy_f_q_t kernel_cpy_f32_q<QK_K,   block_tq2_0,  quantize_tq2_0>;
+
+template<typename T4x4, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
+kernel void kernel_cpy_q_f32(
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig[2];
+    const int32_t i02 = tgpig[1];
+    const int32_t i01 = ntg[1] == 1 ? tgpig[0]%args.ne01 : tgpig[0]*ntg[1] + tpitg.y;
+    const int32_t iw0 = ntg[1] == 1 ? tgpig[0]/args.ne01 : 0;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
+
+    const int32_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int32_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int32_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int32_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
+
+    device const block_q * src_data = (device const block_q *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+    device       T4x4    * dst_data = (device       T4x4    *)(dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1 + i0*args.nb0);
+
+    for (int32_t i00 = iw0*ntg[0] + tpitg.x; i00 < args.nk0;) {
+        T4x4 temp;
+        dequantize_func(src_data + i00/nl, i00%nl, temp);
+        dst_data[i00] = temp;
+
+        break;
+    }
+}
+
+typedef decltype(kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>) cpy_q_f_t;
+
+template [[host_name("kernel_cpy_q1_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q2_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q2_0, 4, dequantize_q2_0>;
+template [[host_name("kernel_cpy_q4_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template [[host_name("kernel_cpy_tq2_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32<float4x4, block_tq2_0, QK_NL, dequantize_tq2_0>;
+
+template [[host_name("kernel_cpy_q1_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q1_0, 8, dequantize_q1_0>;
+template [[host_name("kernel_cpy_q2_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q2_0, 4, dequantize_q2_0>;
+template [[host_name("kernel_cpy_q4_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_cpy_q4_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_cpy_q5_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_cpy_q5_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_cpy_q8_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_q8_0, 2, dequantize_q8_0>;
+
+template [[host_name("kernel_cpy_tq2_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32<half4x4, block_tq2_0, QK_NL, dequantize_tq2_0>;
+
+template<typename T>
+kernel void kernel_concat(
+        constant ggml_metal_kargs_concat & args,
+        device  const char * src0,
+        device  const char * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = ntg.y == 1 ? tgpig.x : tgpig.x*ntg.y + tpitg.y;
+
+    if (i1 >= args.ne1) {
+        return;
+    }
+
+    int o[4] = {0, 0, 0, 0};
+    o[args.dim] = args.dim == 0 ? args.ne00 : (args.dim == 1 ? args.ne01 : (args.dim == 2 ? args.ne02 : args.ne03));
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        device const T * x;
+
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            x = (device const T *)(src0 + (i3       )*args.nb03 + (i2       )*args.nb02 + (i1       )*args.nb01 + (i0       )*args.nb00);
+        } else {
+            x = (device const T *)(src1 + (i3 - o[3])*args.nb13 + (i2 - o[2])*args.nb12 + (i1 - o[1])*args.nb11 + (i0 - o[0])*args.nb10);
+        }
+
+        device T * y = (device T *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
+
+        *y = *x;
+    }
+}
+
+typedef decltype(kernel_concat<float>) kernel_concat_t;
+
+template [[host_name("kernel_concat_f32")]]  kernel kernel_concat_t kernel_concat<float>;
+template [[host_name("kernel_concat_f16")]]  kernel kernel_concat_t kernel_concat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_concat_bf16")]] kernel kernel_concat_t kernel_concat<bfloat>;
+#endif
+template [[host_name("kernel_concat_i8")]]   kernel kernel_concat_t kernel_concat<char>;
+template [[host_name("kernel_concat_i16")]]  kernel kernel_concat_t kernel_concat<short>;
+template [[host_name("kernel_concat_i32")]]  kernel kernel_concat_t kernel_concat<int>;
+template [[host_name("kernel_concat_i64")]]  kernel kernel_concat_t kernel_concat<long>;
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
+kernel void kernel_get_rows_q(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg  [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (device const block_q *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (device      float4x4 *) ((      device char *) dst  + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        float4x4 temp;
+        dequantize_func(psrc + ind/nl, ind%nl, temp);
+        pdst[ind] = temp;
+
+        break;
+    }
+}
+
+template<typename T0, typename T>
+kernel void kernel_get_rows_f(
+        constant ggml_metal_kargs_get_rows & args,
+        device const void * src0,
+        device const void * src1,
+        device       void * dst,
+        uint3               tgpig[[threadgroup_position_in_grid]],
+        ushort              tiitg[[thread_index_in_threadgroup]],
+        ushort3             ntg [[threads_per_threadgroup]]) {
+    const int32_t iw0 = tgpig.x/args.ne10;
+    const int32_t i10 = tgpig.x%args.ne10;
+    const int32_t i11 = tgpig.y;
+    const int32_t i12 = tgpig.z;
+
+    const int32_t r = ((const device int32_t *) ((const device char *) src1 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10))[0];
+
+    const int32_t i02 = i11;
+    const int32_t i03 = i12;
+
+    auto psrc = (const device T0 *) ((const device char *) src0 + i03*args.nb03 + i02*args.nb02 +   r*args.nb01);
+    auto pdst = (      device T  *) ((      device char *)  dst + i12*args.nb3  + i11*args.nb2  + i10*args.nb1);
+
+    for (int ind = iw0*ntg.x + tiitg; ind < args.ne00t;) {
+        pdst[ind] = psrc[ind];
+
+        break;
+    }
+}
+
+typedef decltype(kernel_get_rows_f<float, float>) get_rows_f_t;
+
+template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_f_t kernel_get_rows_f<float, float>;
+template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_f_t kernel_get_rows_f<half,  float>;
+template [[host_name("kernel_get_rows_i32")]]  kernel get_rows_f_t kernel_get_rows_f<int32_t, int32_t>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_get_rows_bf16")]] kernel get_rows_f_t kernel_get_rows_f<bfloat, float>;
+#endif
+
+typedef decltype(kernel_get_rows_q<block_q4_0, 2, dequantize_q4_0>) get_rows_q_t;
+
+template [[host_name("kernel_get_rows_q1_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q1_0,    8, dequantize_q1_0>;
+template [[host_name("kernel_get_rows_q2_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_0,    4, dequantize_q2_0>;
+template [[host_name("kernel_get_rows_q4_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_0,    2, dequantize_q4_0>;
+template [[host_name("kernel_get_rows_q4_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_1,    2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q5_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_0,    2, dequantize_q5_0>;
+template [[host_name("kernel_get_rows_q5_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_1,    2, dequantize_q5_1>;
+template [[host_name("kernel_get_rows_q8_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q8_0,    2, dequantize_q8_0>;
+template [[host_name("kernel_get_rows_mxfp4")]]   kernel get_rows_q_t kernel_get_rows_q<block_mxfp4,   2, dequantize_mxfp4>;
+template [[host_name("kernel_get_rows_q2_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_get_rows_q3_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_get_rows_q4_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_get_rows_q5_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_get_rows_q6_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+template [[host_name("kernel_get_rows_tq2_0")]]   kernel get_rows_q_t kernel_get_rows_q<block_tq2_0,   QK_NL, dequantize_tq2_0>;
+
+template<typename TS, typename TI, short QK, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS      * src_row = (const device TS      *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + QK*ind, dst_row[ind]);
+    }
+}
+
+template<typename TS, typename TI, typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q32(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS      * src_row = (const device TS      *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + 32*ind, dst_row[ind]);
+    }
+}
+
+template<typename TS, typename TI, typename TD>
+kernel void kernel_set_rows_f(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const TI      i1  = ((const device TI *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device TD * dst_row = (      device TD *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device TS * src_row = (const device TS *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        dst_row[ind] = (TD) src_row[ind];
+    }
+}
+
+typedef decltype(kernel_set_rows_f<float, int64_t, float>) set_rows_f_t;
+
+template [[host_name("kernel_set_rows_f32_i64_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, float>;
+template [[host_name("kernel_set_rows_f32_i32_f32")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, float>;
+template [[host_name("kernel_set_rows_f32_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int64_t, half>;
+template [[host_name("kernel_set_rows_f32_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<float, int32_t, half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_f32_i64_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int64_t, bfloat>;
+template [[host_name("kernel_set_rows_f32_i32_bf16")]]  kernel set_rows_f_t kernel_set_rows_f<float, int32_t, bfloat>;
+#endif
+
+template [[host_name("kernel_set_rows_f16_i64_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int64_t, half>;
+template [[host_name("kernel_set_rows_f16_i32_f16")]]   kernel set_rows_f_t kernel_set_rows_f<half, int32_t, half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_set_rows_bf16_i64_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int64_t, bfloat>;
+template [[host_name("kernel_set_rows_bf16_i32_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat, int32_t, bfloat>;
+#endif
+
+typedef decltype(kernel_set_rows_q32<float, int64_t, block_q8_0, quantize_q8_0>) set_rows_q32_t;
+
+template [[host_name("kernel_set_rows_f32_i64_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_f32_i32_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_f32_i64_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_f32_i32_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_f32_i64_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_f32_i32_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_f32_i64_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_f32_i32_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_f32_i64_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_f32_i32_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_f32_i64_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int64_t, block_iq4_nl, quantize_iq4_nl>;
+template [[host_name("kernel_set_rows_f32_i32_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<float, int32_t, block_iq4_nl, quantize_iq4_nl>;
+
+typedef decltype(kernel_set_rows_q<float, int64_t, QK_K, block_tq2_0, quantize_tq2_0>) set_rows_qK_t;
+
+template [[host_name("kernel_set_rows_f32_i64_tq2_0")]]  kernel set_rows_qK_t kernel_set_rows_q<float, int64_t, QK_K, block_tq2_0, quantize_tq2_0>;
+template [[host_name("kernel_set_rows_f32_i32_tq2_0")]]  kernel set_rows_qK_t kernel_set_rows_q<float, int32_t, QK_K, block_tq2_0, quantize_tq2_0>;
+

ggml/src/ggml-metal/kernels/reduce.metal

@@ -0,0 +1,228 @@
+#include "common.h"
+
+kernel void kernel_op_sum_f32(
+        constant ggml_metal_kargs_sum & args,
+        device const float * src0,
+        device       float * dst,
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+
+    if (args.np == 0) {
+        return;
+    }
+
+    // TODO: become function constant
+    const uint nsg = (ntg.x + 31) / 32;
+
+    float sumf = 0;
+
+    for (uint64_t i0 = tpitg.x; i0 < args.np; i0 += ntg.x) {
+        sumf += src0[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float total = 0;
+
+    if (sgitg == 0) {
+        float v = 0;
+
+        if (tpitg.x < nsg) {
+            v = shmem_f32[tpitg.x];
+        }
+
+        total = simd_sum(v);
+
+        if (tpitg.x == 0) {
+            dst[0] = total;
+        }
+    }
+}
+
+constant short FC_sum_rows_op [[function_constant(FC_SUM_ROWS + 0)]];
+
+template <typename T0, typename T>
+kernel void kernel_sum_rows_impl(
+        constant ggml_metal_kargs_sum_rows & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_sum_rows_op
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    threadgroup T0 * shmem_t = (threadgroup T0 *) shmem;
+
+    if (sgitg == 0) {
+        shmem_t[tiisg] = 0.0f;
+    }
+
+    device const T0 * src_row = (device const T0 *) (src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T  * dst_row = (device       T  *) (dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    T0 sumf = T0(0.0f);
+
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
+    }
+
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_t[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_t[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        if (FC_OP == OP_SUM_ROWS_NUM_MEAN) {
+            if (is_same<float4, T0>::value) {
+                dst_row[0] = sum(sumf) / (4*args.ne00);
+            } else {
+                dst_row[0] = sum(sumf) / args.ne00;
+            }
+        } else {
+            dst_row[0] = sum(sumf);
+        }
+    }
+
+#undef FC_OP
+}
+
+typedef decltype(kernel_sum_rows_impl<float, float>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows_f32_f32")]]   kernel kernel_sum_rows_t kernel_sum_rows_impl<float,  float>;
+template [[host_name("kernel_sum_rows_f32_f32_4")]] kernel kernel_sum_rows_t kernel_sum_rows_impl<float4, float>;
+
+template<typename T>
+kernel void kernel_cumsum_blk(
+        constant ggml_metal_kargs_cumsum_blk & args,
+        device const char * src0,
+        device       char * tmp,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * src0_row = (device const float *) (src0 +
+            args.nb01*i01 +
+            args.nb02*i02 +
+            args.nb03*i03);
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+
+    float v = 0.0f;
+
+    if (i00 + tpitg.x < args.ne00) {
+        v = src0_row[i00 + tpitg.x];
+    }
+
+    float s = simd_prefix_inclusive_sum(v);
+
+    if (tiisg == N_SIMDWIDTH - 1) {
+        shmem_f32[sgitg] = s;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = simd_prefix_exclusive_sum(shmem_f32[tiisg]);
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    s += shmem_f32[sgitg];
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] = s;
+    }
+
+    if (args.outb && tpitg.x == ntg.x - 1) {
+        device float * tmp_row = (device float *) tmp +
+            args.net0*i01 +
+            args.net0*args.net1*i02 +
+            args.net0*args.net1*args.net2*i03;
+
+        tmp_row[ib] = s;
+    }
+}
+
+typedef decltype(kernel_cumsum_blk<float>) kernel_cumsum_blk_t;
+
+template [[host_name("kernel_cumsum_blk_f32")]] kernel kernel_cumsum_blk_t kernel_cumsum_blk<float>;
+
+template<typename T>
+kernel void kernel_cumsum_add(
+        constant ggml_metal_kargs_cumsum_add & args,
+        device const char * tmp,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int ib = tgpig[0]/args.ne01;
+
+    if (ib == 0) {
+        return;
+    }
+
+    const int i00 = ib*ntg.x;
+    const int i01 = tgpig[0]%args.ne01;
+    const int i02 = tgpig[1];
+    const int i03 = tgpig[2];
+
+    device const float * tmp_row = (device const float *) (tmp +
+            args.nbt1*i01 +
+            args.nbt2*i02 +
+            args.nbt3*i03);
+
+    device float * dst_row = (device float *) dst +
+        args.ne00*i01 +
+        args.ne00*args.ne01*i02 +
+        args.ne00*args.ne01*args.ne02*i03;
+
+    if (i00 + tpitg.x < args.ne00) {
+        dst_row[i00 + tpitg.x] += tmp_row[ib - 1];
+    }
+}
+
+typedef decltype(kernel_cumsum_add<float>) kernel_cumsum_add_t;
+
+template [[host_name("kernel_cumsum_add_f32")]] kernel kernel_cumsum_add_t kernel_cumsum_add<float>;

ggml/src/ggml-metal/kernels/rope.metal

@@ -0,0 +1,333 @@
+#include "common.h"
+
+constant bool FC_rope_is_imrope [[function_constant(FC_ROPE + 0)]];
+constant bool FC_rope_is_back   [[function_constant(FC_ROPE + 1)]];
+
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+    if (FC_rope_is_back) {
+        *sin_theta *= -1.0f;
+    }
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)));
+}
+
+template<typename T>
+kernel void kernel_rope_norm(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[1];
+
+            dst_data[0] = x0*cos_theta - x1*sin_theta;
+            dst_data[1] = x0*sin_theta + x1*cos_theta;
+        } else {
+            if (args.inplace) {
+                continue;
+            }
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_neox(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + (args.n_offs + ic)*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + (args.n_offs + ic)*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            if (args.inplace) {
+                continue;
+            }
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_multi(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;
+
+            // mrope theta calculations
+            // note: the rest is the same as kernel_rope_neox
+            const int sect_dims = args.sect_0 + args.sect_1 + args.sect_2 + args.sect_3;
+            const int sec_w01   = args.sect_0 + args.sect_1;               // end of section 1
+            const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
+            const int sector    = ic % sect_dims;
+
+            float theta_base;
+            if (FC_rope_is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * args.sect_1) { // h
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * args.sect_2) { // w
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * args.sect_0) { // t
+                    theta_base = (float) pos[i2 + args.ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            } else {
+                if (sector < args.sect_0) {
+                    theta_base = (float) pos[i2];
+                } else if (sector < sec_w01) {
+                    theta_base = (float) pos[i2 + args.ne02 * 1];
+                } else if (sector < sec_w012) {
+                    theta_base = (float) pos[i2 + args.ne02 * 2];
+                } else {
+                    theta_base = (float) pos[i2 + args.ne02 * 3];
+                }
+            }
+            // end of mrope
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + (args.n_offs + ic)*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + (args.n_offs + ic)*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            if (args.inplace) {
+                continue;
+            }
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_vision(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < 2*args.n_dims) { // different from kernel_rope_multi
+            const int ic = i0/2;
+
+            // mrope theta calculations (only support 2 dimensions)
+            const int sect_dims = args.sect_0 + args.sect_1;
+            const int sector    = ic % sect_dims;
+
+            float p;
+            float theta_base;
+            if (sector < args.sect_1) {
+                p = (float) sector;
+                theta_base = (float) pos[i2];
+            } else {
+                p = (float) sector - args.sect_0;
+                theta_base = (float) pos[i2 + args.ne02];
+            }
+
+            const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p);
+            // end of mrope
+
+            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims]; // different from kernel_rope_multi
+
+            dst_data[0]           = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims] = x0*sin_theta + x1*cos_theta; // different from kernel_rope_multi
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+typedef decltype(kernel_rope_norm<float>) kernel_rope_norm_t;
+typedef decltype(kernel_rope_neox<float>) kernel_rope_neox_t;
+typedef decltype(kernel_rope_multi<float>) kernel_rope_multi_t;
+typedef decltype(kernel_rope_vision<float>) kernel_rope_vision_t;
+
+template [[host_name("kernel_rope_norm_f32")]] kernel kernel_rope_norm_t kernel_rope_norm<float>;
+template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_rope_norm<half>;
+
+template [[host_name("kernel_rope_neox_f32")]] kernel kernel_rope_neox_t kernel_rope_neox<float>;
+template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_rope_neox<half>;
+
+template [[host_name("kernel_rope_multi_f32")]] kernel kernel_rope_multi_t kernel_rope_multi<float>;
+template [[host_name("kernel_rope_multi_f16")]] kernel kernel_rope_multi_t kernel_rope_multi<half>;
+
+template [[host_name("kernel_rope_vision_f32")]] kernel kernel_rope_vision_t kernel_rope_vision<float>;
+template [[host_name("kernel_rope_vision_f16")]] kernel kernel_rope_vision_t kernel_rope_vision<half>;

ggml/src/ggml-metal/kernels/softmax.metal

@@ -0,0 +1,223 @@
+#include "common.h"
+
+template<typename T>
+kernel void kernel_soft_max(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float * psrc0 =                (device const float *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const     T * pmask = src1 != src0 ? (device const T *    ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float * psrc2 = src2 != src0 ? (device const float *) (src2)                                                 : nullptr;
+    device       float * pdst  =                (device       float *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+
+    // find the max value in the block
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
+        lsum += exp_psrc0;
+        pdst[i00] = exp_psrc0;
+    }
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
+        pdst[i00] *= inv_sum;
+    }
+}
+
+template<typename T>
+kernel void kernel_soft_max_4(
+        constant ggml_metal_kargs_soft_max & args,
+        device const  char * src0,
+        device const  char * src1,
+        device const  char * src2,
+        device        char * dst,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
+
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float4 * psrc4 =                (device const float4 *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const      T * pmask = src1 != src0 ? (device const T *     ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device const float *  psrc2 = src2 != src0 ? (device const float * ) (src2)                                                 : nullptr;
+    device       float4 * pdst4 =                (device       float4 *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
+
+    float slope = 1.0f;
+
+    if (args.max_bias > 0.0f) {
+        const int32_t h = i02;
+
+        const float base = h < args.n_head_log2 ? args.m0 : args.m1;
+        const int   exp  = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
+    }
+
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+
+    float max_val = simd_max(lmax);
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (tptg.x > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max_val);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
+        pdst4[i00] *= inv_sum;
+    }
+}
+
+typedef decltype(kernel_soft_max<float>)    kernel_soft_max_t;
+typedef decltype(kernel_soft_max_4<float4>) kernel_soft_max_4_t;
+
+template [[host_name("kernel_soft_max_f16")]]   kernel kernel_soft_max_t   kernel_soft_max<half>;
+template [[host_name("kernel_soft_max_f32")]]   kernel kernel_soft_max_t   kernel_soft_max<float>;
+template [[host_name("kernel_soft_max_f16_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<half4>;
+template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<float4>;

ggml/src/ggml-metal/kernels/solve_tri.metal

@@ -0,0 +1,75 @@
+#include "common.h"
+
+constant short FC_solve_tri_nsg [[function_constant(FC_SOLVE_TRI + 0)]];
+constant short FC_solve_tri_n   [[function_constant(FC_SOLVE_TRI + 1)]];
+constant short FC_solve_tri_k   [[function_constant(FC_SOLVE_TRI + 2)]];
+
+kernel void kernel_solve_tri_f32(
+        constant ggml_metal_kargs_solve_tri & args,
+        device   const char * src0,
+        device   const char * src1,
+        device         char * dst,
+        threadgroup    char * shmem [[threadgroup(0)]],
+        ushort3 tgpig[[threadgroup_position_in_grid]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    const short NSG = FC_solve_tri_nsg;
+    const short N   = FC_solve_tri_n;
+    const short K   = FC_solve_tri_k;
+    const short NP  = PAD2(N, NW);
+
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x*NSG + sgitg;
+
+    threadgroup float * sh0 = (threadgroup float *) shmem;
+
+    device const float * src0_ptr = (device const float *)(src0 + i02 * args.nb02 + i03 * args.nb03) + sgitg*N;
+    device const float * src1_ptr = (device const float *)(src1 + i02 * args.nb12 + i03 * args.nb13) + i01;
+    device       float * dst_ptr  = (device       float *)(dst  + i02 * args.nb2  + i03 * args.nb3)  + i01;
+
+    for (short rr = 0; rr < N; rr += NSG) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        {
+            threadgroup float * sh0_cur = sh0 + sgitg*NP;
+
+            for (short t = 0; t*NW < N; ++t) {
+                const short idx = t*NW + tiisg;
+                sh0_cur[idx] = src0_ptr[idx];
+            }
+
+            src0_ptr += NSG*N;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (i01 >= args.ne10) {
+            continue;
+        }
+
+        for (short ir = 0; ir < NSG && rr + ir < N; ++ir) {
+            const short r = rr + ir;
+
+            threadgroup float * sh0_cur = sh0 + ir*NP;
+
+            float sum = 0.0f;
+
+            for (short t = 0; t*NW < r; ++t) {
+                const short idx = t*NW + tiisg;
+                sum += sh0_cur[idx] * dst_ptr[idx*K] * (idx < r);
+            }
+
+            sum = simd_sum(sum);
+
+            if (tiisg == 0) {
+                const float diag = sh0_cur[r];
+
+                dst_ptr[r*K] = (src1_ptr[r*K] - sum) / diag;
+            }
+        }
+    }
+}

ggml/src/ggml-metal/kernels/ssm.metal

@@ -0,0 +1,287 @@
+#include "common.h"
+
+// ref: ggml.c:ggml_compute_forward_ssm_conv_f32
+kernel void kernel_ssm_conv_f32_f32(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+    device       float * x = (device       float *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = args.ne10;
+  //const int64_t ncs = args.ne00;
+  //const int64_t nr  = args.ne01;
+  //const int64_t n_t = args.ne1;
+  //const int64_t n_s = args.ne2;
+
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+    device       float  * x = (device       float  *) ((device       char *) dst  + ir*args.nb0  + i2*args.nb1  + i3*args.nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+constant short FC_ssm_conv_bs   [[function_constant(FC_SSM_CONV + 0)]];
+
+// Batched version: each threadgroup processes multiple tokens for better efficiency
+// Thread layout: each thread handles one token, threadgroup covers BATCH_SIZE tokens
+kernel void kernel_ssm_conv_f32_f32_batched(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_batched_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
+// Optimized version: reduces redundant memory loads by having one thread load shared values
+kernel void kernel_ssm_scan_f32(
+        constant ggml_metal_kargs_ssm_scan & args,
+        device const void * src0,
+        device const void * src1,
+        device const void * src2,
+        device const void * src3,
+        device const void * src4,
+        device const void * src5,
+        device const void * src6,
+        device      float * dst,
+        threadgroup float * shared [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgptg[[simdgroups_per_threadgroup]],
+        uint3    tgpg[[threadgroups_per_grid]]) {
+    constexpr short NW = N_SIMDWIDTH;
+
+    // Shared memory layout:
+    // [0..sgptg*NW-1]: partial sums for reduction (existing)
+    // [sgptg*NW..sgptg*NW+sgptg-1]: pre-computed x_dt values for each token in batch
+    // [sgptg*NW+sgptg..sgptg*NW+2*sgptg-1]: pre-computed dA values for each token in batch
+    threadgroup float * shared_sums = shared;
+    threadgroup float * shared_x_dt = shared + sgptg * NW;
+    threadgroup float * shared_dA   = shared + sgptg * NW + sgptg;
+
+    shared_sums[tpitg.x] = 0.0f;
+
+    const int32_t i0 = tpitg.x;
+    const int32_t i1 = tgpig.x;
+    const int32_t ir = tgpig.y; // current head
+    const int32_t i3 = tgpig.z; // current seq
+
+    const int32_t nc  = args.d_state;
+    const int32_t nr  = args.d_inner;
+    const int32_t nh  = args.n_head;
+    const int32_t ng  = args.n_group;
+    const int32_t n_t = args.n_seq_tokens;
+    const int32_t n_s = args.n_seqs;
+    const int32_t K   = args.K;
+
+    const int32_t s_off = args.s_off;
+
+    device const int32_t * ids = (device const int32_t *) src6;
+
+    device const float * s0_buff = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s_buff  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+
+    const int32_t i = i0 + i1*nc;
+    const int32_t g = ir / (nh / ng); // repeat_interleave
+
+    float s0 = s0_buff[i];
+    float s  = 0.0f;
+
+    device const float * A = (device const float *) ((device const char *) src3 + ir*args.nb31); // {ne30, nh}
+
+    const float A0 = A[i0%args.ne30];
+
+    device const float * x  = (device const float *)((device const char *) src1 + i1*args.nb10  + ir*args.nb11 + i3*args.nb13); // {dim, nh, nt, ns}
+    device const float * dt = (device const float *)((device const char *) src2 + ir*args.nb20  + i3*args.nb22);                // {nh, nt, ns}
+    device const float * B  = (device const float *)((device const char *) src4 +  g*args.nb41  + i3*args.nb43);                // {d_state, ng, nt, ns}
+    device const float * C  = (device const float *)((device const char *) src5 +  g*args.nb51  + i3*args.nb53);                // {d_state, ng, nt, ns}
+
+    device float * y = dst + (i1 + ir*(nr) + i3*(n_t*nh*nr)); // {dim, nh, nt, ns}
+
+    for (int i2 = 0; i2 < n_t; i2 += sgptg) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // Pre-compute x_dt and dA for this batch of tokens
+        // Only first sgptg threads do the loads and expensive math
+        if (i0 < sgptg && i2 + i0 < n_t) {
+            // ns12 and ns21 are element strides (nb12/nb10, nb21/nb20)
+            device const float * x_t  = x  + i0 * args.ns12;
+            device const float * dt_t = dt + i0 * args.ns21;
+
+            const float dt0  = dt_t[0];
+            const float dtsp = dt0 <= 20.0f ? log(1.0f + exp(dt0)) : dt0;
+            shared_x_dt[i0] = x_t[0] * dtsp;
+            shared_dA[i0]   = dtsp;  // Store dtsp, compute exp(dtsp * A0) per-thread since A0 varies
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (int t = 0; t < sgptg && i2 + t < n_t; t++) {
+            const float x_dt = shared_x_dt[t];
+            const float dA   = exp(shared_dA[t] * A0);
+
+            s = (s0 * dA) + (B[i0] * x_dt);
+
+            const float sumf = simd_sum(s * C[i0]);
+
+            if (tiisg == 0) {
+                shared_sums[t*NW + sgitg] = sumf;
+            }
+
+            // recurse
+            s0 = s;
+
+            const int32_t slot = n_t - 1 - (i2 + t);
+            if (slot > 0 && slot < K) {
+                device float * s_snapshot = (device float *) ((device char *) s_buff + (int64_t) slot*n_s*args.nb03);
+                s_snapshot[i] = s;
+            }
+
+            B  += args.ns42;
+            C  += args.ns52;
+        }
+
+        // Advance pointers for next batch
+        x  += sgptg * args.ns12;
+        dt += sgptg * args.ns21;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float sumf = simd_sum(shared_sums[sgitg*NW + tiisg]);
+
+        if (tiisg == 0 && i2 + sgitg < n_t) {
+            y[sgitg*nh*nr] = sumf;
+        }
+
+        y += sgptg*nh*nr;
+    }
+
+    s_buff[i] = s;
+}

ggml/src/ggml-metal/kernels/tri.metal

@@ -0,0 +1,69 @@
+#include "common.h"
+
+template<uint32_t ttype>
+bool _ggml_vec_tri_cmp(const int i, const int r);
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER */ 3>(const int i, const int r) {
+    return i < r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER_DIAG */ 2>(const int i, const int r) {
+    return i <= r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER */ 1>(const int i, const int r) {
+    return i > r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER_DIAG */ 0>(const int i, const int r) {
+    return i >= r;
+}
+
+template<typename T, int ttype>
+kernel void kernel_tri(
+        constant ggml_metal_kargs_tri & args,
+        device const char * src0,
+        device const char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    device const T * src_row = (device const T *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T * dst_row = (device       T *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    // Each thread is a single element of the row if ne00 < max threads per
+    // threadgroup, so this will loop once for each index that this thread is
+    // responsible for
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        // Use the comparison as a mask for branchless
+        dst_row[i0] = static_cast<T>(_ggml_vec_tri_cmp<ttype>(i0, i1)) * src_row[i0];
+    }
+}
+
+typedef decltype(kernel_tri<float, 0>) kernel_tri_t;
+
+template [[host_name("kernel_tri_f32_0")]] kernel kernel_tri_t kernel_tri<float, 0>;
+template [[host_name("kernel_tri_f32_1")]] kernel kernel_tri_t kernel_tri<float, 1>;
+template [[host_name("kernel_tri_f32_2")]] kernel kernel_tri_t kernel_tri<float, 2>;
+template [[host_name("kernel_tri_f32_3")]] kernel kernel_tri_t kernel_tri<float, 3>;
+template [[host_name("kernel_tri_f16_0")]] kernel kernel_tri_t kernel_tri<half, 0>;
+template [[host_name("kernel_tri_f16_1")]] kernel kernel_tri_t kernel_tri<half, 1>;
+template [[host_name("kernel_tri_f16_2")]] kernel kernel_tri_t kernel_tri<half, 2>;
+template [[host_name("kernel_tri_f16_3")]] kernel kernel_tri_t kernel_tri<half, 3>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_tri_bf16_0")]] kernel kernel_tri_t kernel_tri<bfloat, 0>;
+template [[host_name("kernel_tri_bf16_1")]] kernel kernel_tri_t kernel_tri<bfloat, 1>;
+template [[host_name("kernel_tri_bf16_2")]] kernel kernel_tri_t kernel_tri<bfloat, 2>;
+template [[host_name("kernel_tri_bf16_3")]] kernel kernel_tri_t kernel_tri<bfloat, 3>;
+#endif

ggml/src/ggml-metal/kernels/unary.metal

@@ -0,0 +1,374 @@
+#include "common.h"
+
+constant short FC_unary_op [[function_constant(FC_UNARY + 0)]];
+constant bool  FC_unary_cnt[[function_constant(FC_UNARY + 1)]];
+
+template <typename T0, typename T, typename TC>
+kernel void kernel_unary_impl(
+        constant ggml_metal_kargs_unary & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP  FC_unary_op
+#define FC_CNT FC_unary_cnt
+
+    device const T0 * src0_ptr;
+    device       T  * dst_ptr;
+
+    int i0;
+
+    if (FC_CNT) {
+        i0 = tgpig.x;
+
+        src0_ptr = (device const T0 *) (src0);
+        dst_ptr  = (device       T  *) (dst);
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int k0  = tgpig.x/args.ne01;
+        const int i01 = tgpig.x - k0*args.ne01;
+
+        i0 = k0*ntg.x + tpitg.x;
+
+        src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01);
+        dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1 );
+    }
+
+    {
+        //threadgroup_barrier(mem_flags::mem_none);
+
+        if (!FC_CNT) {
+            if (i0 >= args.ne0) {
+                return;
+            }
+        }
+
+        const TC x = (TC) src0_ptr[i0];
+
+        if (FC_OP == OP_UNARY_NUM_SCALE) {
+            dst_ptr[i0] = (T) (args.scale * x + args.bias);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FILL) {
+            dst_ptr[i0] = (T) args.val;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CLAMP) {
+            dst_ptr[i0] = (T) clamp(x, args.min, args.max);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQR) {
+            dst_ptr[i0] = (T) (x * x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SQRT) {
+            dst_ptr[i0] = (T) sqrt(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIN) {
+            dst_ptr[i0] = (T) sin(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_COS) {
+            dst_ptr[i0] = (T) cos(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LOG) {
+            dst_ptr[i0] = (T) log(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_LEAKY_RELU) {
+            dst_ptr[i0] = (T) (TC(x > 0)*x + TC(x <= 0)*(x * args.slope));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TANH) {
+            dst_ptr[i0] = (T) precise::tanh(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_RELU) {
+            dst_ptr[i0] = (T) fmax(0, x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SIGMOID) {
+            dst_ptr[i0] = (T) (1 / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + precise::tanh(SQRT_2_OVER_PI*x*(1 + GELU_COEF_A*x*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_ERF) {
+            dst_ptr[i0] = (T) (0.5*x*(1 + erf_approx(SQRT_2_INV*x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_GELU_QUICK) {
+            dst_ptr[i0] = (T) (x * (1/(1 + exp(GELU_QUICK_COEF*x))));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SILU) {
+            dst_ptr[i0] = (T) (x / (1 + exp(-x)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ELU) {
+            dst_ptr[i0] = (T) elu_approx(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_NEG) {
+            dst_ptr[i0] = (T) -x;
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ABS) {
+            dst_ptr[i0] = (T) fabs(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SGN) {
+            dst_ptr[i0] = T(x > 0) - T(x < 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_STEP) {
+            dst_ptr[i0] = T(x > 0);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSWISH) {
+            dst_ptr[i0] = (T) (x * fmax(0, fmin(1, x/6 + 0.5)));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_HARDSIGMOID) {
+            dst_ptr[i0] = (T) fmax(0, fmin(1, x/6 + 0.5));
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXP) {
+            dst_ptr[i0] = (T) exp(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_SOFTPLUS) {
+            dst_ptr[i0] = (T) select(log(1 + exp(x)), x, x > 20);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_EXPM1) {
+            // TODO: precise implementation
+            dst_ptr[i0] = (T) (exp(x) - 1);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_FLOOR) {
+            dst_ptr[i0] = (T) floor(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_CEIL) {
+            dst_ptr[i0] = (T) ceil(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_ROUND) {
+            dst_ptr[i0] = (T) round(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_TRUNC) {
+            dst_ptr[i0] = (T) trunc(x);
+        }
+
+        if (FC_OP == OP_UNARY_NUM_XIELU) {
+            const TC xi      = x;
+            const TC gate    = TC(xi > TC(0.0f));
+            const TC clamped = fmin(xi, TC(args.val));
+            const TC y_pos   = TC(args.scale) * xi * xi + TC(args.bias) * xi;
+            const TC y_neg   = (exp(clamped) - TC(1.0f) - xi) * TC(args.slope) + TC(args.bias) * xi;
+            dst_ptr[i0] = (T) (gate * y_pos + (TC(1.0f) - gate) * y_neg);
+        }
+    }
+
+#undef FC_OP
+#undef FC_CNT
+}
+
+typedef decltype(kernel_unary_impl<float, float, float>) kernel_unary_t;
+
+template [[host_name("kernel_unary_f32_f32")]]   kernel kernel_unary_t kernel_unary_impl<float,  float,  float>;
+template [[host_name("kernel_unary_f32_f32_4")]] kernel kernel_unary_t kernel_unary_impl<float4, float4, float4>;
+template [[host_name("kernel_unary_f16_f16")]]   kernel kernel_unary_t kernel_unary_impl<half,   half,   float>;
+template [[host_name("kernel_unary_f16_f16_4")]] kernel kernel_unary_t kernel_unary_impl<half4,  half4,  float4>;
+
+kernel void kernel_silu_back_f32(
+        constant ggml_metal_kargs_silu_back & args,
+        device const float * dy,
+        device const float * x,
+        device       float * dx,
+        uint gid [[thread_position_in_grid]]) {
+    if (gid >= args.ne) {
+        return;
+    }
+
+    const float s = 1.0f / (1.0f + exp(-x[gid]));
+    dx[gid] = dy[gid] * s * (1.0f + x[gid] * (1.0f - s));
+}
+
+template<typename T>
+kernel void kernel_reglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        dst_row[i0] = (T)(x0*x1*(x0 > 0.0f));
+    }
+}
+
+typedef decltype(kernel_reglu<float>) kernel_reglu_t;
+
+template [[host_name("kernel_reglu_f32")]] kernel kernel_reglu_t kernel_reglu<float>;
+template [[host_name("kernel_reglu_f16")]] kernel kernel_reglu_t kernel_reglu<half>;
+
+template<typename T>
+kernel void kernel_geglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu = 0.5f*x0*(1.0f + precise::tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = (T)(gelu*x1);
+    }
+}
+
+typedef decltype(kernel_geglu<float>) kernel_geglu_t;
+
+template [[host_name("kernel_geglu_f32")]] kernel kernel_geglu_t kernel_geglu<float>;
+template [[host_name("kernel_geglu_f16")]] kernel kernel_geglu_t kernel_geglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = (T)(silu*x1);
+    }
+}
+
+typedef decltype(kernel_swiglu<float>) kernel_swiglu_t;
+
+template [[host_name("kernel_swiglu_f32")]] kernel kernel_swiglu_t kernel_swiglu<float>;
+template [[host_name("kernel_swiglu_f16")]] kernel kernel_swiglu_t kernel_swiglu<half>;
+
+template<typename T>
+kernel void kernel_swiglu_oai(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        float x0 = src0_row[i0];
+        float x1 = src1_row[i0];
+
+        x0 = min(x0, args.limit);
+        x1 = max(min(x1, args.limit), -args.limit);
+
+        float out_glu = x0 / (1.0f + exp(-x0 * args.alpha));
+        out_glu = out_glu * (1.0f + x1);
+
+        dst_row[i0] = (T)out_glu;
+    }
+}
+
+typedef decltype(kernel_swiglu_oai<float>) kernel_swiglu_oai_t;
+
+template [[host_name("kernel_swiglu_oai_f32")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<float>;
+template [[host_name("kernel_swiglu_oai_f16")]] kernel kernel_swiglu_oai_t kernel_swiglu_oai<half>;
+
+template<typename T>
+kernel void kernel_geglu_erf(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f+erf_approx<float>(x0*SQRT_2_INV));
+
+        dst_row[i0] = (T)(gelu_erf*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_erf<float>) kernel_geglu_erf_t;
+
+template [[host_name("kernel_geglu_erf_f32")]] kernel kernel_geglu_erf_t kernel_geglu_erf<float>;
+template [[host_name("kernel_geglu_erf_f16")]] kernel kernel_geglu_erf_t kernel_geglu_erf<half>;
+
+template<typename T>
+kernel void kernel_geglu_quick(
+        constant ggml_metal_kargs_glu & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = (T)(gelu_quick*x1);
+    }
+}
+
+typedef decltype(kernel_geglu_quick<float>) kernel_geglu_quick_t;
+
+template [[host_name("kernel_geglu_quick_f32")]] kernel kernel_geglu_quick_t kernel_geglu_quick<float>;
+template [[host_name("kernel_geglu_quick_f16")]] kernel kernel_geglu_quick_t kernel_geglu_quick<half>;

ggml/src/ggml-metal/kernels/upscale.metal

@@ -0,0 +1,179 @@
+#include "common.h"
+
+constant bool FC_upscale_aa [[function_constant(FC_UPSCALE + 0)]];
+
+kernel void kernel_upscale_nearest_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3/args.sf3;
+    const int64_t i02 = i2/args.sf2;
+    const int64_t i01 = i1/args.sf1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int64_t i00 = i0/args.sf0;
+
+        device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
+        device       float * dst_ptr  = (device       float *) (dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1  +  i0*args.nb0);
+
+        dst_ptr[0] = src0_ptr[0];
+    }
+}
+
+static inline float bilinear_tri(float x) {
+    return MAX(0.0f, 1.0f - fabs(x));
+}
+
+kernel void kernel_upscale_bilinear_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01  = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01  = MAX(0, MIN(args.ne01 - 1, (int64_t)floor(f01)));
+    const int64_t i01p = MAX(0, MIN(args.ne01 - 1, i01 + 1));
+    const float   fd1  = MAX(0.0f, MIN(1.0f, f01 - (float)i01));
+
+    src0 += i03*args.nb03 + i02*args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1);
+
+    if (FC_upscale_aa) {
+        const float support0  = MAX(1.0f, 1.0f / args.sf0);
+        const float invscale0 = 1.0f / support0;
+        const float support1  = MAX(1.0f, 1.0f / args.sf1);
+        const float invscale1 = 1.0f / support1;
+
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+
+            int64_t x_min = MAX((int64_t)0, (int64_t)floor(f00 - support0 + args.poffs));
+            int64_t x_max = MIN(args.ne00,  (int64_t)ceil (f00 + support0 + args.poffs));
+
+            int64_t y_min = MAX((int64_t)0, (int64_t)floor(f01 - support1 + args.poffs));
+            int64_t y_max = MIN(args.ne01,  (int64_t)ceil (f01 + support1 + args.poffs));
+
+            float sum = 0.0f;
+            float wsum = 0.0f;
+
+            for (int64_t sy = y_min; sy < y_max; ++sy) {
+                const float wy = MAX(0.0f, 1.0f - fabs((float)sy - f01) * invscale1);
+                for (int64_t sx = x_min; sx < x_max; ++sx) {
+                    const float wx = MAX(0.0f, 1.0f - fabs((float)sx - f00) * invscale0);
+                    const float w  = wx * wy;
+                    device const float * src_ptr = (device const float *)(src0 + sy*args.nb01 + sx*args.nb00);
+                    sum  += (*src_ptr) * w;
+                    wsum += w;
+                }
+            }
+
+            const float v = (wsum > 0.0f) ? (sum / wsum) : 0.0f;
+            dst_ptr[i0] = v;
+        }
+    } else {
+        for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+            const float   f00  = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+            const int64_t i00  = MAX(0, MIN(args.ne00 - 1, (int64_t)floor(f00)));
+            const int64_t i00p = MAX(0, MIN(args.ne00 - 1, i00 + 1));
+            const float   fd0  = MAX(0.0f, MIN(1.0f, f00 - (float)i00));
+
+            device const float * src00 = (device const float *)(src0 + i01*args.nb01  + i00*args.nb00);
+            device const float * src10 = (device const float *)(src0 + i01*args.nb01  + i00p*args.nb00);
+            device const float * src01 = (device const float *)(src0 + i01p*args.nb01 + i00*args.nb00);
+            device const float * src11 = (device const float *)(src0 + i01p*args.nb01 + i00p*args.nb00);
+
+            const float v =
+                (*src00) * (1.0f - fd0) * (1.0f - fd1) +
+                (*src10) * fd0          * (1.0f - fd1) +
+                (*src01) * (1.0f - fd0) * fd1 +
+                (*src11) * fd0          * fd1;
+
+            dst_ptr[i0] = v;
+        }
+    }
+}
+
+static inline float bicubic_weight1(float x) {
+    const float a = -0.75f;
+    return ((a + 2) * x - (a + 3)) * x * x + 1;
+}
+
+static inline float bicubic_weight2(float x) {
+    const float a = -0.75f;
+    return ((a * x - 5 * a) * x + 8 * a) * x - 4 * a;
+}
+
+kernel void kernel_upscale_bicubic_f32(
+    constant ggml_metal_kargs_upscale & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 / args.sf3;
+    const int64_t i02 = i2 / args.sf2;
+
+    const float   f01 = ((float)i1 + args.poffs) / args.sf1 - args.poffs;
+    const int64_t i01 = (int64_t)floor(f01);
+    const float   fd1 = f01 - (float)i01;
+
+    const float w_y0 = bicubic_weight2(fd1 + 1.0f);
+    const float w_y1 = bicubic_weight1(fd1);
+    const float w_y2 = bicubic_weight1(1.0f - fd1);
+    const float w_y3 = bicubic_weight2(2.0f - fd1);
+
+    const device char * src_slice = src0 + i03 * args.nb03 + i02 * args.nb02;
+
+    device float * dst_ptr = (device float *)(dst + i3 * args.nb3 + i2 * args.nb2 + i1 * args.nb1);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const float   f00 = ((float)i0 + args.poffs) / args.sf0 - args.poffs;
+        const int64_t i00 = (int64_t)floor(f00);
+        const float   fd0 = f00 - (float)i00;
+
+        const float w_x0 = bicubic_weight2(fd0 + 1.0f);
+        const float w_x1 = bicubic_weight1(fd0);
+        const float w_x2 = bicubic_weight1(1.0f - fd0);
+        const float w_x3 = bicubic_weight2(2.0f - fd0);
+
+        float sum = 0.0f;
+
+        for (int dy = -1; dy <= 2; ++dy) {
+            const int64_t iy = MAX(0, MIN(args.ne01 - 1, i01 + dy));
+            const float wy = (dy == -1) ? w_y0 : (dy == 0) ? w_y1 : (dy == 1) ? w_y2 : w_y3;
+
+            for (int dx = -1; dx <= 2; ++dx) {
+                const int64_t ix = MAX(0, MIN(args.ne00 - 1, i00 + dx));
+                const float wx = (dx == -1) ? w_x0 : (dx == 0) ? w_x1 : (dx == 1) ? w_x2 : w_x3;
+
+                device const float * src_ptr = (device const float *)(src_slice + iy * args.nb01 + ix * args.nb00);
+                sum += (*src_ptr) * wx * wy;
+            }
+        }
+
+        dst_ptr[i0] = sum;
+    }
+}

ggml/src/ggml-metal/kernels/wkv.metal

@@ -0,0 +1,179 @@
+#include "common.h"
+
+kernel void kernel_rwkv_wkv6_f32(
+    device const float * k,
+    device const float * v,
+    device const float * r,
+    device const float * tf,
+    device const float * td,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _k[head_size];
+    threadgroup float _r[head_size];
+    threadgroup float _tf[head_size];
+    threadgroup float _td[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + i * head_size + tid];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    _tf[tid] = tf[head_id * head_size + tid];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0;
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 tf_vec = float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
+            float4 td_vec = float4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            float4 temp = tf_vec * kv + s_vec;
+            y += dot(r_vec, temp);
+
+            s_vec = s_vec * td_vec + kv;
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + i * head_size + tid] = state[i];
+    }
+}
+
+kernel void kernel_rwkv_wkv7_f32(
+    device const float * r,
+    device const float * w,
+    device const float * k,
+    device const float * v,
+    device const float * a,
+    device const float * b,
+    device const float * state_in,
+    device       float * dst,
+    constant    uint & B,
+    constant    uint & T,
+    constant    uint & C,
+    constant    uint & H,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]])  {
+
+    const uint head_size = 64; // TODO: support head_size = 128
+    const uint batch_id = tgpig.x / H;
+    const uint head_id = tgpig.x % H;
+    const uint tid = tpitg.x;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    threadgroup float _r[head_size];
+    threadgroup float _w[head_size];
+    threadgroup float _k[head_size];
+    threadgroup float _a[head_size];
+    threadgroup float _b[head_size];
+
+    float state[head_size];
+
+    for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + tid * head_size + i];
+    }
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        _r[tid] = r[t];
+        _w[tid] = w[t];
+        _k[tid] = k[t];
+        _a[tid] = a[t];
+        _b[tid] = b[t];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        const float v_val = v[t];
+        float y = 0.0, sa = 0.0;
+
+        float4 sa_vec(0.0);
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 a_vec = float4(_a[j], _a[j+1], _a[j+2], _a[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+            sa_vec += a_vec * s_vec;
+        }
+        sa = sa_vec[0] + sa_vec[1] + sa_vec[2] + sa_vec[3];
+
+        for (uint j = 0; j < head_size; j += 4) {
+            float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            float4 w_vec = float4(_w[j], _w[j+1], _w[j+2], _w[j+3]);
+            float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            float4 b_vec = float4(_b[j], _b[j+1], _b[j+2], _b[j+3]);
+            float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            float4 kv = k_vec * v_val;
+
+            s_vec = s_vec * w_vec + kv + sa * b_vec;
+            y += dot(s_vec, r_vec);
+
+            state[j]   = s_vec[0];
+            state[j+1] = s_vec[1];
+            state[j+2] = s_vec[2];
+            state[j+3] = s_vec[3];
+        }
+
+        dst[t] = y;
+    }
+
+    for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + tid * head_size + i] = state[i];
+    }
+}

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

@@ -6018,6 +6018,11 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
                     a->ne[0] > 128 && a->ne[2] == 1 && src0_type == GGML_TYPE_F16) {
                     return false;
                 }
+
+                if (src0_type == GGML_TYPE_TQ2_0) {
+                    return false;
+                }
+
                 return true;
             }
         case GGML_OP_OUT_PROD:
@@ -6068,6 +6073,9 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
 
         case GGML_OP_SET_ROWS:
             {
+                if (op->type == GGML_TYPE_TQ2_0) {
+                    return false;
+                }
                 auto res = (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16 ||
                             op->src[0]->type == GGML_TYPE_BF16) &&
                            (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
@@ -6186,11 +6194,16 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
                         src1_type == GGML_TYPE_IQ3_XXS ||
                         src1_type == GGML_TYPE_IQ3_S ||
                         src1_type == GGML_TYPE_IQ1_S ||
-                        src1_type == GGML_TYPE_IQ1_M) {
+                        src1_type == GGML_TYPE_IQ1_M ||
+                        src1_type == GGML_TYPE_TQ2_0) {
                         return false;
                     }
                 }
 
+                if (src0_type == GGML_TYPE_TQ2_0 || src1_type == GGML_TYPE_TQ2_0) {
+                    return false;
+                }
+
                 return true;
             }
         case GGML_OP_REPEAT_BACK:

ggml/src/ggml-sycl/mmvq.cpp

@@ -2559,7 +2559,7 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
                 }
                 break;
             default:
-                GGML_ABORT("fatal error: unsupport data type=%s\n", ggml_type_name(src0->type));
+                GGML_ABORT("fatal error: unsupport src0 data type %s\n", ggml_type_name(src0->type));
         }
     }
     GGML_UNUSED(src1);

ggml/src/ggml-sycl/set_rows.cpp

@@ -546,7 +546,8 @@ static void set_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor * s
                 stream);
             break;
         default:
-            GGML_ABORT("Unsupported tensor type!");
+            GGML_ABORT("Unsupported tensor type: src0 %s src1 %s dst %s", ggml_type_name(dst->src[0]->type),
+                ggml_type_name(dst->src[1]->type), ggml_type_name(dst->type));
             break;
     }
 }

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -955,6 +955,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_diag[2];
     vk_pipeline pipeline_clamp[2];
     vk_pipeline pipeline_pad_f32;
+    vk_pipeline pipeline_pad_reflect_1d_f32;
     vk_pipeline pipeline_roll_f32;
     vk_pipeline pipeline_repeat_i32, pipeline_repeat_back_f32;
     vk_pipeline pipeline_repeat_i16;
@@ -5630,6 +5631,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_diag[1], "diag_f16", diag_f16_len, diag_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_pad_reflect_1d_f32, "pad_reflect_1d_f32", pad_reflect_1d_f32_len, pad_reflect_1d_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_roll_f32, "roll_f32", roll_f32_len, roll_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -11336,6 +11338,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_pad_f32;
         }
         return nullptr;
+    case GGML_OP_PAD_REFLECT_1D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_pad_reflect_1d_f32;
+        }
+        return nullptr;
     case GGML_OP_ROLL:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_roll_f32;
@@ -12239,6 +12246,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_CLAMP:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_PAD:
+    case GGML_OP_PAD_REFLECT_1D:
     case GGML_OP_ROLL:
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
@@ -13111,6 +13119,17 @@ static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p));
 }
 
+static void ggml_vk_pad_reflect_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t p0 = (uint32_t)dst->op_params[0];
+    const uint32_t p1 = (uint32_t)dst->op_params[1];
+
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
+    memcpy(&p.param1, &p0, sizeof(float));
+    memcpy(&p.param2, &p1, sizeof(float));
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD_REFLECT_1D, std::move(p));
+}
+
 static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const int32_t s0 = ggml_get_op_params_i32(dst, 0);
     const int32_t s1 = ggml_get_op_params_i32(dst, 1);
@@ -15520,6 +15539,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_PAD:
         ggml_vk_pad(ctx, compute_ctx, src0, node);
 
+        break;
+    case GGML_OP_PAD_REFLECT_1D:
+        ggml_vk_pad_reflect_1d(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_ROLL:
         ggml_vk_roll(ctx, compute_ctx, src0, node);
@@ -18446,6 +18469,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SCALE:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ROLL:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_DIAG_MASK_INF:
@@ -19228,6 +19252,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else if (tensor->op == GGML_OP_PAD) {
             tensor_clone = ggml_pad_ext(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3],
                                                                 tensor->op_params[4], tensor->op_params[5], tensor->op_params[6], tensor->op_params[7]);
+        } else if (tensor->op == GGML_OP_PAD_REFLECT_1D) {
+            tensor_clone = ggml_pad_reflect_1d(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1]);
         } else if (tensor->op == GGML_OP_REPEAT) {
             tensor_clone = ggml_repeat(ggml_ctx, src_clone[0], tensor);
         } else if (tensor->op == GGML_OP_REPEAT_BACK) {

ggml/src/ggml-vulkan/vulkan-shaders/pad_reflect_1d.comp

@@ -0,0 +1,43 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_unary_head.glsl" // included to use functions like fastdiv etc.
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint p0 = floatBitsToUint(p.param1);
+    const uint p1 = floatBitsToUint(p.param2);
+
+    const uint i3 = fastdiv(idx, p.ne1_012mp, fastdiv_L(p.ne1_Ls, 0));
+    const uint i3_offset = i3 * p.ne12 * p.ne11 * p.ne10;
+
+    const uint i2 = fastdiv(idx - i3_offset, p.ne1_01mp, fastdiv_L(p.ne1_Ls, 1));
+    const uint i2_offset = i2 * p.ne11 * p.ne10;
+
+    const uint i1 = fastdiv(idx - i3_offset - i2_offset, p.ne1_0mp, fastdiv_L(p.ne1_Ls, 2));
+    const uint i0 = idx - i3_offset - i2_offset - i1 * p.ne10;
+
+    uint src_col;
+
+    if (i0 < p0) {
+        src_col = p0 - i0;    // left pad area
+    } else if (i0 < p0 + p.ne00) {
+        src_col = i0 - p0;    // center area
+    } else {
+        src_col = 2u * p.ne00 - 2u - (i0 - p0);  // right pad area
+    }
+
+    const uint src_idx = i3 * p.nb03 + i2 * p.nb02 + i1 * p.nb01 + src_col * p.nb00;
+    const uint d_idx = i3 * p.nb13 + i2 * p.nb12 + i1 * p.nb11 + i0 * p.nb10;
+
+    // copy the computed value to the destination tensor
+    data_d[get_doffset() + d_idx] = D_TYPE(data_a[get_aoffset() + src_idx]);
+}

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

@@ -896,6 +896,7 @@ void process_shaders() {
     string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
 
     string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    string_to_spv("pad_reflect_1d_f32", "pad_reflect_1d.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
 
     string_to_spv("concat_i8", "concat.comp", {{"A_TYPE", "uint8_t"}, {"B_TYPE", "uint8_t"}, {"D_TYPE", "uint8_t"}});
     string_to_spv("concat_i16", "concat.comp", {{"A_TYPE", "uint16_t"}, {"B_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});

ggml/src/ggml-webgpu/wgsl-shaders/argsort.wgsl

@@ -34,11 +34,9 @@ var<uniform> params: Params;
 var<workgroup> shmem_idx: array<u32, WG_SIZE>;
 
 #if ORDER == 0
-#define EXTREME_VALUE 1e30
 #define SWAP_COMPARE_UP >
 #define SWAP_COMPARE_DOWN <
 #else
-#define EXTREME_VALUE -1e30
 #define SWAP_COMPARE_UP <
 #define SWAP_COMPARE_DOWN >
 #endif
@@ -78,11 +76,9 @@ fn main(@builtin(workgroup_id) wid: vec3<u32>,
                 let dir_up = (lid.x & k) == 0;
                 let a_idx = shmem_idx[lid.x];
                 let b_idx = shmem_idx[ixj];
-                let a_val = select(EXTREME_VALUE, src[row_base + a_idx], a_idx < params.src_ne0);
-                let b_val = select(EXTREME_VALUE, src[row_base + b_idx], b_idx < params.src_ne0);
                 let should_swap = select(
-                    (a_val SWAP_COMPARE_DOWN b_val),
-                    (a_val SWAP_COMPARE_UP b_val),
+                    b_idx >= params.src_ne0 || (a_idx < params.src_ne0 && src[row_base + a_idx] SWAP_COMPARE_DOWN src[row_base + b_idx]),
+                    a_idx >= params.src_ne0 || (b_idx < params.src_ne0 && src[row_base + a_idx] SWAP_COMPARE_UP src[row_base + b_idx]),
                     dir_up);
                 if (should_swap) {
                     shmem_idx[lid.x] = b_idx;

ggml/src/ggml-webgpu/wgsl-shaders/flash_attn.wgsl

@@ -5,10 +5,9 @@ enable subgroups;
 enable chromium_experimental_subgroup_matrix;
 
 #define BYTE_HELPERS
-#include "common_decls.tmpl"
-
 #define FLASH_ATTN_SCALAR_KV
 #include "flash_attn_decls.tmpl"
+#include "common_decls.tmpl"
 
 // Default values
 // The actual values are defined in shader-lib.

ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_tile.wgsl

@@ -2,8 +2,8 @@ enable f16;
 enable subgroups;
 
 #define BYTE_HELPERS
-#include "common_decls.tmpl"
 #include "flash_attn_decls.tmpl"
+#include "common_decls.tmpl"
 
 // Default values
 // The actual values are defined in shader-lib.

ggml/src/ggml-webgpu/wgsl-shaders/flash_attn_vec_split.wgsl

@@ -3,9 +3,9 @@ enable f16;
 enable subgroups;
 
 #define BYTE_HELPERS
-#include "common_decls.tmpl"
 #define FLASH_ATTN_VEC_SPLIT
 #include "flash_attn_decls.tmpl"
+#include "common_decls.tmpl"
 
 // Default values
 // The actual values are defined in shader-lib.

ggml/src/ggml.c

@@ -4042,6 +4042,41 @@ struct ggml_tensor * ggml_diag_mask_zero_inplace(
     return ggml_diag_mask_zero_impl(ctx, a, n_past, true);
 }
 
+// ggml_clamp
+
+static struct ggml_tensor * ggml_clamp_impl(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        float                 min,
+        float                 max,
+        bool                  inplace) {
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    float params[] = { min, max };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_CLAMP;
+    result->src[0] = a;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_clamp(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    float                 min,
+    float                 max) {
+    return ggml_clamp_impl(ctx, a, min, max, false);
+}
+
+struct ggml_tensor * ggml_clamp_inplace(
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    float                 min,
+    float                 max) {
+    return ggml_clamp_impl(ctx, a, min, max, true);
+}
+
 // ggml_soft_max
 
 static struct ggml_tensor * ggml_soft_max_impl(
@@ -4438,25 +4473,6 @@ struct ggml_tensor * ggml_rope_set_offset(
     return a;
 }
 
-// ggml_clamp
-
-struct ggml_tensor * ggml_clamp(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        float                 min,
-        float                 max) {
-    // TODO: when implement backward, fix this:
-    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
-
-    float params[] = { min, max };
-    ggml_set_op_params(result, params, sizeof(params));
-
-    result->op     = GGML_OP_CLAMP;
-    result->src[0] = a;
-
-    return result;
-}
-
 static int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, int d) {
     return (ins + 2 * p - d * (ks - 1) - 1) / s + 1;
 }

gguf-py/gguf/constants.py

@@ -3822,7 +3822,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
         MODEL_TENSOR.FFN_EXP_PROBS_B,
-        # NextN/MTP tensors - preserved but unused
+        # NextN/MTP tensors
         MODEL_TENSOR.NEXTN_EH_PROJ,
         MODEL_TENSOR.NEXTN_EMBED_TOKENS,
         MODEL_TENSOR.NEXTN_ENORM,

gguf-py/gguf/metadata.py

@@ -83,7 +83,7 @@ class Metadata:
             metadata.sampling_xtc_threshold   = gen_config.get("xtc_threshold",   metadata.sampling_xtc_threshold)
             metadata.sampling_temp            = gen_config.get("temperature",     metadata.sampling_temp)
             metadata.sampling_penalty_last_n  = gen_config.get("penalty_last_n",  metadata.sampling_penalty_last_n)
-            metadata.sampling_penalty_repeat  = gen_config.get("penalty_repeat",  metadata.sampling_penalty_repeat)
+            metadata.sampling_penalty_repeat  = gen_config.get("penalty_repeat",  gen_config.get("repetition_penalty", metadata.sampling_penalty_repeat))
             metadata.sampling_mirostat        = gen_config.get("mirostat",        metadata.sampling_mirostat)
             metadata.sampling_mirostat_tau    = gen_config.get("mirostat_tau",    metadata.sampling_mirostat_tau)
             metadata.sampling_mirostat_eta    = gen_config.get("mirostat_eta",    metadata.sampling_mirostat_eta)

include/llama.h

@@ -733,7 +733,7 @@ extern "C" {
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
     // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
-    // seq_id < 0 : match any sequence
+    // seq_id < 0 : match any sequence [TAG_LLAMA_SEQ_ID_NEG]
     // p0 < 0     : [0,  p1]
     // p1 < 0     : [p0, inf)
     LLAMA_API bool llama_memory_seq_rm(

scripts/ccache-clear.sh

@@ -0,0 +1,105 @@
+#!/bin/bash
+# Delete GitHub Actions caches matching a key prefix, oldest first.
+#
+# Usage: ccache-clear.sh --key KEY [--older DURATION] [--min N] [--dry-run]
+#   --key:       cache key prefix to match and delete (without the ccache- prefix)
+#   --older:     only delete caches created more than DURATION ago (e.g. 5m, 1h, 1d);
+#                by default all matching caches are deleted
+#   --min:       stop deleting if fewer than N caches would remain (default: 0)
+#   --dry-run:   only print the caches that would be deleted, without deleting them
+#
+# Env (when running in GitHub Actions):
+#   GH_TOKEN:          token for the gh CLI
+#   GITHUB_REPOSITORY: owner/repo of the caches to manage
+set -euo pipefail
+
+KEY=""
+OLDER=""
+MIN=0
+DRY_RUN=false
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        --key)     [[ $# -ge 2 ]] || { echo "Missing value for $1" >&2; exit 1; }; KEY="$2"; shift 2 ;;
+        --older)   [[ $# -ge 2 ]] || { echo "Missing value for $1" >&2; exit 1; }; OLDER="$2"; shift 2 ;;
+        --min)     [[ $# -ge 2 ]] || { echo "Missing value for $1" >&2; exit 1; }; MIN="$2"; shift 2 ;;
+        --dry-run) DRY_RUN=true; shift ;;
+        *) echo "Unknown argument: $1"; exit 1 ;;
+    esac
+done
+
+command -v gh >/dev/null 2>&1 || { echo "Error: GitHub CLI (gh) is required" >&2; exit 1; }
+[[ -n "${GITHUB_REPOSITORY:-}" ]] || { echo "Error: GITHUB_REPOSITORY not set" >&2; exit 1; }
+[[ -n "$KEY" ]] || { echo "Error: --key is required" >&2; exit 1; }
+[[ "$MIN" =~ ^[0-9]+$ ]] || { echo "Invalid min value: $MIN" >&2; exit 1; }
+
+# Convert a duration (e.g. 90m, 1h, 1d, plain seconds) to seconds
+to_seconds() {
+  local val="$1"
+  [[ "$val" =~ ^[0-9]+$ ]] && { echo "$val"; return 0; }
+  local num="${val%?}" unit="${val: -1}" mult
+  [[ "$num" =~ ^[0-9]+$ ]] || return 1
+  case "$unit" in
+    s) mult=1 ;;
+    m) mult=60 ;;
+    h) mult=3600 ;;
+    d) mult=86400 ;;
+    *) return 1 ;;
+  esac
+  echo $((num * mult))
+}
+
+# Convert an ISO-8601 UTC timestamp (e.g. 2026-08-23T16:51:23.313693Z) to epoch seconds
+to_epoch() {
+  local val="$1" out
+  # GNU date (e.g. Linux)
+  if out=$(date -d "$val" +%s 2>/dev/null) && [[ "$out" =~ ^[0-9]+$ ]]; then
+    echo "$out"
+    return 0
+  fi
+  # BSD date (e.g. macOS); fractional seconds are not needed, TZ forces UTC
+  out=$(TZ=UTC date -j -f "%Y-%m-%dT%H:%M:%S" "${val:0:19}" +%s 2>/dev/null) || return 1
+  [[ "$out" =~ ^[0-9]+$ ]] || return 1
+  echo "$out"
+}
+
+CACHES=$(gh cache list --repo "$GITHUB_REPOSITORY" --key "ccache-$KEY" --json id,key,createdAt --jq '.[] | [.createdAt, .id, .key] | @tsv' | LC_ALL=C sort)
+if [[ -z "$CACHES" ]]; then
+    echo "No caches found with key prefix: $KEY"
+    exit 0
+fi
+
+TOTAL=$(( $(wc -l <<< "$CACHES") ))
+
+echo "Found $TOTAL cache(s) with key prefix: $KEY (oldest first):"
+while IFS=$'\t' read -r CREATED ID CACHE_KEY; do
+    printf '  %s  %s  %s\n' "$CREATED" "$ID" "$CACHE_KEY"
+done <<< "$CACHES"
+
+CUTOFF=""
+if [[ -n "$OLDER" ]]; then
+    OLDER_SECONDS=$(to_seconds "$OLDER") || { echo "Invalid older value: $OLDER (expected e.g. 90m, 1h, 1d)" >&2; exit 1; }
+    CUTOFF=$(( $(date +%s) - OLDER_SECONDS ))
+fi
+
+# Caches are sorted oldest first
+DELETED=0
+while IFS=$'\t' read -r CREATED ID CACHE_KEY; do
+    if [[ -n "$CUTOFF" ]]; then
+        CREATED_SECONDS=$(to_epoch "$CREATED") || { echo "Failed to parse date: $CREATED" >&2; exit 1; }
+        if [[ "$CREATED_SECONDS" -ge "$CUTOFF" ]]; then
+            echo "Rest are not older than $OLDER, stopping"
+            break
+        fi
+    fi
+    if (( TOTAL - DELETED - 1 < MIN )); then
+        echo "Keeping at least $MIN cache(s), stopping"
+        break
+    fi
+    if [[ "$DRY_RUN" == "true" ]]; then
+        echo "Would delete cache: $ID ($CACHE_KEY)"
+    else
+        echo "Deleting cache: $ID ($CACHE_KEY)"
+        gh cache delete --repo "$GITHUB_REPOSITORY" "$ID"
+    fi
+    DELETED=$((DELETED + 1))
+done <<< "$CACHES"

scripts/make-release-summary.txt

@@ -0,0 +1,47 @@
+Take a look at the changelog between the current version and the previous version - use the `./scripts/make-release-desc.sh [current-version]` to obtain it.
+
+Write a summary of the change log in a few sections:
+
+```
+## Overview
+
+[an overview using 1 to 3 sentences (no line breaks)]
+
+### API changes (if applicable)
+
+[summarize any API changes to `/include/*`, `/tools/mtmd/mtmd.h` and `/tools/server`]
+
+### New models (if applicable)
+
+[summarize new models added to the `src/models/` directory]
+
+### Core changes (if applicable)
+
+[summarize the changes to `/src/*`
+
+### Multi-modality changes (if applicable)
+
+[summarize the changes to `/tools/mtmd/`]
+
+### Server changes (if applicable)
+
+[summarize the changes to `/tools/server/`]
+
+### UI changes (if applicable)
+
+[summarize the changes to `/tools/ui/`]
+
+### ggml changes (if applicable)
+
+[if the version has been updated, link to the respective `ggml` releases on Github, f.ex `https://github.com/ggml-org/ggml/releases/tag/v0.22.0`. for each version bump, lookup the release description and copy the summary here]
+
+```
+
+Guidelines:
+
+- All bullet point in the summary should be concise and rarely exceed a single line of 120 characters
+- Avoid repeating `ggml`-specific changes - these should already be covered by the `ggml` release links
+- Provide PR link for each bullet point where possible
+- Don't add bullet point to state that there are no API changes in some module
+
+Output just the summary in a markdown block, without any extra text.

scripts/sync-ggml.last

@@ -1 +1 @@
-8599e0ea3756c4bac4ef813af2241cb1a8bbfb0b
+36da57138425487184aa1da2eee2cde155909c6f

scripts/sync_vendor.py

@@ -27,7 +27,7 @@ vendor = {
     f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/split.py":  "split.py",
     f"https://raw.githubusercontent.com/yhirose/cpp-httplib/{HTTPLIB_VERSION}/LICENSE":   "vendor/cpp-httplib/LICENSE",
 
-    "https://raw.githubusercontent.com/sheredom/subprocess.h/9ce0d701b6fb10f8f8c4445edd31e7c60a1237e3/subprocess.h": "vendor/sheredom/subprocess.h",
+    "https://raw.githubusercontent.com/sheredom/subprocess.h/0dccaa9aa176dd6d7ef8afeca3c18d6e80a32795/subprocess.h": "vendor/sheredom/subprocess.h",
 
     f"https://raw.githubusercontent.com/Cyan4973/xxHash/{XXHASH_COMMIT}/xxhash.c":      "vendor/hash/xxhash/xxhash.c",
     f"https://raw.githubusercontent.com/Cyan4973/xxHash/{XXHASH_COMMIT}/xxhash.h":      "vendor/hash/xxhash/xxhash.h",

skills/add-new-model/SKILL.md

@@ -66,7 +66,7 @@ These recur often enough in review comments on past add-model PRs that they're w
 - Optional hparams that are genuinely absent from some configs (e.g. a shared-expert count) should be read with an explicit optional/fallback accessor, not assumed present.
 - Hparams that are actually load-bearing (the model produces wrong output or crashes without them, e.g. `sliding_window_pattern`, norm-eps) must hard-error if missing, not silently fall back to a default.
 - Don't bake a default chat template into the C++ binary - inject it into the GGUF at conversion time instead, since one `llm_arch` can be reused by multiple fine-tunes with different templates, and a baked-in C++ default fails silently for those.
-- Before writing a dedicated tool-call/output parser, check whether the existing autoparser already handles the template (`llama-debug-template-parser <jinja>` shows what it detects).
+- Before writing a dedicated tool-call/output parser, check whether the existing autoparser already handles the template (`test-chat-auto-parser <jinja>` shows what it detects).
 - Marking a custom EOS/closing-tag token as `eot` at conversion time isn't always sufficient - in long/agentic generations a model can emit the closing sequence as literal text instead of the token, so generation never stops on EOG and raw text leaks past the parser. Verify this case, not just the token path.
 - If reusing or aliasing an existing pre-tokenizer for convenience, justify and test that choice explicitly - silent reuse is an easy source of subtle tokenizer bugs.
 - Watch for excessive graph splits caused by building per-layer view/index tensors inside the layer loop - hoist tensors that don't vary per layer out of the loop (relevant if you hit `GGML_SCHED_MAX_SPLIT_INPUTS`).

src/llama-arch.cpp

@@ -1060,7 +1060,6 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_DEEPSEEK32:
-        case LLM_ARCH_DEEPSEEK4:
         case LLM_ARCH_DOTS3NOTE:
         case LLM_ARCH_GLM_DSA:
         case LLM_ARCH_BITNET:

src/llama-context.cpp

@@ -3218,8 +3218,6 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
 }
 
 size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
-    GGML_UNUSED(seq_id);
-
     if (memory) {
         memory->state_write(io, seq_id, flags);
     }
@@ -3228,8 +3226,6 @@ size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id s
 }
 
 size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
-    GGML_UNUSED(seq_id);
-
     if (memory) {
         memory->state_read(io, seq_id, flags);
     }

src/llama-grammar.cpp

@@ -172,6 +172,7 @@ static std::pair<uint32_t, const char *> parse_char(const char * src) {
             case '"':
             case '[':
             case ']':
+            case '-':
                       return std::make_pair(src[1], src + 2);
             default:
                       throw std::runtime_error(std::string("unknown escape at ") + src);

src/llama-kv-cache-dsv4.cpp

@@ -599,6 +599,33 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_comp_plan(
         }
     }
 
+    if (ratio == DSV4_HCA_RATIO && !plan.state_pos.empty() && plan.state_write_idxs.empty()) {
+        assert(kv_size > 0);
+        // the last slot must not be live, or the dummy write would corrupt it;
+        // a full stream implies a completed block, which implies real writes
+        assert(plan.n_kv < (int64_t) kv_size);
+
+        // Keep the compress/write ops in the graph when no HCA block completes
+        // in this ubatch. The dummy block writes to the last cache slot and is
+        // masked out.
+        uint32_t i = 0;
+        while (i < ubatch.n_tokens && ubatch.pos[i] < 0) {
+            ++i;
+        }
+        assert(i < ubatch.n_tokens);
+
+        const llama_seq_id seq_id = ubatch.seq_id[i][0];
+        const int64_t cache_off = dsv4_stream_offset(n_stream, seq_id, kv_size);
+        const int32_t source_idx = state_source_idx(seq_id, ubatch.pos[i]);
+
+        plan.state_write_idxs.push_back(cache_off + kv_size - 1);
+        plan.state_write_pos .push_back(0);
+
+        for (uint32_t j = 0; j < ratio; ++j) {
+            plan.state_read_idxs.push_back(source_idx);
+        }
+    }
+
     if (overlap) {
         // [ all blocks' prev-window indices | all blocks' cur-window indices ]
         plan.state_read_idxs.reserve(overlap_prev_reads.size() + overlap_cur_reads.size());
@@ -608,7 +635,10 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_comp_plan(
                 overlap_cur_reads.begin(), overlap_cur_reads.end());
     }
 
-    plan.n_kv = GGML_PAD(plan.n_kv, 256u);
+    // Keep the mask (and with it the compressed-attention branch) present even
+    // before the first block is visible, so the graph topology never changes.
+    // Padded slots are masked out; comp cache buffers are zero-initialized.
+    plan.n_kv = std::max<int64_t>(GGML_PAD(plan.n_kv, 256u), 256);
 
     std::sort(persist_rows.begin(), persist_rows.end(),
             [](const persist_row & a, const persist_row & b) {
@@ -620,16 +650,26 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_comp_plan(
         plan.state_persist_dst_idxs.push_back(row.dst);
     }
 
-
     if (n_rs_seq > 0) {
-        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
-            const llama_seq_id seq_id = ubatch.seq_id_unq[s];
-            if (seq_id < 0 || (uint32_t) seq_id >= n_stream) {
-                continue;
+        // Emit restore/snapshot entries for all layout streams so that the
+        // graph tensor sizes do not depend on the ubatch's sequence count.
+        // Streams not present in the ubatch get no-op entries.
+        for (uint32_t stream = 0; stream < n_stream; ++stream) {
+            llama_seq_id seq_id = -1;
+            if (n_stream == 1) {
+                // a unified stream serves any single sequence
+                seq_id = ubatch.n_seqs_unq > 0 ? ubatch.seq_id_unq[0] : -1;
+            } else {
+                for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+                    if (ubatch.seq_id_unq[s] == (llama_seq_id) stream) {
+                        seq_id = ubatch.seq_id_unq[s];
+                        break;
+                    }
+                }
             }
 
-            const int64_t stream_off = dsv4_stream_offset(n_stream, seq_id, state_size);
-            const uint32_t rollback = (uint32_t) seq_id < rs_idx.size() ? rs_idx[seq_id] : 0;
+            const int64_t stream_off = (int64_t) stream*state_size;
+            const uint32_t rollback = seq_id >= 0 && (uint32_t) seq_id < rs_idx.size() ? rs_idx[seq_id] : 0;
             // Keep the restore graph fixed-width when no rollback is pending.
             const int64_t src_plane = rollback > 0 && rollback <= n_rs_seq ? (int64_t) rollback*state_rows : 0;
             for (uint32_t r = 0; r < state_size; ++r) {
@@ -639,35 +679,33 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_comp_plan(
 
             std::vector<uint32_t> token_idxs;
             token_idxs.reserve(ubatch.n_tokens);
-            for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
-                if (dsv4_token_has_seq(ubatch, i, seq_id)) {
-                    token_idxs.push_back(i);
+            if (seq_id >= 0) {
+                for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
+                    if (dsv4_token_has_seq(ubatch, i, seq_id)) {
+                        token_idxs.push_back(i);
+                    }
                 }
             }
-            if (token_idxs.empty()) {
-                continue;
-            }
 
             const uint32_t n_seq_tokens = (uint32_t) token_idxs.size();
             const int64_t scratch_off = (int64_t) state_rows*(1 + n_rs_seq);
             for (uint32_t d = 1; d <= n_rs_seq; ++d) {
                 const int64_t dst_plane = (int64_t) d*state_rows;
+                const uint32_t prefix = d <= n_seq_tokens ? n_seq_tokens - d : 0;
 
                 for (uint32_t r = 0; r < state_size; ++r) {
-                    int32_t src;
-                    if (d <= n_seq_tokens) {
-                        const uint32_t prefix = n_seq_tokens - d;
-                        src = (int32_t) (stream_off + r);
+                    int32_t src = (int32_t) (stream_off + r);
 
-                        for (uint32_t j = 0; j < prefix; ++j) {
-                            const uint32_t i_tok = token_idxs[j];
-                            if (ubatch.pos[i_tok] >= 0 && (uint32_t) (ubatch.pos[i_tok]%state_size) == r) {
-                                src = (int32_t) (scratch_off + i_tok);
-                            }
+                    for (uint32_t j = 0; j < prefix; ++j) {
+                        const uint32_t i_tok = token_idxs[j];
+                        if (ubatch.pos[i_tok] >= 0 && (uint32_t) (ubatch.pos[i_tok]%state_size) == r) {
+                            src = (int32_t) (scratch_off + i_tok);
                         }
-                    } else {
-                        const int64_t src_plane = (int64_t) (d - n_seq_tokens)*state_rows;
-                        src = (int32_t) (src_plane + stream_off + r);
+                    }
+
+                    if (n_seq_tokens == 0) {
+                        // no-op: copy the snapshot plane onto itself
+                        src = (int32_t) (dst_plane + stream_off + r);
                     }
 
                     plan.state_snapshot_src_idxs.push_back(src);
@@ -683,10 +721,16 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_comp_plan(
     }();
 
     if (debug) {
-        LLAMA_LOG_INFO("%s: ratio=%u, n_tokens=%u, state_persist_dst=%s, state_write_pos=%s\n",
-                __func__, ratio, ubatch.n_tokens,
+        LLAMA_LOG_DEBUG("%s: ratio=%u, n_tokens=%u, n_seqs_unq=%u, state_persist_dst=%s, state_write_pos=%s\n",
+                __func__, ratio, ubatch.n_tokens, ubatch.n_seqs_unq,
                 dsv4_plan_positions(plan.state_persist_dst_idxs).c_str(),
                 dsv4_plan_positions(plan.state_write_pos).c_str());
+        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+            const llama_seq_id seq_id = ubatch.seq_id_unq[s];
+            const uint32_t rollback = seq_id >= 0 && (uint32_t) seq_id < rs_idx.size() ? rs_idx[seq_id] : 0;
+            LLAMA_LOG_DEBUG("%s:   seq %d pos [%d, %d] rollback=%u\n", __func__, seq_id,
+                    ubatch.pos[0], ubatch.pos[ubatch.n_tokens - 1], rollback);
+        }
     }
 
     return plan;
@@ -704,8 +748,17 @@ static std::vector<llama_kv_cache_dsv4_context::comp_plan> dsv4_build_comp_plans
     std::vector<llama_kv_cache_dsv4_context::comp_plan> plans;
     plans.reserve(ubatches.size());
 
+    // the first ubatch touching a seq consumes its rollback restore
+    std::vector<uint32_t> rs(rs_idx);
     for (const llama_ubatch & ubatch : ubatches) {
-        plans.push_back(dsv4_build_comp_plan(ubatch, ratio, overlap, state_size, kv_size, n_stream, n_rs_seq, rs_idx));
+        plans.push_back(dsv4_build_comp_plan(ubatch, ratio, overlap, state_size, kv_size, n_stream, n_rs_seq, rs));
+
+        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+            const llama_seq_id seq_id = ubatch.seq_id_unq[s];
+            if (seq_id >= 0 && (size_t) seq_id < rs.size()) {
+                rs[seq_id] = 0;
+            }
+        }
     }
 
     return plans;
@@ -803,16 +856,15 @@ static llama_kv_cache_dsv4_context::comp_plan dsv4_build_reserve_comp_plan(
         return plan;
     }
 
-    const uint32_t n_seqs       = std::max<uint32_t>(1, ubatch.n_seqs);
-    const uint32_t n_seq_tokens = std::max<uint32_t>(1, ubatch.n_seq_tokens);
-    const uint64_t n_blocks_u64 = (uint64_t) n_seqs*((n_seq_tokens + ratio - 1)/ratio);
-    const size_t n_blocks = (size_t) std::max<uint64_t>(1, n_blocks_u64);
-    GGML_ASSERT((uint64_t) n_blocks == std::max<uint64_t>(1, n_blocks_u64));
+    // worst case over every seq split: sum of per-seq ceil(tokens/ratio) is at
+    // most floor(n_tokens/ratio) + n_seqs
+    const uint32_t n_seqs = std::max<uint32_t>(1, ubatch.n_seqs);
+    const size_t n_blocks = (size_t) ubatch.n_tokens/ratio + n_seqs;
 
     const uint64_t state_rows = (uint64_t) state_size*n_stream;
     const size_t n_persist = (size_t) std::min<uint64_t>(ubatch.n_tokens, state_rows);
-    const size_t n_restore = n_rs_seq > 0 ? (size_t) state_size*std::max<uint32_t>(1, ubatch.n_seqs_unq) : 0;
-    const size_t n_snapshot = (size_t) n_rs_seq*state_size*std::max<uint32_t>(1, ubatch.n_seqs_unq);
+    const size_t n_restore = n_rs_seq > 0 ? (size_t) state_size*n_stream : 0;
+    const size_t n_snapshot = (size_t) n_rs_seq*state_size*n_stream;
 
     plan.state_pos .resize(ubatch.n_tokens);
     plan.state_persist_src_idxs.resize(n_persist);
@@ -1356,7 +1408,9 @@ llama_memory_context_ptr llama_kv_cache_dsv4::init_batch(
             if (has_coupled) {
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                ubatch = balloc.split_equal(n_ubatch, raw_per_seq || comp_per_seq, 0);
+                // [TAG_RECURRENT_ROLLBACK_SPLITS]
+                // the trailing (1 + n_rs_seq) tokens of each seq must stay in the same ubatch
+                ubatch = balloc.split_equal(n_ubatch, raw_per_seq || comp_per_seq, n_rs_seq > 0 ? n_rs_seq + 1 : 0);
             }
 
             if (ubatch.n_tokens == 0) {
@@ -1433,6 +1487,11 @@ bool llama_kv_cache_dsv4::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1
             return false;
         }
 
+        // pending rollback is single-use: stacked partial removals don't compose
+        if (rs_idx[seq_id] != 0) {
+            return false;
+        }
+
         const bool res = kv_raw->seq_rm(seq_id, p0, p1);
         if (res) {
             rs_idx[seq_id] = (uint32_t) rollback;
@@ -1594,9 +1653,7 @@ void llama_kv_cache_dsv4::state_read(llama_io_read_i & io, llama_seq_id seq_id,
     kv_raw->state_read(io, seq_id, flags);
 
     if (!partial_only) {
-        kv_csa->clear(true);
-        kv_hca->clear(true);
-        kv_lid->clear(true);
+        clear_compressed(seq_id, true);
 
         dsv4_state_read_k_cache(io, kv_csa.get(), seq_id, flags);
         dsv4_state_read_k_cache(io, kv_hca.get(), seq_id, flags);
@@ -1680,6 +1737,7 @@ void llama_kv_cache_dsv4::clear_compressed(llama_seq_id seq_id, bool data) {
             kv->seq_rm(seq_id, -1, -1);
 
             if (data) {
+                //TODO: do not clear the kv-cache during `seq_rm`, ref: https://github.com/ggml-org/llama.cpp/pull/26490#discussion_r3798143663
                 for (uint32_t il : kv->get_layer_ids()) {
                     dsv4_clear_tensor_stream(kv->get_k_storage(il), (uint32_t) seq_id);
                 }

src/llama-kv-cache.cpp

@@ -383,6 +383,7 @@ bool llama_kv_cache::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
         return true;
     }
 
+    // TODO: fix incosistent handling of `seq_id < 0` and `seq_id == -1` in the codebase [TAG_LLAMA_SEQ_ID_NEG]
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     if (p0 < 0) {
@@ -2043,6 +2044,7 @@ void llama_kv_cache::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama
 
     GGML_UNUSED(flags);
 
+    // TODO: fix incosistent handling of `seq_id < 0` and `seq_id == -1` in the codebase [TAG_LLAMA_SEQ_ID_NEG]
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     uint32_t n_stream_cur;

src/llama-memory-recurrent.cpp

@@ -158,13 +158,14 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
+    if ((uint32_t) seq_id >= this->n_seq_max) {
+        LLAMA_LOG_ERROR("%s: invalid seq_id (%d) - larger than n_seq_max (%d)\n", __func__, seq_id, this->n_seq_max);
+        return false;
+    }
+
     const bool rm_all = p0 == 0 && p1 == std::numeric_limits<llama_pos>::max();
     if (rm_all) {
-        if (seq_id >= 0) {
-            set_rs_idx(seq_id, 0);
-        } else {
-            std::fill(rs_idx.begin(), rs_idx.end(), 0);
-        }
+        set_rs_idx(seq_id, 0);
     }
 
     // models like Mamba or RWKV can't have a state partially erased at the end
@@ -181,7 +182,9 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
             // partial rollback via per-token snapshot index (bounded by n_rs_seq)
             if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
                 const llama_pos rollback = cell.pos - (p0 - 1);
-                if (rollback >= 1 && rollback <= (llama_pos) n_rs_seq) {
+                // pending rollback is single-use
+                const bool pending = rs_idx[seq_id] != 0;
+                if (!pending && rollback >= 1 && rollback <= (llama_pos) n_rs_seq) {
                     set_rs_idx(seq_id, (uint32_t) rollback);
                     cell.pos = p0 - 1;
                     return true;
@@ -390,10 +393,17 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
 }
 
 void llama_memory_recurrent::set_rs_idx(llama_seq_id seq_id, uint32_t idx) {
-    if (seq_id < 0 || (size_t) seq_id >= rs_idx.size()) {
+    if (seq_id < 0) {
+        std::fill(rs_idx.begin(), rs_idx.end(), 0);
         return;
     }
-    rs_idx[seq_id] = (idx > n_rs_seq) ? n_rs_seq : idx;
+
+    assert(n_seq_max == rs_idx.size());
+
+    GGML_ASSERT((uint32_t) seq_id < n_seq_max);
+    GGML_ASSERT(idx <= n_rs_seq);
+
+    rs_idx[seq_id] = idx;
 }
 
 std::map<ggml_backend_buffer_type_t, size_t> llama_memory_recurrent::memory_breakdown() const {
@@ -742,6 +752,7 @@ void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq
     uint32_t cell_range_begin = size;
     for (uint32_t i = 0; i < size; ++i) {
         const auto & cell = cells[i];
+        // TODO: fix incosistent handling of `seq_id < 0` and `seq_id == -1` in the codebase [TAG_LLAMA_SEQ_ID_NEG]
         if ((seq_id == -1 && !cell.is_empty()) || cell.has_seq_id(seq_id)) {
             ++cell_count;
             uint32_t rs_idx_cur = 0;
@@ -827,6 +838,7 @@ void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_i
     }
 
     if (!res) {
+        // TODO: fix incosistent handling of `seq_id < 0` and `seq_id == -1` in the codebase [TAG_LLAMA_SEQ_ID_NEG]
         if (seq_id == -1) {
             clear(true);
         } else {
@@ -836,11 +848,7 @@ void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_i
     }
 
     if (n_rs_seq != 0) {
-        if (seq_id == -1) {
-            std::fill(rs_idx.begin(), rs_idx.end(), 0);
-        } else {
-            set_rs_idx(seq_id, 0);
-        }
+        set_rs_idx(seq_id, 0);
     }
 }
 

src/llama-model-saver.cpp

@@ -293,6 +293,21 @@ void llama_model_saver::add_kv_from_model() {
     add_kv(LLM_KV_ATTENTION_INDEXER_LOCAL_BLOCKS,    hparams.indexer_local_blocks);
     add_kv(LLM_KV_ATTENTION_INDEXER_TYPES,           hparams.is_indexer_full_impl, true);
     add_kv(LLM_KV_ATTENTION_RECURRENT_LAYERS,        hparams.is_recr_impl, true);
+    add_kv(LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT,      hparams.dsv4_o_group_count);
+    add_kv(LLM_KV_ATTENTION_OUTPUT_LORA_RANK,        hparams.dsv4_o_lora_rank);
+    add_kv(LLM_KV_ATTENTION_COMPRESS_ROPE_FREQ_BASE, hparams.dsv4_compress_rope_base);
+    if (model->arch == LLM_ARCH_DEEPSEEK4 || hparams.dsv4_hc_mult > 0) {
+        // the loader requires one compress ratio per layer, including nextn layers
+        const std::vector<uint32_t> compress_ratios(
+                hparams.dsv4_compress_ratios.begin(), hparams.dsv4_compress_ratios.begin() + hparams.n_layer_all);
+        add_kv(LLM_KV_ATTENTION_COMPRESS_RATIOS,         compress_ratios);
+    } else {
+        add_kv(LLM_KV_ATTENTION_COMPRESS_RATIOS,         hparams.dsv4_compress_ratios, true);
+    }
+    add_kv(LLM_KV_HYPER_CONNECTION_COUNT,               hparams.dsv4_hc_mult);
+    add_kv(LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS, hparams.dsv4_hc_sinkhorn_iters);
+    add_kv(LLM_KV_HYPER_CONNECTION_EPSILON,             hparams.dsv4_hc_eps);
+    add_kv(LLM_KV_HASH_LAYER_COUNT,                     hparams.dsv4_hash_layer_count);
 
     const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
 
@@ -417,11 +432,16 @@ void llama_model_saver::add_tensors_from_model() {
     add_tensor(model->output_s);
     add_tensor(model->output_in_s);
     add_tensor(model->output_res_score);
+    add_tensor(model->nextn_proj_pre);
+    add_tensor(model->nextn_proj_post);
     add_tensor(model->cls);
     add_tensor(model->cls_b);
     add_tensor(model->cls_out);
     add_tensor(model->cls_out_b);
     add_tensor(model->cls_norm);
+    add_tensor(model->hc_head_fn);
+    add_tensor(model->hc_head_base);
+    add_tensor(model->hc_head_scale);
 
     for (const struct llama_layer & layer : model->layers) {
         for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {

src/llama-model.cpp

@@ -365,6 +365,8 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     const llama_meta_device_get_split_state_userdata * ud = (const llama_meta_device_get_split_state_userdata *) userdata;
     const llama_hparams & hparams = ud->model->hparams;
     const std::string tensor_name = tensor->name;
+    const bool is_dsv4 = ud->model->arch == LLM_ARCH_DEEPSEEK4 ||
+        (ud->model->arch == LLM_ARCH_DFLASH && hparams.dsv4_hc_mult > 0);
 
     static const std::regex pattern_q_weight        ("blk\\.\\d*\\.attn_q.weight");
     static const std::regex pattern_kv_weight       ("blk\\.\\d*\\.attn_(k|v).weight");
@@ -374,9 +376,13 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     static const std::regex pattern_qkv_bias        ("blk\\.\\d*\\.attn_qkv.bias");
     static const std::regex pattern_qk_norm         ("blk\\.\\d*\\.attn_(q|k)_norm\\.weight");
     static const std::regex pattern_kv_cache        ("cache_(k|v)_l\\d*");
+    static const std::regex pattern_dsv4_state      ("dsv4_(csa|hca|lid)_state_(kv|score)_l\\d*");
     static const std::regex pattern_attn_sinks      ("blk\\.\\d*\\.attn_sinks.weight");
     static const std::regex pattern_attn_out_weight ("blk\\.\\d*\\.attn_output.weight");
     static const std::regex pattern_attn_out_bias   ("blk\\.\\d*\\.attn_output.bias");
+    static const std::regex pattern_attn_out_a_weight("blk\\.\\d*\\.attn_output_a\\.weight");
+    static const std::regex pattern_attn_out_b_weight("blk\\.\\d*\\.attn_output_b\\.weight");
+    static const std::regex pattern_attn_q_b_weight ("blk\\.\\d*\\.attn_q_b\\.weight");
     static const std::regex pattern_attn_gate_weight("blk\\.\\d*\\.attn_gate.weight");
 
     static const std::regex pattern_ssm_dt          ("blk\\.\\d*\\.ssm_dt.bias");
@@ -395,8 +401,11 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     static const std::regex pattern_ffn_gate_bias     ("blk\\.\\d*\\.ffn_gate(_exps)?.bias");
     static const std::regex pattern_ffn_gate_up_weight("blk\\.\\d*\\.ffn_gate_up(_exps)?.weight");
     static const std::regex pattern_ffn_down_weight   ("blk\\.\\d*\\.ffn_down(_exps)?.weight");
-    static const std::regex pattern_ffn_down_bias     ("blk\\.\\d*\\.ffn_down.bias");
-    static const std::regex pattern_ffn_down_exps_bias("blk\\.\\d*\\.ffn_down_exps.bias");
+    static const std::regex pattern_ffn_down_bias         ("blk\\.\\d*\\.ffn_down.bias");
+    static const std::regex pattern_ffn_down_exps_bias    ("blk\\.\\d*\\.ffn_down_exps.bias");
+    static const std::regex pattern_ffn_up_shexp_weight   ("blk\\.\\d*\\.ffn_up_shexp.weight");
+    static const std::regex pattern_ffn_gate_shexp_weight ("blk\\.\\d*\\.ffn_gate_shexp.weight");
+    static const std::regex pattern_ffn_down_shexp_weight ("blk\\.\\d*\\.ffn_down_shexp.weight");
 
     static const std::regex pattern_output_weight("output\\.weight");
     static const std::regex pattern_output_bias  ("output\\.bias");
@@ -453,6 +462,32 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
     };
 
     auto get_tensor_config = [&]() -> tensor_config {
+        if (is_dsv4) {
+            if (std::regex_match(tensor_name, pattern_kv_cache) ||
+                    std::regex_match(tensor_name, pattern_dsv4_state)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+            }
+            if (std::regex_match(tensor_name, pattern_attn_sinks)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0, "attn_output_a.weight");
+            }
+            if (std::regex_match(tensor_name, pattern_attn_q_b_weight)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "attn_output_a.weight");
+            }
+            if (std::regex_match(tensor_name, pattern_attn_out_a_weight)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_2);
+            }
+            if (std::regex_match(tensor_name, pattern_attn_out_b_weight)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0);
+            }
+            if (std::regex_match(tensor_name, pattern_ffn_up_shexp_weight) ||
+                    std::regex_match(tensor_name, pattern_ffn_gate_shexp_weight)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "ffn_down_shexp.weight");
+            }
+            if (std::regex_match(tensor_name, pattern_ffn_down_shexp_weight)) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_0, "ffn_down_shexp.weight");
+            }
+        }
+
         // standard attention
         if (std::regex_match(tensor_name, pattern_q_weight) || std::regex_match(tensor_name, pattern_kv_weight)) {
             return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1, "attn_output.weight", "ssm_out.weight");
@@ -520,11 +555,14 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
             return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_MIRRORED);
         }
         if (std::regex_match(tensor_name, pattern_ffn_down_exps_bias)) {
-            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_PARTIAL);
+            return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_PARTIAL, "ffn_down_exps.weight");
         }
 
         // output
         if (std::regex_match(tensor_name, pattern_output_weight)) {
+            if (is_dsv4) {
+                return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_MIRRORED);
+            }
             return get_tensor_config_impl(GGML_BACKEND_SPLIT_AXIS_1);
         }
         if (std::regex_match(tensor_name, pattern_output_bias)) {
@@ -554,6 +592,9 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
                     GGML_ASSERT(tensor->ne[axis] == 2*key_dim + value_dim);
                     return {{key_dim, 2}, {value_dim, 1}};
                 }
+                if (std::regex_match(tensor_name, pattern_r_cache)) {
+                    return {{key_dim * (hparams.ssm_d_conv - 1), 2}, {value_dim * (hparams.ssm_d_conv - 1), 1}};
+                }
             } else {
                 const int64_t head_ratio = n_v_heads / n_k_heads;
                 if (std::regex_match(tensor_name, pattern_qkv_weight) || std::regex_match(tensor_name, pattern_ssm_conv1d)) {
@@ -642,12 +683,34 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
                 blck_size_perf *= 2;
             }
 
+            const int64_t granularity_q    = std::lcm(n_embd_q, blck_size_perf);
+            const int64_t granularity_head = granularity_q / hparams.n_embd_head_k(il); // for tensors with one value per head
             if (std::regex_match(tensor_name, pattern_attn_sinks)) {
                 GGML_ASSERT(segments.size() == 1);
-                return {std::lcm(n_embd_q, blck_size_perf)/n_embd_q * n_gqa};
+                if (is_dsv4) {
+                    return {hparams.n_head(il) / hparams.dsv4_o_group_count};
+                }
+                return {granularity_head};
             }
 
-            const int64_t granularity_q = std::lcm(n_embd_q, blck_size_perf);
+            if (is_dsv4) {
+                if (std::regex_match(tensor_name, pattern_attn_q_b_weight)) {
+                    GGML_ASSERT(segments.size() == 1);
+                    // the grouped output projection requires each device to hold whole groups of heads
+                    const int64_t n_head_group = hparams.n_head(il) / hparams.dsv4_o_group_count;
+                    return {n_head_group * hparams.n_embd_head_k(il)};
+                }
+                if (std::regex_match(tensor_name, pattern_attn_out_a_weight)) {
+                    GGML_ASSERT(segments.size() == 1);
+                    return {1};
+                }
+                if (std::regex_match(tensor_name, pattern_attn_out_b_weight)) {
+                    GGML_ASSERT(segments.size() == 1);
+                    // the boundaries must align with wo_a's per-group split, so quant blocks must not straddle groups
+                    GGML_ASSERT(hparams.dsv4_o_lora_rank % blck_size == 0);
+                    return {hparams.dsv4_o_lora_rank};
+                }
+            }
             if (std::regex_match(tensor_name, pattern_q_weight) || std::regex_match(tensor_name, pattern_q_bias)) {
                 GGML_ASSERT(segments.size() == 1);
                 // some models have Q gate tensors, for those cases the granularity needs to be doubled:
@@ -660,6 +723,13 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
                 GGML_ASSERT(segments.size() == 1);
                 return {granularity_q};
             }
+            if (std::regex_match(tensor_name, pattern_attn_gate_weight)) {
+                GGML_ASSERT(segments.size() == 1);
+                if (tensor->ne[1] == hparams.n_head(il)) {
+                    return {granularity_head};
+                }
+                return {granularity_q};
+            }
 
             const int64_t granularity_kv = granularity_q / n_gqa;
             if (std::regex_match(tensor_name, pattern_kv_weight) ||
@@ -677,7 +747,11 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
         // FFN
         if (std::regex_match(tensor_name, pattern_ffn_up_weight) || std::regex_match(tensor_name, pattern_ffn_up_bias) ||
                 std::regex_match(tensor_name, pattern_ffn_gate_weight) || std::regex_match(tensor_name, pattern_ffn_gate_bias) ||
-                std::regex_match(tensor_name, pattern_ffn_gate_up_weight) || std::regex_match(tensor_name, pattern_ffn_down_weight)) {
+                std::regex_match(tensor_name, pattern_ffn_gate_up_weight) ||
+                std::regex_match(tensor_name, pattern_ffn_down_weight) ||
+                std::regex_match(tensor_name, pattern_ffn_up_shexp_weight) ||
+                std::regex_match(tensor_name, pattern_ffn_gate_shexp_weight) ||
+                std::regex_match(tensor_name, pattern_ffn_down_shexp_weight)) {
             const int64_t blck_size_perf = std::lcm(blck_size, 128);
             GGML_ASSERT(segments.size() == 1);
             return {blck_size_perf};
@@ -728,6 +802,16 @@ struct ggml_backend_meta_split_state llama_meta_device_get_split_state(const str
         memset(split_state.ne, 0, sizeof(split_state.ne));
         split_state.nr[0] = 1;
         split_state.n_segments = 1;
+        if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
+            GGML_ASSERT(tc.tensor_axis_0 != tensor);
+            const ggml_backend_meta_split_state source_split_state = llama_meta_device_get_split_state(tc.tensor_axis_0, userdata);
+            GGML_ASSERT(source_split_state.axis >= 0 && source_split_state.axis < GGML_MAX_DIMS);
+            for (size_t j = 0; j < ud->n_devices; j++) {
+                for (size_t is = 0; is < source_split_state.n_segments; is++) {
+                    split_state.ne[j] += source_split_state.ne[is*ud->n_devices + j] * source_split_state.nr[is];
+                }
+            }
+        }
     }
     return split_state;
     GGML_UNUSED(userdata);

src/models/deepseek4.cpp

@@ -1,3 +1,4 @@
+#include "llama-hparams.h"
 #include "models.h"
 
 #include "llama-kv-cache-dsv4.h"
@@ -58,6 +59,7 @@ void llama_model_deepseek4::load_arch_hparams(llama_model_loader & ml) {
     if (n_compress_ratios < hparams.n_layer_all) {
         throw std::runtime_error("DeepSeek-V4 compress_ratios is shorter than block_count");
     }
+    GGML_ASSERT(n_compress_ratios <= LLAMA_MAX_LAYERS);
     ml.get_arr(LLM_KV_ATTENTION_COMPRESS_RATIOS, hparams.dsv4_compress_ratios);
 
     ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func);

src/models/dflash.cpp

@@ -117,6 +117,10 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
     output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), { n_embd }, 0); // encoder hidden_norm (after fc)
     output_norm     = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM,    "weight"), { n_embd }, 0); // decoder final norm
 
+    // optional: reduced-vocab drafts ship their own lm head, full-vocab drafts can share the target's via ctx_other
+    // a draft with its own embeddings + head references no target tensors and can run on devices the target does not use (e.g. -devd with a tensor-split target)
+    output   = create_tensor(tn(LLM_TENSOR_OUTPUT,     "weight"), { n_embd, n_vocab_draft }, TENSOR_NOT_REQUIRED);
+
     if (hparams.dsv4_hc_mult > 0) {
         const int64_t q_lora_rank     = hparams.n_lora_q;
         const int64_t n_ff_exp        = hparams.n_ff_exp;
@@ -167,9 +171,6 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
         return;
     }
 
-    // optional: reduced-vocab drafts ship their own, full-vocab drafts share the target's via ctx_other
-    output   = create_tensor(tn(LLM_TENSOR_OUTPUT,     "weight"), { n_embd, n_vocab_draft }, TENSOR_NOT_REQUIRED);
-
     for (int i = 0; i < n_layer; ++i) {
         auto & layer = layers[i];
 

src/models/glm4-moe.cpp

@@ -29,10 +29,19 @@ void llama_model_glm4_moe::load_arch_hparams(llama_model_loader & ml) {
     }
 }
 
-void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {
+void llama_model_glm4_moe::load_arch_tensors(llama_model_loader & ml) {
     LLAMA_LOAD_LOCALS;
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
+    const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);
+    const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
+    const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
+    const int trunk_flags = mtp_only  ? TENSOR_NOT_REQUIRED : 0;
+    int       mtp_flags   = trunk_only ? TENSOR_NOT_REQUIRED : 0;
+
+    if (!ml.load_mtp) {
+        mtp_flags |= TENSOR_SKIP;
+    }
 
     GGML_ASSERT(hparams.n_expert > 0 && "n_expert must be > 0 for GLM4_MOE MoE layers");
     GGML_ASSERT(hparams.n_expert_used > 0 && "n_expert_used must be > 0 for GLM4_MOE MoE layers");
@@ -47,16 +56,9 @@ void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {
         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
     }
 
-    // Load ALL tensors including NextN layer to satisfy total tensor count
-    // but only PROCESS up to last layer (skipping final NextN layer) in forward pass
     for (int i = 0; i < n_layer_all; ++i) {
-        int flags = 0;
-        if (i >= n_layer) {
-            // skip all tensors in the NextN layers
-            flags |= TENSOR_SKIP;
-        }
-
         auto & layer = layers[i];
+        const int flags = i < n_layer ? trunk_flags : mtp_flags;
 
         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
 
@@ -110,24 +112,186 @@ void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {
             layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), { n_embd, n_ff }, flags);
         }
 
-        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+        // NextN/MTP tensors
         if (i >= n_layer) {
             layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
             layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
             layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
 
             // Optional tensors
-            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
-            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
-            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
+            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED | flags);
         }
     }
 }
 
 std::unique_ptr<llm_graph_context> llama_model_glm4_moe::build_arch_graph(const llm_graph_params & params) const {
+    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
+        return std::make_unique<graph_mtp>(*this, params);
+    }
     return std::make_unique<graph>(*this, params);
 }
 
+llama_model_glm4_moe::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)
+    : llm_graph_context(params) {
+    GGML_ASSERT(hparams.n_layer_nextn > 0 && "GLM4_MOE MTP requires n_layer_nextn > 0");
+    GGML_ASSERT(hparams.n_layer_nextn == 1 && "GLM4_MOE MTP currently only supports a single MTP block");
+
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+
+    const int il = hparams.n_layer() + cparams.nextn_layer_offset;
+    GGML_ASSERT(cparams.nextn_layer_offset >= 0 &&
+                cparams.nextn_layer_offset < (int) hparams.n_layer_nextn &&
+                "nextn_layer_offset out of range [0, n_layer_nextn)");
+
+    const auto & layer = model.layers[il];
+
+    GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");
+    GGML_ASSERT(layer.nextn.enorm   && "MTP block missing nextn.enorm");
+    GGML_ASSERT(layer.nextn.hnorm   && "MTP block missing nextn.hnorm");
+    GGML_ASSERT(layer.ffn_gate_inp  && "MTP block missing ffn_gate_inp");
+
+    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);
+
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    ggml_set_input(inp->tokens);
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
+    ggml_set_input(inp->embd);
+
+    ggml_tensor * tok_embd;
+    if (ubatch.token) {
+        ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
+        tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
+    } else {
+        tok_embd = inp->embd;
+    }
+    cb(tok_embd, "mtp_tok_embd", il);
+
+    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
+    ggml_set_input(inp->h);
+    ggml_set_name(inp->h, "mtp_h_input");
+
+    ggml_tensor * h_embd = inp->h;
+
+    res->add_input(std::move(inp));
+
+    ggml_tensor * inp_pos     = build_inp_pos();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
+    cb(h_norm, "mtp_hnorm", il);
+
+    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
+    cb(e_norm, "mtp_enorm", il);
+
+    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, 0);
+    cb(concat, "mtp_concat", il);
+
+    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);
+    cb(cur, "mtp_eh_proj", il);
+
+    ggml_tensor * inpSA = cur;
+
+    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_attn_norm", il);
+
+    auto [Qcur, Kcur, Vcur] = build_qkv(layer, cur,
+            n_embd_head, n_head, n_head_kv, il);
+
+    if (layer.attn_q_norm) {
+        Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il);
+        cb(Qcur, "mtp_Qcur_normed", il);
+    }
+    if (layer.attn_k_norm) {
+        Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il);
+        cb(Kcur, "mtp_Kcur_normed", il);
+    }
+
+    Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
+            rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+
+    Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
+            rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+
+    cb(Qcur, "mtp_Qcur", il);
+    cb(Kcur, "mtp_Kcur", il);
+    cb(Vcur, "mtp_Vcur", il);
+
+    cur = build_attn(inp_attn,
+            layer.wo, nullptr, layer.wo_s,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr,
+            1.0f / sqrtf(float(n_embd_head)), il);
+    cb(cur, "mtp_attn_out", il);
+
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "mtp_ffn_inp", il);
+
+    cur = build_norm(ffn_inp, layer.attn_post_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_post_attn_norm", il);
+
+    ggml_tensor * routed_out = build_moe_ffn(cur,
+            layer.ffn_gate_inp,
+            layer.ffn_up_exps,
+            layer.ffn_gate_exps,
+            layer.ffn_down_exps,
+            layer.ffn_exp_probs_b,
+            n_expert, n_expert_used,
+            LLM_FFN_SILU, hparams.expert_weights_norm,
+            hparams.expert_weights_scale,
+            (llama_expert_gating_func_type) hparams.expert_gating_func,
+            il);
+    cb(routed_out, "mtp_ffn_moe_out", il);
+
+    ggml_tensor * shared_out = build_ffn(cur,
+            layer.ffn_up_shexp,   nullptr, nullptr,
+            layer.ffn_gate_shexp, nullptr, nullptr,
+            layer.ffn_down_shexp, nullptr, nullptr,
+            nullptr,
+            LLM_FFN_SILU, LLM_FFN_PAR, il);
+    cb(shared_out, "mtp_ffn_shexp_out", il);
+
+    cur = ggml_add(ctx0, routed_out, shared_out);
+    cb(cur, "mtp_ffn_out", il);
+
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "mtp_post_ffn", il);
+
+    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
+            ? layer.nextn.shared_head_norm
+            : model.output_norm;
+    GGML_ASSERT(head_norm_w && "GLM4_MOE MTP: missing both nextn.shared_head_norm and output_norm");
+
+    cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+    cb(cur, "mtp_shared_head_norm", -1);
+
+    ggml_tensor * head_w = layer.nextn.shared_head_head
+            ? layer.nextn.shared_head_head
+            : model.output;
+    ggml_tensor * head_s = layer.nextn.shared_head_head
+            ? layer.nextn.shared_head_head_s
+            : model.output_s;
+    GGML_ASSERT(head_w && "GLM4_MOE MTP: missing LM head (nextn.shared_head_head or model.output)");
+
+    cur = build_lora_mm(head_w, cur, head_s);
+    cb(cur, "result_output", -1);
+
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}
+
 llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
@@ -154,8 +318,7 @@ llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_pa
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
-    // Only process up to last layer (skip final NextN layer)
-    // Final layer tensors are loaded but not processed in forward pass
+    // NextN layers are processed by graph_mtp.
     for (int il = 0; il < n_layer; ++il) {
         ggml_tensor * inpSA = inpL;
 
@@ -205,7 +368,7 @@ llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_pa
                     model.layers[il].wo, NULL, model.layers[il].wo_s,
                     Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
         }
-        if (il == n_layer - 1 && inp_out_ids) {
+        if (il == n_layer - 1 && inp_out_ids && (!cparams.embeddings_nextn || cparams.embeddings_nextn_masked)) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }
@@ -265,6 +428,13 @@ llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_pa
     cur = inpL;
     cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
 
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (cparams.embeddings_nextn && !cparams.embeddings_nextn_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 

src/models/mamba-base.cpp

@@ -182,13 +182,14 @@ ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
     ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs);
     conv               = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs);
 
-    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
-    cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs);
-
     // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads
 
-    // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs}
+    // Keep the projection 2D: with a {n_embd, 1, n_seqs} batch the CUDA backend
+    // dispatches a column-batched GEMV for what is a large dense GEMM.
+    // {n_embd, d_in_proj} @ {n_embd, n_tokens} => {d_in_proj, n_tokens}
     ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur, model.layers[il].ssm_in_s);
+    // {d_in_proj, n_tokens} => {d_in_proj, n_seq_tokens, n_seqs}
+    zxBCdt = ggml_reshape_3d(ctx0, zxBCdt, zxBCdt->ne[0], n_seq_tokens, n_seqs);
 
     // split the above in three
     ggml_tensor * z   = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * zxBCdt->nb[0],
@@ -290,15 +291,12 @@ ggml_tensor * llm_build_mamba_base::build_mamba2_layer(llm_graph_input_rs * inp,
             y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il);
         }
 
-        y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs);
+        y = ggml_reshape_2d(ctx0, y, d_inner, n_seq_tokens * n_seqs);
 
-        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
+        // {d_inner, n_embd} @ {d_inner, n_tokens} => {n_embd, n_tokens}
         cur = build_lora_mm(model.layers[il].ssm_out, y, model.layers[il].ssm_out_s);
     }
 
-    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
-    cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs);
     cb(cur, "mamba_out", il);
-
     return cur;
 }

src/models/models.h

@@ -1412,6 +1412,10 @@ struct llama_model_glm4_moe : public llama_model_base {
         graph(const llama_model & model, const llm_graph_params & params);
     };
 
+    struct graph_mtp : public llm_graph_context {
+        graph_mtp(const llama_model & model, const llm_graph_params & params);
+    };
+
     std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
 };
 

tests/CMakeLists.txt

@@ -228,6 +228,15 @@ if (NOT WIN32 OR NOT BUILD_SHARED_LIBS)
     set_tests_properties(test-recurrent-state-rollback-nemotron-h PROPERTIES
         FIXTURES_REQUIRED generate-models
     )
+    llama_test(
+        test-recurrent-state-rollback
+        NAME test-recurrent-state-rollback-dsv4
+        LABEL main
+        ARGS -m "${MODEL_DIR}/deepseek4-moe.gguf"
+    )
+    set_tests_properties(test-recurrent-state-rollback-dsv4 PROPERTIES
+        FIXTURES_REQUIRED generate-models
+    )
 endif()
 
 llama_build_and_test(test-chat-peg-parser.cpp peg-parser/simple-tokenize.cpp)
@@ -235,6 +244,8 @@ llama_build_and_test(test-jinja.cpp)
 llama_test(test-jinja NAME test-jinja-py ARGS -py LABEL python)
 llama_build_and_test(test-chat-auto-parser.cpp WORKING_DIRECTORY ${PROJECT_SOURCE_DIR})
 llama_build_and_test(test-chat-template.cpp)
+# debug tool for chat template differential analysis (not registered as a test, run it manually)
+llama_build(test-chat-analysis.cpp)
 llama_build_and_test(test-log.cpp)
 llama_build_and_test(
     test-peg-parser.cpp

tests/test-backend-ops.cpp

@@ -10573,6 +10573,101 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_from_file(const c
     return test_cases;
 }
 
+// ---- FA vec (Q,NE): forced-config numerical slice (Metal only) ----
+using set_fa_vec_override_t   = void (*)(int, int);
+using clear_fa_vec_override_t = void (*)(void);
+
+// NL = 32/NE must divide both dk/4 and dv/4.
+static std::vector<int> fa_vec_legal_ne(int dk, int dv) {
+    std::vector<int> r;
+    for (int ne : {1, 2, 4}) {
+        const int nl = 32 / ne;
+        if ((dk/4) % nl == 0 && (dv/4) % nl == 0) {
+            r.push_back(ne);
+        }
+    }
+    return r;
+}
+
+static bool op_names_filter_selects(const char * op_names_filter, const char * op_name) {
+    if (!op_names_filter) {
+        return true;
+    }
+    std::string_view filter(op_names_filter);
+    while (!filter.empty()) {
+        auto comma_pos = filter.find_first_of(',');
+        const auto lparen_pos = filter.find_first_of('(');
+        std::string_view entry;
+        if (lparen_pos < comma_pos) {
+            const auto rparen_pos = filter.find_first_of(')');
+            comma_pos = filter.find_first_of(',', rparen_pos);
+            entry = filter.substr(0, lparen_pos);
+        } else {
+            entry = filter.substr(0, comma_pos);
+        }
+        if (entry == op_name) {
+            return true;
+        }
+        filter = comma_pos != std::string_view::npos ? filter.substr(comma_pos + 1) : "";
+    }
+    return false;
+}
+
+// Covers padded rows, sinks, kvpad, multi-SIMDgroup reduction, quantized K/V, and MLA views.
+// The override is backend-global, so this runs after all parallel workers have joined.
+static bool run_fa_vec_slice(ggml_backend_t backend, ggml_backend_t backend_cpu, const char * op_names_filter) {
+    if (!op_names_filter_selects(op_names_filter, "FLASH_ATTN_EXT")) {
+        return true;
+    }
+
+    auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend));
+
+    auto set_ov   = (set_fa_vec_override_t)   ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_set_fa_vec_override");
+    auto clear_ov = (clear_fa_vec_override_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_metal_tuning_clear_fa_vec_override");
+    if (!set_ov || !clear_ov) {
+        return true;  // not the Metal backend: nothing to force
+    }
+
+    struct shape_t { int dk, dv; };
+    const shape_t   shapes[] = { { 128, 128 }, { 576, 512 } };  // mainstream head size + MLA shared K/V view
+    const int       ne01_pts[] = { 1, 3 };                      // decode, and padded rows for Q=2 and Q=4
+    const int       ne11_pts[] = { 512, 4097 };                 // nsg=1, and nsg>=2 together with kvpad
+    const ggml_type types[]    = { GGML_TYPE_F16, GGML_TYPE_Q4_0 };
+
+    int n_run = 0, n_fail = 0;
+    for (auto s : shapes) {
+        for (int ne : fa_vec_legal_ne(s.dk, s.dv)) {
+            for (int Q : { 1, 2, 4 }) {
+                for (ggml_type type_kv : types) {
+                    for (bool sinks : { false, true }) {
+                        for (int ne01 : ne01_pts) {
+                            for (int ne11 : ne11_pts) {
+                                set_ov(Q, ne);
+                                test_flash_attn_ext tc(s.dk, s.dv, /*nh=*/4, { 1, 1 }, /*kv=*/ne11, /*nb=*/ne01,
+                                                       /*mask=*/true, sinks, 0.0f, 0.0f, GGML_PREC_F32,
+                                                       type_kv, type_kv);
+                                auto st = tc.eval(backend, backend_cpu, "FLASH_ATTN_EXT", nullptr);
+                                clear_ov();
+
+                                if (st == test_status_t::FAIL) {
+                                    printf("  FAIL fa_vec slice: dk=%d dv=%d Q=%d ne=%d type=%s ne01=%d ne11=%d sinks=%d\n",
+                                           s.dk, s.dv, Q, ne, ggml_type_name(type_kv), ne01, ne11, (int) sinks);
+                                    n_fail++;
+                                }
+                                n_run++;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    printf("  fa_vec (Q,NE) slice: %d cases run, %d failed\n", n_run, n_fail);
+
+    return n_fail == 0;
+}
+
 static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mode mode, const char * op_names_filter, const char * params_filter,
                          printer * output_printer, const char * test_file_path, int parallel_workers) {
     auto filter_test_cases = [](std::vector<std::unique_ptr<test_case>> & test_cases, const char * params_filter) {
@@ -10710,7 +10805,9 @@ static bool test_backend(ggml_backend_t backend, ggml_backend_dev_t dev, test_mo
         output_printer->print_summary(test_summary_info(n_ok, tests_run, false));
         output_printer->print_failed_tests(failed_tests);
 
-        return n_ok == tests_run;
+        const bool slice_ok = run_fa_vec_slice(backend, backend_cpu.get(), op_names_filter);
+
+        return n_ok == tests_run && slice_ok;
     }
 
     if (mode == MODE_GRAD) {

tests/test-chat-analysis.cpp

@@ -84,11 +84,12 @@ static std::string read_file(const std::string & path) {
 }
 
 static void print_usage(const char * program_name) {
-    LOG_ERR("Usage: %s [options]\n", program_name);
+    LOG_ERR("Debug the auto-parser's differential analysis: render a template with/without tools, reasoning, etc. and show the diffs.\n");
+    LOG_ERR("\nUsage: %s [options]\n", program_name);
     LOG_ERR("\nOptions:\n");
     LOG_ERR("  --template <name>       Analyze specific template from test suite (e.g., 'deepseek' or 'DeepSeek-V3.1')\n");
     LOG_ERR("  --template-file <path>  Analyze custom template file\n");
-    LOG_ERR("  --all                   Analyze all templates from test suite\n");
+    LOG_ERR("  --all                   Analyze all templates from test suite (default when no arguments are given)\n");
     LOG_ERR("\nExamples:\n");
     LOG_ERR("  %s --all\n", program_name);
     LOG_ERR("  %s --template deepseek\n", program_name);
@@ -97,14 +98,17 @@ static void print_usage(const char * program_name) {
 
 static bool parse_options(int argc, char ** argv, analysis_options & opts) {
     if (argc < 2) {
-        print_usage(argv[0]);
-        return false;
+        // default mode: analyze all templates from the test suite
+        opts.analyze_all = true;
     }
 
     for (int i = 1; i < argc; ++i) {
         std::string arg = argv[i];
 
-        if (arg == "--all") {
+        if (arg == "-h" || arg == "--help") {
+            print_usage(argv[0]);
+            return false;
+        } else if (arg == "--all") {
             opts.analyze_all = true;
         } else if (arg == "--template") {
             if (i + 1 >= argc) {

tests/test-chat-auto-parser.cpp

@@ -2,11 +2,18 @@
 #include "chat-auto-parser.h"
 #include "chat-peg-parser.h"
 #include "chat.h"
+#include "gguf.h"
+#include "jinja/runtime.h"
+#include "log.h"
 #include "peg-parser.h"
 #include "testing.h"
 
+#include <cstdlib>
+#include <filesystem>
 #include <fstream>
 #include <iostream>
+#include <iterator>
+#include <optional>
 #include <sstream>
 #include <string>
 
@@ -94,11 +101,447 @@ static void test_bailing_v3_tool_format(testing & t);
 
 static void test_role_markers_all_templates(testing & t);
 
+static json build_tools_definition();
+
+//
+// debug mode: analyze a single template and dump the generated parser and grammar
+//
+
+enum class output_mode {
+    ANALYSIS,  // Only output analysis results (default)
+    TEMPLATE,  // Only output rendered template
+    BOTH       // Output both
+};
+
+enum class input_message_type {
+    NONE,                    // Don't render any message scenarios (only analysis)
+    CONTENT_ONLY,            // Simple assistant message with content
+    REASONING_CONTENT,       // Message with reasoning_content + content
+    TOOL_CALL_ONLY,          // Message with tool_calls only
+    CONTENT_TOOL_CALL,       // Message with content + tool_calls
+    REASONING_TOOL_CALL,     // Message with reasoning_content + tool_calls
+    CONTENT_FAKE_TOOL_CALL,  // Message with content but no actual tool_calls (for testing)
+    ALL                      // Render all scenarios
+};
+
+struct debug_options {
+    std::string        template_path;
+    bool               with_tools        = true;
+    bool               generation_prompt = true;
+    bool               enable_reasoning  = true;
+    bool               debug_jinja       = false;
+    bool               force_tool_call   = false;
+    bool               parallel_tool_calls = true;
+    output_mode        mode              = output_mode::BOTH;
+    input_message_type input_message     = input_message_type::NONE;
+};
+
+static std::string read_file(const std::string & path) {
+    std::ifstream fin(path, std::ios::binary);
+    if (!fin.is_open()) {
+        throw std::runtime_error("Could not open file: " + path);
+    }
+    std::ostringstream buf;
+    buf << fin.rdbuf();
+    return buf.str();
+}
+
+static std::string read_gguf_chat_template(const std::string & path) {
+    struct gguf_init_params params = { /*no_alloc =*/true,  // We only need metadata, not tensor data
+                                       /*ctx=*/nullptr };
+
+    struct gguf_context * ctx = gguf_init_from_file(path.c_str(), params);
+    if (ctx == nullptr) {
+        throw std::runtime_error("Could not open GGUF file: " + path);
+    }
+
+    const char * key    = "tokenizer.chat_template";
+    int64_t      key_id = gguf_find_key(ctx, key);
+
+    if (key_id == -1) {
+        gguf_free(ctx);
+        throw std::runtime_error("GGUF file does not contain chat template key: " + std::string(key));
+    }
+
+    const char * template_str = gguf_get_val_str(ctx, key_id);
+    if (template_str == nullptr) {
+        gguf_free(ctx);
+        throw std::runtime_error("GGUF file contains chat template key but value is null");
+    }
+
+    std::string result = template_str;
+    gguf_free(ctx);
+    return result;
+}
+
+static void print_usage(const char * program_name) {
+    LOG_ERR("Test the chat template auto-parser; also usable as a debug tool that shows the generated PEG parser, GBNF grammar and triggers for a given template.\n");
+    LOG_ERR("\nUsage: %s [filter_regex]                       run the automated tests (default)\n", program_name);
+    LOG_ERR("       %s <template_or_gguf_path> [options]     debug a single template\n", program_name);
+    LOG_ERR("\nDebug mode options:\n");
+    LOG_ERR("  --no-tools              Disable tool definitions\n");
+    LOG_ERR("  --force-tool-call       Set tool calls to forced\n");
+    LOG_ERR("  --parallel-tool-calls=0|1 Set parallel_tool_calls (default: 1)\n");
+    LOG_ERR("  --generation-prompt=0|1 Set add_generation_prompt (default: 1)\n");
+    LOG_ERR("  --enable-reasoning=0|1  Enable reasoning parsing (default: 1)\n");
+    LOG_ERR("  --output=MODE           Output mode: analysis, template, both (default: both)\n");
+    LOG_ERR("  --debug-jinja           Enable Jinja fine-grained debug\n");
+    LOG_ERR("  --input-message=TYPE    Message type to render:\n");
+    LOG_ERR("                          content_only, reasoning_content, tool_call_only,\n");
+    LOG_ERR("                          content_tool_call, reasoning_tool_call,\n");
+    LOG_ERR("                          content_fake_tool_call, all\n");
+    LOG_ERR("\nExamples:\n");
+    LOG_ERR("  %s template.jinja --input-message=all --generation-prompt=1\n", program_name);
+    LOG_ERR("  %s template.jinja --output=template --input-message=tool_call_only\n", program_name);
+}
+
+static bool parse_bool_option(const std::string & value) {
+    return value == "1" || value == "true" || value == "yes";
+}
+
+static bool parse_debug_options(int argc, char ** argv, debug_options & opts) {
+    opts.template_path = argv[1];
+
+    for (int i = 2; i < argc; ++i) {
+        std::string arg = argv[i];
+
+        if (arg == "--force-tool-call") {
+            opts.force_tool_call = true;
+        } else if (arg == "--debug-jinja") {
+            opts.debug_jinja = true;
+        } else if (arg == "--no-tools") {
+            opts.with_tools = false;
+        } else if (arg.rfind("--parallel-tool-calls=", 0) == 0) {
+            opts.parallel_tool_calls = parse_bool_option(arg.substr(22));
+        } else if (arg.rfind("--generation-prompt=", 0) == 0) {
+            opts.generation_prompt = parse_bool_option(arg.substr(20));
+        } else if (arg.rfind("--enable-reasoning=", 0) == 0) {
+            opts.enable_reasoning = parse_bool_option(arg.substr(19));
+        } else if (arg.rfind("--output=", 0) == 0) {
+            std::string mode = arg.substr(9);
+            if (mode == "analysis") {
+                opts.mode = output_mode::ANALYSIS;
+            } else if (mode == "template") {
+                opts.mode = output_mode::TEMPLATE;
+            } else if (mode == "both") {
+                opts.mode = output_mode::BOTH;
+            } else {
+                LOG_ERR("Unknown output mode: %s\n", mode.c_str());
+                return false;
+            }
+        } else if (arg.rfind("--input-message=", 0) == 0) {
+            std::string type = arg.substr(16);
+            if (type == "content_only") {
+                opts.input_message = input_message_type::CONTENT_ONLY;
+            } else if (type == "reasoning_content") {
+                opts.input_message = input_message_type::REASONING_CONTENT;
+            } else if (type == "tool_call_only") {
+                opts.input_message = input_message_type::TOOL_CALL_ONLY;
+            } else if (type == "content_tool_call") {
+                opts.input_message = input_message_type::CONTENT_TOOL_CALL;
+            } else if (type == "reasoning_tool_call") {
+                opts.input_message = input_message_type::REASONING_TOOL_CALL;
+            } else if (type == "content_fake_tool_call") {
+                opts.input_message = input_message_type::CONTENT_FAKE_TOOL_CALL;
+            } else if (type == "all") {
+                opts.input_message = input_message_type::ALL;
+            } else {
+                LOG_ERR("Unknown input message type: %s\n", type.c_str());
+                return false;
+            }
+        } else {
+            LOG_ERR("Unknown option: %s\n", arg.c_str());
+            print_usage(argv[0]);
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static json build_debug_user_message() {
+    return json{
+        { "role",    "user"                               },
+        { "content", "Hello, please help me with a task." }
+    };
+}
+
+static json build_content_only_message() {
+    return json{
+        { "role",    "assistant"                                   },
+        { "content", "Hello! I'm here to help you with your task." }
+    };
+}
+
+static json build_reasoning_content_message() {
+    return json{
+        { "role",              "assistant"                                                               },
+        { "content",           "Hello! I'm here to help you with your task."                             },
+        { "reasoning_content", "The user is greeting me and asking for help. I should respond politely." }
+    };
+}
+
+static json build_tool_call_only_message() {
+    return json{
+        { "role",       "assistant"      },
+        { "content",    nullptr          },
+        { "tool_calls",
+         json::array({ json{
+              { "type", "function" },
+              { "function", json{ { "name", "test_function_name" },
+                                  { "arguments", json::object({ { "param1", "value1" }, { "param2", "value2" } }) } } },
+              { "id", "123456789" } } }) }
+    };
+}
+
+static json build_content_tool_call_message() {
+    return json{
+        { "role",       "assistant"                                                                              },
+        { "content",    "I'll help you by calling a function."                                                   },
+        { "tool_calls",
+         json::array({ json{
+              { "type", "function" },
+              { "function",
+                json{ { "name", "test_function_name" },
+                      { "arguments", json::object({ { "param1", "value1" }, { "param2", "value2" } }) } } } } }) }
+    };
+}
+
+static json build_reasoning_tool_call_message() {
+    return json{
+        { "role",              "assistant"                                                                       },
+        { "content",           nullptr                                                                           },
+        { "reasoning_content", "I need to call a function to help with this task."                               },
+        { "tool_calls",
+         json::array({ json{
+              { "type", "function" },
+              { "function",
+                json{ { "name", "test_function_name" },
+                      { "arguments", json::object({ { "param1", "value1" }, { "param2", "value2" } }) } } } } }) }
+    };
+}
+
+static json build_content_fake_tool_call_message() {
+    // This message has content but NO tool_calls field
+    // It's used to test if a template renders tool definitions but not tool calls
+    return json{
+        { "role",    "assistant"                            },
+        { "content", "I'll help you by calling a function." }
+    };
+}
+
+static void render_scenario(const common_chat_template & tmpl,
+                            const std::string &          scenario_name,
+                            const json &                 messages,
+                            const json &                 tools,
+                            bool                         add_generation_prompt,
+                            bool                         enable_thinking) {
+    LOG_ERR("\n=== Scenario: %s ===\n", scenario_name.c_str());
+    LOG_ERR("add_generation_prompt: %s, enable_thinking: %s\n", add_generation_prompt ? "true" : "false",
+            enable_thinking ? "true" : "false");
+
+    // When add_generation_prompt is true, add a trailing user message to trigger the prompt
+    json final_messages = messages;
+    if (add_generation_prompt && !messages.empty() && messages.back().value("role", "") == "assistant") {
+        final_messages.push_back(json{
+            { "role",    "user"                                       },
+            { "content", "Now please continue with another response." }
+        });
+    }
+
+    LOG_ERR("Messages:\n%s\n", final_messages.dump(2).c_str());
+
+    try {
+        generation_params inputs;
+        inputs.messages                         = final_messages;
+        inputs.add_generation_prompt            = add_generation_prompt;
+        inputs.extra_context["enable_thinking"] = enable_thinking;
+
+        if (!tools.is_null() && tools.is_array() && !tools.empty()) {
+            inputs.tools = tools;
+        }
+
+        std::string output = common_chat_template_direct_apply(tmpl, inputs);
+
+        LOG_ERR("\n--- Rendered Output ---\n");
+        LOG_ERR("%s\n", output.c_str());
+        LOG_ERR("--- End Output (length: %zu) ---\n", output.length());
+    } catch (const std::exception & e) {
+        LOG_ERR("Rendering failed: %s\n", e.what());
+    }
+}
+
+static void render_all_scenarios(const common_chat_template & tmpl,
+                                 const json &                 tools,
+                                 bool                         add_generation_prompt,
+                                 bool                         enable_thinking,
+                                 input_message_type           message_type) {
+    json user_msg = build_debug_user_message();
+
+    auto render_if = [&](input_message_type type, const std::string & name, const json & assistant_msg) {
+        if (message_type == input_message_type::ALL || message_type == type) {
+            json messages = json::array({ user_msg, assistant_msg });
+            render_scenario(tmpl, name, messages, tools, add_generation_prompt, enable_thinking);
+        }
+    };
+
+    render_if(input_message_type::CONTENT_ONLY, "content_only", build_content_only_message());
+    render_if(input_message_type::REASONING_CONTENT, "reasoning_content", build_reasoning_content_message());
+    render_if(input_message_type::TOOL_CALL_ONLY, "tool_call_only", build_tool_call_only_message());
+    render_if(input_message_type::CONTENT_TOOL_CALL, "content_tool_call", build_content_tool_call_message());
+    render_if(input_message_type::REASONING_TOOL_CALL, "reasoning_tool_call", build_reasoning_tool_call_message());
+    render_if(input_message_type::CONTENT_FAKE_TOOL_CALL, "content_fake_tool_call",
+              build_content_fake_tool_call_message());
+
+    // Also render with add_generation_prompt=true to show the prompt ending
+    if (message_type == input_message_type::ALL) {
+        LOG_ERR("\n\n=== Generation Prompt Scenarios (add_generation_prompt=true) ===\n");
+
+        json prompt_messages = json::array({ user_msg });
+        render_scenario(tmpl, "generation_prompt_only", prompt_messages, tools, true, enable_thinking);
+
+        // With enable_thinking toggled
+        render_scenario(tmpl, "generation_prompt_thinking_disabled", prompt_messages, tools, true, false);
+    }
+}
+
+static generation_params prepare_debug_params(const debug_options & opts, const json & tools) {
+    generation_params params;
+    params.messages         = json::array({ build_debug_user_message() });
+    params.reasoning_format = opts.enable_reasoning ? COMMON_REASONING_FORMAT_DEEPSEEK : COMMON_REASONING_FORMAT_NONE;
+    params.enable_thinking  = opts.enable_reasoning;
+    params.add_generation_prompt = opts.generation_prompt;
+
+    if (opts.with_tools) {
+        params.tools       = tools;
+        params.tool_choice = opts.force_tool_call ? COMMON_CHAT_TOOL_CHOICE_REQUIRED : COMMON_CHAT_TOOL_CHOICE_AUTO;
+    } else {
+        params.tools       = json();
+        params.tool_choice = COMMON_CHAT_TOOL_CHOICE_NONE;
+    }
+    params.parallel_tool_calls = opts.parallel_tool_calls;
+    return params;
+}
+
+static int debug_single_template(const debug_options & opts) {
+    std::string template_source;
+    try {
+        // Check if the file is a GGUF file
+        if (opts.template_path.size() >= 5 &&
+            opts.template_path.compare(opts.template_path.size() - 5, 5, ".gguf") == 0) {
+            template_source = read_gguf_chat_template(opts.template_path);
+        } else {
+            template_source = read_file(opts.template_path);
+        }
+    } catch (const std::exception & e) {
+        LOG_ERR("Error reading template: %s\n", e.what());
+        return 1;
+    }
+
+    LOG_ERR("Analyzing template: %s\n", opts.template_path.c_str());
+    LOG_ERR("Options: with_tools=%s, generation_prompt=%s, enable_reasoning=%s\n", opts.with_tools ? "true" : "false",
+            opts.generation_prompt ? "true" : "false", opts.enable_reasoning ? "true" : "false");
+
+    try {
+        common_chat_template chat_template(template_source, "", "");
+
+        json tools = opts.with_tools ? build_tools_definition() : json();
+
+        generation_params params = prepare_debug_params(opts, tools);
+        common_chat_params            parser_data;
+        if (std::optional<common_chat_params> spec_tmpl =
+                common_chat_try_specialized_template(chat_template, template_source, params)) {
+            LOG_ERR("\n");
+            LOG_ERR("This template uses a specialized parser, analysis results will not be available.\n");
+            parser_data = *spec_tmpl;
+        } else {
+            // Render template scenarios if requested
+            if (opts.input_message != input_message_type::NONE &&
+                (opts.mode == output_mode::TEMPLATE || opts.mode == output_mode::BOTH)) {
+                LOG_ERR("\n");
+                LOG_ERR("================================================================================\n");
+                LOG_ERR("                         TEMPLATE RENDERING OUTPUT\n");
+                LOG_ERR("================================================================================\n");
+
+                render_all_scenarios(chat_template, tools, opts.generation_prompt, opts.enable_reasoning,
+                                     opts.input_message);
+            }
+
+            // Output analysis if requested
+            if (opts.mode == output_mode::ANALYSIS || opts.mode == output_mode::BOTH) {
+                LOG_ERR("\n");
+                LOG_ERR("================================================================================\n");
+                LOG_ERR("                           TEMPLATE ANALYSIS\n");
+                LOG_ERR("================================================================================\n");
+
+                struct autoparser analysis;
+                analysis.analyze_template(chat_template);
+
+                // Generate Parser
+                parser_data = peg_generator::generate_parser(chat_template, params, analysis);
+            }
+        }
+
+        if (!std::empty(parser_data.parser)) {
+            LOG_ERR("\n=== Generated Parser ===\n");
+            common_peg_arena arena;
+            arena.load(parser_data.parser);
+            LOG_ERR("%s\n", arena.dump(arena.root()).c_str());
+
+            LOG_ERR("\n=== Generated Grammar ===\n");
+            LOG_ERR("%s\n", parser_data.grammar.c_str());
+
+            LOG_ERR("\n=== Generated Lazy Grammar ===\n");
+            LOG_ERR("%d\n", parser_data.grammar_lazy);
+
+            LOG_ERR("\n=== Generated Grammar Triggers ===\n");
+            for (const common_grammar_trigger & cgt : parser_data.grammar_triggers) {
+                LOG_ERR("Token: %d | Type: %d | Value: %s\n", cgt.token, cgt.type, cgt.value.c_str());
+            }
+
+            LOG_ERR("\n=== Preserved Tokens ===\n");
+            for (const std::string & token : parser_data.preserved_tokens) {
+                LOG_ERR("  '%s'\n", token.c_str());
+            }
+        }
+    } catch (const std::exception & e) {
+        LOG_ERR("Analysis failed: %s\n", e.what());
+        return 1;
+    }
+
+    return 0;
+}
+
 int main(int argc, char * argv[]) {
+    if (argc > 1) {
+        std::string arg = argv[1];
+        if (arg == "-h" || arg == "--help") {
+            common_log_set_verbosity_thold(99);
+            print_usage(argv[0]);
+            return 0;
+        }
+
+        // debug mode: if the first argument is an existing file, analyze that template instead of running the automated tests
+        if (std::filesystem::is_regular_file(arg)) {
+            common_log_set_verbosity_thold(99);
+
+            debug_options opts;
+            if (!parse_debug_options(argc, argv, opts)) {
+                return 1;
+            }
+
+            if (opts.debug_jinja || std::getenv("LLAMA_DEBUG_JINJA") != nullptr) {
+                jinja::enable_debug(true);
+            }
+
+            return debug_single_template(opts);
+        }
+    }
+
     testing t(std::cout);
     t.verbose = true;
 
-    // usage: test-chat-auto-parser-helpers [filter_regex]
+    // usage: test-chat-auto-parser [filter_regex]
 
     if (argc > 1) {
         t.set_filter(argv[1]);

tests/test-chat-template.cpp

@@ -28,6 +28,8 @@ static void run_multiple(const std::string& dir_path, bool stop_on_first_failure
 static void run_single(const std::string& contents, json input, bool use_common = false, bool dump_prog = false, const std::string & output_path = "");
 
 static std::string HELP = R"(
+Test the Jinja engine by rendering chat templates and comparing the output against expected results.
+
 Usage: test-chat-template [OPTIONS] PATH_TO_TEMPLATE
 Options:
   -h, --help               Show this help message and exit.

tests/test-llama-archs.cpp

@@ -101,6 +101,16 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
         n_head = 1;
         n_ff   = 96;
         n_layer = 22; // hparams.n_layer_kv_from_start = 20 is hardcoded
+    } else if (arch == LLM_ARCH_DEEPSEEK4) {
+        // head size 64 so that GPU flash attention kernels support the model
+        n_embd  = 512;
+        n_head  = 8;
+        n_ff    = 1024;
+        n_layer = 4;
+    } else if (arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_LAGUNA) {
+        n_embd = 160; // exercise per-head tensor split granularity with head size 80
+    } else if (arch == LLM_ARCH_QWEN3 || arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_AFMOE) {
+        n_head = 4;
     } else if (arch == LLM_ARCH_DEEPSEEK2
             || arch == LLM_ARCH_DEEPSEEK32
             || arch == LLM_ARCH_GLM_DSA
@@ -120,6 +130,12 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
         n_vocab = 4096; // must be >= the hard-coded codec head size (3072)
     }
 
+    uint32_t n_head_kv = n_head;
+    if (arch == LLM_ARCH_QWEN3) {
+        n_head_kv = 1; // MQA coverage
+    } else if (arch == LLM_ARCH_MUSE_GLIMMER || arch == LLM_ARCH_AFMOE) {
+        n_head_kv = 2; // GQA coverage
+    }
     const uint32_t n_embd_head = n_embd / n_head;
 
     ms.add_kv(LLM_KV_GENERAL_ARCHITECTURE,      llm_arch_name(arch));
@@ -160,11 +176,15 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
         ms.add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV, n_head_per_layer);
     } else {
         ms.add_kv(LLM_KV_ATTENTION_HEAD_COUNT, n_head);
-        ms.add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV, n_head);
+        ms.add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV, arch == LLM_ARCH_DEEPSEEK4 ? uint32_t(1) : n_head_kv);
     }
 
     ms.add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, 8.0f);
-    if (arch == LLM_ARCH_DEEPSEEK2
+    if (arch == LLM_ARCH_DEEPSEEK4) {
+        ms.add_kv(LLM_KV_ATTENTION_KEY_LENGTH,   n_embd_head);
+        ms.add_kv(LLM_KV_ATTENTION_VALUE_LENGTH, n_embd_head);
+        ms.add_kv(LLM_KV_ROPE_DIMENSION_COUNT,   n_embd_head/2);
+    } else if (arch == LLM_ARCH_DEEPSEEK2
             || arch == LLM_ARCH_DEEPSEEK32
             || arch == LLM_ARCH_GLM_DSA
             || arch == LLM_ARCH_DOTS3NOTE
@@ -202,7 +222,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
     ms.add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      1e-5f);
     ms.add_kv(LLM_KV_ATTENTION_GROUPNORM_EPS,          1e-5f);
     ms.add_kv(LLM_KV_ATTENTION_GROUPNORM_GROUPS,       uint32_t(8));
-    ms.add_kv(LLM_KV_ATTENTION_Q_LORA_RANK,            uint32_t(512));
+    ms.add_kv(LLM_KV_ATTENTION_Q_LORA_RANK,            arch == LLM_ARCH_DEEPSEEK4 ? uint32_t(64) : uint32_t(512));
     ms.add_kv(LLM_KV_ATTENTION_KV_LORA_RANK,           uint32_t(512));
     ms.add_kv(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, uint32_t(8));
     ms.add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW,         n_ctx/8);
@@ -229,12 +249,26 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
 
     // MSA requires one indexer head per GQA (KV) head, unlike the DSA archs where the
     // indexer head count is independent of the main attention head count.
-    ms.add_kv(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,   arch == LLM_ARCH_MINIMAX_M3 ? n_head : uint32_t(1));
+    ms.add_kv(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT,   arch == LLM_ARCH_MINIMAX_M3 || arch == LLM_ARCH_DEEPSEEK4 ? n_head : uint32_t(1));
     ms.add_kv(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH,   uint32_t(64));
     ms.add_kv(LLM_KV_ATTENTION_INDEXER_TOP_K,        uint32_t(8));
     ms.add_kv(LLM_KV_ATTENTION_INDEXER_BLOCK_SIZE,   uint32_t(4));
     ms.add_kv(LLM_KV_ATTENTION_INDEXER_LOCAL_BLOCKS, uint32_t(1));
     ms.add_kv(LLM_KV_ROPE_DIMENSION_SECTIONS, std::vector<uint32_t>({n_embd_head/4, n_embd_head/4, n_embd_head/4, n_embd_head/4}));
+
+    if (arch == LLM_ARCH_DEEPSEEK4) {
+        ms.add_kv(LLM_KV_ATTENTION_OUTPUT_GROUP_COUNT,         uint32_t(8));
+        ms.add_kv(LLM_KV_ATTENTION_OUTPUT_LORA_RANK,           uint32_t(32));
+        ms.add_kv(LLM_KV_ATTENTION_COMPRESS_RATIOS,            std::vector<uint32_t>({0, 0, 4, 128}));
+        ms.add_kv(LLM_KV_ATTENTION_COMPRESS_ROPE_FREQ_BASE,    160000.0f);
+        ms.add_kv(LLM_KV_HYPER_CONNECTION_COUNT,               uint32_t(4));
+        ms.add_kv(LLM_KV_HYPER_CONNECTION_SINKHORN_ITERATIONS, uint32_t(2));
+        ms.add_kv(LLM_KV_HYPER_CONNECTION_EPSILON,             1.0e-6f);
+        ms.add_kv(LLM_KV_HASH_LAYER_COUNT,                      uint32_t(0));
+        ms.add_kv(LLM_KV_SWIGLU_CLAMP_EXP,                      10.0f);
+        ms.add_kv(LLM_KV_EXPERT_WEIGHTS_SCALE,                  1.0f);
+        ms.add_kv(LLM_KV_EXPERT_WEIGHTS_NORM,                   true);
+    }
     ms.add_kv(LLM_KV_TOKENIZER_MODEL,         "no_vocab");
     // ms.add_kv(LLM_KV_DENSE_2_FEAT_OUT,     n_embd);
     // ms.add_kv(LLM_KV_DENSE_3_FEAT_IN,      n_embd);
@@ -247,7 +281,7 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
         ms.add_kv(LLM_KV_EXPERT_COUNT,               uint32_t(2));
         ms.add_kv(LLM_KV_EXPERT_USED_COUNT,          uint32_t(1));
         ms.add_kv(LLM_KV_EXPERT_SHARED_COUNT,        uint32_t(1));
-        ms.add_kv(LLM_KV_EXPERT_GATING_FUNC,         uint32_t(2)); // sigmoid
+        ms.add_kv(LLM_KV_EXPERT_GATING_FUNC,         arch == LLM_ARCH_DEEPSEEK4 ? uint32_t(4) : uint32_t(2)); // sqrtsoftplus : sigmoid
         ms.add_kv(LLM_KV_EXPERT_GROUP_SCALE,         1.0f);
         ms.add_kv(LLM_KV_EXPERTS_PER_GROUP,          uint32_t(1));
     }
@@ -385,6 +419,7 @@ static bool moe_mandatory(const llm_arch arch) {
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_DEEPSEEK32:
         case LLM_ARCH_DOTS3NOTE:
+        case LLM_ARCH_DEEPSEEK4:
         case LLM_ARCH_GLM4_MOE:
         case LLM_ARCH_GLM_DSA:
         case LLM_ARCH_EXAONE_MOE:
@@ -470,13 +505,9 @@ static bool arch_supported(const llm_arch arch) {
     if (arch == LLM_ARCH_DEEPSEEK2OCR) {
         return false;
     }
-    if (arch == LLM_ARCH_DEEPSEEK4) {
-        return false;
-    }
-
     // FIXME: these hit scheduler/view-backed-output issues with WebGPU on CI.
 #ifdef GGML_USE_WEBGPU
-    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA) {
+    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_DOTS3NOTE) {
         return false;
     }
 #endif // GGML_USE_WEBGPU

tests/test-recurrent-state-rollback.cpp

@@ -35,6 +35,178 @@ static bool decode_one(llama_context * ctx, llama_token tok, llama_pos pos) {
     return ok;
 }
 
+// Roll back multiple sequences, then replay them in a single batch whose
+// per-seq token count exceeds n_ubatch: each seq's replay spans several
+// ubatches while its rollback restore is still pending. Compared against a
+// reference context that never advanced past the rollback point and decodes
+// the identical replay batch.
+static bool test_multi_seq_split_replay(const common_params & params, llama_model * model, const int n_vocab) {
+    constexpr uint32_t  n_seqs     = 2;
+    constexpr uint32_t  n_ubatch   = 16;
+    constexpr uint32_t  n_prompt   = 19;
+    constexpr uint32_t  n_rollback = 3;
+    constexpr uint32_t  n_replay   = 40; // > n_ubatch so each seq spans multiple ubatches
+    constexpr llama_pos p0         = n_prompt - n_rollback;
+
+    const auto make_ctx_multi = [&]() {
+        auto cparams = common_context_params_to_llama(params);
+        cparams.n_seq_max  = n_seqs;
+        cparams.n_rs_seq   = 8;
+        cparams.n_ctx      = 256;
+        cparams.n_batch    = 256;
+        cparams.n_ubatch   = n_ubatch;
+        cparams.kv_unified = false;
+        return llama_init_from_model(model, cparams);
+    };
+
+    llama_context * ctx_roll = make_ctx_multi();
+    llama_context * ctx_ref  = make_ctx_multi();
+    if (ctx_roll == nullptr || ctx_ref == nullptr) {
+        fprintf(stderr, "%s : failed to init multi-seq contexts\n", __func__);
+        return false;
+    }
+
+    const auto cleanup = [&]() {
+        llama_free(ctx_roll);
+        llama_free(ctx_ref);
+    };
+
+    if (llama_n_rs_seq(ctx_roll) < n_rollback) {
+        fprintf(stderr, "%s : skipping because n_rs_seq is too small\n", __func__);
+        cleanup();
+        return true;
+    }
+
+    const auto tok = [&](uint32_t seq, llama_pos pos) {
+        return (llama_token) ((7*(uint32_t) pos + 31*seq + 1) % (uint32_t) n_vocab);
+    };
+
+    bool ok = true;
+
+    // both contexts decode the identical [0, p0) prefill; only ctx_roll decodes
+    // the tail, which is then rolled back so its restore is pending at replay
+    for (uint32_t s = 0; s < n_seqs && ok; ++s) {
+        llama_batch batch = llama_batch_init(n_prompt, 0, 1);
+        for (llama_pos pos = 0; pos < (llama_pos) p0; ++pos) {
+            common_batch_add(batch, tok(s, pos), pos, { (llama_seq_id) s }, false);
+        }
+        ok = ok && llama_decode(ctx_roll, batch) == 0;
+        ok = ok && llama_decode(ctx_ref,  batch) == 0;
+
+        common_batch_clear(batch);
+        for (llama_pos pos = p0; pos < (llama_pos) n_prompt; ++pos) {
+            common_batch_add(batch, tok(s, pos), pos, { (llama_seq_id) s }, false);
+        }
+        ok = ok && llama_decode(ctx_roll, batch) == 0;
+        llama_batch_free(batch);
+
+        ok = ok && llama_memory_seq_rm(llama_get_memory(ctx_roll), (llama_seq_id) s, p0, -1);
+
+        // a second partial removal while one is pending must be refused
+        ok = ok && !llama_memory_seq_rm(llama_get_memory(ctx_roll), (llama_seq_id) s, p0 - 1, -1);
+    }
+    if (!ok) {
+        fprintf(stderr, "%s : multi-seq prefill/rollback failed\n", __func__);
+        cleanup();
+        return false;
+    }
+
+    llama_batch batch = llama_batch_init(n_seqs*n_replay, 0, 1);
+    for (uint32_t s = 0; s < n_seqs; ++s) {
+        for (uint32_t i = 0; i < n_replay; ++i) {
+            const llama_pos pos = p0 + (llama_pos) i;
+            common_batch_add(batch, tok(s, pos), pos, { (llama_seq_id) s }, true);
+        }
+    }
+    ok = llama_decode(ctx_roll, batch) == 0;
+    ok = ok && llama_decode(ctx_ref, batch) == 0;
+    llama_batch_free(batch);
+    if (!ok) {
+        fprintf(stderr, "%s : multi-seq replay decode failed\n", __func__);
+        cleanup();
+        return false;
+    }
+
+    // identical ubatch shapes from bit-exact states: a correct implementation
+    // matches bitwise, so eps only allows backend scheduling noise
+    constexpr float eps = 1e-7f;
+
+    float    diff_max  = 0.0f;
+    uint32_t seq_first = 0;
+    int32_t  pos_first = -1;
+    for (uint32_t i = 0; i < n_seqs*n_replay; ++i) {
+        const float * l_roll = llama_get_logits_ith(ctx_roll, i);
+        const float * l_ref  = llama_get_logits_ith(ctx_ref,  i);
+        if (l_roll == nullptr || l_ref == nullptr) {
+            fprintf(stderr, "%s : missing multi-seq logits at index %u\n", __func__, i);
+            cleanup();
+            return false;
+        }
+        for (int t = 0; t < n_vocab; ++t) {
+            const float diff = std::fabs(l_roll[t] - l_ref[t]);
+            if (diff > eps && pos_first < 0) {
+                seq_first = i/n_replay;
+                pos_first = p0 + (int32_t) (i%n_replay);
+            }
+            diff_max = std::max(diff_max, diff);
+        }
+    }
+
+    if (diff_max > eps) {
+        fprintf(stderr, "%s : multi-seq split replay logits mismatch (max diff %g, first at seq %u pos %d)\n",
+                __func__, (double) diff_max, seq_first, pos_first);
+        cleanup();
+        return false;
+    }
+
+    fprintf(stderr, "%s : multi-seq split replay matched (max diff %g)\n", __func__, (double) diff_max);
+
+    // seq-1-only decodes must be independent of seq 0's content: diverge seq 0
+    // in ctx_ref only, then compare identical seq-1-only continuations bitwise
+    constexpr uint32_t n_tail = 4;
+
+    {
+        llama_batch batch_tail = llama_batch_init(n_tail, 0, 1);
+        for (uint32_t i = 0; i < n_tail; ++i) {
+            const llama_pos pos = p0 + (llama_pos) (n_replay + i);
+            common_batch_add(batch_tail, tok(0, pos + 7), pos, { 0 }, false);
+        }
+        ok = llama_decode(ctx_ref, batch_tail) == 0;
+        llama_batch_free(batch_tail);
+    }
+
+    float diff_tail = 0.0f;
+    for (uint32_t i = 0; i < n_tail && ok; ++i) {
+        const llama_pos pos = p0 + (llama_pos) (n_replay + i);
+        llama_batch batch_one = llama_batch_init(1, 0, 1);
+        common_batch_add(batch_one, tok(1, pos), pos, { 1 }, true);
+        ok = llama_decode(ctx_roll, batch_one) == 0;
+        ok = ok && llama_decode(ctx_ref, batch_one) == 0;
+        llama_batch_free(batch_one);
+        if (!ok) {
+            break;
+        }
+
+        const float * l_roll = llama_get_logits_ith(ctx_roll, 0);
+        const float * l_ref  = llama_get_logits_ith(ctx_ref,  0);
+        ok = l_roll != nullptr && l_ref != nullptr;
+        for (int t = 0; ok && t < n_vocab; ++t) {
+            diff_tail = std::max(diff_tail, std::fabs(l_roll[t] - l_ref[t]));
+        }
+    }
+
+    if (!ok || diff_tail > eps) {
+        fprintf(stderr, "%s : seq-1-only decode leaked seq 0 state (ok=%d, max diff %g)\n",
+                __func__, ok ? 1 : 0, (double) diff_tail);
+        cleanup();
+        return false;
+    }
+
+    fprintf(stderr, "%s : seq-1-only decode independent of seq 0 (max diff %g)\n", __func__, (double) diff_tail);
+    cleanup();
+    return true;
+}
+
 int main(int argc, char ** argv) {
     std::setlocale(LC_NUMERIC, "C");
 
@@ -220,5 +392,10 @@ int main(int argc, char ** argv) {
     llama_free(ctx_src);
     llama_free(ctx_dst);
     llama_free(ctx_dirty);
+
+    if (!test_multi_seq_split_replay(params, model, n_vocab)) {
+        return 1;
+    }
+
     return 0;
 }