sync : ggml

* hmx: extract shared interleave headers and unify matmul batched

* hmx: add HMX-accelerated flash attention for prefill

* hmx: replace asm wrappers with Q6_ intrinsics in hmx-utils.h

Switches three single-instruction helpers from inline asm to the matching
Q6_ intrinsics, matching the style established by aizip f8737609a and used
by the upstream PR #21554 hmx-matmul-ops.c rewrite:

  hmx_set_output_scales       asm "bias=mxmem2"  -> Q6_bias_mxmem2_A
  hmx_load_tile_pair_fp16     asm packet         -> Q6_activation_hf_mxmem_RR
                                                    + Q6_weight_hf_mxmem_RR
  hmx_consume_accumulator_fp16 asm "mxmem=acc"   -> Q6_mxmem_AR_after_hf

hmx_load_tiles_fp16 stays on inline asm: it uses ":deep" activation
streaming, and the mixed Q6_activation_hf_mxmem_RR_deep + non-deep
Q6_weight_hf_mxmem_RR pair fails the HMX backend constraint check
("activate weight pair (1) exceeds limit (1)"). The asm bundle keeps
both halves in one VLIW packet and avoids the diagnostic.

Functionally equivalent — same instructions emitted; the Q6_ intrinsics
just give the compiler more visibility for scheduling.

* hmx: drop the duplicate interleave_fp16_weight_chunk_to_tiles

* hmx:  apply upstream optimization to hmx-flash-attn-ops.c
apply restrict, __builtin_assume, and pointer accumulation to the three HMX workers (qk_dot, o_update, o_norm) and the matching inline HMX loops in op_hmx_flash_attn_ext.

* hmx: unify interleave helper

* hmx: multi-thread Q load / O store and enable prefill FA dispatch

Extract inline Q-load and O-store loops into worker_pool-parallel helpers
(fa_phase_q_load, fa_phase_o_store) so HVX threads split the F32↔F16
conversion work across row ranges.  Also relax the softmax threading
gate from n_row_vec_cnt >= n_threads to >= 2, which was unnecessarily
forcing single-thread fallback when n_rows_g < 512.

On the dispatch side, remove the ne[2] != 1 guard that blocked multi-head
(prefill) FA from reaching the HTP backend — GQA is already handled
internally by both the HMX and HVX flash-attention paths.

* hmx: relax matmul pipeline gate to cover k > n shapes (e.g. FFN_down)

* hmx: optimize FA softmax mask phase (no-ALiBi fast path + GQA dedup)

* hmx: Add an asm memory clobber at the phase boundary to prevent reorder bug

* [experimental]: fp16 softmax (EXP2_HF) to accelerate fa

Bake log2(e) into qk_scale and use hvx_exp2_hf directly for P and m_diff
(base-2 consistent, matches htp-ops-lib). ~22 ALU ops for 64 lanes vs
~44 for the F32 round-trip path.

* hmx flash-attn: refine cost model coefficients based on profiling data

* hmx flash-attn: replace asm clobber with targeted volatile reads on vtcm_d_tiles

* hmx flash-attn: fix prefill correctness (dst indexing, softmax reduce, V stride)

* hmx flash-attn: fix p_tiles dual-tile OOB race; enable MT + pipeline

* hmx flash-attn: preserve additive mask bias in no-ALiBi fast path

The no-ALiBi fast path (max_bias==0) was skipping mask add entirely on
the assumption that mask values are only {0, -inf}.  This is wrong when
the mask carries additive positional bias — those terms were silently
dropped.  Keep the slope-mul skip (slope≡1.0) but add mask back so the
bias survives; vmux still clamps below -16 to -inf.

Also add HMX FA coverage to test-backend-ops: prefill shapes (nb=64,
nb=32) × {mask on/off} × {ALiBi on/off} × {softcap on/off}, F16 KV,
hs ∈ {64, 128}.

* hmx: fix softcap+EXP2_HF interaction, tighten matmul pipeline gate, add FA tests

- flash-attn: when EXP2_HF is on AND logit_softcap is active, fold
  log2(e) into the post-tanh multiplier (v_cap) instead of pre-baking
  it into qk_scale.  Pre-baking shifted the tanh knee from x≈c to
  x≈c/log2(e) and produced numerically wrong softcapped outputs
  whenever both knobs were enabled.
- flash-attn softmax (fa_softmax_thread): replace the union+memcpy
  scalar extract pattern with HVX vmux-based per-row accumulators on
  rowmax/rowsum.  Add hvx_vec_get_f16 helper in hvx-base.h.  Functional
  parity, less scalar code, clearer hf/qf16 lane-format contract.
- matmul (hmx_mat_mul_permuted_qk_0_d16a32): pick pipeline vs sequential
  layout based on whether the chunker actually yields >=2 n-chunks,
  instead of the static (m>=128 && n>=256) gate.  Avoids paying for
  output double-buffer + worker dispatch when there is no HMX/HVX
  overlap to gain (e.g. shapes that collapse to one n-chunk).
- tests: add HMX flash-attention coverage over the
  {mask, ALiBi (max_bias), logit_softcap} cross-product for the prefill
  path — head_dim 64/128, GQA 4×4, kv=512/nb=64 plus a kv=113/nb=32
  non-aligned case.

* [Help Wanted]: refactor D matrix computation into separate function for clarity and maintainability

* format code

* hexagon: looks like -O3 is causing issues with the large code base, switch to -O2 and -flto instead

* hexagon: use hex_ prefix for swap_ptr

* hexagon: move vtcm_seq_alloc into vtcm-utils.h

More vtcm allocator updates are coming so it makes sense to start the separate hdr for it.

* hmx-utils: add hmx_prefix for layout converters

* hmx-mm: move main hmx_mm functions to the end, remove unused fwd decls, etc

* hmx-mm: remove unused qweight_fetch_task_state_t and minor alignment fixes

* hmx-fa: minor alignment fixes

* hmx-fa: move hmx_flash_atten into hmx-ops.h

* hmx-fa: remove redundant workpool pointer in the hmx_fa_ctx, plus minor alignment updates

* hmx-fa: minor alignment and simplifications

* hexagon: move FA_EXP_F16 option to hostside CMake file

* hmx-fa: use hvx_vec_splat_f16 instead of fp16_to_bits

* hmx-fa: add hvx_splat_u16/u8 and use that in the fa instead custom hvx_fill

* hmx-fa: some more alignment updates in the core fa function

* hmx-fa: keep slopes in vtcm in fp16

Saves malloc/free and removes the need for float -> fp16 downcast on every use.

* hexagon: consistent noinline usage (after static)

* hex-hmx: consistent use FARF_HIGH to enable debug output

* hmx-utils: no need for always_inline attr

* hex-hmx: consistent noinline usage (static noinline ...)

* hex-hmx: simplify init_col_scales

* hexagon: fix editorconfig errors

* hmx-mm: minor alignment fixes

---------

Co-authored-by: Max Krasnyansky <maxk@qti.qualcomm.com>

.devops/intel.Dockerfile

@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2025.3.2-0-devel-ubuntu24.04
+ARG ONEAPI_VERSION=2025.3.3-0-devel-ubuntu24.04
 
 ## Build Image
 

.devops/openvino.Dockerfile

@@ -2,7 +2,19 @@ ARG OPENVINO_VERSION_MAJOR=2026.0
 ARG OPENVINO_VERSION_FULL=2026.0.0.20965.c6d6a13a886
 ARG UBUNTU_VERSION=24.04
 
-# Optional proxy build arguments - empty by default
+# Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
+ARG IGC_VERSION=v2.30.1
+ARG IGC_VERSION_FULL=2_2.30.1+20950
+ARG COMPUTE_RUNTIME_VERSION=26.09.37435.1
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.09.37435.1-0
+ARG IGDGMM_VERSION=22.9.0
+
+# Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
+ARG NPU_DRIVER_VERSION=v1.32.0
+ARG NPU_DRIVER_FULL=v1.32.0.20260402-23905121947
+ARG LIBZE1_VERSION=1.27.0-1~24.04~ppa2
+
+# Optional proxy build arguments
 ARG http_proxy=
 ARG https_proxy=
 
@@ -78,13 +90,47 @@ ARG http_proxy
 ARG https_proxy
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 libtbb12 curl \
+    && apt-get install -y libgomp1 libtbb12 curl wget ocl-icd-libopencl1 \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \
     && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
     && find /var/cache -type f -delete
 
+# Install GPU drivers
+ARG IGC_VERSION
+ARG IGC_VERSION_FULL
+ARG COMPUTE_RUNTIME_VERSION
+ARG COMPUTE_RUNTIME_VERSION_FULL
+ARG IGDGMM_VERSION
+RUN mkdir /tmp/neo/ && cd /tmp/neo/ \
+    && wget https://github.com/intel/intel-graphics-compiler/releases/download/${IGC_VERSION}/intel-igc-core-${IGC_VERSION_FULL}_amd64.deb \
+    && wget https://github.com/intel/intel-graphics-compiler/releases/download/${IGC_VERSION}/intel-igc-opencl-${IGC_VERSION_FULL}_amd64.deb \
+    && wget https://github.com/intel/compute-runtime/releases/download/${COMPUTE_RUNTIME_VERSION}/intel-ocloc-dbgsym_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.ddeb \
+    && wget https://github.com/intel/compute-runtime/releases/download/${COMPUTE_RUNTIME_VERSION}/intel-ocloc_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.deb \
+    && wget https://github.com/intel/compute-runtime/releases/download/${COMPUTE_RUNTIME_VERSION}/intel-opencl-icd-dbgsym_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.ddeb \
+    && wget https://github.com/intel/compute-runtime/releases/download/${COMPUTE_RUNTIME_VERSION}/intel-opencl-icd_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.deb \
+    && wget https://github.com/intel/compute-runtime/releases/download/${COMPUTE_RUNTIME_VERSION}/libigdgmm12_${IGDGMM_VERSION}_amd64.deb \
+    && wget https://github.com/intel/compute-runtime/releases/download/${COMPUTE_RUNTIME_VERSION}/libze-intel-gpu1-dbgsym_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.ddeb \
+    && wget https://github.com/intel/compute-runtime/releases/download/${COMPUTE_RUNTIME_VERSION}/libze-intel-gpu1_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.deb \
+    && dpkg --install *.deb \
+    && rm -rf /tmp/neo/
+
+# Install NPU drivers
+ARG NPU_DRIVER_VERSION
+ARG NPU_DRIVER_FULL
+ARG LIBZE1_VERSION
+RUN mkdir /tmp/npu/ && cd /tmp/npu/ \
+    && wget https://github.com/intel/linux-npu-driver/releases/download/${NPU_DRIVER_VERSION}/linux-npu-driver-${NPU_DRIVER_FULL}-ubuntu2404.tar.gz \
+    && tar -xf linux-npu-driver-${NPU_DRIVER_FULL}-ubuntu2404.tar.gz \
+    && dpkg --install *.deb \
+    && rm -rf /tmp/npu/
+
+RUN cd /tmp \
+    && wget https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260324T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb \
+    && dpkg --install libze1_${LIBZE1_VERSION}_amd64.deb \
+    && rm libze1_${LIBZE1_VERSION}_amd64.deb
+
 COPY --from=build /app/lib/ /app/
 
 ### Full (all binaries)

.github/pull_request_template.md

@@ -6,7 +6,7 @@
 
 <!-- You can provide more details and link related discussions here. Delete this section if not applicable -->
 
-# Requirements
+## Requirements
 
 <!-- IMPORTANT: Please do NOT delete this section, otherwise your PR may be rejected -->
 

.github/workflows/build-and-test-snapdragon.yml

@@ -0,0 +1,116 @@
+name: CI (snapdragon)
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - master
+    paths:
+      - '.github/workflows/build-and-test-snapdragon.yml'
+      - 'ggml/include/ggml-hexagon.h'
+      - 'ggml/src/ggml-hexagon/**'
+      - 'docs/backend/snapdragon/**'
+      - 'scripts/snapdragon/**'
+      - 'CMakePresets.json'
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths:
+      - '.github/workflows/build-and-test-snapdragon.yml'
+      - 'ggml/include/ggml-hexagon.h'
+      - 'ggml/src/ggml-hexagon/**'
+      - 'docs/backend/snapdragon/**'
+      - 'scripts/snapdragon/**'
+      - 'CMakePresets.json'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  android-ndk-snapdragon:
+    runs-on: ubuntu-latest
+    container:
+      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.3'
+    defaults:
+      run:
+        shell: bash
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          lfs: false
+
+      - name: Build Llama.CPP for Snapdragon Android
+        id: build_llama_cpp_snapdragon_android
+        run: |
+          cp docs/backend/snapdragon/CMakeUserPresets.json .
+          cmake --preset arm64-android-snapdragon-release -B build
+          cmake --build build
+          cmake --install build --prefix pkg-snapdragon/llama.cpp
+
+      - name: Upload Llama.CPP Snapdragon Android Build Artifact
+        if: ${{ always() && steps.build_llama_cpp_snapdragon_android.outcome == 'success' }}
+        uses: actions/upload-artifact@v6
+        with:
+          name: llama-cpp-android-arm64-snapdragon
+          path: pkg-snapdragon/llama.cpp
+
+  test-snapdragon-qdc:
+    name: Test on QDC Android Device (${{ matrix.device }})
+    needs: [android-ndk-snapdragon]
+    runs-on: ubuntu-slim
+    strategy:
+      fail-fast: false
+      matrix:
+        device: [SM8750, SM8650, SM8850]
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+
+      - name: Download build artifact
+        uses: actions/download-artifact@v7
+        with:
+          name: llama-cpp-android-arm64-snapdragon
+          path: pkg-snapdragon/llama.cpp
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.x'
+          cache: pip
+
+      - name: Install system dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y curl unzip
+
+      - name: Install QDC SDK wheel
+        run: |
+          curl -fSL -o qdc_sdk.zip https://softwarecenter.qualcomm.com/api/download/software/tools/Qualcomm_Device_Cloud_SDK/All/0.2.3/qualcomm_device_cloud_sdk-0.2.3.zip
+          unzip qdc_sdk.zip -d qdc_sdk
+          pip install qdc_sdk/qualcomm_device_cloud_sdk-0.2.3-py3-none-any.whl
+
+      - name: Check QDC API key
+        id: check_secret
+        env:
+          QDC_API_KEY: ${{ secrets.QDC_API_KEY }}
+        run: echo "has-qdc-key=${{ env.QDC_API_KEY != '' }}" >> "$GITHUB_OUTPUT"
+
+      - name: Run QDC tests (${{ matrix.device }})
+        if: steps.check_secret.outputs.has-qdc-key == 'true'
+        run: |
+          python scripts/snapdragon/qdc/run_qdc_jobs.py \
+              --test       all \
+              --pkg-dir    pkg-snapdragon/llama.cpp \
+              --model-url  "https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf" \
+              --device     ${{ matrix.device }}
+        env:
+          QDC_API_KEY: ${{ secrets.QDC_API_KEY }}
+
+      - name: Cleanup
+        if: always()
+        run: rm -rf pkg-snapdragon qdc_sdk qdc_sdk.zip

.github/workflows/build-android.yml

@@ -1,26 +1,24 @@
 name: CI (android)
 
 on:
-  workflow_dispatch: # allows manual triggering
+  workflow_dispatch:
   push:
     branches:
       - master
-    paths: [
-      '.github/workflows/build-android.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp'
-    ]
+    paths:
+      - '.github/workflows/build-android.yml'
+      - '**/CMakeLists.txt'
+      - '**/.cmake'
+      - '**/*.h'
+      - '**/*.hpp'
+      - '**/*.c'
+      - '**/*.cpp'
 
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-android.yml',
-      'examples/llama.android/**'
-    ]
+    paths:
+      - '.github/workflows/build-android.yml'
+      - 'examples/llama.android/**'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -67,35 +65,24 @@ jobs:
     defaults:
       run:
         shell: bash
-    strategy:
-      matrix:
-        include:
-          - build: 'arm64-cpu'
-            defines: '-D ANDROID_ABI=arm64-v8a -D ANDROID_PLATFORM=android-31 -D CMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -D GGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm -G Ninja -D LLAMA_OPENSSL=OFF -D GGML_OPENMP=OFF'
-          - build: 'arm64-snapdragon'
-            defines: '--preset arm64-android-snapdragon-release'
 
     steps:
       - name: Clone
-        id: checkout
         uses: actions/checkout@v6
         with:
           fetch-depth: 0
           lfs: false
 
-      - name: Build Llama.CPP for Hexagon Android
-        id: build_llama_cpp_hexagon_android
+      - name: Build
+        id: ndk_build
         run: |
-          if [[ "${{ matrix.build }}" == "arm64-snapdragon" ]]; then
-            cp docs/backend/snapdragon/CMakeUserPresets.json .
-          fi
-          cmake ${{ matrix.defines }} -B build
+          cmake -D ANDROID_ABI=arm64-v8a -D ANDROID_PLATFORM=android-31 -D CMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -D GGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm -G Ninja -D LLAMA_OPENSSL=OFF -D GGML_OPENMP=OFF -B build
           cmake --build build
           cmake --install build --prefix pkg-adb/llama.cpp
 
-      - name: Upload Llama.CPP Hexagon Android Build Artifact
-        if: ${{ always() && steps.build_llama_cpp_hexagon_android.outcome == 'success' }}
+      - name: Upload Android Build Artifact
+        if: ${{ always() && steps.ndk_build.outcome == 'success' }}
         uses: actions/upload-artifact@v6
         with:
-          name: llama-cpp-android-${{ matrix.build }}
+          name: llama-cpp-android-arm64-cpu
           path: pkg-adb/llama.cpp

.github/workflows/build-cross.yml

@@ -246,6 +246,7 @@ jobs:
           apt-get install -y --no-install-recommends \
                   build-essential \
                   glslc \
+                  spirv-headers \
                   gcc-14-loongarch64-linux-gnu \
                   g++-14-loongarch64-linux-gnu \
                   libvulkan-dev:loong64

.github/workflows/build-openvino.yml

@@ -0,0 +1,120 @@
+name: CI (openvino)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-openvino.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-openvino.yml',
+      'ggml/src/ggml-openvino/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  ubuntu-24-openvino:
+    name: ubuntu-24-openvino-${{ matrix.openvino_device }}
+
+    concurrency:
+      group: openvino-${{ matrix.variant }}-${{ github.head_ref || github.ref }}
+      cancel-in-progress: false
+
+    strategy:
+      matrix:
+        include:
+          - variant: cpu
+            runner: '"ubuntu-24.04"'
+            openvino_device: "CPU"
+          - variant: gpu
+            runner: '["self-hosted","Linux","Intel","OpenVINO"]'
+            openvino_device: "GPU"
+
+    runs-on: ${{ fromJSON(matrix.runner) }}
+
+    env:
+      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      OPENVINO_VERSION_MAJOR: "2026.0"
+      OPENVINO_VERSION_FULL: "2026.0.0.20965.c6d6a13a886"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        if: runner.environment == 'github-hosted'
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: ubuntu-24-openvino-${{ matrix.variant }}-no-preset-v1
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y build-essential libssl-dev libtbb12 cmake ninja-build python3-pip
+          sudo apt-get install -y ocl-icd-opencl-dev opencl-headers opencl-clhpp-headers intel-opencl-icd
+
+      - name: Use OpenVINO Toolkit Cache
+        if: runner.environment == 'github-hosted'
+        uses: actions/cache@v5
+        id: cache-openvino
+        with:
+          path: ./openvino_toolkit
+          key: openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
+
+      - name: Setup OpenVINO Toolkit
+        if: steps.cache-openvino.outputs.cache-hit != 'true'
+        uses: ./.github/actions/linux-setup-openvino
+        with:
+          path: ./openvino_toolkit
+          version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
+          version_full: ${{ env.OPENVINO_VERSION_FULL }}
+
+      - name: Install OpenVINO dependencies
+        run: |
+          cd ./openvino_toolkit
+          chmod +x ./install_dependencies/install_openvino_dependencies.sh
+          echo "Y" | sudo -E ./install_dependencies/install_openvino_dependencies.sh
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source ./openvino_toolkit/setupvars.sh
+          cmake -B build/ReleaseOV -G Ninja \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_OPENVINO=ON
+          time cmake --build build/ReleaseOV --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        # TODO: fix and re-enable the `test-llama-archs` test below
+        run: |
+          cd ${{ github.workspace }}
+          if [ "${{ matrix.openvino_device }}" = "GPU" ]; then
+            export GGML_OPENVINO_DEVICE=GPU
+          fi
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000

.github/workflows/build-self-hosted.yml

@@ -265,6 +265,10 @@ jobs:
   ggml-ci-intel-openvino-gpu-low-perf:
     runs-on: [self-hosted, Linux, Intel, OpenVINO]
 
+    concurrency:
+      group: openvino-gpu-${{ github.head_ref || github.ref }}
+      cancel-in-progress: false
+
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
       OPENVINO_VERSION_MAJOR: "2026.0"

.github/workflows/build-sycl.yml

@@ -0,0 +1,142 @@
+name: CI (sycl)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-sycl.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-sycl.yml',
+      'ggml/src/ggml-sycl/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+
+  ubuntu-24-sycl:
+    strategy:
+      matrix:
+        build: [fp32, fp16]
+        include:
+          - build: fp32
+            fp16: OFF
+          - build: fp16
+            fp16: ON
+
+    runs-on: ubuntu-24.04
+
+    env:
+      ONEAPI_ROOT: /opt/intel/oneapi/
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+
+    continue-on-error: true
+
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Use oneAPI Installation Cache
+        uses: actions/cache@v5
+        id: cache-sycl
+        with:
+          path: ${{ env.ONEAPI_ROOT }}
+          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Download & Install oneAPI
+        shell: bash
+        if: steps.cache-sycl.outputs.cache-hit != 'true'
+        run: |
+          cd /tmp
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
+          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: ubuntu-24-sycl-${{ matrix.build }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_SYCL=ON \
+            -DCMAKE_C_COMPILER=icx \
+            -DCMAKE_CXX_COMPILER=icpx \
+            -DLLAMA_OPENSSL=OFF \
+            -DGGML_NATIVE=OFF \
+            -DGGML_SYCL_F16=${{ matrix.fp16 }}
+          time cmake --build build --config Release -j $(nproc)
+
+  windows-latest-sycl:
+    runs-on: windows-2022
+
+    defaults:
+      run:
+        shell: bash
+
+    env:
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
+      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
+      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Use oneAPI Installation Cache
+        uses: actions/cache@v5
+        id: cache-sycl
+        with:
+          path: ${{ env.ONEAPI_ROOT }}
+          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Download & Install oneAPI
+        shell: bash
+        if: steps.cache-sycl.outputs.cache-hit != 'true'
+        run: |
+          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: windows-latest-sycl
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
+
+      - name: Build
+        id: cmake_build
+        run:  examples/sycl/win-build-sycl.bat

.github/workflows/build.yml

@@ -555,186 +555,6 @@ jobs:
             -DGGML_MUSA=ON
           time cmake --build build --config Release -j $(nproc)
 
-  ubuntu-22-sycl:
-    runs-on: ubuntu-22.04
-
-    continue-on-error: true
-
-    steps:
-      - uses: actions/checkout@v6
-
-      - name: add oneAPI to apt
-        shell: bash
-        run: |
-          cd /tmp
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
-
-      - name: install oneAPI dpcpp compiler
-        shell: bash
-        run: |
-          sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libssl-dev
-
-      - name: install oneAPI MKL library
-        shell: bash
-        run: |
-          sudo apt install intel-oneapi-mkl-devel
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: ubuntu-22-sycl
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build
-        id: cmake_build
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          cmake -B build \
-            -DGGML_SYCL=ON \
-            -DCMAKE_C_COMPILER=icx \
-            -DCMAKE_CXX_COMPILER=icpx
-          time cmake --build build --config Release -j $(nproc)
-
-  ubuntu-22-sycl-fp16:
-    runs-on: ubuntu-22.04
-
-    continue-on-error: true
-
-    steps:
-      - uses: actions/checkout@v6
-
-      - name: add oneAPI to apt
-        shell: bash
-        run: |
-          cd /tmp
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
-
-      - name: install oneAPI dpcpp compiler
-        shell: bash
-        run: |
-          sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libssl-dev ninja-build
-
-      - name: install oneAPI MKL library
-        shell: bash
-        run: |
-          sudo apt install intel-oneapi-mkl-devel
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: ubuntu-22-sycl-fp16
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build
-        id: cmake_build
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          cmake -B build \
-            -G "Ninja" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_SYCL=ON \
-            -DCMAKE_C_COMPILER=icx \
-            -DCMAKE_CXX_COMPILER=icpx \
-            -DGGML_SYCL_F16=ON
-          time cmake --build build --config Release -j $(nproc)
-
-  ubuntu-24-openvino:
-      name: ubuntu-24-openvino-${{ matrix.openvino_device }}
-      strategy:
-        matrix:
-          include:
-            - variant: cpu
-              runner: '"ubuntu-24.04"'
-              openvino_device: "CPU"
-            - variant: gpu
-              runner: '["self-hosted","Linux","X64","Intel"]'
-              openvino_device: "GPU"
-
-      runs-on: ${{ fromJSON(matrix.runner) }}
-
-      env:
-        # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-        OPENVINO_VERSION_MAJOR: "2026.0"
-        OPENVINO_VERSION_FULL: "2026.0.0.20965.c6d6a13a886"
-
-      steps:
-        - name: Clone
-          id: checkout
-          uses: actions/checkout@v6
-
-        - name: ccache
-          if: runner.environment == 'github-hosted'
-          uses: ggml-org/ccache-action@v1.2.21
-          with:
-            key: ubuntu-24-openvino-${{ matrix.variant }}-no-preset-v1
-            evict-old-files: 1d
-            save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-        - name: Dependencies
-          id: depends
-          run: |
-            sudo apt-get update
-            sudo apt-get install -y build-essential libssl-dev libtbb12 cmake ninja-build python3-pip
-            sudo apt-get install -y ocl-icd-opencl-dev opencl-headers opencl-clhpp-headers intel-opencl-icd
-
-        - name: Use OpenVINO Toolkit Cache
-          if: runner.environment == 'github-hosted'
-          uses: actions/cache@v5
-          id: cache-openvino
-          with:
-            path: ./openvino_toolkit
-            key: openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
-
-        - name: Setup OpenVINO Toolkit
-          if: steps.cache-openvino.outputs.cache-hit != 'true'
-          uses: ./.github/actions/linux-setup-openvino
-          with:
-            path: ./openvino_toolkit
-            version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
-            version_full: ${{ env.OPENVINO_VERSION_FULL }}
-
-        - name: Install OpenVINO dependencies
-          run: |
-            cd ./openvino_toolkit
-            chmod +x ./install_dependencies/install_openvino_dependencies.sh
-            echo "Y" | sudo -E ./install_dependencies/install_openvino_dependencies.sh
-
-        - name: Build
-          id: cmake_build
-          run: |
-            source ./openvino_toolkit/setupvars.sh
-            cmake -B build/ReleaseOV -G Ninja \
-              -DCMAKE_BUILD_TYPE=Release \
-              -DGGML_OPENVINO=ON
-            time cmake --build build/ReleaseOV --config Release -j $(nproc)
-
-        - name: Test
-          id: cmake_test
-          # TODO: fix and re-enable the `test-llama-archs` test below
-          run: |
-            cd ${{ github.workspace }}
-            if [ "${{ matrix.openvino_device }}" = "GPU" ]; then
-              export GGML_OPENVINO_DEVICE=GPU
-            fi
-            ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000
 
   windows-latest:
     runs-on: windows-2025
@@ -943,39 +763,6 @@ jobs:
           cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release
 
-  windows-latest-sycl:
-    runs-on: windows-2022
-
-    defaults:
-      run:
-        shell: bash
-
-    env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
-      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
-      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: windows-latest-sycl
-          variant: ccache
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Install
-        run:  |
-          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
-
-      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
-
-      - name: Build
-        id: cmake_build
-        run:  examples/sycl/win-build-sycl.bat
 
   windows-latest-hip:
     runs-on: windows-2022

.github/workflows/python-type-check.yml

@@ -31,7 +31,7 @@ jobs:
         uses: actions/setup-python@v6
         with:
           python-version: "3.11"
-          pip-install: -r requirements/requirements-all.txt ty==0.0.26
+          pip-install: -r requirements/requirements-all.txt ty==0.0.33
       # - name: Type-check with Pyright
       #   uses: jakebailey/pyright-action@v2
       #   with:

.github/workflows/release.yml

@@ -598,15 +598,29 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v6
 
+      - name: Use oneAPI Installation Cache
+        uses: actions/cache@v5
+        id: cache-sycl
+        with:
+          path: ${{ env.ONEAPI_ROOT }}
+          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Download & Install oneAPI
+        shell: bash
+        if: steps.cache-sycl.outputs.cache-hit != 'true'
+        run: |
+          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
@@ -614,10 +628,6 @@ jobs:
           variant: ccache
           evict-old-files: 1d
 
-      - name: Install
-        run:  |
-          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
-
       - name: Build
         id: cmake_build
         shell: cmd
@@ -670,6 +680,82 @@ jobs:
           path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
+  ubuntu-24-sycl:
+    strategy:
+      matrix:
+        build: [fp32, fp16]
+        include:
+          - build: fp32
+            fp16: OFF
+          - build: fp16
+            fp16: ON
+
+    runs-on: ubuntu-24.04
+
+    env:
+      ONEAPI_ROOT: /opt/intel/oneapi/
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: Use oneAPI Installation Cache
+        uses: actions/cache@v5
+        id: cache-sycl
+        with:
+          path: ${{ env.ONEAPI_ROOT }}
+          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Download & Install oneAPI
+        shell: bash
+        if: steps.cache-sycl.outputs.cache-hit != 'true'
+        run: |
+          cd /tmp
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
+          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: ubuntu-24-sycl-${{ matrix.build }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_SYCL=ON \
+            -DCMAKE_C_COMPILER=icx \
+            -DCMAKE_CXX_COMPILER=icpx \
+            -DLLAMA_OPENSSL=OFF \
+            -DGGML_NATIVE=OFF \
+            -DGGML_SYCL_F16=${{ matrix.fp16 }}
+          time cmake --build build --config Release -j $(nproc)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
+          name: llama-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
+
   ubuntu-22-rocm:
     runs-on: ubuntu-22.04
 
@@ -1045,6 +1131,7 @@ jobs:
       - ubuntu-cpu
       - ubuntu-vulkan
       - ubuntu-24-openvino
+      - ubuntu-24-sycl
       - android-arm64
       - macOS-cpu
       - ios-xcode-build
@@ -1133,6 +1220,8 @@ jobs:
             - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
             - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
             - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
+            - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
+            - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
 
             **Android:**
             - [Android arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz)

.gitignore

@@ -34,7 +34,6 @@
 /.vscode/
 /nppBackup
 
-
 # Coverage
 
 /gcovr-report/
@@ -74,6 +73,7 @@
 !/models/templates
 
 # Zig
+
 /zig-out/
 /zig-cache/
 
@@ -93,6 +93,7 @@
 !/examples/sycl/*.sh
 
 # Server Web UI temporary files
+
 /tools/server/webui/node_modules
 /tools/server/webui/dist
 # we no longer use gz for index.html
@@ -106,9 +107,11 @@ __pycache__/
 poetry.toml
 
 # Nix
+
 /result
 
 # Test binaries
+
 /tests/test-backend-ops
 /tests/test-double-float
 /tests/test-grad0
@@ -124,6 +127,7 @@ poetry.toml
 /tests/test-tokenizer-1-spm
 
 # Scripts
+
 !/scripts/install-oneapi.bat
 
 # Generated by scripts
@@ -132,16 +136,24 @@ poetry.toml
 /wikitext-2-raw/
 
 # Test models for lora adapters
+
 /lora-tests
 
 # Local scripts
+
 /run-vim.sh
 /run-chat.sh
 /run-spec.sh
 /.ccache/
 
 # IDE
+
 /*.code-workspace
 /.windsurf/
 # emscripten
 a.out.*
+
+# AGENTS
+
+AGENTS.local.md
+.pi/SYSTEM.md

.pi/gg/SYSTEM.md

@@ -0,0 +1,33 @@
+You are a coding agent. Here are some very important rules that you must follow:
+
+General:
+- By very precise and concise when writing code, comments, explanations, etc.
+- PR and commit titles format: `<module> : <title>`. Lookup recents for examples
+- Don't try to build or run the code unless you are explicitly asked to do so
+
+Coding:
+- When in doubt, always refer to the CONTRIBUTING.md file of the project
+- When referencing issues or PRs in comments, use the format:
+  - C/C++ code: `// ref: <url>`
+  - Other (CMake, etc.): `# ref: <url>`
+
+Pull requests (PRs):
+- New branch names are prefixed with "gg/"
+- Before opening a pull request, ask the user to confirm the description
+- When creating a pull request, look for the repository's PR template and follow it
+- For the AI usage disclosure section, write "YES. llama.cpp + pi"
+- Always create the pull requests in draft mode
+
+Commits:
+- On every commit that you make, include a "Assisted-by: llama.cpp:local pi" tag
+- Do not explicitly set the git author in commits - rely on the default git config
+
+Resources (read on demand):
+- [CONTRIBUTING.md](CONTRIBUTING.md)
+- [Build documentation](docs/build.md)
+- [Server usage documentation](tools/server/README.md)
+- [Server development documentation](tools/server/README-dev.md)
+- [PEG parser](docs/development/parsing.md)
+- [Auto parser](docs/autoparser.md)
+- [Jinja engine](common/jinja/README.md)
+- [PR template](.github/pull_request_template.md)

CODEOWNERS

@@ -23,6 +23,7 @@
 /ci/                                    @ggerganov
 /cmake/                                 @ggerganov
 /common/                                @ggml-org/llama-common
+/common/fit.*                           @JohannesGaessler
 /common/jinja/                          @CISC
 /common/ngram-map.*                     @srogmann
 /convert_*.py                           @CISC
@@ -52,28 +53,29 @@
 /examples/speculative/                  @ggerganov
 /ggml/cmake/                            @ggerganov
 /ggml/include/                          @ggerganov
+/ggml/src/ggml-backend-meta.cpp         @JohannesGaessler
 /ggml/src/ggml-cann/                    @ggml-org/ggml-cann
 /ggml/src/ggml-common.h                 @ggerganov
 /ggml/src/ggml-cpu/                     @ggerganov
 /ggml/src/ggml-cpu/spacemit/            @alex-spacemit
 /ggml/src/ggml-cuda/                    @ggml-org/ggml-cuda
-/ggml/src/ggml-cuda/fattn-wmma*         @IMbackK
-/ggml/src/ggml-hip/                     @IMbackK
 /ggml/src/ggml-cuda/vendors/hip.h       @IMbackK
+/ggml/src/ggml-cuda/fattn-wmma*         @IMbackK
+/ggml/src/ggml-hexagon/                 @ggml-org/ggml-hexagon
+/ggml/src/ggml-hip/                     @IMbackK
 /ggml/src/ggml-impl.h                   @ggerganov
 /ggml/src/ggml-metal/                   @ggml-org/ggml-metal
 /ggml/src/ggml-opencl/                  @ggml-org/ggml-opencl
-/ggml/src/ggml-hexagon/                 @ggml-org/ggml-hexagon
+/ggml/src/ggml-openvino/                @cavusmustafa @wine99
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
 /ggml/src/ggml-rpc/                     @ggml-org/ggml-rpc
 /ggml/src/ggml-sycl/                    @ggml-org/ggml-sycl
 /ggml/src/ggml-threading.*              @ggerganov
-/ggml/src/ggml-vulkan/                  @ggml-org/ggml-vulkan
 /ggml/src/ggml-virtgpu/                 @kpouget
+/ggml/src/ggml-vulkan/                  @ggml-org/ggml-vulkan
 /ggml/src/ggml-webgpu/                  @ggml-org/ggml-webgpu
 /ggml/src/ggml-zdnn/                    @ggml-org/ggml-zdnn @Andreas-Krebbel @AlekseiNikiforovIBM
-/ggml/src/ggml-openvino/                @cavusmustafa @wine99
 /ggml/src/ggml.c                        @ggerganov
 /ggml/src/ggml.cpp                      @ggerganov
 /ggml/src/gguf.cpp                      @JohannesGaessler @Green-Sky

common/CMakeLists.txt

@@ -73,6 +73,8 @@ add_library(${TARGET}
     debug.h
     download.cpp
     download.h
+    fit.cpp
+    fit.h
     hf-cache.cpp
     hf-cache.h
     http.h

common/arg.h

@@ -25,7 +25,8 @@ struct common_arg {
     const char * value_hint_2 = nullptr; // for second arg value
     const char * env          = nullptr;
     std::string help;
-    bool is_sparam = false; // is current arg a sampling param?
+    bool is_sampling = false; // is current arg a sampling param?
+    bool is_spec = false; // is current arg a speculative decoding param?
     bool is_preset_only = false; // is current arg preset-only (not treated as CLI arg)
     void (*handler_void)   (common_params & params) = nullptr;
     void (*handler_string) (common_params & params, const std::string &) = nullptr;
@@ -74,7 +75,8 @@ struct common_arg {
     common_arg & set_examples(std::initializer_list<enum llama_example> examples);
     common_arg & set_excludes(std::initializer_list<enum llama_example> excludes);
     common_arg & set_env(const char * env);
-    common_arg & set_sparam();
+    common_arg & set_sampling();
+    common_arg & set_spec();
     common_arg & set_preset_only();
     bool in_example(enum llama_example ex);
     bool is_exclude(enum llama_example ex);

common/chat-auto-parser-generator.cpp

@@ -443,14 +443,14 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
     if (!format.per_call_start.empty()) {
         auto wrapped_call = format.per_call_start + p.space() + tool_choice + p.space() + format.per_call_end;
         if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call));
+            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call) + p.space());
         } else {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call);
+            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.space());
         }
         if (!format.section_start.empty()) {
             tool_calls = p.trigger_rule("tool-calls",
                                         p.literal(format.section_start) + p.space() + tool_calls + p.space() +
-                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end)));
+                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end) + p.space()));
         }
     } else {
         std::string separator = ", ";  // Default

common/chat-diff-analyzer.cpp

@@ -296,7 +296,7 @@ void analyze_reasoning::compare_reasoning_presence() {
             return p.literal(reasoning_content) + p.space() + p.optional(p.tag("post", (p.marker() + p.space())) + p.rest());
         });
         auto parser_wrapped = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
-            return p.tag("pre", p.marker() + p.space()) + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
+            return p.tag("pre", p.marker() + p.space()) + p.literal(reasoning_content) + p.tag("post", (p.space() + p.marker() + p.space())) + p.rest();
         });
         // try the more aggressive parse first, if it fails, fall back to the delimiter one
         auto result = parser_wrapped.parse_anywhere_and_extract(comparison->output_B);
@@ -306,11 +306,11 @@ void analyze_reasoning::compare_reasoning_presence() {
         if (result.result.success()) {
             if (!result.tags["pre"].empty() && !result.tags["post"].empty()) {
                 mode = reasoning_mode::TAG_BASED;
-                start = trim_leading_whitespace(result.tags["pre"]);
-                end   = trim_trailing_whitespace(result.tags["post"]);
+                start = result.tags["pre"];
+                end   = result.tags["post"];
             } else if (!result.tags["post"].empty()) {
                 mode = reasoning_mode::TAG_BASED;
-                end = trim_trailing_whitespace(result.tags["post"]);
+                end = result.tags["post"];
             }
         }
     }

common/chat.cpp

@@ -397,6 +397,25 @@ json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msg
     return render_message_to_json(msgs, c);
 }
 
+json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools) {
+    if (tools.empty()) {
+        return json();
+    }
+
+    auto result = json::array();
+    for (const auto & tool : tools) {
+        result.push_back({
+            { "type",     "function" },
+            { "function", {
+                { "name", tool.name },
+                { "description", tool.description },
+                { "parameters", json::parse(tool.parameters) },
+            }},
+        });
+    }
+    return result;
+}
+
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {
     std::vector<common_chat_tool> result;
 
@@ -432,56 +451,6 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
     return result;
 }
 
-json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools) {
-    if (tools.empty()) {
-        return json();
-    }
-
-    auto result = json::array();
-    for (const auto & tool : tools) {
-        result.push_back({
-            { "type",     "function" },
-            { "function",
-             {
-                  { "name", tool.name },
-                  { "description", tool.description },
-                  { "parameters", json::parse(tool.parameters) },
-              }                      },
-        });
-    }
-    return result;
-}
-
-json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
-    json delta = json::object();
-    if (!diff.reasoning_content_delta.empty()) {
-        delta["reasoning_content"] = diff.reasoning_content_delta;
-    }
-    if (!diff.content_delta.empty()) {
-        delta["content"] = diff.content_delta;
-    }
-    if (diff.tool_call_index != std::string::npos) {
-        json tool_call;
-        tool_call["index"] = diff.tool_call_index;
-        if (!diff.tool_call_delta.id.empty()) {
-            tool_call["id"]   = diff.tool_call_delta.id;
-            tool_call["type"] = "function";
-        }
-        if (!diff.tool_call_delta.name.empty() || !diff.tool_call_delta.arguments.empty()) {
-            json function = json::object();
-            if (!diff.tool_call_delta.name.empty()) {
-                function["name"] = diff.tool_call_delta.name;
-            }
-            if (!diff.tool_call_delta.arguments.empty()) {
-                function["arguments"] = diff.tool_call_delta.arguments;
-            }
-            tool_call["function"] = function;
-        }
-        delta["tool_calls"] = json::array({ tool_call });
-    }
-    return delta;
-}
-
 bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) {
     if (use_jinja) {
         try {
@@ -575,6 +544,26 @@ bool common_chat_templates_was_explicit(const struct common_chat_templates * tmp
     return tmpls->has_explicit_template;
 }
 
+// LFM2 format detection: template uses <|tool_list_start|>[...]<|tool_list_end|> around the tool list
+// and <|tool_call_start|>[...]<|tool_call_end|> around each tool call
+static bool is_lfm2_template(const std::string & src) {
+    return src.find("<|tool_list_start|>") != std::string::npos &&
+           src.find("<|tool_list_end|>")   != std::string::npos;
+}
+
+common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates) {
+    common_chat_prompt_preset asr_preset;
+    asr_preset.system = "";
+    asr_preset.user   = "Transcribe audio to text";
+
+    if (chat_templates && chat_templates->template_default && is_lfm2_template(chat_templates->template_default->source())) {
+        asr_preset.system = "Perform ASR.";
+        asr_preset.user   = "";
+    }
+
+    return asr_preset;
+}
+
 std::string common_chat_templates_source(const struct common_chat_templates * tmpls, const std::string & variant) {
     if (!variant.empty()) {
         if (variant == "tool_use") {
@@ -2084,10 +2073,7 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_kimi_k2(tmpl, params);
     }
 
-    // LFM2 format detection: template uses <|tool_list_start|>[...]<|tool_list_end|> around the tool list
-    // and <|tool_call_start|>[...]<|tool_call_end|> around each tool call
-    if (src.find("<|tool_list_start|>") != std::string::npos &&
-        src.find("<|tool_list_end|>") != std::string::npos) {
+    if (is_lfm2_template(src)) {
         LOG_DBG("Using specialized template: LFM2\n");
         return common_chat_params_init_lfm2(tmpl, params);
     }
@@ -2334,7 +2320,7 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena &          src_pars
         ? input
         : params.generation_prompt + input;
 
-    LOG_DBG("Parsing PEG input with format %s: %s\n", common_chat_format_name(params.format), effective_input.c_str());
+    //LOG_DBG("Parsing PEG input with format %s: %s\n", common_chat_format_name(params.format), effective_input.c_str());
 
     common_peg_parse_flags flags = COMMON_PEG_PARSE_FLAG_LENIENT;
     if (params.debug) {
@@ -2396,4 +2382,3 @@ std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_tem
     GGML_ASSERT(chat_templates->template_default != nullptr);
     return chat_templates->template_default->caps.to_map();
 }
-

common/chat.h

@@ -256,14 +256,13 @@ bool common_chat_templates_support_enable_thinking(const common_chat_templates *
 // Parses a JSON array of messages in OpenAI's chat completion API format.
 std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
 
+std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
+
 // DEPRECATED: only used in tests
 nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
-std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
 nlohmann::ordered_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
-nlohmann::ordered_json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);
-
 // get template caps, useful for reporting to server /props endpoint
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
 
@@ -275,3 +274,11 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         const common_chat_template &          tmpl,
         const std::string &                   src,
         autoparser::generation_params & params);
+
+// specialized per-task preset
+struct common_chat_prompt_preset {
+    std::string system;
+    std::string user;
+};
+
+common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates);

common/common.cpp

@@ -3,6 +3,7 @@
 
 #include "build-info.h"
 #include "common.h"
+#include "fit.h"
 #include "log.h"
 #include "llama.h"
 #include "sampling.h"
@@ -69,7 +70,7 @@ common_time_meas::~common_time_meas() {
 // CPU utils
 //
 
-int32_t cpu_get_num_physical_cores() {
+int32_t common_cpu_get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
     std::unordered_set<std::string> siblings;
@@ -184,11 +185,11 @@ static int cpu_count_math_cpus(int n_cpu) {
 /**
  * Returns number of CPUs on system that are useful for math.
  */
-int32_t cpu_get_num_math() {
+int32_t common_cpu_get_num_math() {
 #if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
     int n_cpu = sysconf(_SC_NPROCESSORS_ONLN);
     if (n_cpu < 1) {
-        return cpu_get_num_physical_cores();
+        return common_cpu_get_num_physical_cores();
     }
     if (is_hybrid_cpu()) {
         cpu_set_t affinity;
@@ -201,7 +202,7 @@ int32_t cpu_get_num_math() {
         }
     }
 #endif
-    return cpu_get_num_physical_cores();
+    return common_cpu_get_num_physical_cores();
 }
 
 // Helper for setting process priority
@@ -262,7 +263,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
 //
 
 
-void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) {
+void postprocess_cpu_params(common_cpu_params & cpuparams, const common_cpu_params * role_model) {
     int32_t n_set = 0;
 
     if (cpuparams.n_threads < 0) {
@@ -270,7 +271,7 @@ void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model)
         if (role_model != nullptr) {
             cpuparams = *role_model;
         } else {
-            cpuparams.n_threads = cpu_get_num_math();
+            cpuparams.n_threads = common_cpu_get_num_math();
         }
     }
 
@@ -1147,7 +1148,7 @@ common_init_result::common_init_result(common_params & params) :
 
     if (params.fit_params) {
         LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
-        llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
+        common_fit_params(params.model.path.c_str(), &mparams, &cparams,
             params.tensor_split,
             params.tensor_buft_overrides.data(),
             params.fit_params_target.data(),
@@ -1382,7 +1383,7 @@ common_init_result_ptr common_init_from_params(common_params & params) {
 
 common_init_result::~common_init_result() = default;
 
-std::string get_model_endpoint() {
+std::string common_get_model_endpoint() {
     const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
     // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
     const char * hf_endpoint_env = getenv("HF_ENDPOINT");
@@ -1397,6 +1398,42 @@ std::string get_model_endpoint() {
     return model_endpoint;
 }
 
+common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx) {
+    auto * mem = llama_get_memory(ctx);
+    if (mem == nullptr) {
+        return COMMON_CONTEXT_SEQ_RM_TYPE_NO;
+    }
+
+    common_context_seq_rm_type res = COMMON_CONTEXT_SEQ_RM_TYPE_PART;
+
+    llama_memory_clear(mem, true);
+
+    // eval 2 tokens to check if the context is compatible
+    std::vector<llama_token> tmp;
+    tmp.push_back(0);
+    tmp.push_back(0);
+
+    int ret = llama_decode(ctx, llama_batch_get_one(tmp.data(), tmp.size()));
+    if (ret != 0) {
+        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
+        res = COMMON_CONTEXT_SEQ_RM_TYPE_NO;
+        goto done;
+    }
+
+    // try to remove the last tokens
+    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
+        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
+        res = COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+        goto done;
+    }
+
+done:
+    llama_memory_clear(mem, true);
+    llama_synchronize(ctx);
+
+    return res;
+}
+
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
     std::vector<llama_adapter_lora *> loras;
     std::vector<float> scales;
@@ -1484,7 +1521,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     return cparams;
 }
 
-struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params) {
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const common_cpu_params & params) {
     struct ggml_threadpool_params tpp;
 
     ggml_threadpool_params_init(&tpp, params.n_threads); // setup the defaults

common/common.h

@@ -11,7 +11,6 @@
 #include <sstream>
 #include <string>
 #include <string_view>
-#include <variant>
 #include <vector>
 #include <map>
 
@@ -55,7 +54,7 @@ struct common_control_vector_load_info;
 // CPU utils
 //
 
-struct cpu_params {
+struct common_cpu_params {
     int      n_threads                   = -1;
     bool     cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
     bool     mask_valid                  = false;   // Default: any CPU
@@ -64,8 +63,8 @@ struct cpu_params {
     uint32_t poll                        = 50;      // Polling (busywait) level (0 - no polling, 100 - mostly polling)
 };
 
-int32_t cpu_get_num_physical_cores();
-int32_t cpu_get_num_math();
+int32_t common_cpu_get_num_physical_cores();
+int32_t common_cpu_get_num_math();
 
 //
 // Common params
@@ -275,6 +274,7 @@ struct common_params_sampling {
     std::vector<llama_token> reasoning_budget_start;           // start tag token sequence
     std::vector<llama_token> reasoning_budget_end;             // end tag token sequence
     std::vector<llama_token> reasoning_budget_forced;          // forced sequence (message + end tag)
+    std::string              reasoning_budget_message;         // message injected before end tag when budget exhausted
 
     bool backend_sampling = false;
 
@@ -297,34 +297,19 @@ struct common_params_model {
 
 struct common_ngram_mod;
 
-struct common_params_speculative {
-    common_speculative_type type = COMMON_SPECULATIVE_TYPE_NONE; // type of speculative decoding
+// draft-model-based speculative decoding parameters
+struct common_params_speculative_draft {
+    int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding
+    int32_t n_min = 0;  // minimum number of draft tokens to use for speculative decoding
 
-    // general-purpose speculative decoding parameters
+    float p_split = 0.1f;  // speculative decoding split probability
+    float p_min   = 0.75f; // minimum speculative decoding probability (greedy)
 
-    int32_t n_max   = 16; // maximum number of tokens to draft during speculative decoding
-    int32_t n_min   = 0; // minimum number of draft tokens to use for speculative decoding
-    float   p_split = 0.1f; // speculative decoding split probability
-    float   p_min   = 0.75f; // minimum speculative decoding probability (greedy)
+    common_params_model mparams;
 
-    // ngram-based speculative decoding
+    llama_model * model = nullptr; // a llama_model that can be shared by multiple speculative contexts
 
-    uint16_t ngram_size_n     = 12; // ngram size for lookup
-    uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
-    uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed
-
-    std::shared_ptr<common_ngram_mod> ngram_mod;
-
-    std::string lookup_cache_static;  // path of static ngram cache file for lookup decoding           // NOLINT
-    std::string lookup_cache_dynamic; // path of dynamic ngram cache file for lookup decoding          // NOLINT
-
-    // draft-model speculative decoding
-
-    struct common_params_model mparams_dft;
-
-    llama_model * model_dft = nullptr; // a llama_model that can be shared by multiple speculative contexts
-
-    llama_context_params cparams_dft; // these are the parameters for the draft llama_context
+    llama_context_params cparams; // these are the parameters for the draft llama_context
 
     int32_t n_ctx        = 0;  // draft context size
     int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
@@ -332,25 +317,60 @@ struct common_params_speculative {
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
 
-    struct cpu_params cpuparams;
-    struct cpu_params cpuparams_batch;
+    common_cpu_params cpuparams;
+    common_cpu_params cpuparams_batch;
 
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
     std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
     std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
+};
+
+struct common_params_speculative_ngram_mod {
+    int32_t n_match = 24;
+
+    int32_t n_max = 64;
+    int32_t n_min = 48;
+
+    // shared instance of the ngram container for all speculative decoding contexts
+    std::shared_ptr<common_ngram_mod> obj;
+};
+
+struct common_params_speculative_ngram_map {
+    uint16_t size_n   = 12; // ngram size for lookup
+    uint16_t size_m   = 48; // mgram size for speculative tokens
+    uint16_t min_hits = 1;  // minimum hits at ngram/mgram lookup for mgram to be proposed
+};
+
+struct common_params_speculative_ngram_cache {
+    std::string lookup_cache_static;  // path of static ngram cache file for lookup decoding
+    std::string lookup_cache_dynamic; // path of dynamic ngram cache file for lookup decoding
+};
+
+struct common_params_speculative {
+    // TODO: become a vector in order to support "chains of speculators"
+    common_speculative_type type = COMMON_SPECULATIVE_TYPE_NONE;
+
+    common_params_speculative_draft draft;
+
+    common_params_speculative_ngram_mod ngram_mod;
+    common_params_speculative_ngram_map ngram_simple;
+    common_params_speculative_ngram_map ngram_map_k;
+    common_params_speculative_ngram_map ngram_map_k4v;
+
+    common_params_speculative_ngram_cache ngram_cache;
 
     bool has_dft() const {
-        return !mparams_dft.path.empty() || !mparams_dft.hf_repo.empty();
+        return !draft.mparams.path.empty() || !draft.mparams.hf_repo.empty();
     }
 };
 
 struct common_params_vocoder {
     struct common_params_model model;
 
-    std::string speaker_file = ""; // speaker file path                                      // NOLINT
+    std::string speaker_file; // speaker file path
 
-    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
+    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy
 };
 
 struct common_params_diffusion {
@@ -421,19 +441,20 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers       = -1;   // number of layers to store in VRAM, -1 is auto, <= -2 is all
-    int32_t main_gpu           = 0;    // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]  = {0};  // how split tensors should be distributed across GPUs
-    bool    fit_params         = true; // whether to fit unset model/context parameters to free device memory
-    int32_t fit_params_min_ctx = 4096; // minimum context size to set when trying to reduce memory use
+    int32_t n_gpu_layers       = -1;    // number of layers to store in VRAM, -1 is auto, <= -2 is all
+    int32_t main_gpu           = 0;     // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]  = {0};   // how split tensors should be distributed across GPUs
+    bool    fit_params         = true;  // whether to fit unset model/context parameters to free device memory
+    bool    fit_params_print   = false; // print the estimated required memory to run the model
+    int32_t fit_params_min_ctx = 4096;  // minimum context size to set when trying to reduce memory use
 
     // margin per device in bytes for fitting parameters to free memory:
     std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
-    struct cpu_params cpuparams;
-    struct cpu_params cpuparams_batch;
+    common_cpu_params cpuparams;
+    common_cpu_params cpuparams_batch;
 
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;
@@ -567,7 +588,7 @@ struct common_params {
     int32_t n_threads_http      = -1;    // number of threads to process HTTP requests (TODO: support threadpool)
     int32_t n_cache_reuse       = 0;     // min chunk size to reuse from the cache via KV shifting
     bool    cache_prompt        = true;  // whether to enable prompt caching
-    bool    clear_idle          = true;  // save and clear idle slots upon starting a new task
+    bool    cache_idle_slots    = true;  // save and clear idle slots upon starting a new task
     int32_t n_ctx_checkpoints   = 32;    // max number of context checkpoints per slot
     int32_t checkpoint_every_nt = 8192;  // make a checkpoint every n tokens during prefill
     int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
@@ -581,8 +602,6 @@ struct common_params {
     bool force_pure_content_parser = false;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int enable_reasoning = -1; // -1 = auto, 0 = disable, 1 = enable
-    int reasoning_budget = -1;
-    std::string reasoning_budget_message; // message injected before end tag when budget exhausted
     bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
     int sleep_idle_seconds = -1;   // if >0, server will sleep after this many seconds of idle time
 
@@ -679,7 +698,7 @@ std::string common_params_get_system_info(const common_params & params);
 
 bool parse_cpu_range(const std::string & range, bool(&boolmask)[GGML_MAX_N_THREADS]);
 bool parse_cpu_mask(const std::string & mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
-void postprocess_cpu_params(cpu_params & cpuparams, const cpu_params * role_model = nullptr);
+void postprocess_cpu_params(common_cpu_params & cpuparams, const common_cpu_params * role_model = nullptr);
 bool set_process_priority(enum ggml_sched_priority prio);
 
 //
@@ -747,6 +766,11 @@ inline bool string_starts_with(std::string_view str, std::string_view prefix) {
            str.compare(0, prefix.size(), prefix) == 0;
 }
 
+// remove when moving to c++20
+inline bool string_starts_with(std::string_view str, char prefix) {
+    return !str.empty() && str.front() == prefix;
+}
+
 // remove when moving to c++20
 inline bool string_ends_with(std::string_view str, std::string_view suffix) {
     return str.size() >= suffix.size() &&
@@ -842,12 +866,28 @@ common_init_result_ptr common_init_from_params(common_params & params);
 
 struct llama_model_params     common_model_params_to_llama  (      common_params & params);
 struct llama_context_params   common_context_params_to_llama(const common_params & params);
-struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const common_cpu_params & params);
 
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
 
-std::string                   get_model_endpoint();
+// model endpoint from env
+std::string common_get_model_endpoint();
+
+//
+// Context utils
+//
+
+enum common_context_seq_rm_type {
+    COMMON_CONTEXT_SEQ_RM_TYPE_NO   = 0, // seq_rm not supported (e.g. no memory module)
+    COMMON_CONTEXT_SEQ_RM_TYPE_PART = 1, // can seq_rm partial sequences
+    COMMON_CONTEXT_SEQ_RM_TYPE_FULL = 2, // can seq_rm full sequences only
+};
+
+// check if the llama_context can remove sequences
+// note: clears the memory of the context
+common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx);
+
 
 //
 // Batch utils

common/debug.cpp

@@ -1,9 +1,38 @@
 #include "debug.h"
 
+#include "common.h"
 #include "log.h"
 
 #include <cmath>
+#include <regex>
 #include <string>
+#include <vector>
+
+struct common_debug_cb_user_data::impl {
+    std::vector<uint8_t>    data;
+    std::vector<std::regex> tensor_filters;
+    bool                    abort_on_nan{false};
+};
+
+common_debug_cb_user_data::common_debug_cb_user_data() : pimpl(std::make_unique<impl>()) {}
+common_debug_cb_user_data::~common_debug_cb_user_data() = default;
+
+common_debug_cb_user_data::common_debug_cb_user_data(common_params & params, const std::vector<std::string> & filter_patterns, bool abort_on_nan)
+    : pimpl(std::make_unique<impl>())
+{
+    for (const auto & pattern : filter_patterns) {
+        try {
+            std::string anchored_pattern = "^" + pattern;
+            pimpl->tensor_filters.emplace_back(anchored_pattern, std::regex::optimize);
+        } catch (const std::regex_error & e) {
+            throw std::runtime_error("Invalid regex pattern '" + pattern + "': " + e.what());
+        }
+    }
+    pimpl->abort_on_nan = abort_on_nan;
+
+    params.cb_eval           = common_debug_cb_eval;
+    params.cb_eval_user_data = this;
+}
 
 static std::string common_ggml_ne_string(const ggml_tensor * t) {
     std::string str;
@@ -47,8 +76,7 @@ static float common_ggml_get_float_value(const uint8_t * data,
 
 #define INDENT "    "
 
-template <bool abort>
-void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
+static void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n, bool abort_on_nan) {
     GGML_ASSERT(n > 0);
     float sum = 0;
     for (int64_t i3 = 0; i3 < ne[3]; i3++) {
@@ -94,7 +122,7 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
         LOG(INDENT "sum = %f\n", sum);
     }
 
-    if constexpr (abort) {
+    if (abort_on_nan) {
         if (std::isnan(sum)) {
             LOG("encountered NaN - aborting\n");
             exit(0);
@@ -112,8 +140,9 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
  * @param user_data user data to pass at each call back
  * @return true to receive data or continue the graph, false otherwise
  */
-template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
-    auto * cb_data = (base_callback_data *) user_data;
+bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
+    auto * cb_data = (common_debug_cb_user_data *) user_data;
+    auto * pimpl = cb_data->pimpl.get();
 
     const struct ggml_tensor * src0 = t->src[0];
     const struct ggml_tensor * src1 = t->src[1];
@@ -122,10 +151,10 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
         return true;  // Always retrieve data
     }
 
-    bool matches_filter = cb_data->tensor_filters.empty();
+    bool matches_filter = pimpl->tensor_filters.empty();
 
     if (!matches_filter) {
-        for (const auto & filter : cb_data->tensor_filters) {
+        for (const auto & filter : pimpl->tensor_filters) {
             if (std::regex_search(t->name, filter)) {
                 matches_filter = true;
                 break;
@@ -148,20 +177,14 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
 
     if (!is_host) {
         auto n_bytes = ggml_nbytes(t);
-        cb_data->data.resize(n_bytes);
-        ggml_backend_tensor_get(t, cb_data->data.data(), 0, n_bytes);
+        pimpl->data.resize(n_bytes);
+        ggml_backend_tensor_get(t, pimpl->data.data(), 0, n_bytes);
     }
 
     if (!ggml_is_quantized(t->type) && matches_filter) {
-        uint8_t * data = is_host ? (uint8_t *) t->data : cb_data->data.data();
-        common_debug_print_tensor<abort_on_nan>(data, t->type, t->ne, t->nb, 3);
+        uint8_t * data = is_host ? (uint8_t *) t->data : pimpl->data.data();
+        common_debug_print_tensor(data, t->type, t->ne, t->nb, 3, pimpl->abort_on_nan);
     }
 
     return true;
 }
-
-// Explicit template instantiations
-template bool common_debug_cb_eval<false>(ggml_tensor *, bool, void *);
-template bool common_debug_cb_eval<true>(ggml_tensor *, bool, void *);
-template void common_debug_print_tensor<false>(uint8_t *, ggml_type, const int64_t *, const size_t *, int64_t);
-template void common_debug_print_tensor<true>(uint8_t *, ggml_type, const int64_t *, const size_t *, int64_t);

common/debug.h

@@ -1,43 +1,31 @@
 #pragma once
-#include "common.h"
+
+#include <memory>
 #include <string>
 #include <vector>
-#include <regex>
 
 // common debug functions and structs
 
-// Print a tensor's detailed data
-// data - the tensor's data in byte format
-// type - the tensor's quantization type
-// ne   - the tensor dimensions array
-// nb   - the tensor strides array
-// n    - the number of rows/columns to fully print
-template <bool abort_on_nan> void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n);
+struct common_params;
 
 // Intended to use as callback for ggml_backend_sched_eval_callback
 // prints tensors that are processed in the computation graph
-// by default prints all tensors, but can be configured by creating a `base_callback_data` instance with
-// non-empty filter_patterns. See examples/debug.ccp for possible usage patterns
-// The template parameter determines whether an error should be thrown whenever a NaN is encountered
+// by default prints all tensors, but can be configured by creating a `common_debug_cb_user_data` instance with
+// non-empty filter_patterns. See examples/debug.cpp for possible usage patterns
+// `common_debug_cb_user_data` contains `abort_on_nan` flag that determines whether an error should be thrown whenever a NaN is encountered
 // in a tensor (useful for stopping debug sessions on first erroneous tensor)
 // The callback data will be passed as the third parameter (user_data)
-template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data);
-struct base_callback_data {
-    std::vector<uint8_t>    data;
-    std::vector<std::regex> tensor_filters;
+bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data);
 
-    base_callback_data() = default;
+struct common_debug_cb_user_data {
+    struct impl;
+    std::unique_ptr<impl> pimpl;
 
-    base_callback_data(common_params & params, const std::vector<std::string> & filter_patterns) {
-        for (const auto & pattern : filter_patterns) {
-            try {
-                std::string anchored_pattern = "^" + pattern;
-                tensor_filters.emplace_back(anchored_pattern, std::regex::optimize);
-            } catch (const std::regex_error & e) {
-                throw std::runtime_error("Invalid regex pattern '" + pattern + "': " + e.what());
-            }
-        }
-        params.cb_eval           = common_debug_cb_eval<false>;
-        params.cb_eval_user_data = this;
-    }
+    common_debug_cb_user_data();
+    ~common_debug_cb_user_data();
+
+    common_debug_cb_user_data(const common_debug_cb_user_data &) = delete;
+    common_debug_cb_user_data & operator=(const common_debug_cb_user_data &) = delete;
+
+    common_debug_cb_user_data(common_params & params, const std::vector<std::string> & filter_patterns, bool abort_on_nan = false);
 };

common/download.cpp

@@ -627,7 +627,7 @@ static hf_cache::hf_file find_best_model(const hf_cache::hf_files & files,
     if (!tag.empty()) {
         tags.push_back(tag);
     } else {
-        tags = {"Q4_K_M", "Q4_0"};
+        tags = {"Q4_K_M", "Q8_0"};
     }
 
     for (const auto & t : tags) {

common/fit.cpp

@@ -0,0 +1,951 @@
+#include "fit.h"
+
+#include "log.h"
+
+#include "../src/llama-ext.h"
+
+#include <array>
+#include <cassert>
+#include <stdexcept>
+#include <cinttypes>
+#include <set>
+#include <string>
+#include <vector>
+
+// this enum is only used in llama_params_fit_impl but needs to be defined outside of it to fix a Windows compilation issue
+// enum to identify part of a layer for distributing its tensors:
+enum common_layer_fraction_t {
+    LAYER_FRACTION_NONE = 0, // nothing
+    LAYER_FRACTION_ATTN = 1, // attention
+    LAYER_FRACTION_UP   = 2, // attention + up
+    LAYER_FRACTION_GATE = 3, // attention + up + gate
+    LAYER_FRACTION_MOE  = 4, // everything but sparse MoE weights
+};
+
+class common_params_fit_exception : public std::runtime_error {
+    using std::runtime_error::runtime_error;
+};
+
+static std::vector<llama_device_memory_data> common_get_device_memory_data(
+        const char * path_model,
+        const llama_model_params * mparams,
+        const llama_context_params * cparams,
+        std::vector<ggml_backend_dev_t> & devs,
+        uint32_t & hp_ngl,
+        uint32_t & hp_n_ctx_train,
+        uint32_t & hp_n_expert,
+        ggml_log_level log_level) {
+    struct user_data_t {
+        struct {
+            ggml_log_callback callback;
+            void * user_data;
+        } original_logger;
+        ggml_log_level min_level; // prints below this log level go to debug log
+    };
+    user_data_t ud;
+    llama_log_get(&ud.original_logger.callback, &ud.original_logger.user_data);
+    ud.min_level = log_level;
+
+    llama_log_set([](ggml_log_level level, const char * text, void * user_data) {
+        const user_data_t * ud = (const user_data_t *) user_data;
+        const ggml_log_level level_eff = level >= ud->min_level ? level : GGML_LOG_LEVEL_DEBUG;
+        ud->original_logger.callback(level_eff, text, ud->original_logger.user_data);
+    }, &ud);
+
+    llama_model_params mparams_copy = *mparams;
+    mparams_copy.no_alloc  = true;
+    mparams_copy.use_mmap  = false;
+    mparams_copy.use_mlock = false;
+
+    llama_model * model = llama_model_load_from_file(path_model, mparams_copy);
+    if (model == nullptr) {
+        llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+        throw std::runtime_error("failed to load model");
+    }
+
+    llama_context * ctx = llama_init_from_model(model, *cparams);
+    if (ctx == nullptr) {
+        llama_model_free(model);
+        llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+        throw std::runtime_error("failed to create llama_context from model");
+    }
+
+    const size_t nd = llama_model_n_devices(model);
+    std::vector<llama_device_memory_data> ret(nd + 1);
+
+    llama_memory_breakdown memory_breakdown = llama_get_memory_breakdown(ctx);
+
+    for (const auto & [buft, mb] : memory_breakdown) {
+        if (ggml_backend_buft_is_host(buft)) {
+            ret.back().mb.model   += mb.model;
+            ret.back().mb.context += mb.context;
+            ret.back().mb.compute += mb.compute;
+            continue;
+        }
+
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (!dev) {
+            continue;
+        }
+        for (size_t i = 0; i < nd; i++) {
+            if (dev == llama_model_get_device(model, i)) {
+                ret[i].mb.model   += mb.model;
+                ret[i].mb.context += mb.context;
+                ret[i].mb.compute += mb.compute;
+                break;
+            }
+        }
+    }
+
+    {
+        ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (cpu_dev == nullptr) {
+            throw std::runtime_error("no CPU backend found");
+        }
+        size_t free;
+        size_t total;
+        ggml_backend_dev_memory(cpu_dev, &free, &total);
+        ret.back().free  = free;
+        ret.back().total = total;
+    }
+    for (size_t i = 0; i < nd; i++) {
+        size_t free;
+        size_t total;
+        ggml_backend_dev_memory(llama_model_get_device(model, i), &free, &total);
+
+        // devices can return 0 bytes for free and total memory if they do not
+        // have any to report. in this case, we will use the host memory as a fallback
+        // fixes: https://github.com/ggml-org/llama.cpp/issues/18577
+        if (free == 0 && total == 0) {
+            free  = ret.back().free;
+            total = ret.back().total;
+        }
+        ret[i].free  = free;
+        ret[i].total = total;
+    }
+
+    devs.clear();
+    for (int i = 0; i < llama_model_n_devices(model); i++) {
+        devs.push_back(llama_model_get_device(model, i));
+    }
+
+    hp_ngl         = llama_model_n_layer(model);
+    hp_n_ctx_train = llama_model_n_ctx_train(model);
+    hp_n_expert    = llama_model_n_expert(model);
+
+    common_memory_breakdown_print(ctx);
+
+    llama_free(ctx);
+    llama_model_free(model);
+    llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+
+    return ret;
+}
+
+static void common_params_fit_impl(
+        const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
+        float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
+        size_t * margins_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+    if (mparams->split_mode == LLAMA_SPLIT_MODE_TENSOR) {
+        throw common_params_fit_exception("llama_params_fit is not implemented for SPLIT_MODE_TENSOR, abort");
+    }
+    constexpr int64_t MiB = 1024*1024;
+    typedef std::vector<llama_device_memory_data> dmds_t;
+    const llama_model_params default_mparams = llama_model_default_params();
+
+    std::vector<ggml_backend_dev_t> devs;
+    uint32_t hp_ngl = 0; // hparams.n_gpu_layers
+    uint32_t hp_nct = 0; // hparams.n_ctx_train
+    uint32_t hp_nex = 0; // hparams.n_expert
+
+    // step 1: get data for default parameters and check whether any changes are necessary in the first place
+
+    LOG_INF("%s: getting device memory data for initial parameters:\n", __func__);
+    const dmds_t dmds_full = common_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+    const size_t nd = devs.size(); // number of devices
+
+    std::vector<int64_t> margins; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
+    margins.reserve(nd);
+    if (nd == 0) {
+        margins.push_back(margins_s[0]);
+    } else {
+        for (size_t id = 0; id < nd; id++) {
+            margins.push_back(margins_s[id]);
+        }
+    }
+
+    std::vector<std::string> dev_names;
+    {
+        dev_names.reserve(nd);
+        size_t max_length = 0;
+        for (const auto & dev : devs) {
+            std::string name = ggml_backend_dev_name(dev);
+            name += " (";
+            name += ggml_backend_dev_description(dev);
+            name += ")";
+            dev_names.push_back(name);
+            max_length = std::max(max_length, name.length());
+        }
+        for (std::string & dn : dev_names) {
+            dn.insert(dn.end(), max_length - dn.length(), ' ');
+        }
+    }
+
+    int64_t sum_free            = 0;
+    int64_t sum_projected_free  = 0;
+    int64_t sum_projected_used  = 0;
+    int64_t sum_projected_model = 0;
+    std::vector<int64_t> projected_free_per_device;
+    projected_free_per_device.reserve(nd);
+
+    if (nd == 0) {
+        sum_projected_used = dmds_full.back().mb.total();
+        sum_free           = dmds_full.back().total;
+        sum_projected_free = sum_free - sum_projected_used;
+        LOG_INF("%s: projected to use %" PRId64 " MiB of host memory vs. %" PRId64 " MiB of total host memory\n",
+            __func__, sum_projected_used/MiB, sum_free/MiB);
+        if (sum_projected_free >= margins[0]) {
+            LOG_INF("%s: will leave %" PRId64 " >= %" PRId64 " MiB of system memory, no changes needed\n",
+                __func__, sum_projected_free/MiB, margins[0]/MiB);
+            return;
+        }
+    } else {
+        if (nd > 1) {
+            LOG_INF("%s: projected memory use with initial parameters [MiB]:\n", __func__);
+        }
+        for (size_t id = 0; id < nd; id++) {
+            const llama_device_memory_data & dmd = dmds_full[id];
+
+            const int64_t projected_used = dmd.mb.total();
+            const int64_t projected_free = dmd.free - projected_used;
+            projected_free_per_device.push_back(projected_free);
+
+            sum_free            += dmd.free;
+            sum_projected_used  += projected_used;
+            sum_projected_free  += projected_free;
+            sum_projected_model += dmd.mb.model;
+
+            if (nd > 1) {
+                LOG_INF("%s:   - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " free vs. target of %6" PRId64 "\n",
+                    __func__, dev_names[id].c_str(), dmd.total/MiB, projected_used/MiB, projected_free/MiB, margins[id]/MiB);
+            }
+        }
+        assert(sum_free >= 0 && sum_projected_used >= 0);
+        LOG_INF("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n",
+            __func__, sum_projected_used/MiB, sum_free/MiB);
+        if (nd == 1) {
+            if (projected_free_per_device[0] >= margins[0]) {
+                LOG_INF("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
+                    __func__, projected_free_per_device[0]/MiB, margins[0]/MiB);
+                return;
+            }
+        } else {
+            bool changes_needed = false;
+            for (size_t id = 0; id < nd; id++) {
+                if (projected_free_per_device[id] < margins[id]) {
+                    changes_needed = true;
+                    break;
+                }
+            }
+            if (!changes_needed) {
+                LOG_INF("%s: targets for free memory can be met on all devices, no changes needed\n", __func__);
+                return;
+            }
+        }
+    }
+
+    // step 2: try reducing memory use by reducing the context size
+
+    {
+        int64_t global_surplus = sum_projected_free;
+        if (nd == 0) {
+            global_surplus -= margins[0];
+        } else {
+            for (size_t id = 0; id < nd; id++) {
+                global_surplus -= margins[id];
+            }
+        }
+        if (global_surplus < 0) {
+            if (nd <= 1) {
+                LOG_INF("%s: cannot meet free memory target of %" PRId64 " MiB, need to reduce device memory by %" PRId64 " MiB\n",
+                    __func__, margins[0]/MiB, -global_surplus/MiB);
+            } else {
+                LOG_INF(
+                    "%s: cannot meet free memory targets on all devices, need to use %" PRId64 " MiB less in total\n",
+                    __func__, -global_surplus/MiB);
+            }
+            if (cparams->n_ctx == 0) {
+                if (hp_nct > n_ctx_min) {
+                    int64_t sum_used_target = sum_free;
+                    if (nd == 0) {
+                        sum_used_target -= margins[0];
+                    } else {
+                        for (size_t id = 0; id < nd; id++) {
+                            sum_used_target -= margins[id];
+                        }
+                    }
+                    if (nd > 1) {
+                        // for multiple devices we need to be more conservative in terms of how much context we think can fit:
+                        //   - for dense models only whole layers can be assigned to devices
+                        //   - for MoE models only whole tensors can be assigned to devices, which we estimate to be <= 1/3 of a layer
+                        //   - on average we expect a waste of 0.5 layers/tensors per device
+                        //   - use slightly more than the expected average for nd devices to be safe
+                        const int64_t model_per_layer = sum_projected_model / std::min(uint32_t(mparams->n_gpu_layers), hp_ngl);
+                        sum_used_target -= (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6);
+                    }
+
+                    int64_t sum_projected_used_min_ctx = 0;
+                    cparams->n_ctx = n_ctx_min;
+                    const dmds_t dmds_min_ctx = common_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+                    if (nd == 0) {
+                        sum_projected_used_min_ctx = dmds_min_ctx.back().mb.total();
+                    } else {
+                        for (size_t id = 0; id < nd; id++) {
+                            sum_projected_used_min_ctx += dmds_min_ctx[id].mb.total();
+                        }
+                    }
+                    if (sum_used_target > sum_projected_used_min_ctx) {
+                        // linear interpolation between minimum and maximum context size:
+                        cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
+                            / (sum_projected_used - sum_projected_used_min_ctx);
+                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
+
+                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min);
+                        const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx;
+                        LOG_INF("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
+                        if (nd <= 1) {
+                            LOG_INF("%s: entire model can be fit by reducing context\n", __func__);
+                            return;
+                        }
+                        LOG_INF("%s: entire model should be fit across devices by reducing context\n", __func__);
+                    } else {
+                        const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
+                        LOG_INF("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
+                    }
+                } else {
+                    if (n_ctx_min == UINT32_MAX) {
+                        LOG_INF("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
+                    } else {
+                        LOG_INF("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
+                            __func__, hp_nct, n_ctx_min);
+                    }
+                }
+            } else {
+                LOG_INF("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);
+            }
+        }
+    }
+    if (nd == 0) {
+        throw common_params_fit_exception("was unable to fit model into system memory by reducing context, abort");
+    }
+
+    if (mparams->n_gpu_layers != default_mparams.n_gpu_layers) {
+        throw common_params_fit_exception("n_gpu_layers already set by user to " + std::to_string(mparams->n_gpu_layers) + ", abort");
+    }
+    if (nd > 1) {
+        if (!tensor_split) {
+            throw common_params_fit_exception("did not provide a buffer to write the tensor_split to, abort");
+        }
+        if (mparams->tensor_split) {
+            for (size_t id = 0; id < nd; id++) {
+                if (mparams->tensor_split[id] != 0.0f) {
+                    throw common_params_fit_exception("model_params::tensor_split already set by user, abort");
+                }
+            }
+        }
+        if (mparams->split_mode == LLAMA_SPLIT_MODE_ROW) {
+            throw common_params_fit_exception("changing weight allocation for LLAMA_SPLIT_MODE_ROW not implemented, abort");
+        }
+    }
+    if (!tensor_buft_overrides) {
+        throw common_params_fit_exception("did not provide buffer to set tensor_buft_overrides, abort");
+    }
+    if (mparams->tensor_buft_overrides && (mparams->tensor_buft_overrides->pattern || mparams->tensor_buft_overrides->buft)) {
+        throw common_params_fit_exception("model_params::tensor_buft_overrides already set by user, abort");
+    }
+
+    // step 3: iteratively fill the back to front with "dense" layers
+    //   - for a dense model simply fill full layers, giving each device a contiguous slice of the model
+    //   - for a MoE model, same as dense model but with all MoE tensors in system memory
+
+    // utility function that returns a static C string matching the tensors for a specific layer index and layer fraction:
+    auto get_overflow_pattern = [&](const size_t il, const common_layer_fraction_t lf) -> const char * {
+        constexpr size_t n_strings = 1000;
+        if (il >= n_strings) {
+            throw std::runtime_error("at most " + std::to_string(n_strings) + " model layers are supported");
+        }
+        switch (lf) {
+            case LAYER_FRACTION_ATTN: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(gate|up|gate_up|down).*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_UP: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(gate|gate_up|down).*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_GATE: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_down.*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_MOE: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|down|gate_up|gate)_(ch|)exps";
+                }
+                return patterns[il].c_str();
+            }
+            default:
+                GGML_ABORT("fatal error");
+        }
+    };
+
+    struct ngl_t {
+        uint32_t n_layer = 0; // number of total layers
+        uint32_t n_part  = 0; // number of partial layers, <= n_layer
+
+        // for the first partial layer varying parts can overflow, all further layers use LAYER_FRACTION_MOE:
+        common_layer_fraction_t overflow_type = LAYER_FRACTION_MOE;
+
+        uint32_t n_full() const {
+            assert(n_layer >= n_part);
+            return n_layer - n_part;
+        }
+    };
+
+    const size_t ntbo = llama_max_tensor_buft_overrides();
+
+    // utility function to set n_gpu_layers and tensor_split
+    auto set_ngl_tensor_split_tbo = [&](
+            const std::vector<ngl_t> & ngl_per_device,
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts,
+            llama_model_params & mparams) {
+        mparams.n_gpu_layers = 0;
+        for (size_t id = 0; id < nd; id++) {
+            mparams.n_gpu_layers += ngl_per_device[id].n_layer;
+            if (nd > 1) {
+                tensor_split[id] = ngl_per_device[id].n_layer;
+            }
+        }
+        assert(uint32_t(mparams.n_gpu_layers) <= hp_ngl + 1);
+        uint32_t il0 = hp_ngl + 1 - mparams.n_gpu_layers; // start index for tensor buft overrides
+
+        mparams.tensor_split = tensor_split;
+
+        size_t itbo = 0;
+        for (size_t id = 0; id < nd; id++) {
+            il0 += ngl_per_device[id].n_full();
+            for (uint32_t il = il0; il < il0 + ngl_per_device[id].n_part; il++) {
+                if (itbo + 1 >= ntbo) {
+                    tensor_buft_overrides[itbo].pattern = nullptr;
+                    tensor_buft_overrides[itbo].buft    = nullptr;
+                    itbo++;
+                    mparams.tensor_buft_overrides = tensor_buft_overrides;
+                    throw common_params_fit_exception("llama_max_tensor_buft_overrides() == "
+                        + std::to_string(ntbo) + " is insufficient for model");
+                }
+                tensor_buft_overrides[itbo].pattern = get_overflow_pattern(il, il == il0 ? ngl_per_device[id].overflow_type : LAYER_FRACTION_MOE);
+                tensor_buft_overrides[itbo].buft = il == il0 ? overflow_bufts[id] : ggml_backend_cpu_buffer_type();
+                itbo++;
+            }
+            il0 += ngl_per_device[id].n_part;
+        }
+        tensor_buft_overrides[itbo].pattern = nullptr;
+        tensor_buft_overrides[itbo].buft    = nullptr;
+        itbo++;
+        mparams.tensor_buft_overrides = tensor_buft_overrides;
+    };
+
+    // utility function that returns the memory use per device for given numbers of layers per device
+    auto get_memory_for_layers = [&](
+            const char * func_name,
+            const std::vector<ngl_t> & ngl_per_device,
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts) -> std::vector<int64_t> {
+        llama_model_params mparams_copy = *mparams;
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy);
+
+        const dmds_t dmd_nl = common_get_device_memory_data(
+            path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+
+        LOG_INF("%s: memory for test allocation by device:\n", func_name);
+        for (size_t id = 0; id < nd; id++) {
+            const ngl_t & n = ngl_per_device[id];
+            LOG_INF(
+                "%s: id=%zu, n_layer=%2" PRIu32 ", n_part=%2" PRIu32 ", overflow_type=%d, mem=%6" PRId64 " MiB\n",
+                func_name, id, n.n_layer, n.n_part, int(n.overflow_type), dmd_nl[id].mb.total()/MiB);
+        }
+
+        std::vector<int64_t> ret;
+        ret.reserve(nd);
+        for (size_t id = 0; id < nd; id++) {
+            ret.push_back(dmd_nl[id].mb.total());
+        }
+        return ret;
+    };
+
+    int64_t global_surplus_cpu_moe = 0;
+    if (hp_nex > 0) {
+        const static std::string pattern_moe_all = "blk\\.\\d+\\.ffn_(up|down|gate_up|gate)_(ch|)exps"; // matches all MoE tensors
+        ggml_backend_buffer_type_t cpu_buft = ggml_backend_cpu_buffer_type();
+        tensor_buft_overrides[0] = {pattern_moe_all.c_str(), cpu_buft};
+        tensor_buft_overrides[1] = {nullptr, nullptr};
+        mparams->tensor_buft_overrides = tensor_buft_overrides;
+
+        LOG_INF("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
+        const dmds_t dmds_cpu_moe = common_get_device_memory_data(
+            path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+
+        for (size_t id = 0; id < nd; id++) {
+            global_surplus_cpu_moe += dmds_cpu_moe[id].free;
+            global_surplus_cpu_moe -= int64_t(dmds_cpu_moe[id].mb.total()) + margins[id];
+        }
+
+        if (global_surplus_cpu_moe > 0) {
+            LOG_INF("%s: with only dense weights in device memory there is a total surplus of %" PRId64 " MiB\n",
+                __func__, global_surplus_cpu_moe/MiB);
+        } else {
+            LOG_INF("%s: with only dense weights in device memory there is still a total deficit of %" PRId64 " MiB\n",
+                __func__, -global_surplus_cpu_moe/MiB);
+        }
+
+        // reset
+        tensor_buft_overrides[0] = {nullptr, nullptr};
+        mparams->tensor_buft_overrides = tensor_buft_overrides;
+    }
+
+    std::vector<int64_t> targets; // maximum acceptable memory use per device
+    targets.reserve(nd);
+    for (size_t id = 0; id < nd; id++) {
+        targets.push_back(dmds_full[id].free - margins[id]);
+        LOG_INF("%s: id=%zu, target=%" PRId64 " MiB\n", __func__, id, targets[id]/MiB);
+    }
+
+    std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the first partial layer of a device overflows to:
+    overflow_bufts.reserve(nd);
+    for (size_t id = 0; id < nd; id++) {
+        overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
+    }
+
+    std::vector<ngl_t> ngl_per_device(nd);
+    std::vector<int64_t> mem = get_memory_for_layers(__func__, ngl_per_device, overflow_bufts);
+
+    // optimize the number of layers per device using the method of false position:
+    //   - ngl_per_device has 0 layers for each device, lower bound
+    //   - try a "high" configuration where a device is given all unassigned layers
+    //   - interpolate the memory use / layer between low and high linearly to get a guess where it meets our target
+    //   - check memory use of our guess, replace either the low or high bound
+    //   - once we only have a difference of a single layer, stop and return the lower bound that just barely still fits
+    //   - the last device has the output layer, which cannot be a partial layer
+    if (hp_nex == 0) {
+        LOG_INF("%s: filling dense layers back-to-front:\n", __func__);
+    } else {
+        LOG_INF("%s: filling dense-only layers back-to-front:\n", __func__);
+    }
+    for (int id = nd - 1; id >= 0; id--) {
+        uint32_t n_unassigned = hp_ngl + 1;
+        for (size_t jd = id + 1; jd < nd; ++jd) {
+            assert(n_unassigned >= ngl_per_device[jd].n_layer);
+            n_unassigned -= ngl_per_device[jd].n_layer;
+        }
+
+        std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
+        ngl_per_device_high[id].n_layer = n_unassigned;
+        if (hp_nex > 0) {
+            ngl_per_device_high[id].n_part = size_t(id) < nd - 1 ? ngl_per_device_high[id].n_layer : ngl_per_device_high[id].n_layer - 1;
+        }
+        if (ngl_per_device_high[id].n_layer > 0) {
+            std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts);
+            if (mem_high[id] > targets[id]) {
+                assert(ngl_per_device_high[id].n_layer > ngl_per_device[id].n_layer);
+                uint32_t delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                LOG_INF("%s: start filling device %" PRIu32 ", delta=%" PRIu32 "\n", __func__, id, delta);
+                while (delta > 1) {
+                    uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
+                    step_size = std::max(step_size, uint32_t(1));
+                    step_size = std::min(step_size, delta - 1);
+
+                    std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+                    ngl_per_device_test[id].n_layer += step_size;
+                    if (hp_nex) {
+                        ngl_per_device_test[id].n_part += size_t(id) == nd - 1 && ngl_per_device_test[id].n_part == 0 ?
+                            step_size - 1 : step_size; // the first layer is the output layer which must always be full
+                    }
+                    const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
+
+                    if (mem_test[id] <= targets[id]) {
+                        ngl_per_device = ngl_per_device_test;
+                        mem            = mem_test;
+                        LOG_INF("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+                    } else {
+                        ngl_per_device_high = ngl_per_device_test;
+                        mem_high            = mem_test;
+                        LOG_INF("%s: set ngl_per_device_high[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device_high[id].n_layer);
+                    }
+                    delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                }
+            } else {
+                assert(ngl_per_device_high[id].n_layer == n_unassigned);
+                ngl_per_device = ngl_per_device_high;
+                mem            = mem_high;
+                LOG_INF("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+            }
+        }
+
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LOG_INF(
+            "%s:   - %s: %2" PRIu32 " layers, %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, mem[id]/MiB, projected_margin/MiB);
+    }
+    if (hp_nex == 0 || global_surplus_cpu_moe <= 0) {
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams);
+        return;
+    }
+
+    // step 4: for a MoE model where all dense tensors fit,
+    //     convert the dense-only layers in the back to full layers in the front until all devices are full
+    // essentially the same procedure as for the dense-only layers except front-to-back
+    // also, try fitting at least part of one more layer to reduce waste for "small" GPUs with e.g. 24 GiB VRAM
+
+    size_t id_dense_start = nd;
+    for (int id = nd - 1; id >= 0; id--) {
+        if (ngl_per_device[id].n_layer > 0) {
+            id_dense_start = id;
+            continue;
+        }
+        break;
+    }
+    assert(id_dense_start < nd);
+
+    LOG_INF("%s: converting dense-only layers to full layers and filling them front-to-back with overflow to next device/system memory:\n", __func__);
+    for (size_t id = 0; id <= id_dense_start && id_dense_start < nd; id++) {
+        std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
+        for (size_t jd = id_dense_start; jd < nd; jd++) {
+            const uint32_t n_layer_move = jd < nd - 1 ? ngl_per_device_high[jd].n_layer : ngl_per_device_high[jd].n_layer - 1;
+            ngl_per_device_high[id].n_layer += n_layer_move;
+            ngl_per_device_high[jd].n_layer -= n_layer_move;
+            ngl_per_device_high[jd].n_part = 0;
+        }
+        size_t id_dense_start_high = nd - 1;
+        std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts);
+
+        if (mem_high[id] > targets[id]) {
+            assert(ngl_per_device_high[id].n_full() >= ngl_per_device[id].n_full());
+            uint32_t delta = ngl_per_device_high[id].n_full() - ngl_per_device[id].n_full();
+            while (delta > 1) {
+                uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
+                step_size = std::max(step_size, uint32_t(1));
+                step_size = std::min(step_size, delta - 1);
+
+                std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+                size_t id_dense_start_test = id_dense_start;
+                uint32_t n_converted_test = 0;
+                for (;id_dense_start_test < nd; id_dense_start_test++) {
+                    const uint32_t n_convert_jd = std::min(step_size - n_converted_test, ngl_per_device_test[id_dense_start_test].n_part);
+                    ngl_per_device_test[id_dense_start_test].n_layer -= n_convert_jd;
+                    ngl_per_device_test[id_dense_start_test].n_part -= n_convert_jd;
+                    ngl_per_device_test[id].n_layer += n_convert_jd;
+                    n_converted_test += n_convert_jd;
+
+                    if (ngl_per_device_test[id_dense_start_test].n_part > 0) {
+                        break;
+                    }
+                }
+                const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
+
+                if (mem_test[id] <= targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LOG_INF("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                } else {
+                    ngl_per_device_high = ngl_per_device_test;
+                    mem_high            = mem_test;
+                    id_dense_start_high = id_dense_start_test;
+                    LOG_INF("%s: set ngl_per_device_high[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start_high=%zu\n",
+                        __func__, id, ngl_per_device_high[id].n_layer, ngl_per_device_high[id].n_part, id_dense_start_high);
+                }
+                assert(ngl_per_device_high[id].n_full() >= ngl_per_device[id].n_full());
+                delta = ngl_per_device_high[id].n_full() - ngl_per_device[id].n_full();
+            }
+        } else {
+            ngl_per_device = ngl_per_device_high;
+            mem            = mem_high;
+            id_dense_start = id_dense_start_high;
+            LOG_INF("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
+                __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+        }
+
+        // try to fit at least part of one more layer
+        if (ngl_per_device[id_dense_start].n_layer > (id < nd - 1 ? 0 : 1)) {
+            std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+            size_t id_dense_start_test = id_dense_start;
+            ngl_per_device_test[id_dense_start_test].n_layer--;
+            ngl_per_device_test[id_dense_start_test].n_part--;
+            ngl_per_device_test[id].n_layer++;
+            ngl_per_device_test[id].n_part++;
+            if (ngl_per_device_test[id_dense_start_test].n_part == 0) {
+                id_dense_start_test++;
+            }
+            ngl_per_device_test[id].overflow_type = LAYER_FRACTION_UP;
+            std::vector<ggml_backend_buffer_type_t> overflow_bufts_test = overflow_bufts;
+            if (id < nd - 1) {
+                overflow_bufts_test[id] = ggml_backend_dev_buffer_type(devs[id + 1]);
+            }
+            LOG_INF("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_UP\n", __func__);
+            std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+            if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
+                ngl_per_device = ngl_per_device_test;
+                overflow_bufts = overflow_bufts_test;
+                mem            = mem_test;
+                id_dense_start = id_dense_start_test;
+                LOG_INF("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", UP), id_dense_start=%zu\n",
+                    __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+
+                ngl_per_device_test[id].overflow_type = LAYER_FRACTION_GATE;
+                LOG_INF("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_GATE\n", __func__);
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+                if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
+                    ngl_per_device = ngl_per_device_test;
+                    overflow_bufts = overflow_bufts_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LOG_INF("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", GATE), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                }
+            } else {
+                ngl_per_device_test[id].overflow_type = LAYER_FRACTION_ATTN;
+                LOG_INF("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_ATTN\n", __func__);
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+                if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
+                    ngl_per_device = ngl_per_device_test;
+                    overflow_bufts = overflow_bufts_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LOG_INF("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", ATTN), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                }
+            }
+        }
+
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LOG_INF(
+            "%s:   - %s: %2" PRIu32 " layers (%2" PRIu32 " overflowing), %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
+    }
+
+    // print info for devices that were not changed during the conversion from dense only to full layers:
+    for (size_t id = id_dense_start + 1; id < nd; id++) {
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LOG_INF(
+            "%s:   - %s: %2" PRIu32 " layers (%2" PRIu32 " overflowing), %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
+    }
+
+    set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams);
+}
+
+enum common_params_fit_status common_fit_params(
+        const char * path_model,
+        llama_model_params * mparams,
+        llama_context_params * cparams,
+        float * tensor_split,
+        llama_model_tensor_buft_override * tensor_buft_overrides,
+        size_t * margins,
+        uint32_t n_ctx_min,
+        ggml_log_level log_level) {
+    const int64_t t0_us = llama_time_us();
+    common_params_fit_status status = COMMON_PARAMS_FIT_STATUS_SUCCESS;
+    try {
+        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, log_level);
+        LOG_INF("%s: successfully fit params to free device memory\n", __func__);
+    } catch (const common_params_fit_exception & e) {
+        LOG_WRN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());
+        status = COMMON_PARAMS_FIT_STATUS_FAILURE;
+    } catch (const std::runtime_error & e) {
+        LOG_ERR("%s: encountered an error while trying to fit params to free device memory: %s\n", __func__, e.what());
+        status = COMMON_PARAMS_FIT_STATUS_ERROR;
+    }
+    const int64_t t1_us = llama_time_us();
+    LOG_INF("%s: fitting params to free memory took %.2f seconds\n", __func__, (t1_us - t0_us) * 1e-6);
+    return status;
+}
+
+void common_memory_breakdown_print(const struct llama_context * ctx) {
+    //const auto & devices = ctx->get_model().devices;
+    const auto * model = llama_get_model(ctx);
+
+    std::vector<ggml_backend_dev_t> devices;
+    for (int i = 0; i < llama_model_n_devices(model); i++) {
+        devices.push_back(llama_model_get_device(model, i));
+    }
+
+    llama_memory_breakdown memory_breakdown = llama_get_memory_breakdown(ctx);
+
+    std::vector<std::array<std::string, 9>> table_data;
+    table_data.reserve(devices.size());
+    const std::string template_header = "%s: | %s | %s   %s    %s   %s   %s   %s    %s |\n";
+    const std::string template_gpu    = "%s: | %s | %s = %s + (%s = %s + %s + %s) + %s |\n";
+    const std::string template_other  = "%s: | %s | %s   %s    %s = %s + %s + %s    %s |\n";
+
+    table_data.push_back({template_header, "memory breakdown [MiB]", "total", "free", "self", "model", "context", "compute", "unaccounted"});
+
+    constexpr size_t MiB = 1024 * 1024;
+    const std::vector<std::string> desc_prefixes_strip = {"NVIDIA ", "GeForce ", "Tesla ", "AMD ", "Radeon ", "Instinct "};
+
+    // track seen buffer types to avoid double counting:
+    std::set<ggml_backend_buffer_type_t> seen_buffer_types;
+
+    // accumulative memory breakdown for each device and for host:
+    std::vector<llama_memory_breakdown_data> mb_dev(devices.size());
+    llama_memory_breakdown_data              mb_host;
+
+    for (const auto & buft_mb : memory_breakdown) {
+        ggml_backend_buffer_type_t          buft = buft_mb.first;
+        const llama_memory_breakdown_data & mb   = buft_mb.second;
+        if (ggml_backend_buft_is_host(buft)) {
+            mb_host.model   += mb.model;
+            mb_host.context += mb.context;
+            mb_host.compute += mb.compute;
+            seen_buffer_types.insert(buft);
+            continue;
+        }
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (dev) {
+            int i_dev = -1;
+            for (size_t i = 0; i < devices.size(); i++) {
+                if (devices[i] == dev) {
+                    i_dev = i;
+                    break;
+                }
+            }
+            if (i_dev != -1) {
+                mb_dev[i_dev].model   += mb.model;
+                mb_dev[i_dev].context += mb.context;
+                mb_dev[i_dev].compute += mb.compute;
+                seen_buffer_types.insert(buft);
+                continue;
+            }
+        }
+    }
+
+    // print memory breakdown for each device:
+    for (size_t i = 0; i < devices.size(); i++) {
+        ggml_backend_dev_t dev = devices[i];
+        llama_memory_breakdown_data mb = mb_dev[i];
+
+        const std::string name = ggml_backend_dev_name(dev);
+        std::string desc = ggml_backend_dev_description(dev);
+        for (const std::string & prefix : desc_prefixes_strip) {
+            if (desc.length() >= prefix.length() && desc.substr(0, prefix.length()) == prefix) {
+                desc = desc.substr(prefix.length());
+            }
+        }
+
+        size_t free, total;
+        ggml_backend_dev_memory(dev, &free, &total);
+
+        const size_t self = mb.model + mb.context + mb.compute;
+        const int64_t unaccounted = static_cast<int64_t>(total) - static_cast<int64_t>(free) - static_cast<int64_t>(self);
+
+        table_data.push_back({
+            template_gpu,
+            "  - " + name + " (" + desc + ")",
+            std::to_string(total / MiB),
+            std::to_string(free / MiB),
+            std::to_string(self / MiB),
+            std::to_string(mb.model / MiB),
+            std::to_string(mb.context / MiB),
+            std::to_string(mb.compute / MiB),
+            std::to_string(unaccounted / static_cast<int64_t>(MiB))});
+    }
+
+    // print memory breakdown for host:
+    {
+        const size_t self = mb_host.model + mb_host.context + mb_host.compute;
+        table_data.push_back({
+            template_other,
+            "  - Host",
+            "", // total
+            "", // free
+            std::to_string(self / MiB),
+            std::to_string(mb_host.model / MiB),
+            std::to_string(mb_host.context / MiB),
+            std::to_string(mb_host.compute / MiB),
+            ""}); // unaccounted
+    }
+
+    // print memory breakdown for all remaining buffer types:
+    for (const auto & buft_mb : memory_breakdown) {
+        ggml_backend_buffer_type_t          buft = buft_mb.first;
+        const llama_memory_breakdown_data & mb   = buft_mb.second;
+        if (seen_buffer_types.count(buft) == 1) {
+            continue;
+        }
+        const std::string name = ggml_backend_buft_name(buft);
+        const size_t self = mb.model + mb.context + mb.compute;
+        table_data.push_back({
+            template_other,
+            "  - " + name,
+            "", // total
+            "", // free
+            std::to_string(self / MiB),
+            std::to_string(mb.model / MiB),
+            std::to_string(mb.context / MiB),
+            std::to_string(mb.compute / MiB),
+            ""}); // unaccounted
+        seen_buffer_types.insert(buft);
+    }
+
+    for (size_t j = 1; j < table_data[0].size(); j++) {
+        size_t max_len = 0;
+        for (const auto & td : table_data) {
+            max_len = std::max(max_len, td[j].length());
+        }
+        for (auto & td : table_data) {
+            td[j].insert(j == 1 ? td[j].length() : 0, max_len - td[j].length(), ' ');
+        }
+    }
+    for (const auto & td : table_data) {
+        LOG_INF(td[0].c_str(),
+            __func__, td[1].c_str(), td[2].c_str(), td[3].c_str(), td[4].c_str(), td[5].c_str(),
+            td[6].c_str(), td[7].c_str(), td[8].c_str());
+    }
+}
+
+void common_fit_print(
+        const char * path_model,
+        llama_model_params * mparams,
+        llama_context_params * cparams) {
+    std::vector<ggml_backend_dev_t> devs;
+    uint32_t hp_ngl = 0; // hparams.n_gpu_layers
+    uint32_t hp_nct = 0; // hparams.n_ctx_train
+    uint32_t hp_nex = 0; // hparams.n_expert
+
+    auto dmd = common_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, GGML_LOG_LEVEL_ERROR);
+    GGML_ASSERT(dmd.size() == devs.size() + 1);
+
+    for (size_t id = 0; id < devs.size(); id++) {
+        printf("%s ",  ggml_backend_dev_name(devs[id]));
+        printf("%zu ", dmd[id].mb.model/1024/1024);
+        printf("%zu ", dmd[id].mb.context/1024/1024);
+        printf("%zu ", dmd[id].mb.compute/1024/1024);
+        printf("\n");
+    }
+
+    printf("Host ");
+    printf("%zu ", dmd.back().mb.model/1024/1024);
+    printf("%zu ", dmd.back().mb.context/1024/1024);
+    printf("%zu ", dmd.back().mb.compute/1024/1024);
+    printf("\n");
+}

common/fit.h

@@ -0,0 +1,32 @@
+#pragma once
+
+#include "ggml.h"
+
+enum common_params_fit_status {
+    COMMON_PARAMS_FIT_STATUS_SUCCESS = 0, // found allocations that are projected to fit
+    COMMON_PARAMS_FIT_STATUS_FAILURE = 1, // could not find allocations that are projected to fit
+    COMMON_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occurred, e.g. because no model could be found at the specified path
+};
+
+// fits mparams and cparams to free device memory (assumes system memory is unlimited)
+//   - returns true if the parameters could be successfully modified to fit device memory
+//   - this function is NOT thread safe because it modifies the global llama logger state
+//   - only parameters that have the same value as in llama_default_model_params are modified
+//     with the exception of the context size which is modified if and only if equal to 0
+enum common_params_fit_status common_fit_params(
+                               const char   * path_model,
+                struct llama_model_params   * mparams,
+                struct llama_context_params * cparams,
+                                      float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements
+    struct llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
+                                     size_t * margins,               // margins of memory to leave per device in bytes
+                                   uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
+                        enum ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
+
+// print estimated memory to stdout
+void common_fit_print(
+                               const char   * path_model,
+                struct llama_model_params   * mparams,
+                struct llama_context_params * cparams);
+
+void common_memory_breakdown_print(const struct llama_context * ctx);

common/hf-cache.cpp

@@ -57,7 +57,7 @@ static fs::path get_cache_directory() {
 #ifndef _WIN32
         const struct passwd * pw = getpwuid(getuid());
 
-        if (pw->pw_dir && *pw->pw_dir) {
+        if (pw && pw->pw_dir && *pw->pw_dir) {
             return fs::path(pw->pw_dir) / ".cache" / "huggingface" / "hub";
         }
 #endif
@@ -230,7 +230,7 @@ static nl::json api_get(const std::string & url,
 static std::string get_repo_commit(const std::string & repo_id,
                                    const std::string & token) {
     try {
-        auto endpoint = get_model_endpoint();
+        auto endpoint = common_get_model_endpoint();
         auto json = api_get(endpoint + "api/models/" + repo_id + "/refs", token);
 
         if (!json.is_object() ||
@@ -308,7 +308,7 @@ hf_files get_repo_files(const std::string & repo_id,
     hf_files files;
 
     try {
-        auto endpoint = get_model_endpoint();
+        auto endpoint = common_get_model_endpoint();
         auto json = api_get(endpoint + "api/models/" + repo_id + "/tree/" + commit + "?recursive=true", token);
 
         if (!json.is_array()) {

common/jinja/caps.cpp

@@ -1,4 +1,3 @@
-#include "log.h"
 #include "value.h"
 #include "runtime.h"
 #include "caps.h"

common/jinja/runtime.h

@@ -106,10 +106,16 @@ struct statement {
     size_t pos; // position in source, for debugging
     virtual ~statement() = default;
     virtual std::string type() const { return "Statement"; }
+
     // execute_impl must be overridden by derived classes
-    virtual value execute_impl(context &) { throw std::runtime_error("cannot exec " + type()); }
+    virtual value execute_impl(context &) { throw_exec_error(); }
     // execute is the public method to execute a statement with error handling
     value execute(context &);
+
+private:
+    [[noreturn]] void throw_exec_error() const {
+        throw std::runtime_error("cannot exec " + type());
+    }
 };
 
 // Type Checking Utilities
@@ -143,7 +149,7 @@ struct program : public statement {
     program() = default;
     explicit program(statements && body) : body(std::move(body)) {}
     std::string type() const override { return "Program"; }
-    value execute_impl(context &) override {
+    [[noreturn]] value execute_impl(context &) override {
         throw std::runtime_error("Cannot execute program directly, use jinja::runtime instead");
     }
 };
@@ -195,7 +201,7 @@ struct break_statement : public statement {
         }
     };
 
-    value execute_impl(context &) override {
+    [[noreturn]] value execute_impl(context &) override {
         throw break_statement::signal();
     }
 };
@@ -209,7 +215,7 @@ struct continue_statement : public statement {
         }
     };
 
-    value execute_impl(context &) override {
+    [[noreturn]] value execute_impl(context &) override {
         throw continue_statement::signal();
     }
 };
@@ -509,7 +515,7 @@ struct slice_expression : public expression {
         chk_type<expression>(this->step_expr);
     }
     std::string type() const override { return "SliceExpression"; }
-    value execute_impl(context &) override {
+    [[noreturn]] value execute_impl(context &) override {
         throw std::runtime_error("must be handled by MemberExpression");
     }
 };

common/jinja/value.cpp

@@ -590,6 +590,10 @@ static bool string_endswith(const std::string & str, const std::string & suffix)
     return str.compare(str.length() - suffix.length(), suffix.length(), suffix) == 0;
 }
 
+[[noreturn]] static value string_join_not_implemented(const func_args &) {
+    throw not_implemented_exception("String join builtin not implemented");
+}
+
 const func_builtins & value_string_t::get_builtins() const {
     static const func_builtins builtins = {
         {"default", default_value},
@@ -851,9 +855,7 @@ const func_builtins & value_string_t::get_builtins() const {
             res->val_str.mark_input_based_on(val_input->as_string());
             return res;
         }},
-        {"join", [](const func_args &) -> value {
-            throw not_implemented_exception("String join builtin not implemented");
-        }},
+        {"join", string_join_not_implemented},
     };
     return builtins;
 }
@@ -884,6 +886,9 @@ const func_builtins & value_bool_t::get_builtins() const {
     return builtins;
 }
 
+[[noreturn]] static value array_unique_not_implemented(const func_args &) {
+    throw not_implemented_exception("Array unique builtin not implemented");
+}
 
 const func_builtins & value_array_t::get_builtins() const {
     static const func_builtins builtins = {
@@ -1084,13 +1089,14 @@ const func_builtins & value_array_t::get_builtins() const {
             std::reverse(arr.begin(), arr.end());
             return is_val<value_tuple>(val) ? mk_val<value_tuple>(std::move(arr)) : mk_val<value_array>(std::move(arr));
         }},
-        {"unique", [](const func_args &) -> value {
-            throw not_implemented_exception("Array unique builtin not implemented");
-        }},
+        {"unique", array_unique_not_implemented},
     };
     return builtins;
 }
 
+[[noreturn]] static value object_join_not_implemented(const func_args &) {
+    throw not_implemented_exception("object join not implemented");
+}
 
 const func_builtins & value_object_t::get_builtins() const {
     if (!has_builtins) {
@@ -1183,9 +1189,7 @@ const func_builtins & value_object_t::get_builtins() const {
             });
             return result;
         }},
-        {"join", [](const func_args &) -> value {
-            throw not_implemented_exception("object join not implemented");
-        }},
+        {"join", object_join_not_implemented},
     };
     return builtins;
 }

common/jinja/value.h

@@ -129,27 +129,25 @@ struct value_t {
     // Note: only for debugging and error reporting purposes
     virtual std::string type() const { return ""; }
 
-    virtual int64_t as_int() const { throw std::runtime_error(type() + " is not an int value"); }
-    virtual double as_float() const { throw std::runtime_error(type() + " is not a float value"); }
-    virtual string as_string() const { throw std::runtime_error(type() + " is not a string value"); }
-    virtual bool as_bool() const { throw std::runtime_error(type() + " is not a bool value"); }
-    virtual const std::vector<value> & as_array() const { throw std::runtime_error(type() + " is not an array value"); }
-    virtual const std::vector<std::pair<value, value>> & as_ordered_object() const { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value invoke(const func_args &) const { throw std::runtime_error(type() + " is not a function value"); }
+    virtual int64_t as_int() const { throw_type_error("is not an int value"); }
+    virtual double as_float() const { throw_type_error("is not a float value"); }
+    virtual string as_string() const { throw_type_error("is not a string value"); }
+    virtual bool as_bool() const { throw_type_error("is not a bool value"); }
+    virtual const std::vector<value> & as_array() const { throw_type_error("is not an array value"); }
+    virtual const std::vector<std::pair<value, value>> & as_ordered_object() const { throw_type_error("is not an object value"); }
+    virtual value invoke(const func_args &) const { throw_type_error("is not a function value"); }
     virtual bool is_none() const { return false; }
     virtual bool is_undefined() const { return false; }
-    virtual const func_builtins & get_builtins() const {
-        throw std::runtime_error("No builtins available for type " + type());
-    }
+    virtual const func_builtins & get_builtins() const { throw_type_error("has no builtins"); }
 
-    virtual bool has_key(const value &) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual void insert(const value & /* key */, const value & /* val */) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value & at(const value & /* key */, value & /* default_val */) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value & at(const value & /* key */) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value & at(const std::string & /* key */, value & /* default_val */) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value & at(const std::string & /* key */) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value & at(int64_t /* idx */, value & /* default_val */) { throw std::runtime_error(type() + " is not an array value"); }
-    virtual value & at(int64_t /* idx */) { throw std::runtime_error(type() + " is not an array value"); }
+    virtual bool has_key(const value &) { throw_type_error("is not an object value"); }
+    virtual void insert(const value & /* key */, const value & /* val */) { throw_type_error("is not an object value"); }
+    virtual value & at(const value & /* key */, value & /* default_val */) { throw_type_error("is not an object value"); }
+    virtual value & at(const value & /* key */) { throw_type_error("is not an object value"); }
+    virtual value & at(const std::string & /* key */, value & /* default_val */) { throw_type_error("is not an object value"); }
+    virtual value & at(const std::string & /* key */) { throw_type_error("is not an object value"); }
+    virtual value & at(int64_t /* idx */, value & /* default_val */) { throw_type_error("is not an array value"); }
+    virtual value & at(int64_t /* idx */) { throw_type_error("is not an array value"); }
 
     virtual bool is_numeric() const { return false; }
     virtual bool is_hashable() const { return false; }
@@ -163,6 +161,11 @@ struct value_t {
     // Note: only for debugging purposes
     virtual std::string as_repr() const { return as_string().str(); }
 
+private:
+    [[noreturn]] void throw_type_error(const char* expected) const {
+        throw std::runtime_error(type() + " " + expected);
+    }
+
 protected:
     virtual bool equivalent(const value_t &) const = 0;
     virtual bool nonequal(const value_t & other) const { return !equivalent(other); }

common/log.cpp

@@ -49,7 +49,7 @@ enum common_log_col : int {
 };
 
 // disable colors by default
-static std::vector<const char *> g_col = {
+static const char* g_col[] = {
     "",
     "",
     "",
@@ -247,7 +247,6 @@ public:
 
             entries = std::move(new_entries);
         }
-
         cv.notify_one();
     }
 
@@ -265,7 +264,6 @@ public:
                 {
                     std::unique_lock<std::mutex> lock(mtx);
                     cv.wait(lock, [this]() { return head != tail; });
-
                     cur = entries[head];
 
                     head = (head + 1) % entries.size();
@@ -301,7 +299,6 @@ public:
 
                 tail = (tail + 1) % entries.size();
             }
-
             cv.notify_one();
         }
 
@@ -338,7 +335,7 @@ public:
             g_col[COMMON_LOG_COL_CYAN]    = LOG_COL_CYAN;
             g_col[COMMON_LOG_COL_WHITE]   = LOG_COL_WHITE;
         } else {
-            for (size_t i = 0; i < g_col.size(); i++) {
+            for (size_t i = 0; i < std::size(g_col); i++) {
                 g_col[i] = "";
             }
         }
@@ -368,14 +365,20 @@ struct common_log * common_log_init() {
 }
 
 struct common_log * common_log_main() {
-    static struct common_log log;
+    // We intentionally leak (i.e. do not delete) the logger singleton because
+    // common_log destructor called at DLL teardown phase will cause hanging on Windows.
+    // OS will release resources anyway so it should not be a significant issue,
+    // though this design may cause logs to be lost if not flushed before the program exits.
+    // Refer to https://github.com/ggml-org/llama.cpp/issues/22142 for details.
+    static struct common_log * log;
     static std::once_flag    init_flag;
     std::call_once(init_flag, [&]() {
+        log = new common_log;
         // Set default to auto-detect colors
-        log.set_colors(tty_can_use_colors());
+        log->set_colors(tty_can_use_colors());
     });
 
-    return &log;
+    return log;
 }
 
 void common_log_pause(struct common_log * log) {

common/log.h

@@ -49,7 +49,11 @@ void common_log_default_callback(enum ggml_log_level level, const char * text, v
 struct common_log;
 
 struct common_log * common_log_init();
-struct common_log * common_log_main(); // singleton, automatically destroys itself on exit
+
+// Singleton, intentionally leaked to avoid Windows teardown hangs.
+// Call common_log_flush() before exit if you want to ensure all logs are flushed.
+struct common_log * common_log_main();
+
 void                common_log_pause (struct common_log * log); // pause  the worker thread, not thread-safe
 void                common_log_resume(struct common_log * log); // resume the worker thread, not thread-safe
 void                common_log_free  (struct common_log * log);

common/ngram-map.cpp

@@ -208,7 +208,7 @@ void common_ngram_map_begin(
                 count_keys, count_keys_del, count_values_del, count_map_entries_upd);
     }
 
-    map.idx_last_check = (map.size_last_begin > 0) ? map.size_last_begin - 1 : 0;
+    map.idx_last_check = size_begin;
     map.size_last_begin = size_begin;
 }
 
@@ -231,7 +231,7 @@ void common_ngram_map_draft(common_ngram_map & map,
         GGML_ABORT("%s: cur_len exceeds UINT32_MAX: %zu", __func__, cur_len);
     }
 
-    if (map.idx_last_check  > cur_len) {
+    if (map.idx_last_check > cur_len) {
         // Should not happen because of common_ngram_map_begin().
         GGML_ABORT("%s: map.idx_last_check > cur_len: %zu > %zu", __func__, map.idx_last_check, cur_len);
     }
@@ -386,7 +386,7 @@ void common_ngram_map_draft(common_ngram_map & map,
         LOG_DBG("%s: key_idx = %zu, key_offset = %zu, key_num = %d, draft.size = %zu\n", __func__,
                 curr_key.key_idx, key_offset, curr_key.key_num, draft.size());
 
-        map.last_draft_created   = false;
+        map.last_draft_created   = true;
         map.last_draft_key_idx   = key_offset;
         map.last_draft_value_idx = 0; // value 0 is used for simple mode
         return;
@@ -524,7 +524,7 @@ void common_ngram_map_accept(common_ngram_map & map, uint16_t n_accepted) {
     struct common_ngram_map_value & curr_value = curr_key.values[val_idx]; // value used for draft generation.
 
     // update the value statistics
-    LOG_INF("common_ngram_map_send_accepted: n_accepted = %d, prev value_num = %d\n",
+    LOG_DBG("common_ngram_map_send_accepted: n_accepted = %d, prev value_num = %d\n",
             n_accepted, curr_value.n_accepted);
     curr_value.n_accepted = n_accepted;
 }

common/preset.cpp

@@ -43,7 +43,7 @@ static std::set<std::string> get_remote_preset_whitelist(const std::map<std::str
     for (const auto & it : key_to_opt) {
         const std::string & key = it.first;
         const common_arg & opt = it.second;
-        if (allowed_options.find(key) != allowed_options.end() || opt.is_sparam) {
+        if (allowed_options.find(key) != allowed_options.end() || opt.is_sampling) {
             allowed_keys.insert(key);
             // also add variant keys (args without leading dashes and env vars)
             for (const auto & arg : opt.get_args()) {

common/reasoning-budget.cpp

@@ -122,6 +122,20 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to
             }
             break;
         case REASONING_BUDGET_DONE:
+            // Re-arm on a new start tag: some models emit multiple <think> blocks
+            // per response, and each should get a fresh budget window.
+            if (ctx->start_matcher.advance(token)) {
+                ctx->state = REASONING_BUDGET_COUNTING;
+                ctx->remaining = ctx->budget;
+                ctx->end_matcher.reset();
+                LOG_INF("reasoning-budget: re-activated on new start tag, budget=%d tokens\n", ctx->budget);
+
+                if (ctx->remaining <= 0) {
+                    ctx->state = REASONING_BUDGET_FORCING;
+                    ctx->force_pos = 0;
+                    LOG_INF("reasoning-budget: budget=0, forcing immediately\n");
+                }
+            }
             break;
     }
 }
@@ -218,34 +232,6 @@ static struct llama_sampler * common_reasoning_budget_init_state(
     );
 }
 
-struct llama_sampler * common_reasoning_budget_init(
-        const struct llama_vocab       * vocab,
-        const std::vector<llama_token> & start_tokens,
-        const std::vector<llama_token> & end_tokens,
-        const std::vector<llama_token> & forced_tokens,
-        int32_t                          budget,
-        const std::vector<llama_token> & prefill_tokens) {
-    // Determine initial state from prefill: COUNTING if the prefill begins with
-    // the start sequence but does not also contain the end sequence after it.
-    common_reasoning_budget_state initial_state = REASONING_BUDGET_IDLE;
-    if (!prefill_tokens.empty() && !start_tokens.empty() &&
-            prefill_tokens.size() >= start_tokens.size() &&
-            std::equal(start_tokens.begin(), start_tokens.end(), prefill_tokens.begin())) {
-        initial_state = REASONING_BUDGET_COUNTING;
-        // If the end sequence also follows the start in the prefill, reasoning
-        // was opened and immediately closed — stay IDLE.
-        if (!end_tokens.empty() &&
-                prefill_tokens.size() >= start_tokens.size() + end_tokens.size()) {
-            auto end_start = prefill_tokens.end() - (ptrdiff_t) end_tokens.size();
-            if (end_start >= prefill_tokens.begin() + (ptrdiff_t) start_tokens.size() &&
-                    std::equal(end_tokens.begin(), end_tokens.end(), end_start)) {
-                initial_state = REASONING_BUDGET_IDLE;
-            }
-        }
-    }
-    return common_reasoning_budget_init_state(vocab, start_tokens, end_tokens, forced_tokens, budget, initial_state);
-}
-
 struct llama_sampler * common_reasoning_budget_init(
         const struct llama_vocab       * vocab,
         const std::vector<llama_token> & start_tokens,

common/reasoning-budget.h

@@ -29,10 +29,7 @@ enum common_reasoning_budget_state {
 //   end_tokens     - token sequence for natural deactivation
 //   forced_tokens  - token sequence forced when budget expires
 //   budget         - max tokens allowed in the reasoning block
-//   prefill_tokens - tokens already present in the prompt (generation prompt);
-//                    used to determine the initial state: COUNTING if they begin
-//                    with start_tokens (but don't also end with end_tokens),
-//                    IDLE otherwise. COUNTING with budget <= 0 is promoted to FORCING.
+//   initial_state  - initial state
 //
 struct llama_sampler * common_reasoning_budget_init(
         const struct llama_vocab       * vocab,
@@ -40,16 +37,6 @@ struct llama_sampler * common_reasoning_budget_init(
         const std::vector<llama_token> & end_tokens,
         const std::vector<llama_token> & forced_tokens,
         int32_t                          budget,
-        const std::vector<llama_token> & prefill_tokens = {});
-
-// Variant that takes an explicit initial state (used by tests and clone).
-// COUNTING with budget <= 0 is promoted to FORCING.
-struct llama_sampler * common_reasoning_budget_init(
-        const struct llama_vocab       * vocab,
-        const std::vector<llama_token> & start_tokens,
-        const std::vector<llama_token> & end_tokens,
-        const std::vector<llama_token> & forced_tokens,
-        int32_t                          budget,
-        common_reasoning_budget_state    initial_state);
+        common_reasoning_budget_state    initial_state = REASONING_BUDGET_IDLE);
 
 common_reasoning_budget_state common_reasoning_budget_get_state(const struct llama_sampler * smpl);

common/sampling.cpp

@@ -1,10 +1,12 @@
 #include "sampling.h"
 
 #include "common.h"
-#include "ggml.h"
+#include "fit.h"
 #include "log.h"
 #include "reasoning-budget.h"
 
+#include "ggml.h"
+
 #include <algorithm>
 #include <cctype>
 #include <climits>
@@ -258,32 +260,35 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
         }
     }
 
+    // Compute prefill tokens from the generation prompt
+    std::vector<llama_token> prefill_tokens;
+    if (!params.generation_prompt.empty()) {
+        GGML_ASSERT(vocab != nullptr);
+        auto tokens = common_tokenize(vocab, params.generation_prompt, false, true);
+        for (size_t i = 0; i < tokens.size(); i++) {
+            std::string piece = common_token_to_piece(vocab, tokens[i], true);
+            if (i == 0 && std::isspace(piece[0]) && !std::isspace(params.generation_prompt[0])) {
+                // Some tokenizers will add a space before the first special token, need to exclude
+                continue;
+            }
+            LOG_DBG("%s: prefill token: %d = %s\n", __func__, tokens[i], piece.c_str());
+            prefill_tokens.push_back(tokens[i]);
+        }
+    }
+
     // Feed generation prompt tokens to the grammar sampler so it advances past
     // tokens the template already placed in the prompt.
     // Only applies to output-format and tool-call grammars; user-supplied grammars must not be prefilled.
-    std::vector<llama_token> prefill_tokens;
-    if (!params.generation_prompt.empty() && common_grammar_needs_prefill(params.grammar)) {
-        GGML_ASSERT(vocab != nullptr);
-        prefill_tokens = common_tokenize(vocab, params.generation_prompt, false, true);
-        if (!prefill_tokens.empty()) {
-            std::string first_token = common_token_to_piece(vocab, prefill_tokens[0], true);
-            if (std::isspace(first_token[0]) && !std::isspace(params.generation_prompt[0])) {
-                // Some tokenizers will add a space before the first special token, need to remove
-                prefill_tokens = std::vector<llama_token>(prefill_tokens.begin() + 1, prefill_tokens.end());
-            }
-        }
-
-        if (grmr && !params.grammar_lazy) {
-            try {
-                for (const auto & token : prefill_tokens) {
-                    llama_sampler_accept(grmr, token);
-                    LOG_DBG("%s: accepted prefill token (%d)\n", __func__, token);
-                }
-            } catch (std::exception &e) {
-                LOG_ERR("%s: error initializing grammar sampler for grammar:\n%s\n\nGeneration prompt:\n'%s'\n", __func__,
-                    common_grammar_value(params.grammar).c_str(), params.generation_prompt.c_str());
-                throw e;
+    if (grmr && !params.grammar_lazy && common_grammar_needs_prefill(params.grammar)) {
+        try {
+            for (const auto & token : prefill_tokens) {
+                llama_sampler_accept(grmr, token);
+                LOG_DBG("%s: grammar accepted prefill token (%d)\n", __func__, token);
             }
+        } catch (std::exception &e) {
+            LOG_ERR("%s: error initializing grammar sampler for grammar:\n%s\n\nGeneration prompt:\n'%s'\n", __func__,
+                common_grammar_value(params.grammar).c_str(), params.generation_prompt.c_str());
+            throw e;
         }
     }
 
@@ -294,8 +299,12 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
             params.reasoning_budget_start,
             params.reasoning_budget_end,
             params.reasoning_budget_forced,
-            params.reasoning_budget_tokens < 0 ? INT_MAX : params.reasoning_budget_tokens,
-            prefill_tokens);
+            params.reasoning_budget_tokens < 0 ? INT_MAX : params.reasoning_budget_tokens);
+
+        for (const auto & token : prefill_tokens) {
+            llama_sampler_accept(rbudget, token);
+            LOG_DBG("%s: reasoning-budget accepted prefill token (%d)\n", __func__, token);
+        }
     }
 
     if (params.has_logit_bias()) {
@@ -429,7 +438,7 @@ static bool grammar_should_apply(struct common_sampler * gsmpl) {
     return true;
 }
 
-void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
+void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool is_generated) {
     if (!gsmpl) {
         return;
     }
@@ -437,9 +446,11 @@ void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, boo
     const auto tm = gsmpl->tm();
 
     // grammar_should_apply() checks the reasoning budget state, so calculate this before we accept
-    accept_grammar = accept_grammar && grammar_should_apply(gsmpl);
+    const auto accept_grammar = is_generated && grammar_should_apply(gsmpl);
 
-    llama_sampler_accept(gsmpl->rbudget, token);
+    if (gsmpl->rbudget && is_generated) {
+        llama_sampler_accept(gsmpl->rbudget, token);
+    }
 
     if (gsmpl->grmr && accept_grammar) {
         llama_sampler_accept(gsmpl->grmr, token);
@@ -511,7 +522,7 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
         LOG_INF("%s: unaccounted time = %10.2f ms / %5.1f %%      (total - sampling - prompt eval - eval) / (total)\n", __func__, t_unacc_ms, t_unacc_pc);
         LOG_INF("%s:    graphs reused = %10d\n", __func__, data.n_reused);
 
-        llama_memory_breakdown_print(ctx);
+        common_memory_breakdown_print(ctx);
     }
 }
 

common/sampling.h

@@ -41,8 +41,8 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
 
 void common_sampler_free(struct common_sampler * gsmpl);
 
-// if accept_grammar is true, the token is accepted both by the sampling chain and the grammar
-void                    common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar);
+// if is_generated is true, the token is accepted by the sampling chain, the reasoning budget sampler, and the grammar sampler
+void                    common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool is_generated);
 void                    common_sampler_reset (struct common_sampler * gsmpl);
 struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
 

common/speculative.cpp

@@ -13,6 +13,7 @@
 #include <cstring>
 #include <iomanip>
 #include <map>
+#include <cinttypes>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
@@ -60,18 +61,26 @@ static bool common_speculative_are_compatible(
     LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
 
     if (vocab_type_tgt != vocab_type_dft) {
-        LOG_DBG("%s: draft model vocab type must match target model to use speculation but ", __func__);
-        LOG_DBG("vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
+        LOG_WRN("%s: draft model vocab type must match target model to use speculation but "
+                "vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
         return false;
     }
 
-    if (
-        llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
-        llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
-        llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft) ||
-        llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)
-    ) {
-        LOG_DBG("%s: draft model special tokens must match target model to use speculation\n", __func__);
+    if (llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
+        (llama_vocab_get_add_bos(vocab_tgt) && llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft))) {
+        LOG_WRN("%s: draft model bos tokens must match target model to use speculation. add: %d - %d, id: %d - %d)\n",
+                __func__,
+                llama_vocab_get_add_bos(vocab_tgt), llama_vocab_get_add_bos(vocab_dft),
+                llama_vocab_bos(vocab_tgt), llama_vocab_bos(vocab_dft));
+        return false;
+    }
+
+    if (llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
+        (llama_vocab_get_add_eos(vocab_tgt) && llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft))) {
+        LOG_WRN("%s: draft model eos tokens must match target model to use speculation. add: %d - %d, id: %d - %d)\n",
+                __func__,
+                llama_vocab_get_add_eos(vocab_tgt), llama_vocab_get_add_eos(vocab_dft),
+                llama_vocab_eos(vocab_tgt), llama_vocab_eos(vocab_dft));
         return false;
     }
 
@@ -142,12 +151,31 @@ struct common_speculative_state {
             llama_tokens & result) = 0;
 
     virtual void accept(uint16_t n_accepted) = 0;
+
+    virtual int32_t n_max(const common_params_speculative & params) const = 0;
+    virtual int32_t n_min(const common_params_speculative & params) const = 0;
+};
+
+struct common_speculative_checkpoint {
+    llama_pos pos_min  = 0;
+    llama_pos pos_max  = 0;
+
+    int64_t   n_tokens = 0;
+
+    std::vector<uint8_t> data;
+
+    size_t size() const {
+        return data.size();
+    }
 };
 
 struct common_speculative_state_draft : public common_speculative_state {
     llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
     llama_context * ctx_dft;
 
+    bool use_ckpt = false;
+    common_speculative_checkpoint ckpt;
+
     common_sampler * smpl;
 
     llama_batch  batch;
@@ -160,10 +188,12 @@ struct common_speculative_state_draft : public common_speculative_state {
             enum common_speculative_type type,
             llama_context * ctx_tgt,
             llama_context * ctx_dft,
-            const std::vector<std::pair<std::string, std::string>> & replacements)
+            const std::vector<std::pair<std::string, std::string>> & replacements,
+            bool use_ckpt)
         : common_speculative_state(type)
         , ctx_tgt(ctx_tgt)
         , ctx_dft(ctx_dft)
+        , use_ckpt(use_ckpt)
     {
         batch = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
         smpl = nullptr;
@@ -217,8 +247,39 @@ struct common_speculative_state_draft : public common_speculative_state {
         llama_batch_free(batch);
     }
 
-    void begin(const llama_tokens & prompt) override {
-        GGML_UNUSED(prompt);
+    void begin(const llama_tokens & /*prompt*/) override {
+    }
+
+    size_t create_checkpoint(int n_tokens_prompt) {
+        int slot_id = 0;
+        const size_t checkpoint_size = llama_state_seq_get_size_ext(ctx_dft, slot_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+
+        ckpt.pos_min  = llama_memory_seq_pos_min(llama_get_memory(ctx_dft), slot_id);
+        ckpt.pos_max  = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), slot_id);
+        ckpt.n_tokens = n_tokens_prompt;
+        ckpt.data.resize(checkpoint_size);
+
+        const size_t n = llama_state_seq_get_data_ext(ctx_dft, ckpt.data.data(), checkpoint_size, slot_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+        if (n != checkpoint_size) {
+            GGML_ABORT("checkpoint size mismatch: expected %zu, got %zu\n", checkpoint_size, n);
+        }
+
+        LOG_DBG("%s: pos_min = %d, pos_max = %d, size = %.3f MiB\n", __func__,
+                ckpt.pos_min, ckpt.pos_max, (float) ckpt.data.size() / 1024 / 1024);
+        return n;
+    }
+
+    size_t restore_checkpoint() {
+        int slot_id = 0;
+        LOG_DBG("%s: pos_min = %d, pos_max = %d\n", __func__, ckpt.pos_min, ckpt.pos_max);
+        const size_t n = llama_state_seq_set_data_ext(ctx_dft, ckpt.data.data(), ckpt.size(), slot_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+        if (n != ckpt.size()) {
+            GGML_ABORT("%s: failed to restore context checkpoint (pos_min=%d, pos_max=%d, size=%zu",
+                        __func__, ckpt.pos_min, ckpt.pos_max, ckpt.size());
+        }
+        llama_memory_seq_rm(llama_get_memory(ctx_dft), slot_id, ckpt.pos_max + 1, -1);
+
+        return n;
     }
 
     void draft(
@@ -226,6 +287,8 @@ struct common_speculative_state_draft : public common_speculative_state {
             const llama_tokens & prompt_tgt,
             llama_token id_last,
             llama_tokens & result) override {
+        const auto & sparams = params.draft;
+
         auto * spec = this;
 
         auto & batch      = spec->batch;
@@ -236,10 +299,10 @@ struct common_speculative_state_draft : public common_speculative_state {
 
         auto * mem_dft = llama_get_memory(ctx_dft);
 
-        int reuse_i = 0;
-        int reuse_n = 0;
+        int reuse_i = 0; // index of part to be reused in prompt_dft
+        int reuse_n = 0; // length of part to be reused in prompt_dft
 
-        const int n_ctx = llama_n_ctx(ctx_dft) - params.n_max;
+        const int n_ctx = llama_n_ctx(ctx_dft) - sparams.n_max;
 
         llama_tokens prompt_cnv;
         if (!spec->vocab_cmpt) {
@@ -271,13 +334,18 @@ struct common_speculative_state_draft : public common_speculative_state {
 
         const int i_start = std::max<int>(0, (int) prompt_cur.size() - n_ctx);
 
+        if (use_ckpt && i_start > 0) {
+            LOG_WRN("%s: context shift is not supported with checkpoint-based contexts - skipping\n", __func__);
+            return;
+        }
+
         // reuse as much as possible from the old draft context
         // ideally, the draft context should be as big as the target context and we will always reuse the entire prompt
         for (int i = 0; i < (int) prompt_dft.size(); ++i) {
             int cur = 0;
             while (i_start + cur < (int) prompt_cur.size() &&
-                    i       + cur < (int) prompt_dft.size() &&
-                    prompt_cur[i_start + cur] == prompt_dft[i + cur]) {
+                   i       + cur < (int) prompt_dft.size() &&
+                   prompt_cur[i_start + cur] == prompt_dft[i + cur]) {
                 cur++;
             }
 
@@ -285,24 +353,37 @@ struct common_speculative_state_draft : public common_speculative_state {
                 reuse_i = i;
                 reuse_n = cur;
             }
+
+            if (use_ckpt) {
+                break;
+            }
         }
 
-        LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size());
+        LOG_DBG("%s: reuse_i = %d, reuse_n = %d, #prompt_dft = %zu, #prompt_cur = %zu\n",
+                __func__, reuse_i, reuse_n, prompt_dft.size(), prompt_cur.size());
+        if (use_ckpt && ckpt.n_tokens > reuse_n) {
+            LOG_DBG("%s: checkpoint (n_tokens = %d) is outdated -> delete it\n", __func__, (int) ckpt.n_tokens);
+
+            reuse_i = 0;
+            reuse_n = 0;
+
+            ckpt = {};
+        }
 
         result.clear();
-        result.reserve(params.n_max);
+        result.reserve(sparams.n_max);
 
-        if (reuse_n == 0) {
+        if (reuse_n == 0 || (use_ckpt && reuse_i > 0)) {
             llama_memory_clear(mem_dft, false);
             prompt_dft.clear();
         } else {
             // this happens when a previous draft has been discarded (for example, due to being too small), but the
             // target model agreed with it. in this case, we simply pass back the previous results to save compute
-            if (reuse_i + reuse_n < (int) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
+            if (reuse_i + reuse_n < (int64_t) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
                 for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) {
                     result.push_back(prompt_dft[i]);
 
-                    if (params.n_max <= (int) result.size()) {
+                    if (sparams.n_max <= (int) result.size()) {
                         break;
                     }
                 }
@@ -311,15 +392,34 @@ struct common_speculative_state_draft : public common_speculative_state {
             }
 
             if (reuse_i > 0) {
-                llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
+                GGML_ASSERT(!use_ckpt);
+
+                bool is_removed = llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
+                if (!is_removed) {
+                    LOG_ERR("%s: llama_memory_seq_rm failed, reuse_i=%d\n", __func__, reuse_i);
+                    return;
+                }
                 llama_memory_seq_add(mem_dft, 0, reuse_i, -1, -reuse_i);
 
                 prompt_dft.erase(prompt_dft.begin(), prompt_dft.begin() + reuse_i);
             }
 
             if (reuse_n < (int) prompt_dft.size()) {
-                llama_memory_seq_rm (mem_dft, 0, reuse_n, -1);
-                prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
+                if (use_ckpt) {
+                    if (ckpt.n_tokens > 0) {
+                        LOG_DBG("%s: restoring checkpoint, reuse_n=%d, prompt_dft.size=%zu\n", __func__, reuse_n, prompt_dft.size());
+                        restore_checkpoint();
+                        reuse_n = ckpt.n_tokens;
+                        prompt_dft.resize(reuse_n);
+                    }
+                } else {
+                    const bool is_removed = llama_memory_seq_rm(mem_dft, 0, reuse_n, -1);
+                    if (!is_removed) {
+                        LOG_ERR("%s: llama_memory_seq_rm failed, reuse_n=%d, prompt_dft.size=%zu\n", __func__, reuse_n, prompt_dft.size());
+                        return;
+                    }
+                    prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
+                }
             }
         }
 
@@ -336,8 +436,17 @@ struct common_speculative_state_draft : public common_speculative_state {
         // we should rarely end-up here during normal decoding
         if (batch.n_tokens > 0) {
             //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
+            LOG_DBG("%s: draft prompt batch: %d tokens\n", __func__, batch.n_tokens);
 
-            llama_decode(ctx_dft, batch);
+            int ret = llama_decode(ctx_dft, batch);
+            if (ret != 0 && ret != 1) {
+                LOG_WRN("%s: llama_decode returned %d, prompt_cur.size=%zu\n",
+                        __func__, ret, prompt_cur.size());
+            }
+
+            if (use_ckpt) {
+                create_checkpoint(prompt_dft.size());
+            }
         }
 
         const llama_pos n_past = prompt_dft.size();
@@ -349,14 +458,18 @@ struct common_speculative_state_draft : public common_speculative_state {
 
         prompt_dft.push_back(id_last);
 
-        LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str());
+        //LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str());
 
-        llama_decode(ctx_dft, batch);
+        int ret = llama_decode(ctx_dft, batch);
+        if (ret != 0 && ret != 1) {
+            LOG_WRN("%s: llama_decode returned %d, prompt_cur.size=%zu, prompt_dft.size=%zu\n",
+                    __func__, ret, prompt_cur.size(), prompt_dft.size());
+        }
 
         common_sampler_reset(smpl);
 
         // sample n_draft tokens from the draft model
-        for (int i = 0; i < params.n_max; ++i) {
+        for (int i = 0; i < sparams.n_max; ++i) {
             common_batch_clear(batch);
 
             common_sampler_sample(smpl, ctx_dft, 0, true);
@@ -373,21 +486,25 @@ struct common_speculative_state_draft : public common_speculative_state {
 
             common_sampler_accept(smpl, id, true);
 
-            result.push_back(id);
-
-            if (params.n_max <= (int) result.size()) {
+            // only collect very high-confidence draft tokens
+            if (cur_p->data[0].p < sparams.p_min) {
                 break;
             }
 
-            // only collect very high-confidence draft tokens
-            if (cur_p->data[0].p < params.p_min) {
+            result.push_back(id);
+
+            if (sparams.n_max <= (int) result.size()) {
                 break;
             }
 
             common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
 
             // evaluate the drafted tokens on the draft model
-            llama_decode(ctx_dft, batch);
+            ret = llama_decode(ctx_dft, batch);
+            if (ret != 0) {
+                LOG_WRN("%s: llama_decode[%d] returned %d, prompt_cur.size=%zu, prompt_dft.size=%zu\n",
+                        __func__, i, ret, prompt_cur.size(), prompt_dft.size());
+            }
 
             prompt_dft.push_back(id);
         }
@@ -397,10 +514,14 @@ struct common_speculative_state_draft : public common_speculative_state {
             detokenized = replace_to_tgt(detokenized);
             LOG_DBG("draft->main detokenized string: '%s'\n", detokenized.c_str());
             result = common_tokenize(ctx_tgt, detokenized, false, true);
-            if (result.size() > (size_t)params.n_max) {
-                result.resize(params.n_max);
+            if (result.size() > (size_t) sparams.n_max) {
+                result.resize(sparams.n_max);
             }
         }
+
+        if (result.size() < (size_t) sparams.n_min) {
+            result.clear();
+        }
     }
 
     void accept(uint16_t n_accepted) override {
@@ -408,6 +529,14 @@ struct common_speculative_state_draft : public common_speculative_state {
         GGML_UNUSED(n_accepted);
     }
 
+    int32_t n_max(const common_params_speculative & params) const override {
+        return params.draft.n_max;
+    }
+
+    int32_t n_min(const common_params_speculative & params) const override {
+        return params.draft.n_min;
+    }
+
     std::string replace_to_dft(const std::string & input) const {
         std::string result = input;
 
@@ -460,6 +589,14 @@ struct common_speculative_state_eagle3 : public common_speculative_state {
         // noop
         GGML_UNUSED(n_accepted);
     }
+
+    int32_t n_max(const common_params_speculative & params) const override {
+        return params.draft.n_max;
+    }
+
+    int32_t n_min(const common_params_speculative & params) const override {
+        return params.draft.n_min;
+    }
 };
 
 // state of self-speculation (simple implementation, not ngram-map)
@@ -489,19 +626,27 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
         // noop
         GGML_UNUSED(n_accepted);
     }
+
+    int32_t n_max(const common_params_speculative & /*params*/) const override {
+        return config.size_mgram;
+    }
+
+    int32_t n_min(const common_params_speculative & /*params*/) const override {
+        return config.size_mgram;
+    }
 };
 
 struct common_speculative_state_ngram_map_k : public common_speculative_state {
     // draft ngram map for speculative decoding without draft model
-    common_ngram_map map;
+    common_ngram_map config;
 
     common_speculative_state_ngram_map_k(
             enum common_speculative_type type,
-            common_ngram_map map)
-        : common_speculative_state(type), map(std::move(map)) {}
+            common_ngram_map config)
+        : common_speculative_state(type), config(std::move(config)) {}
 
     void begin(const llama_tokens & prompt) override {
-        common_ngram_map_begin(map, prompt);
+        common_ngram_map_begin(config, prompt);
     }
 
     void draft(
@@ -509,12 +654,20 @@ struct common_speculative_state_ngram_map_k : public common_speculative_state {
             const llama_tokens & prompt_tgt,
             llama_token id_last,
             llama_tokens & result) override {
-        common_ngram_map_draft(map, prompt_tgt, id_last, result);
+        common_ngram_map_draft(config, prompt_tgt, id_last, result);
         GGML_UNUSED(params);
     }
 
     void accept(uint16_t n_accepted) override {
-        common_ngram_map_accept(map, n_accepted);
+        common_ngram_map_accept(config, n_accepted);
+    }
+
+    int32_t n_max(const common_params_speculative & /*params*/) const override {
+        return config.size_value;
+    }
+
+    int32_t n_min(const common_params_speculative & /*params*/) const override {
+        return config.size_value;
     }
 };
 
@@ -571,7 +724,7 @@ struct common_speculative_state_ngram_mod : public common_speculative_state {
             const llama_tokens & prompt_tgt,
             llama_token id_last,
             llama_tokens & result) override {
-        GGML_UNUSED(params);
+        const auto & sparams = params.ngram_mod;
 
         n_draft_last = 0;
 
@@ -591,16 +744,16 @@ struct common_speculative_state_ngram_mod : public common_speculative_state {
             i_last = cur_len - n;
         }
 
-        result.resize(n + params.n_max);
+        result.resize(n + sparams.n_max);
         for (size_t i = 0; i < n - 1; ++i) {
             result[i] = prompt_tgt[cur_len - n + 1 + i];
         }
         result[n - 1] = id_last;
 
-        for (int i = 0; i < params.n_max; ++i) {
+        for (int i = 0; i < sparams.n_max; ++i) {
             const llama_token token = mod.get(result.data() + i);
             if (token == common_ngram_mod::EMPTY) {
-                if (i < params.n_min) {
+                if (i < sparams.n_min) {
                     result.clear();
                     return;
                 }
@@ -622,26 +775,33 @@ struct common_speculative_state_ngram_mod : public common_speculative_state {
     }
 
     void accept(uint16_t n_accepted) override {
-        if (verbose) {
-            LOG_INF("%s: accepted %d tokens from %zu drafted tokens\n", __func__, n_accepted, n_draft_last);
-        }
-
         // compute acceptance fraction if we have a recorded draft length
         if (n_draft_last > 0) {
             const double f_acc = (double)n_accepted / (double)n_draft_last;
             if (f_acc < 0.5) {
                 n_low++;
                 if (n_low >= 3) {
-                    LOG_WRN("%s: low acceptance streak (%d) – resetting ngram_mod\n", __func__, n_low);
+                    if (verbose) {
+                        LOG_WRN("%s: low acceptance streak (%d) – resetting ngram_mod\n", __func__, n_low);
+                    }
 
                     mod.reset();
                     n_low = 0;
+                    i_last = 0;
                 }
             } else {
                 n_low = 0;
             }
         }
     }
+
+    int32_t n_max(const common_params_speculative & params) const override {
+        return params.ngram_mod.n_max;
+    }
+
+    int32_t n_min(const common_params_speculative & params) const override {
+        return params.ngram_mod.n_min;
+    }
 };
 
 struct common_speculative_state_ngram_cache : public common_speculative_state {
@@ -735,18 +895,29 @@ struct common_speculative_state_ngram_cache : public common_speculative_state {
         // TODO: noop
         GGML_UNUSED(n_accepted);
     }
+
+    int32_t n_max(const common_params_speculative & /*params*/) const override {
+        return n_draft;
+    }
+
+    int32_t n_min(const common_params_speculative & /*params*/) const override {
+        return 0;
+    }
 };
 
 struct common_speculative {
     std::vector<std::unique_ptr<common_speculative_state>> impls; // list of implementations to use and their states
+
     common_speculative_state * curr_impl = nullptr; // current implementation in use (for stats)
 };
 
-static common_ngram_map get_common_ngram_map(const common_speculative_config & config) {
-    uint16_t size_key   = config.params.ngram_size_n;
-    uint16_t size_value = config.params.ngram_size_m;
-    bool     key_only   = (config.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
-    uint16_t min_hits   = config.params.ngram_min_hits;
+static common_ngram_map get_common_ngram_map(
+        common_speculative_type type,
+        const common_params_speculative_ngram_map & config) {
+    uint16_t size_key   = config.size_n;
+    uint16_t size_value = config.size_m;
+    bool     key_only   = type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K;
+    uint16_t min_hits   = config.min_hits;
 
     return common_ngram_map(size_key, size_value, key_only, min_hits);
 }
@@ -798,50 +969,14 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
     return it->second;
 }
 
-bool common_speculative_is_compat(llama_context * ctx_tgt) {
-    auto * mem = llama_get_memory(ctx_tgt);
-    if (mem == nullptr) {
-        return false;
-    }
-
-    bool res = true;
-
-    llama_memory_clear(mem, true);
-
-    // eval 2 tokens to check if the context is compatible
-    std::vector<llama_token> tmp;
-    tmp.push_back(0);
-    tmp.push_back(0);
-
-    int ret = llama_decode(ctx_tgt, llama_batch_get_one(tmp.data(), tmp.size()));
-    if (ret != 0) {
-        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
-        res = false;
-        goto done;
-    }
-
-    // try to remove the last tokens
-    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
-        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
-        res = false;
-        goto done;
-    }
-
-done:
-    llama_memory_clear(mem, true);
-    llama_synchronize(ctx_tgt);
-
-    return res;
-}
-
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(
         common_params_speculative & params,
         llama_context             * ctx_tgt) {
     llama_context * ctx_dft = nullptr;
-    if (params.model_dft) {
-        ctx_dft = llama_init_from_model(params.model_dft, params.cparams_dft);
+    if (params.draft.model) {
+        ctx_dft = llama_init_from_model(params.draft.model, params.draft.cparams);
         if (ctx_dft == nullptr) {
             LOG_ERR("%s", "failed to create draft context\n");
             return nullptr;
@@ -851,7 +986,7 @@ common_speculative * common_speculative_init(
     // Compute the implementations to use based on the config and their order of preference
     std::vector<common_speculative_config> configs = {}; // list of speculative configs to try
     {
-        bool has_draft = !params.mparams_dft.path.empty();
+        bool has_draft = !params.draft.mparams.path.empty();
         bool has_draft_eagle3 = false; // TODO PR-18039: if params.speculative.eagle3
 
         bool has_ngram_cache   = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_CACHE);
@@ -874,16 +1009,17 @@ common_speculative * common_speculative_init(
             configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, params));
         }
         if (has_ngram_mod) {
-            // shared instance for all speculative decoding contexts
-            if (!params.ngram_mod) {
-                params.ngram_mod = std::make_shared<common_ngram_mod>(params.ngram_size_n, 4*1024*1024);
+            auto & sparams = params.ngram_mod;
 
-                LOG_INF("%s: initialized ngram_mod with n=%d, size=%zu (%.3f MB)\n", __func__,
-                        params.ngram_size_n, params.ngram_mod->size(),
-                        (float)(params.ngram_mod->size_bytes())/1024/1024);
+            if (!sparams.obj) {
+                sparams.obj = std::make_shared<common_ngram_mod>(sparams.n_match, 4*1024*1024);
 
-                if (params.ngram_size_n < 16) {
-                    LOG_WRN("%s: ngram_mod n=%d is too small - poor quality is possible, see: https://github.com/ggml-org/llama.cpp/pull/19164\n", __func__, params.ngram_size_n);
+                LOG_INF("%s: initialized ngram_mod with n_match=%d, size=%zu (%.3f MB)\n", __func__,
+                        sparams.n_match, sparams.obj->size(), (float)(sparams.obj->size_bytes())/1024/1024);
+
+                if (sparams.n_match < 16) {
+                    LOG_WRN("%s: ngram_mod n_match=%d is too small - poor quality is possible, "
+                            "see: https://github.com/ggml-org/llama.cpp/pull/19164\n", __func__, sparams.n_match);
                 }
             }
 
@@ -908,10 +1044,13 @@ common_speculative * common_speculative_init(
             case COMMON_SPECULATIVE_TYPE_NONE:
                 break;
             case COMMON_SPECULATIVE_TYPE_DRAFT: {
+                const bool use_ckpt = common_context_can_seq_rm(ctx_dft) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+
                 impls.push_back(std::make_unique<common_speculative_state_draft>(config.type,
                     /* .ctx_tgt      = */ ctx_tgt,
                     /* .ctx_dft      = */ ctx_dft,
-                    /* .replacements = */ params.replacements
+                    /* .replacements = */ params.draft.replacements,
+                    /* .use_ckpt     = */ use_ckpt
                 ));
                 break;
             }
@@ -920,18 +1059,18 @@ common_speculative * common_speculative_init(
                 break;
             }
             case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: {
-                common_ngram_map ngram_map = get_common_ngram_map(config);
+                common_ngram_map ngram_map = get_common_ngram_map(config.type, config.params.ngram_simple);
 
                 uint16_t ngram_size_key   = ngram_map.size_key;
                 uint16_t mgram_size_value = ngram_map.size_value;
 
                 auto config_simple = common_ngram_simple_config {
-                    /* .size_ngram      = */ ngram_size_key,
-                    /* .size_mgram      = */ mgram_size_value
+                    /* .size_ngram = */ ngram_size_key,
+                    /* .size_mgram = */ mgram_size_value
                 };
                 auto state = std::make_unique<common_speculative_state_ngram_simple>(
-                    /* .type            = */ config.type,
-                    /* .state           = */ config_simple
+                    /* .type  = */ config.type,
+                    /* .state = */ config_simple
                 );
                 impls.push_back(std::move(state));
                 break;
@@ -940,18 +1079,17 @@ common_speculative * common_speculative_init(
             case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: {
                 impls.push_back(std::make_unique<common_speculative_state_ngram_map_k>(
                     (config.type),
-                    get_common_ngram_map(config)
+                    get_common_ngram_map(config.type, config.params.ngram_map_k)
                 ));
                 break;
             }
             case COMMON_SPECULATIVE_TYPE_NGRAM_MOD: {
-                GGML_ASSERT(config.params.ngram_mod);
-                impls.push_back(std::make_unique<common_speculative_state_ngram_mod>(config.type, *config.params.ngram_mod));
+                GGML_ASSERT(config.params.ngram_mod.obj);
+                impls.push_back(std::make_unique<common_speculative_state_ngram_mod>(config.type, *config.params.ngram_mod.obj));
                 break;
             }
             case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE: {
-                auto state = create_state_ngram_cache(
-                        params.lookup_cache_static, params.lookup_cache_dynamic, config);
+                auto state = create_state_ngram_cache(params.ngram_cache.lookup_cache_static, params.ngram_cache.lookup_cache_dynamic, config);
                 impls.push_back(std::make_unique<common_speculative_state_ngram_cache>(state));
                 break;
             }
@@ -966,7 +1104,8 @@ common_speculative * common_speculative_init(
     }
 
     auto * result = new common_speculative {
-        /* .impls = */ std::move(impls)
+        /* .impls     = */ std::move(impls),
+        /* .curr_impl = */ nullptr,
     };
 
     return result;
@@ -1008,6 +1147,15 @@ llama_tokens common_speculative_draft(
             impl->n_call_draft++;
         }
 
+        {
+            const int n_min = impl->n_min(params);
+
+            if (!result.empty() && (int) result.size() < n_min) {
+                LOG_DBG("%s: ignoring small draft: %d < %d\n", __func__, (int) result.size(), n_min);
+                result.clear();
+            }
+        }
+
         if (!result.empty()) {
             LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__,
                     common_speculative_type_to_str(impl.get()->type).c_str(), prompt_tgt.size(),
@@ -1017,7 +1165,7 @@ llama_tokens common_speculative_draft(
             impl->n_gen_drafts++;
             impl->n_gen_tokens += result.size();
 
-            break; // We have a draft, so break out of the loop and return it.
+            break; // we have a draft, so break out of the loop and return it.
         }
     }
 
@@ -1045,6 +1193,32 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) {
     }
 }
 
+int32_t common_speculative_n_max(const common_speculative * spec, const common_params_speculative & params) {
+    if (spec == nullptr) {
+        return 0;
+    }
+
+    int32_t n_max = 0;
+    for (const auto & impl : spec->impls) {
+        n_max = std::max(n_max, impl->n_max(params));
+    }
+
+    return n_max;
+}
+
+int32_t common_speculative_n_min(const common_speculative * spec, const common_params_speculative & params) {
+    if (spec == nullptr) {
+        return 0;
+    }
+
+    int32_t n_min = 0;
+    for (const auto & impl : spec->impls) {
+        n_min = std::max(n_min, impl->n_min(params));
+    }
+
+    return n_min;
+}
+
 void common_speculative_print_stats(const common_speculative * spec) {
     if (spec == nullptr) {
         return;

common/speculative.h

@@ -14,10 +14,6 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
 // convert type to string
 std::string common_speculative_type_to_str(enum common_speculative_type type);
 
-// check if the llama_context is compatible for speculative decoding
-// note: clears the memory of the context
-bool common_speculative_is_compat(llama_context * ctx_tgt);
-
 common_speculative * common_speculative_init(
         common_params_speculative & params,
         llama_context             * ctx_tgt);
@@ -37,5 +33,14 @@ llama_tokens common_speculative_draft(
 // informs the speculative decoder that n_accepted tokens were accepted by the target model
 void common_speculative_accept(common_speculative * spec, uint16_t n_accepted);
 
+int32_t common_speculative_n_max(const common_speculative * spec, const common_params_speculative & params);
+int32_t common_speculative_n_min(const common_speculative * spec, const common_params_speculative & params);
+
 // print statistics about the speculative decoding
 void common_speculative_print_stats(const common_speculative * spec);
+
+struct common_speculative_deleter {
+    void operator()(common_speculative * s) { common_speculative_free(s); }
+};
+
+typedef std::unique_ptr<common_speculative, common_speculative_deleter> common_speculative_ptr;

convert_hf_to_gguf.py

@@ -272,6 +272,22 @@ class ModelBase:
 
         return tensors
 
+    @staticmethod
+    def _scale_is_trivial(scale: Tensor) -> bool:
+        return scale.numel() <= 1 and abs(float(scale.float().sum()) - 1.0) < 1e-6
+
+    def _write_scale_tensor(self, scale_name: str, scale: Tensor):
+        if not self._scale_is_trivial(scale):
+            scale_f32 = scale.float().numpy().flatten()
+            logger.info(f"  + {scale_name} (per-tensor scale, shape [{scale_f32.size}])")
+            self.gguf_writer.add_tensor(scale_name, scale_f32)
+
+    def _write_scales_tensor(self, scale_name: str, scales: list[float]):
+        if not np.allclose(scales, 1.0, atol=1e-6):
+            scale_vals = np.array(scales, dtype=np.float32)
+            logger.info(f"  + {scale_name} (per-expert scale, shape [{len(scales)}])")
+            self.gguf_writer.add_tensor(scale_name, scale_vals)
+
     def dequant_model(self):
         # If all quantized tensors were already handled (e.g. pure NVFP4), skip
         if self._is_nvfp4 and not any(k.endswith((".weight_scale", ".weight_scale_inv")) for k in self.model_tensors):
@@ -494,7 +510,7 @@ class ModelBase:
                         s = self.model_tensors[name]
                         self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), None)
                         tensors_to_remove.append(name)
-                    if name.endswith((".k_scale", ".v_scale")):
+                    if name.endswith((".input_scale", ".k_scale", ".v_scale")):
                         tensors_to_remove.append(name)
             elif quant_method is not None:
                 raise NotImplementedError(f"Quant method is not yet supported: {quant_method!r}")
@@ -602,10 +618,6 @@ class ModelBase:
         raw = np.concatenate([d_grouped, qs_grouped], axis=-1).reshape(out_features, n_super * 36)
         return raw, [out_features, n_super * 64]
 
-    @staticmethod
-    def _nvfp4_scale2_is_trivial(scale2: Tensor) -> bool:
-        return scale2.numel() <= 1 and abs(float(scale2.float().sum()) - 1.0) < 1e-6
-
     def _repack_nvfp4(self, name: str, weight: Tensor, scale: Tensor, scale2: Tensor, input_scale: Tensor):
         if "language_model." in name:
             name = name.replace("language_model.", "")
@@ -616,19 +628,8 @@ class ModelBase:
         logger.info(f"Repacked {new_name} with shape {shape} and quantization NVFP4")
         self.gguf_writer.add_tensor(new_name, raw, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
 
-        # Emit per-tensor scale2 as a separate F32 tensor when non-trivial
-        if not self._nvfp4_scale2_is_trivial(scale2):
-            scale2_f32 = scale2.float().numpy().flatten()
-            scale_name = new_name.replace(".weight", ".scale")
-            logger.info(f"  + {scale_name} (per-tensor NVFP4 scale2, shape [{scale2_f32.size}])")
-            self.gguf_writer.add_tensor(scale_name, scale2_f32)
-
-        # Emit per-tensor input_scale as a separate F32 tensor when non-trivial
-        if not self._nvfp4_scale2_is_trivial(input_scale):
-            input_scale_f32 = input_scale.float().numpy().flatten()
-            input_scale_name = new_name.replace(".weight", ".input_scale")
-            logger.info(f"  + {input_scale_name} (per-tensor NVFP4 input_scale, shape [{input_scale_f32.size}])")
-            self.gguf_writer.add_tensor(input_scale_name, input_scale_f32)
+        self._write_scale_tensor(new_name.replace(".weight", ".scale"), scale2)
+        self._write_scale_tensor(new_name.replace(".weight", ".input_scale"), input_scale)
 
     def _generate_nvfp4_tensors(self):
         # Per-layer expert merging to avoid holding all experts in memory
@@ -719,24 +720,17 @@ class ModelBase:
         logger.info(f"Repacked {new_name} with shape [{len(experts)}, {shape[0]}, {shape[1]}] and quantization NVFP4")
         self.gguf_writer.add_tensor(new_name, merged, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
 
-        # Emit per-expert scale2 tensor if any expert has non-trivial scale2
         scales.sort(key=lambda x: x[0])
-        scale_vals = np.array([s[1] for s in scales], dtype=np.float32)
-        if not np.allclose(scale_vals, 1.0, atol=1e-6):
-            scale_name = new_name.replace(".weight", ".scale")
-            logger.info(f"  + {scale_name} (per-expert NVFP4 scale2, shape [{len(scales)}])")
-            self.gguf_writer.add_tensor(scale_name, scale_vals)
+        self._write_scales_tensor(new_name.replace(".weight", ".scale"), [s[1] for s in scales])
 
-        # Emit per-expert input_scale tensor if any expert has non-trivial input_scale
         input_scales.sort(key=lambda x: x[0])
-        input_scale_vals = np.array([s[1] for s in input_scales], dtype=np.float32)
-        if not np.allclose(input_scale_vals, 1.0, atol=1e-6):
-            input_scale_name = new_name.replace(".weight", ".input_scale")
-            logger.info(f"  + {input_scale_name} (per-expert NVFP4 input_scale, shape [{len(input_scales)}])")
-            self.gguf_writer.add_tensor(input_scale_name, input_scale_vals)
+        self._write_scales_tensor(new_name.replace(".weight", ".input_scale"), [s[1] for s in input_scales])
 
         del experts, merged
 
+    def _needs_nvfp4_processing(self) -> bool:
+        return True
+
     def prepare_tensors(self):
         # detect NVFP4 quantization (ModelOpt format)
         quant_algo = (self.hparams.get("quantization_config") or {}).get("quant_algo")
@@ -746,7 +740,12 @@ class ModelBase:
 
         if (not quant_algo or not quant_layers) and quant_config_file.is_file():
             with open(quant_config_file, "r", encoding="utf-8") as f:
-                quant_config = json.load(f).get("quantization") or {}
+                hf_quant_config = json.load(f)
+                quant_config = hf_quant_config.get("quantization") or {}
+                producer = hf_quant_config.get("producer") or {}
+                producer_name = (producer.get("name") or "").lower()
+                if quant_method is None:
+                    self.hparams.setdefault("quantization_config", {})["quant_method"] = producer_name
                 quant_algo = quant_config.get("quant_algo", quant_algo)
                 quant_layers = quant_config.get("quantized_layers", quant_layers) or {}
 
@@ -762,7 +761,7 @@ class ModelBase:
         # NVFP4 weights are repacked and written directly to gguf_writer.
         # This must run before dequant_model so NVFP4 tensors are removed
         # from model_tensors, leaving only non-NVFP4 (e.g. FP8) for dequant.
-        if self._is_nvfp4:
+        if self._is_nvfp4 and self._needs_nvfp4_processing():
             self._generate_nvfp4_tensors()
 
         self.dequant_model()
@@ -1850,20 +1849,28 @@ class TextModel(ModelBase):
             with open(module_path, encoding="utf-8") as f:
                 modules = json.load(f)
             for mod in modules:
-                if mod["type"] == "sentence_transformers.models.Pooling":
+                if mod["type"].endswith("Pooling"):
                     pooling_path = mod["path"]
                     break
 
+        mode_mapping = {
+            "mean": gguf.PoolingType.MEAN,
+            "cls": gguf.PoolingType.CLS,
+            "lasttoken": gguf.PoolingType.LAST,
+        }
+
         # get pooling type
         if pooling_path is not None:
             with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
                 pooling = json.load(f)
-            if pooling["pooling_mode_mean_tokens"]:
+            if pooling.get("pooling_mode_mean_tokens"):
                 pooling_type = gguf.PoolingType.MEAN
-            elif pooling["pooling_mode_cls_token"]:
+            elif pooling.get("pooling_mode_cls_token"):
                 pooling_type = gguf.PoolingType.CLS
-            elif pooling["pooling_mode_lasttoken"]:
+            elif pooling.get("pooling_mode_lasttoken"):
                 pooling_type = gguf.PoolingType.LAST
+            elif (pooling_mode := pooling.get("pooling_mode")) in mode_mapping:
+                pooling_type = mode_mapping[pooling_mode]
             else:
                 raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
             self.gguf_writer.add_pooling_type(pooling_type)
@@ -2186,6 +2193,10 @@ class MmprojModel(ModelBase):
                 # merge configs
                 self.preprocessor_config = {**self.preprocessor_config, **cfg}
 
+    def _needs_nvfp4_processing(self) -> bool:
+        # nvfp4 quantization applies to the text model only.
+        return False
+
     def get_vision_config(self) -> dict[str, Any] | None:
         config_name = "vision_config" if not self.is_mistral_format else "vision_encoder"
         return self.global_config.get(config_name)
@@ -4446,6 +4457,12 @@ class NemotronNanoV2VLModel(MmprojModel):
         }
         return vision_config
 
+    def dequant_model(self):
+        if self._is_nvfp4:
+            # Skip nvfp4 quantization for vision/audio model.
+            return
+        super().dequant_model()
+
     def set_gguf_parameters(self):
         if "image_mean" not in self.preprocessor_config:
             self.preprocessor_config["image_mean"] = [0.485, 0.456, 0.406]
@@ -4469,6 +4486,10 @@ class NemotronNanoV2VLModel(MmprojModel):
         if "input_conditioner" in name:
             return
 
+        # mtmd does not support video yet so skip tensors related to video.
+        if "radio_model.model.patch_generator.video_embedder" in name:
+            return
+
         # RADIO's pos_embed doesn't have .weight suffix, but clip.cpp expects it
         if "patch_generator.pos_embed" in name:
             if not name.endswith(".weight"):
@@ -6637,7 +6658,7 @@ class BertModel(TextModel):
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
-        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size  # ty: ignore[invalid-assignment]
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
 
         if isinstance(tokenizer, SentencePieceProcessor):
             for token_id in range(tokenizer.vocab_size()):
@@ -7180,7 +7201,7 @@ class EmbeddingGemma(Gemma3Model):
                 with open(modules_file, encoding="utf-8") as modules_json_file:
                     mods = json.load(modules_json_file)
                 for mod in mods:
-                    if mod["type"] == "sentence_transformers.models.Dense":
+                    if mod["type"].endswith("Dense"):
                         mod_path = mod["path"]
                         # check if model.safetensors file for Dense layer exists
                         model_tensors_file = self.dir_model / mod_path / "model.safetensors"
@@ -10816,7 +10837,11 @@ class NemotronHModel(GraniteHybridModel):
         # uses self.model_arch to build the tensor name map, and all MoE-specific
         # mappings would be missed if it were called with the default non-MoE arch.
         hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
-        if "num_experts_per_tok" in hparams:
+        has_moe_params = (
+            "num_experts_per_tok" in hparams
+            or (isinstance(hparams.get("llm_config"), dict) and "num_experts_per_tok" in hparams["llm_config"])
+        )
+        if has_moe_params:
             self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
             self.is_moe = True
 
@@ -10963,6 +10988,11 @@ class NemotronHModel(GraniteHybridModel):
         if name.startswith(("vision_model.", "mlp1.")):
             return
 
+        if name.startswith(("sound_encoder.")):
+            return
+        if name.startswith(("sound_projection.")):
+            return
+
         # Strip language_model. prefix for VLM models (e.g., Nemotron Nano 12B v2 VL)
         if name.startswith("language_model."):
             name = name[len("language_model."):]
@@ -11847,7 +11877,7 @@ class LLaDAMoEModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("HunYuanDenseV1ForCausalLM", "HunYuanVLForConditionalGeneration")
+@ModelBase.register("HunYuanDenseV1ForCausalLM")
 class HunYuanModel(TextModel):
     model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE
 
@@ -11986,28 +12016,58 @@ class HunYuanModel(TextModel):
 
 
 @ModelBase.register("HunYuanVLForConditionalGeneration")
-class HunyuanOCRVisionModel(MmprojModel):
+class HunyuanVLVisionModel(MmprojModel):
+    # Handles both HunyuanOCR and HunyuanVL, which share the HF architecture name
+    # "HunYuanVLForConditionalGeneration" and the `vit.perceive.*` vision layout.
+    # Each variant maps to a different projector type in clip.cpp so image
+    # preprocessing follows the correct code path.
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         assert self.hparams_vision is not None
-        # HunyuanOCR uses max_image_size instead of image_size
+        # HunyuanOCR / HunyuanVL uses max_image_size instead of image_size
         if "image_size" not in self.hparams_vision:
             self.hparams_vision["image_size"] = self.hparams_vision.get("max_image_size", 2048)
 
+    @staticmethod
+    def is_ocr_variant(hparams: dict) -> bool:
+        """Return True for HunyuanOCR, False for HunyuanVL.
+
+        The projector's output dim must equal the text model's hidden_size by
+        construction (that's what "projector" means). HunyuanOCR pairs a 1B text
+        backbone (hidden=1024); HunyuanVL pairs a 4B one (hidden=3072). So the
+        ViT -> LLM projection dim is a hard architectural signature, not a
+        magic number.
+        """
+        vision_out = int((hparams.get("vision_config") or {}).get("out_hidden_size", 0))
+        return vision_out == 1024
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         assert self.hparams_vision is not None
-        hparams = self.hparams_vision
-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.HUNYUANOCR)
-        self.gguf_writer.add_vision_use_gelu(True)
-        self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("rms_norm_eps", 1e-5))
-        self.gguf_writer.add_vision_spatial_merge_size(hparams.get("spatial_merge_size", 2))
-        self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])
-        self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])
+        vcfg = self.hparams_vision
+
+        if self.is_ocr_variant(self.global_config):
+            # --- HunyuanOCR ---
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.HUNYUANOCR)
+            self.gguf_writer.add_vision_use_gelu(True)
+            self.gguf_writer.add_vision_attention_layernorm_eps(vcfg.get("rms_norm_eps", 1e-5))
+            self.gguf_writer.add_vision_spatial_merge_size(vcfg.get("spatial_merge_size", 2))
+            self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])
+            self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])
+            return
+
+        # --- HunyuanVL ---
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.HUNYUANVL)
+        self.gguf_writer.add_vision_use_gelu(str(vcfg["hidden_act"]).lower() == "gelu")
+        self.gguf_writer.add_vision_attention_layernorm_eps(float(vcfg["rms_norm_eps"]))
+        self.gguf_writer.add_vision_spatial_merge_size(int(vcfg["spatial_merge_size"]))
+        self.gguf_writer.add_vision_min_pixels(int(self.preprocessor_config["min_pixels"]))
+        self.gguf_writer.add_vision_max_pixels(int(self.preprocessor_config["max_pixels"]))
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if not name.startswith("vit."):
-            return  # skip text tensors
+            return
         # strip CLS token (row 0) from position embeddings so resize_position_embeddings works
         if "position_embedding" in name:
             data_torch = data_torch[1:]  # [n_patches+1, n_embd] -> [n_patches, n_embd]
@@ -12015,11 +12075,66 @@ class HunyuanOCRVisionModel(MmprojModel):
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         # force conv weights to F32 or F16 to avoid BF16 IM2COL issues on Metal
+        # Both HunyuanOCR and HunyuanVL emit the ViT -> LLM projection as mm.0/mm.2.
         if ("mm.0." in new_name or "mm.2." in new_name) and new_name.endswith(".weight"):
             return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
 
+@ModelBase.register("HunYuanVLForConditionalGeneration")
+class HunyuanVLTextModel(HunYuanModel):
+    # The "HunYuanVLForConditionalGeneration" HF architecture covers both HunyuanOCR
+    # and HunyuanVL. HunyuanOCR reuses the HunYuan-Dense text backbone (standard RoPE),
+    # while HunyuanVL introduces a new LLM arch with XD-RoPE. Detect the variant from
+    # the config and pick the matching GGUF architecture.
+    model_arch = gguf.MODEL_ARCH.HUNYUAN_VL
+
+    @staticmethod
+    def _is_ocr_config(hparams: dict) -> bool:
+        # OCR pairs a 1B text backbone (hidden=1024) with a ViT projector that
+        # outputs 1024-d; HunyuanVL uses 3072-d. Keep in sync with
+        # HunyuanVLVisionModel.is_ocr_variant.
+        return int((hparams.get("vision_config") or {}).get("out_hidden_size", 0)) == 1024
+
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        raw_hparams = kwargs.get("hparams") or ModelBase.load_hparams(dir_model, is_mistral_format=False)
+        if self._is_ocr_config(raw_hparams):
+            self.model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE
+        else:
+            self.model_arch = gguf.MODEL_ARCH.HUNYUAN_VL
+        super().__init__(dir_model, *args, **kwargs)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # Only emit XD-RoPE metadata for the HunyuanVL backbone; HunyuanOCR uses
+        # the HunYuan-Dense arch which already handles standard rope in super().
+        if self.model_arch != gguf.MODEL_ARCH.HUNYUAN_VL:
+            return
+
+        if self.rope_parameters.get("rope_type") != "xdrope":
+            return
+
+        # defaults for HunyuanVL. The C++ side later computes:
+        #   freq_base = rope_theta * alpha ** (head_dim / (head_dim - 2))
+        self.gguf_writer.add_rope_freq_base(float(self.rope_parameters["rope_theta"]))
+        self.gguf_writer.add_rope_scaling_alpha(float(self.rope_parameters["alpha"]))
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        self.gguf_writer.add_rope_scaling_factor(float(self.rope_parameters.get("factor", 1)))
+
+        ctx_len = int(self.hparams["max_position_embeddings"])
+        self.gguf_writer.add_rope_scaling_orig_ctx_len(ctx_len)
+        self.gguf_writer.add_context_length(ctx_len)
+
+        self.gguf_writer.add_rope_dimension_sections(list(self.rope_parameters["xdrope_section"]))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision tensors — they are written by HunyuanVLVisionModel
+        if name.startswith("vit."):
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("SmolLM3ForCausalLM")
 class SmolLM3Model(LlamaModel):
     model_arch = gguf.MODEL_ARCH.SMOLLM3

docs/backend/OPENVINO.md

@@ -244,7 +244,6 @@ build\ReleaseOV\bin\llama-cli.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf"
 - `-fa 1` is required when running llama-bench with the OpenVINO backend.
   - `GGML_OPENVINO_STATEFUL_EXECUTION=1 GGML_OPENVINO_DEVICE=GPU ./llama-bench -fa 1`
 - `llama-server` with OpenVINO backend supports only one chat session/thread, when `GGML_OPENVINO_STATEFUL_EXECUTION=1` is enabled.
-- For Intel GPU, NPU detection in containers, GPU, NPU user-space drivers/libraries must be present inside the image. We will include in a future PR. Until then, you can use this reference Dockerfile: [openvino.Dockerfile](https://github.com/ravi9/llama.cpp/blob/ov-docker-update/.devops/openvino.Dockerfile)
 
 > [!NOTE]
 > The OpenVINO backend is actively under development. Fixes are underway, and this document will continue to be updated as issues are resolved.
@@ -274,8 +273,6 @@ docker build --build-arg http_proxy=$http_proxy --build-arg https_proxy=$https_p
 Run llama.cpp with OpenVINO backend Docker container.
 Save sample models in `~/models` as [shown above](#3-download-sample-model). It will be mounted to the container in the examples below.
 
-> [!NOTE]
-> Intel GPU, NPU detection in containers will be included in a future PR. Until then, you can use this reference Dockerfile: [openvino.Dockerfile](https://github.com/ravi9/llama.cpp/blob/ov-docker-update/.devops/openvino.Dockerfile).
 
 ```bash
 #  Run Docker container

docs/backend/SYCL.md

@@ -31,6 +31,8 @@ SYCL cross-platform capabilities enable support for other vendor GPUs as well.
 
 ## Recommended Release
 
+### Windows
+
 The following releases are verified and recommended:
 
 |Commit ID|Tag|Release|Verified  Platform| Update date|
@@ -39,9 +41,22 @@ The following releases are verified and recommended:
 |3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc A770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
 |fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc A770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
 
+### Ubuntu 24.04
+
+The release packages for Ubuntu 24.04 x64 (FP32/FP16) only include the binary files of the llama.cpp SYCL backend. They require the target machine to have pre-installed Intel GPU drivers and oneAPI packages that are the same version as the build package. To get the version and installation info, refer to release.yml: ubuntu-24-sycl -> Download & Install oneAPI.
+
+It is recommended to use them with Intel Docker.
+
+The packages for FP32 and FP16 would have different accuracy and performance on LLMs. Please choose it acording to the test result.
 
 ## News
 
+- 2026.04
+
+  - Optimize mul_mat by reorder feature for data type: Q4_K, Q5_K, Q_K, Q8_0.
+  - Fused MoE.
+  - Upgrate CI and built package for oneAPI 2025.3.3, support Ubuntu 24.04 built package.
+
 - 2026.03
   - Support Flash-Attention: less memory usage, performance impact depends on LLM.
 
@@ -229,6 +244,7 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
 
 |Verified release|
 |-|
+|2025.3.3 |
 |2025.2.1|
 |2025.1|
 |2024.1|
@@ -339,6 +355,12 @@ Choose one of following methods to run.
 ./examples/sycl/test.sh
 ```
 
+- Run llama-server:
+
+```sh
+./examples/sycl/start-svr.sh -m PATH/MODEL_FILE
+```
+
 2. Command line
 Launch inference
 
@@ -627,10 +649,18 @@ Choose one of following methods to run.
 
 1. Script
 
+- Run test:
+
 ```
 examples\sycl\win-test.bat
 ```
 
+- Run llama-server:
+
+```
+examples\sycl\win-start-svr.bat -m PATH\MODEL_FILE
+```
+
 2. Command line
 
 Launch inference

docs/backend/snapdragon/README.md

@@ -249,18 +249,27 @@ build: 6a8cf8914 (6733)
   ```
 
 - `GGML_HEXAGON_PROFILE=1`
-  Generates a host-side profile for the ggml-hexagon Ops.
+  Enables Op profiling:
 
-- `GGML_HEXAGON_OPMASK=0x0`
-  Allows enabling specific stages of the processing pipeline:
+  - `1` Basic profile with per-op `usecs` and `cycles` counters
+  - `2` Extended profile with per-op `usecs`, `cycles` and default PMU counter data
+  - `0x1,...,0x8` Extended profile with per-op `usecs`, `cycles` and custom PMU counter data
+
+  The logging output can be either saved into a file for post-processing or it can be piped directly into the post-processing tool to generate the report.
+  Examples:
+
+      `GGML_HEXAGON_PROFILE=1 llama-completion ... |& ./scripts/snapdragon/ggml-hexagon-profile.py -`
+
+- `GGML_HEXAGON_OPSTAGE=0x0`
+  Allows enabling specific stages of the Op processing pipeline:
 
   - `0x1` Enable Op Queue (i.e., queuing Ops into NPU)
   - `0x2` Enable Op Compute (MUL_MAT, etc.)
 
   Examples:
 
-      `GGML_HEXAGON_OPMASK=0x1 llama-completion ...` - Ops are enqueued but NPU-side processing is stubbed out
-      `GGML_HEXAGON_OPMASK=0x3 llama-completion ...` - Full queuing and processing of Ops (default)
+      `GGML_HEXAGON_OPSTAGE=0x1 llama-completion ...` - Ops are enqueued to the NPU but dma & compute are disabled
+      `GGML_HEXAGON_OPSTAGE=0x3 llama-completion ...` - Full queuing and processing of Ops (default)
 
 - `GGML_HEXAGON_OPFILTER=regex`
   Allows filtering (disabling) Ops that match the regex pattern:

docs/ops.md

@@ -26,7 +26,7 @@ Legend:
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -60,7 +60,7 @@ Legend:
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
@@ -105,7 +105,7 @@ Legend:
 |                              SQR | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              SUM | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ❌ | ❌ |

examples/debug/debug.cpp

@@ -202,10 +202,14 @@ static bool run(llama_context * ctx, const common_params & params) {
     print_tokenized_prompt(ctx, tokens, params.prompt);
 
     if (params.save_logits) {
-        output_data output {ctx, model, params};
-        std::filesystem::path model_path{params.model.path};
-        std::string model_name{model_path.stem().string()};
-        save_output_data(output, model_name, params.logits_output_dir);
+        try {
+            output_data output {ctx, model, params};
+            std::filesystem::path model_path{params.model.path};
+            std::string model_name{model_path.stem().string()};
+            save_output_data(output, model_name, params.logits_output_dir);
+        } catch (const std::exception & e) {
+            LOG_ERR("%s : error saving logits: %s\n", __func__, e.what());
+        }
     }
 
     return true;
@@ -223,7 +227,7 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    std::optional<base_callback_data> cb_data;
+    std::optional<common_debug_cb_user_data> cb_data;
     if (!params.save_logits) {
         cb_data.emplace(params, params.tensor_filter);
     }

examples/eval-callback/eval-callback.cpp

@@ -3,7 +3,6 @@
 #include "debug.h"
 #include "log.h"
 #include "llama.h"
-#include "llama-cpp.h"
 
 #include <clocale>
 #include <string>
@@ -38,7 +37,7 @@ static bool run(llama_context * ctx, const common_params & params) {
 int main(int argc, char ** argv) {
     std::setlocale(LC_NUMERIC, "C");
 
-    base_callback_data cb_data;
+    common_debug_cb_user_data cb_data;
 
     common_params params;
 
@@ -53,7 +52,7 @@ int main(int argc, char ** argv) {
 
     // pass the callback to the backend scheduler
     // it will be executed for each node during the graph computation
-    params.cb_eval = common_debug_cb_eval<false>;
+    params.cb_eval = common_debug_cb_eval;
     params.cb_eval_user_data = &cb_data;
     params.warmup = false;
 

examples/gen-docs/gen-docs.cpp

@@ -73,12 +73,12 @@ static void write_help(std::ostringstream & ss, const md_file & md) {
     auto ctx_arg = common_params_parser_init(params, md.ex);
 
     std::vector<common_arg *> common_options;
-    std::vector<common_arg *> sparam_options;
+    std::vector<common_arg *> sampling_options;
     std::vector<common_arg *> specific_options;
     for (auto & opt : ctx_arg.options) {
         // in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example
-        if (opt.is_sparam) {
-            sparam_options.push_back(&opt);
+        if (opt.is_sampling) {
+            sampling_options.push_back(&opt);
         } else if (opt.in_example(ctx_arg.ex)) {
             specific_options.push_back(&opt);
         } else {
@@ -93,7 +93,7 @@ static void write_help(std::ostringstream & ss, const md_file & md) {
     ss << "### Common params\n\n";
     write_table(ss, common_options);
     ss << "\n\n### Sampling params\n\n";
-    write_table(ss, sparam_options);
+    write_table(ss, sampling_options);
     ss << "\n\n### " << md.specific_section_header << "\n\n";
     write_table(ss, specific_options);
 

examples/lookup/lookup-create.cpp

@@ -37,9 +37,9 @@ int main(int argc, char ** argv){
 
     common_ngram_cache ngram_cache;
     common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
-    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.speculative.lookup_cache_static.c_str());
+    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.speculative.ngram_cache.lookup_cache_static.c_str());
 
-    common_ngram_cache_save(ngram_cache, params.speculative.lookup_cache_static);
+    common_ngram_cache_save(ngram_cache, params.speculative.ngram_cache.lookup_cache_static);
 
     return 0;
 }

examples/lookup/lookup-stats.cpp

@@ -24,7 +24,7 @@ int main(int argc, char ** argv){
         return 1;
     }
 
-    const int n_draft = params.speculative.n_max;
+    const int n_draft = params.speculative.draft.n_max;
 
     // init llama.cpp
     llama_backend_init();
@@ -49,18 +49,18 @@ int main(int argc, char ** argv){
     {
         const int64_t t_start_draft_us = ggml_time_us();
 
-        if (!params.speculative.lookup_cache_static.empty()) {
+        if (!params.speculative.ngram_cache.lookup_cache_static.empty()) {
             try {
-                ngram_cache_static = common_ngram_cache_load(params.speculative.lookup_cache_static);
+                ngram_cache_static = common_ngram_cache_load(params.speculative.ngram_cache.lookup_cache_static);
             } catch (std::ifstream::failure const &) {
-                LOG_ERR("failed to open static lookup cache: %s", params.speculative.lookup_cache_static.c_str());
+                LOG_ERR("failed to open static lookup cache: %s", params.speculative.ngram_cache.lookup_cache_static.c_str());
                 exit(1);
             }
         }
 
-        if (!params.speculative.lookup_cache_dynamic.empty()) {
+        if (!params.speculative.ngram_cache.lookup_cache_dynamic.empty()) {
             try {
-                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.lookup_cache_dynamic);
+                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.ngram_cache.lookup_cache_dynamic);
             } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
         }
 

examples/lookup/lookup.cpp

@@ -25,7 +25,7 @@ int main(int argc, char ** argv){
     }
 
     // max. number of additional tokens to draft if match is found
-    const int n_draft = params.speculative.n_max;
+    const int n_draft = params.speculative.draft.n_max;
 
     // init llama.cpp
     llama_backend_init();
@@ -54,18 +54,18 @@ int main(int argc, char ** argv){
         const int64_t t_start_draft_us = ggml_time_us();
         common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
 
-        if (!params.speculative.lookup_cache_static.empty()) {
+        if (!params.speculative.ngram_cache.lookup_cache_static.empty()) {
             try {
-                ngram_cache_static = common_ngram_cache_load(params.speculative.lookup_cache_static);
+                ngram_cache_static = common_ngram_cache_load(params.speculative.ngram_cache.lookup_cache_static);
             } catch (std::ifstream::failure const &) {
-                LOG_ERR("failed to open static lookup cache: %s", params.speculative.lookup_cache_static.c_str());
+                LOG_ERR("failed to open static lookup cache: %s", params.speculative.ngram_cache.lookup_cache_static.c_str());
                 exit(1);
             }
         }
 
-        if (!params.speculative.lookup_cache_dynamic.empty()) {
+        if (!params.speculative.ngram_cache.lookup_cache_dynamic.empty()) {
             try {
-                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.lookup_cache_dynamic);
+                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.ngram_cache.lookup_cache_dynamic);
             } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
         }
 
@@ -213,7 +213,7 @@ int main(int argc, char ** argv){
 
     // Update dynamic ngram cache with context ngram cache and save it to disk:
     common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
-    common_ngram_cache_save(ngram_cache_dynamic, params.speculative.lookup_cache_dynamic);
+    common_ngram_cache_save(ngram_cache_dynamic, params.speculative.ngram_cache.lookup_cache_dynamic);
 
     LOG("\n\n");
 

examples/model-conversion/scripts/causal/convert-model.sh

@@ -25,7 +25,11 @@ MODEL_NAME="${MODEL_NAME:-$(basename "$MODEL_PATH")}"
 OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
 TYPE="${OUTTYPE:-f16}"
 METADATA_OVERRIDE="${METADATA_OVERRIDE:-}"
-CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+if [[ -n "$MMPROJ" ]]; then
+    CONVERTED_MODEL="${OUTPUT_DIR}/mmproj-${MODEL_NAME}.gguf"
+else
+    CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+fi
 
 echo "Model path: ${MODEL_PATH}"
 echo "Model name: ${MODEL_NAME}"
@@ -38,6 +42,7 @@ if [[ -n "$DEBUG" ]]; then
 else
     CMD_ARGS=("python")
 fi
+
 CMD_ARGS+=("../../convert_hf_to_gguf.py" "--verbose")
 CMD_ARGS+=("${MODEL_PATH}")
 CMD_ARGS+=("--outfile" "${CONVERTED_MODEL}")
@@ -50,7 +55,3 @@ CMD_ARGS+=("--outtype" "${TYPE}")
 echo ""
 echo "The environment variable CONVERTED_MODEL can be set to this path using:"
 echo "export CONVERTED_MODEL=$(realpath ${CONVERTED_MODEL})"
-if [[ -n "$MMPROJ" ]]; then
-    mmproj_file="${OUTPUT_DIR}/mmproj-$(basename "${CONVERTED_MODEL}")"
-    echo "The mmproj model was created in $(realpath "$mmproj_file")"
-fi

examples/speculative-simple/speculative-simple.cpp

@@ -8,8 +8,24 @@
 #include <clocale>
 #include <cstdio>
 #include <cstring>
+#include <cinttypes>
 #include <string>
 #include <vector>
+#include <utility>
+
+struct spec_checkpoint {
+    int64_t n_tokens = 0;
+
+    std::vector<uint8_t> data;
+
+    size_t size() const {
+        return data.size();
+    }
+
+    bool empty() const {
+        return data.empty();
+    }
+};
 
 int main(int argc, char ** argv) {
     std::setlocale(LC_NUMERIC, "C");
@@ -27,7 +43,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    if (params.speculative.mparams_dft.path.empty()) {
+    if (params.speculative.draft.mparams.path.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;
     }
@@ -46,6 +62,14 @@ int main(int argc, char ** argv) {
     model_tgt = llama_init_tgt->model();
     ctx_tgt   = llama_init_tgt->context();
 
+    // check if the context supports partial sequence removal
+    const auto ctx_seq_rm = common_context_can_seq_rm(ctx_tgt);
+    const bool use_ckpt = (ctx_seq_rm == COMMON_CONTEXT_SEQ_RM_TYPE_FULL);
+
+    if (use_ckpt) {
+        LOG_INF("speculative decoding will use checkpoints (context does not support partial sequence removal)\n");
+    }
+
     const llama_vocab * vocab = llama_model_get_vocab(model_tgt);
 
     // load the draft model
@@ -53,7 +77,7 @@ int main(int argc, char ** argv) {
 
     // TODO: simplify this logic
     {
-        const auto & params_spec = params.speculative;
+        const auto & params_spec = params.speculative.draft;
 
         auto params_dft = params;
 
@@ -61,15 +85,15 @@ int main(int argc, char ** argv) {
         params_dft.n_ctx        = params_spec.n_ctx;
         params_dft.n_batch      = llama_n_ctx_seq(ctx_tgt);
         params_dft.devices      = params_spec.devices;
-        params_dft.model        = params_spec.mparams_dft;
+        params_dft.model        = params_spec.mparams;
         params_dft.n_gpu_layers = params_spec.n_gpu_layers;
 
         if (params_spec.cpuparams.n_threads > 0) {
-            params_dft.cpuparams.n_threads       = params.speculative.cpuparams.n_threads;
-            params_dft.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
+            params_dft.cpuparams.n_threads       = params.speculative.draft.cpuparams.n_threads;
+            params_dft.cpuparams_batch.n_threads = params.speculative.draft.cpuparams_batch.n_threads;
         }
 
-        params_dft.tensor_buft_overrides = params.speculative.tensor_buft_overrides;
+        params_dft.tensor_buft_overrides = params.speculative.draft.tensor_buft_overrides;
 
         auto mparams_dft = common_model_params_to_llama(params_dft);
 
@@ -79,8 +103,8 @@ int main(int argc, char ** argv) {
             return 1;
         }
 
-        params.speculative.model_dft = model_dft.get();
-        params.speculative.cparams_dft = common_context_params_to_llama(params_dft);
+        params.speculative.draft.model = model_dft.get();
+        params.speculative.draft.cparams = common_context_params_to_llama(params_dft);
     }
 
     // Tokenize the prompt
@@ -119,7 +143,7 @@ int main(int argc, char ** argv) {
     const auto t_enc_start = ggml_time_us();
 
     // target model sampling context
-    struct common_sampler * smpl = common_sampler_init(model_tgt, params.sampling);
+    common_sampler_ptr smpl(common_sampler_init(model_tgt, params.sampling));
 
     // eval the prompt
     llama_decode(ctx_tgt, llama_batch_get_one(inp.data(), inp.size() - 1));
@@ -142,21 +166,49 @@ int main(int argc, char ** argv) {
 
     llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
 
+    size_t n_draft = 0;
+
+    llama_tokens draft;
+    spec_checkpoint spec_ckpt;
+
     const auto t_enc_end = ggml_time_us();
 
     const auto t_dec_start = ggml_time_us();
 
     while (true) {
-        // optionally, generate draft tokens that can be appended to the target batch
+        // generate or reuse draft tokens
         //
         // this is the most important part of the speculation. the more probable tokens that are provided here
         // the better the performance will be. in theory, this computation can be performed asynchronously and even
         // offloaded to a remote device. it doesn't even have to be based on an LLM. instead, it can provide tokens
         // from a cache or lookup tables.
         //
-        llama_tokens draft = common_speculative_draft(spec, params_spec, prompt_tgt, id_last);
+        if (draft.empty()) {
+            // generate a new draft
+            draft = common_speculative_draft(spec, params_spec, prompt_tgt, id_last);
 
-        //LOG_DBG("draft: %s\n", string_from(ctx_dft, draft).c_str());
+            // save the original draft size
+            n_draft = draft.size();
+
+            // save a checkpoint of the target context before evaluating the draft
+            // this allows us to restore the state if partial draft acceptance occurs
+            if (!draft.empty() && use_ckpt) {
+                const size_t ckpt_size = llama_state_seq_get_size_ext(ctx_tgt, 0, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                spec_ckpt.data.resize(ckpt_size);
+
+                const size_t n = llama_state_seq_get_data_ext(ctx_tgt, spec_ckpt.data.data(), ckpt_size, 0, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                GGML_ASSERT(n == ckpt_size);
+
+                spec_ckpt.n_tokens = (int64_t) prompt_tgt.size();
+                LOG_DBG("created speculative checkpoint (n_tokens = %" PRId64 ", size = %.3f MiB)\n",
+                        spec_ckpt.n_tokens, (float) spec_ckpt.data.size() / 1024 / 1024);
+            }
+        } else {
+            // we have a previous (partial) draft to reuse from checkpoint restoration
+            if (use_ckpt) {
+                GGML_ASSERT(!spec_ckpt.empty());
+            }
+        }
 
         // always have a token to evaluate from before - id_last
         common_batch_clear(batch_tgt);
@@ -164,11 +216,6 @@ int main(int argc, char ** argv) {
 
         // evaluate the target model on [id_last, draft0, draft1, ..., draftN-1]
         {
-            // do not waste time on small drafts
-            if (draft.size() < (size_t) params_spec.n_min) {
-                draft.clear();
-            }
-
             for (size_t i = 0; i < draft.size(); ++i) {
                 common_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
             }
@@ -178,6 +225,12 @@ int main(int argc, char ** argv) {
             llama_decode(ctx_tgt, batch_tgt);
         }
 
+        // only save the sampler sampler state if we use checkpoints
+        common_sampler_ptr smpl_save;
+        if (use_ckpt) {
+            smpl_save.reset(common_sampler_clone(smpl.get()));
+        }
+
         // sample from the full target batch and return the accepted tokens based on the target sampler
         //
         // for each token to be accepted, the sampler would have to sample that same token
@@ -185,14 +238,38 @@ int main(int argc, char ** argv) {
         // available logits from the batch and sample the next token until we run out of logits or the sampler
         // disagrees with the draft
         //
-        const auto ids = common_sampler_sample_and_accept_n(smpl, ctx_tgt, draft);
+        auto ids = common_sampler_sample_and_accept_n(smpl.get(), ctx_tgt, draft);
 
         //LOG_DBG("ids: %s\n", string_from(ctx_tgt, ids).c_str());
 
         GGML_ASSERT(ids.size() > 0); // there will always be at least one accepted token
 
+        // check for partial draft acceptance:
+        // if the context doesn't support partial sequence removal, restore the checkpoint
+        // and make the accepted tokens the new partial draft for the next iteration
+        if (use_ckpt && ids.size() - 1 < draft.size()) {
+            LOG_DBG("partial acceptance: %zu < %zu, restoring checkpoint\n", ids.size() - 1, draft.size());
+
+            draft = std::move(ids);
+
+            const size_t n = llama_state_seq_set_data_ext(ctx_tgt, spec_ckpt.data.data(), spec_ckpt.size(), 0, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+            GGML_ASSERT(n == spec_ckpt.size());
+
+            llama_memory_seq_rm(llama_get_memory(ctx_tgt), 0, spec_ckpt.n_tokens, -1);
+
+            prompt_tgt.resize(spec_ckpt.n_tokens);
+            smpl = std::move(smpl_save);
+
+            n_past = (int) prompt_tgt.size();
+
+            continue;
+        }
+
+        common_speculative_accept(spec, ids.size() - 1);
+
+        // full acceptance: consume the draft and commit accepted tokens
         n_past    += ids.size() - 1;
-        n_drafted += draft.size(); // note: we ignore the discarded small drafts
+        n_drafted += n_draft; // note: we ignore the discarded small drafts
         n_accept  += ids.size() - 1;
         n_predict += ids.size();
 
@@ -222,6 +299,9 @@ int main(int argc, char ** argv) {
 
         LOG_DBG("accepted %d/%d draft tokens, the last target token is: (%d)\n", (int) ids.size() - 1, (int) draft.size(), id_last);
 
+        // clear the draft since it has been consumed
+        draft.clear();
+
         {
             LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
 
@@ -243,7 +323,7 @@ int main(int argc, char ** argv) {
     LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
 
     LOG_INF("\n");
-    LOG_INF("n_draft   = %d\n", params_spec.n_max);
+    LOG_INF("n_draft   = %d\n", params_spec.draft.n_max);
     LOG_INF("n_predict = %d\n", n_predict);
     LOG_INF("n_drafted = %d\n", n_drafted);
     LOG_INF("n_accept  = %d\n", n_accept);
@@ -254,11 +334,10 @@ int main(int argc, char ** argv) {
 
     LOG_INF("\n");
     LOG_INF("target:\n\n");
-    common_perf_print(ctx_tgt, smpl);
+    common_perf_print(ctx_tgt, smpl.get());
 
     llama_batch_free(batch_tgt);
 
-    common_sampler_free(smpl);
     common_speculative_free(spec);
 
     llama_backend_free();

examples/speculative/speculative.cpp

@@ -49,7 +49,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    if (params.speculative.mparams_dft.path.empty()) {
+    if (params.speculative.draft.mparams.path.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;
     }
@@ -58,7 +58,7 @@ int main(int argc, char ** argv) {
     const int n_seq_dft = params.n_parallel;
 
     // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
-    const float p_draft_split = params.speculative.p_split;
+    const float p_draft_split = params.speculative.draft.p_split;
 
     std::default_random_engine rng(params.sampling.seed == LLAMA_DEFAULT_SEED ? std::random_device()() : params.sampling.seed);
     std::uniform_real_distribution<> u_dist;
@@ -80,15 +80,15 @@ int main(int argc, char ** argv) {
     ctx_tgt   = llama_init_tgt->context();
 
     // load the draft model
-    params.devices = params.speculative.devices;
-    params.model = params.speculative.mparams_dft;
-    params.n_gpu_layers = params.speculative.n_gpu_layers;
-    if (params.speculative.cpuparams.n_threads > 0) {
-        params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
+    params.devices = params.speculative.draft.devices;
+    params.model = params.speculative.draft.mparams;
+    params.n_gpu_layers = params.speculative.draft.n_gpu_layers;
+    if (params.speculative.draft.cpuparams.n_threads > 0) {
+        params.cpuparams.n_threads = params.speculative.draft.cpuparams.n_threads;
     }
 
-    params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
-    params.tensor_buft_overrides     = params.speculative.tensor_buft_overrides;
+    params.cpuparams_batch.n_threads = params.speculative.draft.cpuparams_batch.n_threads;
+    params.tensor_buft_overrides     = params.speculative.draft.tensor_buft_overrides;
 
     auto llama_init_dft = common_init_from_params(params);
 
@@ -110,13 +110,21 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    if (
-        llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
-        llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
-        llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft) ||
-        llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)
-    ) {
-        LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
+    if (llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
+        (llama_vocab_get_add_bos(vocab_tgt) && llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft))) {
+        LOG_ERR("%s: draft model bos tokens must match target model to use speculation. add: %d - %d, id: %d - %d)\n",
+                __func__,
+                llama_vocab_get_add_bos(vocab_tgt), llama_vocab_get_add_bos(vocab_dft),
+                llama_vocab_bos(vocab_tgt), llama_vocab_bos(vocab_dft));
+        return 1;
+    }
+
+    if (llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
+        (llama_vocab_get_add_eos(vocab_tgt) && llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft))) {
+        LOG_ERR("%s: draft model eos tokens must match target model to use speculation. add: %d - %d, id: %d - %d)\n",
+                __func__,
+                llama_vocab_get_add_eos(vocab_tgt), llama_vocab_get_add_eos(vocab_dft),
+                llama_vocab_eos(vocab_tgt), llama_vocab_eos(vocab_dft));
         return 1;
     }
 
@@ -137,11 +145,12 @@ int main(int argc, char ** argv) {
         for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
             const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i);
             const char * token_text_dft = llama_vocab_get_text(vocab_dft, i);
+
             if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
                 LOG_ERR("%s: draft model vocab must match target model to use speculation but ", __func__);
                 LOG_ERR("token %d content differs - target '%s', draft '%s'\n", i,
-                        common_token_to_piece(ctx_tgt, i).c_str(),
-                        common_token_to_piece(ctx_dft, i).c_str());
+                        common_token_to_piece(vocab_tgt, i).c_str(),
+                        common_token_to_piece(vocab_dft, i).c_str());
                 return 1;
             }
         }
@@ -183,7 +192,7 @@ int main(int argc, char ** argv) {
     //GGML_ASSERT(n_vocab == llama_vocab_n_tokens(model_dft));
 
     // how many tokens to draft each time
-    int n_draft = params.speculative.n_max;
+    int n_draft = params.speculative.draft.n_max;
 
     int n_predict = 0;
     int n_drafted = 0;

examples/sycl/start-svr.sh

@@ -0,0 +1,124 @@
+#!/bin/bash
+
+#  MIT license
+#  Copyright (C) 2024 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+Help() {
+  cat << EOF
+Usage: $(basename "$0") [OPTIONS]
+
+This script processes files with specified options.
+
+Options:
+  -h, --help    Display this help message and exit.
+  -c, --context <value>    Set context length. Bigger need more memory.
+  -p, --promote <value>    Prompt to start generation with.
+  -m, --model   <value>    Full model file path.
+  -mg,--main-gpu <value>   Set main GPU ID (0 - n) for single GPU mode.
+  -sm,--split-mode <value> How to split the model across multiple GPUs, one of:
+                            - none: use one GPU only
+                            - layer (default): split layers and KV across GPUs
+                            - row: split rows across GPUs
+  -ngl,--n-gpu-layers <value>  Max. number of layers to store in VRAM (default: -1)
+  -lv,--log-verbosity <value>  Set the verbosity threshold. Messages with a higher verbosity will be
+                               ignored. Values:
+                                - 0: generic output
+                                - 1: error
+                                - 2: warning
+                                - 3: info
+                                - 4: debug
+
+
+EOF
+}
+
+BIN_FILE=./build/bin/llama-server
+SEED=0
+GPUS_SETTING=""
+
+MODEL_FILE=../models/Qwen3.5-4B-Q4_0.gguf
+NGL=99
+CONTEXT=4096
+GGML_SYCL_DEVICE=-1
+SPLIT_MODE=layer
+LOG_VERBOSE=3
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        -c|--context)
+            CONTEXT=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -m|--model)
+            MODEL_FILE="$2"
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -mg|--main-gpu)
+            GGML_SYCL_DEVICE=$2
+            SPLIT_MODE=none
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -sm|--split-mode)
+            SPLIT_MODE=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -ngl|--n-gpu-layers)
+            NGL=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -lv|--log-verbosity)
+            LOG_VERBOSE=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -h|--help)
+            Help
+            exit 0
+            ;;
+        *)
+            # Handle unknown options or stop processing options
+            echo "Invalid option: $1"
+            # Optional: exit script or shift to treat remaining as positional args
+            exit 1
+            ;;
+    esac
+done
+
+
+
+source /opt/intel/oneapi/setvars.sh
+
+#export GGML_SYCL_DEBUG=1
+
+#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
+
+#support malloc device memory more than 4GB.
+export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
+echo "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=${UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS}"
+
+if [ $GGML_SYCL_DEVICE -ne -1 ]; then
+    echo "Use $GGML_SYCL_DEVICE as main GPU"
+    #use signle GPU only
+    GPUS_SETTING="-mg $GGML_SYCL_DEVICE -sm ${SPLIT_MODE}"
+    export ONEAPI_DEVICE_SELECTOR="level_zero:${$GGML_SYCL_DEVICE}"
+    echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
+else
+    echo "Use all Intel GPUs, including iGPU & dGPU"
+    GPUS_SETTING="-sm ${SPLIT_MODE}"
+ fi
+
+echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE}  --mmap "
+ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --mmap --host 0.0.0.0 --port 8000
+
+

examples/sycl/test.sh

@@ -38,7 +38,7 @@ SEED=0
 GPUS_SETTING=""
 
 INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
-MODEL_FILE=models/llama-2-7b.Q4_0.gguf
+MODEL_FILE=../models/llama-2-7b.Q4_0.gguf
 NGL=99
 CONTEXT=4096
 GGML_SYCL_DEVICE=-1
@@ -122,9 +122,10 @@ if [ $GGML_SYCL_DEVICE -ne -1 ]; then
     export ONEAPI_DEVICE_SELECTOR="level_zero:${$GGML_SYCL_DEVICE}"
     echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
 else
-   echo "Use all Intel GPUs, including iGPU & dGPU"
+    echo "Use all Intel GPUs, including iGPU & dGPU"
+    GPUS_SETTING="-sm ${SPLIT_MODE}"
  fi
 
-echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE}  --mmap "
-ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --mmap
+echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE}  --mmap "
+ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --mmap
 

examples/sycl/win-start-svr.bat

@@ -0,0 +1,179 @@
+::  MIT license
+::  Copyright (C) 2024 Intel Corporation
+::  SPDX-License-Identifier: MIT
+
+@echo off
+setlocal EnableExtensions EnableDelayedExpansion
+
+set "BIN_FILE=.\build\bin\llama-server.exe"
+set "SEED=0"
+set "GPUS_SETTING="
+
+set "MODEL_FILE=..\models\Qwen3.5-4B-Q4_0.gguf"
+set "NGL=99"
+set "CONTEXT=4096"
+set "GGML_SYCL_DEVICE=-1"
+set "SPLIT_MODE=layer"
+set "LOG_VERBOSE=3"
+
+if "%~1"=="" goto after_args
+
+:parse_args
+if "%~1"=="" goto after_args
+
+if /I "%~1"=="-c" (
+  if "%~2"=="" goto missing_value
+  set "CONTEXT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--context" (
+  if "%~2"=="" goto missing_value
+  set "CONTEXT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-m" (
+  if "%~2"=="" goto missing_value
+  set "MODEL_FILE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--model" (
+  if "%~2"=="" goto missing_value
+  set "MODEL_FILE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-mg" (
+  if "%~2"=="" goto missing_value
+  set "GGML_SYCL_DEVICE=%~2"
+  set "SPLIT_MODE=none"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--main-gpu" (
+  if "%~2"=="" goto missing_value
+  set "GGML_SYCL_DEVICE=%~2"
+  set "SPLIT_MODE=none"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-sm" (
+  if "%~2"=="" goto missing_value
+  set "SPLIT_MODE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--split-mode" (
+  if "%~2"=="" goto missing_value
+  set "SPLIT_MODE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-ngl" (
+  if "%~2"=="" goto missing_value
+  set "NGL=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--n-gpu-layers" (
+  if "%~2"=="" goto missing_value
+  set "NGL=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-lv" (
+  if "%~2"=="" goto missing_value
+  set "LOG_VERBOSE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--log-verbosity" (
+  if "%~2"=="" goto missing_value
+  set "LOG_VERBOSE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-h" goto help
+if /I "%~1"=="--help" goto help
+
+echo Invalid option: %~1
+exit /b 1
+
+:missing_value
+echo Missing value for option: %~1
+exit /b 1
+
+:help
+echo Usage: %~n0 [OPTIONS]
+echo.
+echo This script processes files with specified options.
+echo.
+echo Options:
+echo   -h, --help    Display this help message and exit.
+echo   -c, --context ^<value^>    Set context length. Bigger need more memory.
+echo   -m, --model   ^<value^>    Full model file path.
+echo   -mg,--main-gpu ^<value^>   Set main GPU ID (0 - n) for single GPU mode.
+echo   -sm,--split-mode ^<value^> How to split the model across multiple GPUs, one of:
+echo                             - none: use one GPU only
+echo                             - layer (default): split layers and KV across GPUs
+echo                             - row: split rows across GPUs
+echo   -ngl,--n-gpu-layers ^<value^>  Max. number of layers to store in VRAM (default: -1)
+echo   -lv,--log-verbosity ^<value^>  Set the verbosity threshold. Messages with a higher verbosity will be
+echo                                ignored. Values:
+echo                                 - 0: generic output
+echo                                 - 1: error
+echo                                 - 2: warning
+echo                                 - 3: info
+echo                                 - 4: debug
+exit /b 0
+
+:after_args
+
+REM In Windows CMD, source is not available; call oneAPI setvars if present.
+if exist "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" (
+  call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" >nul
+) else (
+  echo Warning: oneAPI setvars.bat not found. Continuing without environment setup.
+)
+
+REM Support malloc device memory more than 4GB.
+set "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1"
+echo UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=%UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS%
+
+if not "%GGML_SYCL_DEVICE%"=="-1" (
+  echo Use %GGML_SYCL_DEVICE% as main GPU
+  REM Use single GPU only.
+  set "GPUS_SETTING=-mg %GGML_SYCL_DEVICE% -sm %SPLIT_MODE%"
+  set "ONEAPI_DEVICE_SELECTOR=level_zero:%GGML_SYCL_DEVICE%"
+  echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%
+) else (
+  echo Use all Intel GPUs, including iGPU ^& dGPU
+  set "GPUS_SETTING=-sm %SPLIT_MODE%"
+)
+
+echo run cmd: ZES_ENABLE_SYSMAN=1 %BIN_FILE% -m "%MODEL_FILE%" -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap --host 0.0.0.0 --port 8000
+set "ZES_ENABLE_SYSMAN=1"
+%BIN_FILE% -m "%MODEL_FILE%" -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap --host 0.0.0.0 --port 8000
+
+endlocal
+

examples/sycl/win-test.bat

@@ -2,10 +2,200 @@
 ::  Copyright (C) 2024 Intel Corporation
 ::  SPDX-License-Identifier: MIT
 
-set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
-@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
-:: support malloc device memory more than 4GB.
-set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
-set LOAD_MODE="--mmap"
-.\build\bin\llama-completion.exe -m models\llama-2-7b.Q4_0.gguf -no-cnv -p %INPUT2% -n 400 -e -ngl 99 -s 0 %LOAD_MODE%
+@echo off
+setlocal EnableExtensions EnableDelayedExpansion
+
+REM MIT license
+REM Copyright (C) 2024 Intel Corporation
+REM SPDX-License-Identifier: MIT
+
+set "BIN_FILE=.\build\bin\llama-completion.exe"
+set "SEED=0"
+set "GPUS_SETTING="
+
+set "INPUT_PROMPT=Building a website can be done in 10 simple steps:^nStep 1:"
+set "MODEL_FILE=..\models\llama-2-7b.Q4_0.gguf"
+set "NGL=99"
+set "CONTEXT=4096"
+set "GGML_SYCL_DEVICE=-1"
+set "SPLIT_MODE=layer"
+set "LOG_VERBOSE=3"
+
+if "%~1"=="" goto after_args
+
+:parse_args
+if "%~1"=="" goto after_args
+
+if /I "%~1"=="-c" (
+  if "%~2"=="" goto missing_value
+  set "CONTEXT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--context" (
+  if "%~2"=="" goto missing_value
+  set "CONTEXT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-p" (
+  if "%~2"=="" goto missing_value
+  set "INPUT_PROMPT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--promote" (
+  if "%~2"=="" goto missing_value
+  set "INPUT_PROMPT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-m" (
+  if "%~2"=="" goto missing_value
+  set "MODEL_FILE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--model" (
+  if "%~2"=="" goto missing_value
+  set "MODEL_FILE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-mg" (
+  if "%~2"=="" goto missing_value
+  set "GGML_SYCL_DEVICE=%~2"
+  set "SPLIT_MODE=none"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--main-gpu" (
+  if "%~2"=="" goto missing_value
+  set "GGML_SYCL_DEVICE=%~2"
+  set "SPLIT_MODE=none"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-sm" (
+  if "%~2"=="" goto missing_value
+  set "SPLIT_MODE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--split-mode" (
+  if "%~2"=="" goto missing_value
+  set "SPLIT_MODE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-ngl" (
+  if "%~2"=="" goto missing_value
+  set "NGL=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--n-gpu-layers" (
+  if "%~2"=="" goto missing_value
+  set "NGL=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-lv" (
+  if "%~2"=="" goto missing_value
+  set "LOG_VERBOSE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--log-verbosity" (
+  if "%~2"=="" goto missing_value
+  set "LOG_VERBOSE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-h" goto help
+if /I "%~1"=="--help" goto help
+
+echo Invalid option: %~1
+exit /b 1
+
+:missing_value
+echo Missing value for option: %~1
+exit /b 1
+
+:help
+echo Usage: %~n0 [OPTIONS]
+echo.
+echo This script processes files with specified options.
+echo.
+echo Options:
+echo   -h, --help    Display this help message and exit.
+echo   -c, --context ^<value^>    Set context length. Bigger need more memory.
+echo   -p, --promote ^<value^>    Prompt to start generation with.
+echo   -m, --model   ^<value^>    Full model file path.
+echo   -mg,--main-gpu ^<value^>   Set main GPU ID (0 - n) for single GPU mode.
+echo   -sm,--split-mode ^<value^> How to split the model across multiple GPUs, one of:
+echo                             - none: use one GPU only
+echo                             - layer (default): split layers and KV across GPUs
+echo                             - row: split rows across GPUs
+echo   -ngl,--n-gpu-layers ^<value^>  Max. number of layers to store in VRAM (default: -1)
+echo   -lv,--log-verbosity ^<value^>  Set the verbosity threshold. Messages with a higher verbosity will be
+echo                                ignored. Values:
+echo                                 - 0: generic output
+echo                                 - 1: error
+echo                                 - 2: warning
+echo                                 - 3: info
+echo                                 - 4: debug
+exit /b 0
+
+:after_args
+
+REM In Windows CMD, source is not available; call oneAPI setvars if present.
+if exist "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" (
+  call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" >nul
+) else (
+  echo Warning: oneAPI setvars.bat not found. Continuing without environment setup.
+)
+
+REM Support malloc device memory more than 4GB.
+set "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1"
+echo UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=%UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS%
+
+if not "%GGML_SYCL_DEVICE%"=="-1" (
+  echo Use %GGML_SYCL_DEVICE% as main GPU
+  REM Use single GPU only.
+  set "GPUS_SETTING=-mg %GGML_SYCL_DEVICE% -sm %SPLIT_MODE%"
+  set "ONEAPI_DEVICE_SELECTOR=level_zero:%GGML_SYCL_DEVICE%"
+  echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%
+) else (
+  echo Use all Intel GPUs, including iGPU ^& dGPU
+  set "GPUS_SETTING=-sm %SPLIT_MODE%"
+)
+
+echo run cmd: ZES_ENABLE_SYSMAN=1 %BIN_FILE% -m %MODEL_FILE% -no-cnv -p "%INPUT_PROMPT%" -n 200 -e -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap
+set "ZES_ENABLE_SYSMAN=1"
+%BIN_FILE% -m "%MODEL_FILE%" -no-cnv -p "%INPUT_PROMPT%" -n 200 -e -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap
+
+endlocal
+

ggml/CMakeLists.txt

@@ -1,17 +1,11 @@
 cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
 
-# ref: https://cmake.org/cmake/help/latest/policy/CMP0194.html
-# MSVC is not a valid assembler for the ASM language.
-# Set to NEW to avoid a warning on CMake 4.1+ with MSVC.
-if (POLICY CMP0194)
-    cmake_policy(SET CMP0194 NEW)
-endif()
 project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 9)
-set(GGML_VERSION_PATCH 11)
+set(GGML_VERSION_MINOR 10)
+set(GGML_VERSION_PATCH 1)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -219,7 +213,7 @@ set   (GGML_CUDA_COMPRESSION_MODE "size" CACHE STRING
 set_property(CACHE GGML_CUDA_COMPRESSION_MODE PROPERTY STRINGS "none;speed;balance;size")
 
 option(GGML_HIP                             "ggml: use HIP"                                   OFF)
-option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
+option(GGML_HIP_GRAPHS                      "ggml: use HIP graph"                              ON)
 option(GGML_HIP_RCCL                        "ggml: use ROCm Collective Comm. Library"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)

ggml/src/CMakeLists.txt

@@ -470,11 +470,10 @@ endforeach()
 
 target_link_libraries(ggml-base PRIVATE Threads::Threads)
 
-find_library(MATH_LIBRARY m)
-if (MATH_LIBRARY)
-    if (NOT WIN32 OR NOT DEFINED ENV{ONEAPI_ROOT})
-        target_link_libraries(ggml-base PRIVATE m)
-    endif()
+if (DEFINED MATH_LIBRARY)
+    target_link_libraries(ggml-base PRIVATE ${MATH_LIBRARY})
+elseif (NOT WIN32 AND NOT DEFINED ENV{ONEAPI_ROOT})
+    target_link_libraries(ggml-base PRIVATE m)
 endif()
 
 if (CMAKE_SYSTEM_NAME MATCHES "Android")

ggml/src/ggml-backend-meta.cpp

@@ -1133,7 +1133,7 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
         if (t_ij->view_src != nullptr && ggml_backend_buffer_is_meta(t_ij->view_src->buffer)) {
             t_ij->view_src = ggml_backend_meta_buffer_simple_tensor(tensor->view_src, j);
             if (t_ij->view_offs > 0 && split_dim >= 0 && split_dim < GGML_MAX_DIMS) {
-                GGML_ASSERT(ne[split_dim] != 0 && tensor->ne[split_dim] != 0);
+                GGML_ASSERT(tensor->ne[split_dim] != 0);
                 const int split_dim_view_src = ggml_backend_meta_get_split_state(tensor->view_src, /*assume_sync =*/ true).axis;
                 GGML_ASSERT(split_dim_view_src >= 0 && split_dim_view_src < GGML_MAX_DIMS);
 
@@ -1170,6 +1170,28 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
 
         simple_tensors.push_back(t_ij);
     }
+
+    // If one of the sources has a zero-sized slice, disable the computation:
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        if (tensor->src[i] == nullptr || !ggml_backend_buffer_is_meta(tensor->src[i]->buffer)) {
+            continue;
+        }
+
+        const ggml_backend_meta_split_state split_state_src = ggml_backend_meta_get_split_state(tensor->src[i], /*assume_sync =*/ true);
+        if (split_state_src.axis < 0 || split_state_src.axis >= GGML_MAX_DIMS) {
+            continue;
+        }
+        for (size_t j = 0; j < n_simple_bufs; j++) {
+            int64_t ne_sum = 0;
+            for (size_t s = 0; s < split_state_src.n_segments; s++) {
+                ne_sum += split_state_src.ne[s*n_simple_bufs + j];
+            }
+            if (ne_sum == 0) {
+                simple_tensors[j]->flags &= ~GGML_TENSOR_FLAG_COMPUTE;
+            }
+        }
+    }
+
     buf_ctx->simple_tensors[tensor] = simple_tensors;
 
     return GGML_STATUS_SUCCESS;
@@ -1183,40 +1205,57 @@ static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, gg
 
     if (split_state.n_segments != 1) {
         GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
-        GGML_ASSERT(offset == 0);
-        GGML_ASSERT(size == ggml_nbytes(tensor));
         GGML_ASSERT(tensor->ne[3] == 1);
+
         size_t offset_data = 0;
         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 r_start = offset / row_stride;
+            const int64_t r_count = size   / row_stride;
+            GGML_ASSERT(r_start + r_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 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];
-                    ggml_backend_tensor_set_2d(simple_tensor, (const char *) data + offset_data, simple_offsets[j], nbytes,
-                        tensor->ne[1], simple_tensor->nb[1], tensor->nb[1]);
+                    ggml_backend_tensor_set_2d(simple_tensor, (const char *) data + offset_data,
+                        simple_offsets[j] + r_start * simple_tensor->nb[1], nbytes,
+                        r_count, simple_tensor->nb[1], tensor->nb[1]);
                     offset_data       += nbytes;
                     simple_offsets[j] += nbytes;
                 }
             }
-            GGML_ASSERT(offset_data*tensor->ne[1] == size);
+            GGML_ASSERT(offset_data*r_count == size);
             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 r_start = offset / row_stride;
+        const int64_t r_count = size   / row_stride;
+        GGML_ASSERT(r_start + r_count <= tensor->ne[2]);
+
         for (size_t s = 0; s < split_state.n_segments; s++) {
             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];
-                ggml_backend_tensor_set_2d(simple_tensor, (const char *) data + offset_data, simple_offsets[j], nbytes,
-                    tensor->ne[2], simple_tensor->nb[2], tensor->nb[2]);
+                ggml_backend_tensor_set_2d(simple_tensor, (const char *) data + offset_data,
+                    simple_offsets[j] + r_start * simple_tensor->nb[2], nbytes,
+                    r_count, simple_tensor->nb[2], tensor->nb[2]);
                 offset_data       += nbytes;
                 simple_offsets[j] += nbytes;
             }
         }
-        GGML_ASSERT(offset_data*tensor->ne[2] == size);
+        GGML_ASSERT(offset_data*r_count == size);
         return;
     }
 
@@ -1273,40 +1312,57 @@ static void ggml_backend_meta_buffer_get_tensor(ggml_backend_buffer_t buffer, co
 
     if (split_state.n_segments != 1) {
         GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
-        GGML_ASSERT(offset == 0);
-        GGML_ASSERT(size == ggml_nbytes(tensor));
         GGML_ASSERT(tensor->ne[3] == 1);
+
         size_t offset_data = 0;
         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 r_start = offset / row_stride;
+            const int64_t r_count = size   / row_stride;
+            GGML_ASSERT(r_start + r_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 j = 0; j < n_bufs; j++) {
                     const 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];
-                    ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data, simple_offsets[j], nbytes,
-                        tensor->ne[1], simple_tensor->nb[1], tensor->nb[1]);
+                    ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data,
+                        simple_offsets[j] + r_start * simple_tensor->nb[1], nbytes,
+                        r_count, simple_tensor->nb[1], tensor->nb[1]);
                     offset_data       += nbytes;
                     simple_offsets[j] += nbytes;
                 }
             }
-            GGML_ASSERT(offset_data*tensor->ne[1] == size);
+            GGML_ASSERT(offset_data*r_count == size);
             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 r_start = offset / row_stride;
+        const int64_t r_count = size   / row_stride;
+        GGML_ASSERT(r_start + r_count <= tensor->ne[2]);
+
         for (size_t s = 0; s < split_state.n_segments; s++) {
             for (size_t j = 0; j < n_bufs; j++) {
                 const 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];
-                ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data, simple_offsets[j], nbytes,
-                    tensor->ne[2], simple_tensor->nb[2], tensor->nb[2]);
+                ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data,
+                    simple_offsets[j] + r_start * simple_tensor->nb[2], nbytes,
+                    r_count, simple_tensor->nb[2], tensor->nb[2]);
                 offset_data       += nbytes;
                 simple_offsets[j] += nbytes;
             }
         }
-        GGML_ASSERT(offset_data*tensor->ne[2] == size);
+        GGML_ASSERT(offset_data*r_count == size);
         return;
     }
 
@@ -1442,26 +1498,32 @@ struct ggml_backend_meta_context {
     struct backend_config {
         ggml_backend_t backend;
 
-        std::vector<cgraph_config> cgraphs;
-        std::vector<ggml_tensor *> nodes;
-        ggml_backend_buffer_ptr    buf;
+        std::vector<cgraph_config>           cgraphs;
+        std::vector<ggml_tensor *>           nodes;
+        std::vector<ggml_backend_buffer_ptr> bufs;
 
-        backend_config(ggml_backend_t backend) : backend(backend) {}
+        backend_config(ggml_backend_t backend, const size_t n_reduce_steps) : backend(backend) {
+            bufs.resize(n_reduce_steps);
+        }
     };
     std::string                 name;
     std::vector<backend_config> backend_configs;
     ggml_context_ptr            ctx;
     std::vector<ggml_cgraph *>  cgraphs_aux;
     std::vector<ggml_tensor *>  nodes_aux;
+    size_t                      n_reduce_steps;
     int                         max_nnodes    = 0;
     size_t                      max_tmp_size  = 0;
     size_t                      max_subgraphs = 0;
+    size_t                      n_subgraphs   = 0;
+    uint64_t                    uid           = 0;
 
     void *                               comm_ctx       = nullptr;
     ggml_backend_comm_allreduce_tensor_t comm_allreduce = nullptr;
 
     ggml_backend_meta_context(ggml_backend_dev_t meta_dev, const char * params) {
         const size_t n_devs = ggml_backend_meta_dev_n_devs(meta_dev);
+        n_reduce_steps = std::ceil(std::log2(n_devs));
         name = "Meta(";
         std::vector<ggml_backend_t> simple_backends;
         backend_configs.reserve(n_devs);
@@ -1473,7 +1535,7 @@ struct ggml_backend_meta_context {
             }
             name += ggml_backend_dev_name(simple_dev);
             simple_backends.push_back(ggml_backend_dev_init(simple_dev, params));
-            backend_configs.emplace_back(simple_backends.back());
+            backend_configs.emplace_back(simple_backends.back(), n_reduce_steps);
         }
         name += ")";
 
@@ -1503,10 +1565,6 @@ struct ggml_backend_meta_context {
             ggml_backend_free(bc.backend);
         }
     }
-
-    size_t n_reduce_steps() const {
-        return std::ceil(std::log2(backend_configs.size()));
-    }
 };
 
 static const char * ggml_backend_meta_get_name(ggml_backend_t backend) {
@@ -1616,6 +1674,9 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
     const size_t n_backends = ggml_backend_meta_n_backends(backend);
     ggml_backend_meta_context * backend_ctx = (ggml_backend_meta_context *) backend->context;
 
+    // If the previous cgraph had a defined UID it can be used to skip rebuilding the subgraphs per simple backend.
+    const bool needs_rebuild = (cgraph->uid == 0) || (cgraph->uid != backend_ctx->uid);
+
     bool max_nnodes_raised = false;
     if (cgraph->n_nodes > backend_ctx->max_nnodes) {
         for (size_t j = 0; j < n_backends; j++) {
@@ -1625,173 +1686,233 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
         }
         backend_ctx->max_nnodes = cgraph->n_nodes;
         max_nnodes_raised = true;
+        assert(needs_rebuild);
     }
-    for (size_t j = 0; j < n_backends; j++) {
-        auto & bcj = backend_ctx->backend_configs[j];
 
-        for (int i = 0; i < cgraph->n_nodes; i++) {
-            ggml_tensor * node = cgraph->nodes[i];
-            if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
-                // FIXME s_copy_main is on the CPU and its view seems to be incorrectly added to the graph nodes.
-                // For regular usage this doesn't matter since it's a noop but trying to call ggml_backend_meta_buffer_simple_tensor results in a crash.
-                bcj.nodes[i] = node;
-                continue;
+    if (needs_rebuild) {
+        size_t n_subgraphs  = 0;
+        size_t max_tmp_size = 0;
+
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+                if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
+                    // FIXME s_copy_main is on the CPU and its view seems to be incorrectly added to the graph nodes.
+                    // For regular usage this doesn't matter since it's a noop but trying to call ggml_backend_meta_buffer_simple_tensor results in a crash.
+                    bcj.nodes[i] = node;
+                    continue;
+                }
+                bcj.nodes[i] = ggml_backend_meta_buffer_simple_tensor(node, j);
+                GGML_ASSERT(bcj.nodes[i]);
             }
-            bcj.nodes[i] = ggml_backend_meta_buffer_simple_tensor(node, j);
-            GGML_ASSERT(bcj.nodes[i]);
         }
-    }
 
-    size_t n_subgraphs  = 0;
-    size_t max_tmp_size = 0;
-    {
-        // 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 {
-            int id = i; // i_delayed
-            int idr = i; // i_delayed return, last safe return value
+        {
+            // 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 {
+                int id = i; // i_delayed
+                int idr = i; // i_delayed return, last safe return value
 
-            ggml_tensor * node = cgraph->nodes[id];
-            int32_t n_used = ggml_node_get_use_count(cgraph, id);
-            if (id + 1 >= cgraph->n_nodes) {
-                return idr;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op == GGML_OP_ADD_ID && next->src[0] == node &&
-                        ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL &&
-                        ggml_backend_meta_get_split_state(next->src[2], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-                    node = next;
+                ggml_tensor * node = cgraph->nodes[id];
+                int32_t n_used = ggml_node_get_use_count(cgraph, id);
+
+                // Skip MIRRORED nodes that don't consume node
+                auto skip_unrelated = [&]() {
+                    while (id + 1 < cgraph->n_nodes) {
+                        ggml_tensor * next = cgraph->nodes[id+1];
+                        if (ggml_backend_meta_get_split_state(next, false).axis != GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                            break;
+                        }
+                        bool safe = true;
+                        for (int s = 0; s < GGML_MAX_SRC; s++) {
+                            if (next->src[s] == nullptr) {
+                                continue;
+                            }
+                            if (next->src[s] == node) {
+                                safe = false;
+                                break;
+                            }
+                            if (ggml_backend_meta_get_split_state(next->src[s], false).axis != GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                                safe = false;
+                                break;
+                            }
+                        }
+                        if (!safe) {
+                            break;
+                        }
+                        id++;
+                    }
+                };
+
+                skip_unrelated();
+                if (id + 1 >= cgraph->n_nodes) {
+                    return idr;
+                }
+                {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op == GGML_OP_ADD_ID && next->src[0] == node &&
+                            ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL &&
+                            ggml_backend_meta_get_split_state(next->src[2], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        node = next;
+                        id++;
+                        idr = id;
+                        n_used = ggml_node_get_use_count(cgraph, id);
+                    }
+                }
+                // Chain of MULs with MIRRORED src[1]
+                while (true) {
+                    skip_unrelated();
+                    if (id + 1 >= cgraph->n_nodes) {
+                        return idr;
+                    }
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op == GGML_OP_MUL && next->src[0] == node &&
+                            ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        node = next;
+                        id++;
+                        idr = id;
+                        n_used = ggml_node_get_use_count(cgraph, id);
+                    } else {
+                        break;
+                    }
+                }
+
+                if (n_used != node->ne[1] || id + 2*n_used-1 >= cgraph->n_nodes) {
+                    return idr;
+                }
+                for (int32_t k = 0; k < n_used; k++) {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_VIEW || next->view_src != node || next->view_offs != k*node->nb[1] ||
+                            next->ne[0] != node->ne[0] || next->ne[1] != node->ne[2] || next->nb[1] != node->nb[2] ||
+                            ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
                     id++;
-                    idr = id;
-                    n_used = ggml_node_get_use_count(cgraph, id);
                 }
-            }
-            if (id + 1 >= cgraph->n_nodes) {
-                return idr;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op == GGML_OP_MUL && next->src[0] == node &&
-                        ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-                    node = next;
+                {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id - (n_used-1)] ||
+                            next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
                     id++;
-                    idr = id;
-                    n_used = ggml_node_get_use_count(cgraph, id);
                 }
-            }
-
-            if (n_used != node->ne[1] || id + 2*n_used-1 >= cgraph->n_nodes) {
+                for (int32_t k = 0; k < n_used - 2; k++) {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id] ||
+                            next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
+                    id++;
+                }
+                idr = id;
                 return idr;
-            }
-            for (int32_t k = 0; k < n_used; k++) {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_VIEW || next->view_src != node || next->view_offs != k*node->nb[1] ||
-                        next->ne[0] != node->ne[0] || next->ne[1] != node->ne[2] || next->nb[1] != node->nb[2] ||
-                        ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id - (n_used-1)] ||
-                        next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            for (int32_t k = 0; k < n_used - 2; k++) {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id] ||
-                        next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            idr = id;
-            return idr;
-        };
+            };
 
-        int i_start = 0;
-        for (int i = 0; i < cgraph->n_nodes; i++) {
-            ggml_tensor * node = cgraph->nodes[i];
-            if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
-                continue;
-            }
-            const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(node, /*assume_sync =*/ false);
-            if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
-                max_tmp_size = std::max(max_tmp_size, ggml_nbytes(node));
-            }
-            const bool new_subgraph = i + 1 == cgraph->n_nodes || split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL;
-            if (!new_subgraph) {
-                continue;
-            }
+            int i_start = 0;
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+                if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
+                    continue;
+                }
+                const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(node, /*assume_sync =*/ false);
+                if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
+                    max_tmp_size = std::max(max_tmp_size, ggml_nbytes(node));
+                }
+                const bool new_subgraph = i + 1 == cgraph->n_nodes || split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL;
+                if (!new_subgraph) {
+                    continue;
+                }
 
-            i = get_i_delayed(i);
+                const int i_delayed = get_i_delayed(i);
 
+                // If we can delay the AllReduce we need to consider the interaction with zero-sized tensor slices.
+                // A backend with such a slice would normally have valid data after participating in the AllReduce with a node that has
+                //     its compute flag disabled and thus gets its data zeroed out.
+                // If the AllReduce is delayed then the nodes until that point also need to have their compute flag disabled.
+                if (i_delayed > i) {
+                    for (size_t j = 0; j < n_backends; j++) {
+                        auto & bcj = backend_ctx->backend_configs[j];
+                        if ((bcj.nodes[i]->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                            for (int ii = i + 1; ii <= i_delayed; ii++) {
+                                bcj.nodes[ii]->flags &= ~GGML_TENSOR_FLAG_COMPUTE;
+                            }
+                        }
+                    }
+                }
+
+                i = i_delayed;
+
+                for (size_t j = 0; j < n_backends; j++) {
+                    auto & bcj = backend_ctx->backend_configs[j];
+                    bcj.cgraphs[n_subgraphs].offset = i_start;
+                }
+                n_subgraphs++;
+                i_start = i + 1;
+            }
+            GGML_ASSERT(i_start == cgraph->n_nodes);
+        }
+
+        backend_ctx->uid         = cgraph->uid;
+        backend_ctx->n_subgraphs = n_subgraphs;
+
+        if (max_tmp_size > backend_ctx->max_tmp_size) {
             for (size_t j = 0; j < n_backends; j++) {
                 auto & bcj = backend_ctx->backend_configs[j];
-                bcj.cgraphs[n_subgraphs].offset = i_start;
+                for (size_t i = 0; i < backend_ctx->n_reduce_steps; i++) {
+                    bcj.bufs[i].reset(ggml_backend_alloc_buffer(bcj.backend, max_tmp_size));
+                }
             }
-            n_subgraphs++;
-            i_start = i + 1;
+            backend_ctx->max_tmp_size = max_tmp_size;
         }
-        GGML_ASSERT(i_start == cgraph->n_nodes);
-    }
 
-    if (max_tmp_size > backend_ctx->max_tmp_size) {
-        for (size_t j = 0; j < n_backends; j++) {
-            auto & bcj = backend_ctx->backend_configs[j];
-            bcj.buf.reset(ggml_backend_alloc_buffer(bcj.backend, max_tmp_size));
-        }
-        backend_ctx->max_tmp_size = max_tmp_size;
-    }
-
-
-    if (max_nnodes_raised || n_subgraphs > backend_ctx->max_subgraphs) {
-        backend_ctx->max_subgraphs = std::max(backend_ctx->max_subgraphs, n_subgraphs);
-        const size_t n_reduce_steps = backend_ctx->n_reduce_steps();
-        const size_t n_nodes_per_device = 2 * n_reduce_steps; // tmp + ADD per step
-        const size_t n_cgraphs_per_device = n_reduce_steps;    // 1 ADD graph per step
-        const size_t mem_per_device_graphs_main = backend_ctx->max_subgraphs*ggml_graph_overhead_custom(backend_ctx->max_nnodes, cgraph->grads);
-        const size_t mem_per_device_graphs_aux = n_cgraphs_per_device*backend_ctx->max_subgraphs*ggml_graph_overhead_custom(1, cgraph->grads);
-        const size_t mem_per_device_nodes_aux = n_nodes_per_device*backend_ctx->max_subgraphs*ggml_tensor_overhead();
-        ggml_init_params params = {
-            /*.mem_size   =*/ n_backends * (mem_per_device_graphs_main + mem_per_device_graphs_aux + mem_per_device_nodes_aux),
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        backend_ctx->ctx.reset(ggml_init(params));
-        for (size_t j = 0; j < n_backends; j++) {
-            auto & bcj = backend_ctx->backend_configs[j];
-            for (size_t i = 0; i < n_subgraphs; i++) {
-                bcj.cgraphs[i].cgraph_main = ggml_new_graph_custom(backend_ctx->ctx.get(), cgraph->n_nodes, /*grads =*/ false);
+        if (max_nnodes_raised || n_subgraphs > backend_ctx->max_subgraphs) {
+            backend_ctx->max_subgraphs = std::max(backend_ctx->max_subgraphs, n_subgraphs);
+            const size_t n_nodes_per_device = 3 * backend_ctx->n_reduce_steps; // tmp + ADD (+zeroing) graph per step and device
+            const size_t n_cgraphs_per_device = 2 * backend_ctx->n_reduce_steps; // ADD ( + zeroing) graph per step and device
+            const size_t mem_per_device_graphs_main = backend_ctx->max_subgraphs*ggml_graph_overhead_custom(backend_ctx->max_nnodes, cgraph->grads);
+            const size_t mem_per_device_graphs_aux = n_cgraphs_per_device*backend_ctx->max_subgraphs*ggml_graph_overhead_custom(1, cgraph->grads);
+            const size_t mem_per_device_nodes_aux = n_nodes_per_device*backend_ctx->max_subgraphs*ggml_tensor_overhead();
+            ggml_init_params params = {
+                /*.mem_size   =*/ n_backends * (mem_per_device_graphs_main + mem_per_device_graphs_aux + mem_per_device_nodes_aux),
+                /*.mem_buffer =*/ nullptr,
+                /*.no_alloc   =*/ true,
+            };
+            backend_ctx->ctx.reset(ggml_init(params));
+            for (size_t j = 0; j < n_backends; j++) {
+                auto & bcj = backend_ctx->backend_configs[j];
+                for (size_t i = 0; i < n_subgraphs; i++) {
+                    bcj.cgraphs[i].cgraph_main = ggml_new_graph_custom(backend_ctx->ctx.get(), cgraph->n_nodes, /*grads =*/ false);
+                }
+            }
+            backend_ctx->cgraphs_aux.resize(n_backends*n_cgraphs_per_device*backend_ctx->max_subgraphs);
+            for (size_t k = 0; k < backend_ctx->cgraphs_aux.size(); k++) {
+                backend_ctx->cgraphs_aux[k] = ggml_new_graph_custom(backend_ctx->ctx.get(), 1, cgraph->grads);
+            }
+            backend_ctx->nodes_aux.resize(n_backends*n_nodes_per_device*backend_ctx->max_subgraphs);
+            for (size_t k = 0; k < backend_ctx->nodes_aux.size(); k++) {
+                backend_ctx->nodes_aux[k] = ggml_new_tensor_1d(backend_ctx->ctx.get(), GGML_TYPE_F32, 1);
             }
         }
-        backend_ctx->cgraphs_aux.resize(n_backends*n_cgraphs_per_device*backend_ctx->max_subgraphs);
-        for (size_t k = 0; k < backend_ctx->cgraphs_aux.size(); k++) {
-            backend_ctx->cgraphs_aux[k] = ggml_new_graph_custom(backend_ctx->ctx.get(), 1, cgraph->grads);
-        }
-        backend_ctx->nodes_aux.resize(n_backends*n_nodes_per_device*backend_ctx->max_subgraphs);
-        for (size_t k = 0; k < backend_ctx->nodes_aux.size(); k++) {
-            backend_ctx->nodes_aux[k] = ggml_new_tensor_1d(backend_ctx->ctx.get(), GGML_TYPE_F32, 1);
-        }
-    }
 
-    for (size_t j = 0; j < n_backends; j++) {
-        auto & bcj = backend_ctx->backend_configs[j];
-        for (size_t i_graph = 0; i_graph < n_subgraphs; i_graph++) {
-            ggml_cgraph * cgraph_ij = bcj.cgraphs[i_graph].cgraph_main;
-            const size_t i_node_start = bcj.cgraphs[i_graph].offset;
-            const size_t i_node_stop = i_graph + 1 < n_subgraphs ? bcj.cgraphs[i_graph + 1].offset : cgraph->n_nodes;
-            cgraph_ij->n_nodes = i_node_stop - i_node_start;
-            ggml_hash_set_reset(&cgraph_ij->visited_hash_set);
-            for (size_t i_node = i_node_start; i_node < i_node_stop; i_node++) {
-                ggml_tensor * node_ij = bcj.nodes[i_node];
-                cgraph_ij->nodes[i_node - i_node_start] = node_ij;
-                const size_t hash_pos_orig = ggml_hash_find(&cgraph->visited_hash_set, cgraph->nodes[i_node]);
-                const size_t hash_pos_ij = ggml_hash_insert(&cgraph_ij->visited_hash_set, node_ij);
-                cgraph_ij->use_counts[hash_pos_ij] = cgraph->use_counts[hash_pos_orig];
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+            for (size_t i_graph = 0; i_graph < n_subgraphs; i_graph++) {
+                ggml_cgraph * cgraph_ij = bcj.cgraphs[i_graph].cgraph_main;
+                const size_t i_node_start = bcj.cgraphs[i_graph].offset;
+                const size_t i_node_stop = i_graph + 1 < n_subgraphs ? bcj.cgraphs[i_graph + 1].offset : cgraph->n_nodes;
+                cgraph_ij->n_nodes = i_node_stop - i_node_start;
+                ggml_hash_set_reset(&cgraph_ij->visited_hash_set);
+                for (size_t i_node = i_node_start; i_node < i_node_stop; i_node++) {
+                    ggml_tensor * node_ij = bcj.nodes[i_node];
+                    cgraph_ij->nodes[i_node - i_node_start] = node_ij;
+                    const size_t hash_pos_orig = ggml_hash_find(&cgraph->visited_hash_set, cgraph->nodes[i_node]);
+                    const size_t hash_pos_ij = ggml_hash_insert(&cgraph_ij->visited_hash_set, node_ij);
+                    cgraph_ij->use_counts[hash_pos_ij] = cgraph->use_counts[hash_pos_orig];
+                }
+                cgraph_ij->uid = ggml_graph_next_uid();
             }
         }
     }
@@ -1799,11 +1920,6 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
     size_t iga = 0; // i graph aux
     size_t ina = 0; // i node aux
 
-    // FIXME usage_counts
-    auto get_cgraph_aux = [&]() -> ggml_cgraph * {
-        ggml_cgraph * ret = backend_ctx->cgraphs_aux[iga++];
-        return ret;
-    };
     auto get_node_aux = [&](ggml_tensor * t) -> ggml_tensor * {
         ggml_tensor * ret = backend_ctx->nodes_aux[ina++];
         memset(ret, 0, sizeof(ggml_tensor));
@@ -1815,75 +1931,110 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
         }
         return ret;
     };
+    auto set_tmp_data = [&](ggml_tensor * tensor, const size_t j, const size_t i_buf) {
+        auto & bcj = backend_ctx->backend_configs[j];
+        ggml_backend_buffer_ptr & buf_ptr = bcj.bufs[i_buf];
+        if (!buf_ptr || ggml_backend_buffer_get_size(buf_ptr.get()) < backend_ctx->max_tmp_size) {
+            buf_ptr.reset(ggml_backend_alloc_buffer(bcj.backend, backend_ctx->max_tmp_size));
+        }
+        tensor->buffer = buf_ptr.get();
+        tensor->data   = ggml_backend_buffer_get_base(buf_ptr.get());
+    };
+    // FIXME usage_counts
+    auto get_cgraph_aux = [&]() -> ggml_cgraph * {
+        ggml_cgraph * ret = backend_ctx->cgraphs_aux[iga++];
+        return ret;
+    };
 
     // Preferentially use backend-specific allreduce_tensor_async (e.g. NCCL for CUDA), use a generic fallback if unavailable:
     auto allreduce_fallback = [&](size_t i) -> ggml_status {
         std::vector<ggml_cgraph *> step_cgraphs(n_backends, nullptr);
 
-        for (size_t offset_j = 1; offset_j < n_backends; offset_j *= 2) {
+        // Zero out nodes that were disabled due to having a zero-sized slice:
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+            ggml_tensor * node = bcj.cgraphs[i].cgraph_main->nodes[bcj.cgraphs[i].cgraph_main->n_nodes - 1];
+            if (node->flags & GGML_TENSOR_FLAG_COMPUTE) {
+                continue;
+            }
+            ggml_tensor * node_zero = get_node_aux(node);
+            node_zero->op = GGML_OP_SCALE; // FIXME 0.0f * NaN == NaN
+            node_zero->src[0] = node;
+            ggml_set_op_params_f32(node_zero, 0, 0.0f);
+            node_zero->data = node->data;
+            node_zero->flags |= GGML_TENSOR_FLAG_COMPUTE;
+
+            step_cgraphs[j] = get_cgraph_aux();
+            step_cgraphs[j]->nodes[0] = node_zero;
+            step_cgraphs[j]->n_nodes = 1;
+            const ggml_status status = ggml_backend_graph_compute_async(bcj.backend, step_cgraphs[j]);
+            if (status != GGML_STATUS_SUCCESS) {
+                return status;
+            }
+        }
+        std::fill(step_cgraphs.begin(), step_cgraphs.end(), nullptr);
+
+        auto push_data = [&](const size_t j_src, const size_t j_dst, const size_t i_buf) {
+            assert(step_cgraphs[j_dst] == nullptr);
+            auto & bcj_src = backend_ctx->backend_configs[j_src];
+            auto & bcj_dst = backend_ctx->backend_configs[j_dst];
+
+            ggml_tensor * node_src = bcj_src.cgraphs[i].cgraph_main->nodes[bcj_src.cgraphs[i].cgraph_main->n_nodes - 1];
+            ggml_tensor * node_dst = bcj_dst.cgraphs[i].cgraph_main->nodes[bcj_dst.cgraphs[i].cgraph_main->n_nodes - 1];
+            GGML_ASSERT(ggml_is_contiguous(node_src));
+            GGML_ASSERT(ggml_is_contiguous(node_dst));
+
+            ggml_tensor * node_tmp = get_node_aux(node_dst);
+            set_tmp_data(node_tmp, j_dst, i_buf);
+
+            ggml_backend_tensor_copy_async(bcj_src.backend, bcj_dst.backend, node_src, node_tmp);
+
+            ggml_tensor * node_red = get_node_aux(node_dst);
+            node_red->view_src = node_dst->view_src == nullptr ? node_dst : node_dst->view_src;
+            node_red->view_offs = node_dst->view_offs;
+            node_red->op = GGML_OP_ADD;
+            node_red->src[0] = node_dst;
+            node_red->src[1] = node_tmp;
+            node_red->flags |= GGML_TENSOR_FLAG_COMPUTE;
+            ggml_backend_view_init(node_red);
+
+            ggml_cgraph * cgraph_aux = get_cgraph_aux();
+            cgraph_aux->nodes[0] = node_red;
+            cgraph_aux->n_nodes = 1;
+            step_cgraphs[j_dst] = cgraph_aux;
+        };
+
+        size_t offset_j = n_backends/2;
+        while ((offset_j & (offset_j - 1)) != 0) {
+            offset_j--;
+        }
+        const size_t offset_j_max = offset_j;
+        size_t i_buf = 0;
+
+        // If n_backends is not a power of 2, fold in the excess prior to butterfly reduction:
+        for (size_t j_src = 2*offset_j_max; j_src < n_backends; j_src++) {
+            const size_t j_dst = j_src - 2*offset_j_max;
+            push_data(j_src, j_dst, i_buf);
+            const ggml_status status = ggml_backend_graph_compute_async(backend_ctx->backend_configs[j_dst].backend, step_cgraphs[j_dst]);
+            if (status != GGML_STATUS_SUCCESS) {
+                return status;
+            }
+            i_buf = 1;
+        }
+
+        // Butterfly reduction:
+        for (; offset_j >= 1; offset_j /= 2) {
             std::fill(step_cgraphs.begin(), step_cgraphs.end(), nullptr);
 
-            for (size_t j = 0; j < n_backends; j++) {
+            for (size_t j = 0; j < 2*offset_j_max; j++) {
                 const size_t j_other = j ^ offset_j;
-                if (j_other > j) {
+                if (j_other >= n_backends) {
                     continue;
                 }
-
-                auto & bcj1 = backend_ctx->backend_configs[j];
-                auto & bcj2 = backend_ctx->backend_configs[j_other];
-
-                ggml_tensor * node1 = bcj1.cgraphs[i].cgraph_main->nodes[bcj1.cgraphs[i].cgraph_main->n_nodes - 1];
-                ggml_tensor * node2 = bcj2.cgraphs[i].cgraph_main->nodes[bcj2.cgraphs[i].cgraph_main->n_nodes - 1];
-                GGML_ASSERT(ggml_is_contiguous(node1));
-                GGML_ASSERT(ggml_is_contiguous(node2));
-
-                // Tmp tensors to receive P2P copies
-                ggml_tensor * node_tmp_1 = get_node_aux(node1);
-                node_tmp_1->buffer = bcj1.buf.get();
-                node_tmp_1->data = ggml_backend_buffer_get_base(bcj1.buf.get());
-
-                ggml_tensor * node_tmp_2 = get_node_aux(node2);
-                node_tmp_2->buffer = bcj2.buf.get();
-                node_tmp_2->data = ggml_backend_buffer_get_base(bcj2.buf.get());
-
-                // 2 P2P copies: exchange full buffers
-                ggml_backend_tensor_copy_async(bcj1.backend, bcj2.backend, node1, node_tmp_2);
-                ggml_backend_tensor_copy_async(bcj2.backend, bcj1.backend, node2, node_tmp_1);
-
-                // Local ADD: node1 += tmp1 (in-place via view)
-                ggml_tensor * node_red_1 = get_node_aux(node1);
-                node_red_1->view_src = node1->view_src == nullptr ? node1 : node1->view_src;
-                node_red_1->view_offs = node1->view_offs;
-                node_red_1->op = GGML_OP_ADD;
-                node_red_1->src[0] = node1;
-                node_red_1->src[1] = node_tmp_1;
-                node_red_1->flags |= GGML_TENSOR_FLAG_COMPUTE;
-                ggml_backend_view_init(node_red_1);
-
-                // Local ADD: node2 += tmp2 (in-place via view)
-                ggml_tensor * node_red_2 = get_node_aux(node2);
-                node_red_2->view_src = node2->view_src == nullptr ? node2 : node2->view_src;
-                node_red_2->view_offs = node2->view_offs;
-                node_red_2->op = GGML_OP_ADD;
-                node_red_2->src[0] = node2;
-                node_red_2->src[1] = node_tmp_2;
-                node_red_2->flags |= GGML_TENSOR_FLAG_COMPUTE;
-                ggml_backend_view_init(node_red_2);
-
-                // Build 1-node cgraphs for the ADD ops
-                ggml_cgraph * cgraph_aux_1 = get_cgraph_aux();
-                cgraph_aux_1->nodes[0] = node_red_1;
-                cgraph_aux_1->n_nodes = 1;
-                step_cgraphs[j] = cgraph_aux_1;
-
-                ggml_cgraph * cgraph_aux_2 = get_cgraph_aux();
-                cgraph_aux_2->nodes[0] = node_red_2;
-                cgraph_aux_2->n_nodes = 1;
-                step_cgraphs[j_other] = cgraph_aux_2;
+                push_data(j, j_other, i_buf);
             }
 
-            // Execute local ADDs for this step
-            for (size_t j = 0; j < n_backends; j++) {
+            for (size_t j = 0; j < 2*offset_j_max; j++) {
                 if (step_cgraphs[j] == nullptr) {
                     continue;
                 }
@@ -1893,12 +2044,25 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
                     return status;
                 }
             }
+            i_buf++;
         }
+        assert(i_buf == backend_ctx->n_reduce_steps);
+
+        // If n_backends is not a power of 2, copy back the reduced tensors to the excess:
+        for (size_t j = 2*offset_j_max; j < n_backends; j++) {
+            auto & bcj_src = backend_ctx->backend_configs[j - 2*offset_j_max];
+            auto & bcj_dst = backend_ctx->backend_configs[j];
+
+            ggml_tensor * node_src = bcj_src.cgraphs[i].cgraph_main->nodes[bcj_src.cgraphs[i].cgraph_main->n_nodes - 1];
+            ggml_tensor * node_dst = bcj_dst.cgraphs[i].cgraph_main->nodes[bcj_dst.cgraphs[i].cgraph_main->n_nodes - 1];
+            ggml_backend_tensor_copy_async(bcj_src.backend, bcj_dst.backend, node_src, node_dst);
+        }
+
         return GGML_STATUS_SUCCESS;
     };
 
 
-    for (size_t i = 0; i < n_subgraphs; i++) {
+    for (size_t i = 0; i < backend_ctx->n_subgraphs; i++) {
         for (size_t j = 0; j < n_backends; j++) {
             auto & bcj = backend_ctx->backend_configs[j];
             const ggml_status status = ggml_backend_graph_compute_async(bcj.backend, bcj.cgraphs[i].cgraph_main);
@@ -1907,7 +2071,7 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
             }
         }
 
-        if (n_backends > 1 && i < n_subgraphs - 1) {
+        if (n_backends > 1 && i < backend_ctx->n_subgraphs - 1) {
             bool backend_allreduce_success = false;
             if (backend_ctx->comm_ctx) {
                 std::vector<ggml_tensor *> nodes;
@@ -1936,8 +2100,8 @@ static const ggml_backend_i ggml_backend_meta_i = {
     /* .free                    = */ ggml_backend_meta_free,
     /* .set_tensor_async        = */ ggml_backend_meta_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_meta_get_tensor_async,
-    /* .get_tensor_2d_async     = */ nullptr,
     /* .set_tensor_2d_async     = */ nullptr,
+    /* .get_tensor_2d_async     = */ nullptr,
     /* .cpy_tensor_async        = */ nullptr,
     /* .synchronize             = */ ggml_backend_meta_synchronize,
     /* .graph_plan_create       = */ nullptr,

ggml/src/ggml-backend-reg.cpp

@@ -181,6 +181,12 @@ struct ggml_backend_registry {
             return;
         }
 
+        for (auto & entry : backends) {
+            if (entry.reg == reg) {
+                return;
+            }
+        }
+
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
             __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
@@ -192,6 +198,12 @@ struct ggml_backend_registry {
     }
 
     void register_device(ggml_backend_dev_t device) {
+        for (auto & dev : devices) {
+            if (dev == device) {
+                return;
+            }
+        }
+
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
 #endif

ggml/src/ggml-blas/ggml-blas.cpp

@@ -262,9 +262,9 @@ static struct ggml_backend_i blas_backend_i = {
     /* .get_name                = */ ggml_backend_blas_get_name,
     /* .free                    = */ ggml_backend_blas_free,
     /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_2d_async     = */ NULL,
-    /* .set_tensor_2d_async     = */ NULL,
     /* .get_tensor_async        = */ NULL,
+    /* .set_tensor_2d_async     = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
     /* .synchronize             = */ NULL,
     /* .graph_plan_create       = */ NULL,

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -25,6 +25,7 @@
 #include "ggml-impl.h"
 #include "ggml.h"
 
+
 #include <aclnnop/aclnn_add.h>
 #include <aclnnop/aclnn_add_rms_norm.h>
 #include <aclnnop/aclnn_addcdiv.h>
@@ -45,7 +46,9 @@
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
 #include <aclnnop/aclnn_ger.h>
 #include <aclnnop/aclnn_group_norm.h>
+#include <aclnnop/aclnn_gather_v2.h>
 #include <aclnnop/aclnn_grouped_matmul_v3.h>
+#include <aclnnop/aclnn_scatter.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_im2col.h>
 #include <aclnnop/aclnn_index_copy.h>
@@ -62,6 +65,7 @@
 #include <aclnnop/aclnn_permute.h>
 #include <aclnnop/aclnn_pow.h>
 #include <aclnnop/aclnn_pow_tensor_tensor.h>
+#include <aclnnop/aclnn_recurrent_gated_delta_rule.h>
 #include <aclnnop/aclnn_reduce_sum.h>
 #include <aclnnop/aclnn_reflection_pad1d.h>
 #include <aclnnop/aclnn_repeat.h>
@@ -69,11 +73,15 @@
 #include <aclnnop/aclnn_rms_norm.h>
 #include <aclnnop/aclnn_roll.h>
 #include <aclnnop/aclnn_softmax.h>
+#include <aclnnop/aclnn_softmax_cross_entropy_with_logits.h>
 #include <aclnnop/aclnn_sub.h>
 #include <aclnnop/aclnn_sum.h>
 #include <aclnnop/aclnn_threshold.h>
 #include <aclnnop/aclnn_tril.h>
+#include <aclnnop/aclnn_triangular_solve.h>
 #include <aclnnop/aclnn_triu.h>
+#include <aclnnop/aclnn_logical_not.h>
+#include <aclnnop/aclnn_masked_fill_scalar.h>
 #include <aclnnop/aclnn_upsample_nearest_2d.h>
 #include <aclnnop/aclnn_weight_quant_batch_matmul_v2.h>
 #include <aclnnop/aclnn_zero.h>
@@ -151,6 +159,107 @@ void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, ac
     GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, acl_dst.get(), acl_src1.get());
 }
 
+// Fused SwiGLU using aclnnSwiGlu: splits input along innermost dim, applies
+// SiLU to left half, multiplies by right half.
+//
+// Falls back to the generic two-kernel path when src[1] != nullptr (two
+// independent halves) or swapped != 0 (reversed activation order), as
+// aclnnSwiGlu only handles the single interleaved tensor in standard order.
+//
+// CANN tiling for SwiGlu requires (storageShapeDim + viewDims) to be even.
+// aclCreateTensor always uses storageShapeDim=1, so viewDims must be odd.
+// We use a 3D view (1+3=4, even) to satisfy this constraint while preserving
+// correct split semantics along the innermost (ne[0]) dimension.
+void ggml_cann_swiglu(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    auto silu_fn = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) {
+        GGML_CANN_CALL_ACLNN_OP(ctx, Silu, acl_src, acl_dst);
+    };
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+    if (dst->src[1] != nullptr || swapped != 0) {
+        ggml_cann_op_unary_gated(silu_fn, ctx, dst);
+        return;
+    }
+
+    // aclnnSwiGlu requires the split dim (src->ne[0]) to be even; fall back otherwise.
+    if (dst->src[0]->ne[0] % 2 != 0) {
+        ggml_cann_op_unary_gated(silu_fn, ctx, dst);
+        return;
+    }
+
+    ggml_tensor * src0 = dst->src[0];
+    size_t elem_size = ggml_element_size(src0);
+
+    // src0 GGML: [2*ne0, ne1, ne2, ne3] → 3D view [2*ne0, ne1, ne2*ne3]
+    // CANN reversed: [ne2*ne3, ne1, 2*ne0], split along CANN dim 2 (last).
+    int64_t ne0_x2   = src0->ne[0];
+    int64_t ne1      = src0->ne[1];
+    int64_t ne23     = src0->ne[2] * src0->ne[3];
+    int64_t src3d_ne[] = { ne0_x2, ne1, ne23 };
+    size_t  src3d_nb[] = { (size_t)src0->nb[0], (size_t)src0->nb[1], (size_t)src0->nb[2] };
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src0->data, ggml_cann_type_mapping(src0->type),
+                                                     elem_size, src3d_ne, src3d_nb, 3);
+
+    // dst GGML: [ne0, ne1, ne2, ne3] → 3D view [ne0, ne1, ne2*ne3]
+    int64_t ne0      = dst->ne[0];
+    int64_t dst3d_ne[] = { ne0, ne1, ne23 };
+    size_t  dst3d_nb[] = { (size_t)dst->nb[0], (size_t)dst->nb[1], (size_t)dst->nb[2] };
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst->data, ggml_cann_type_mapping(dst->type),
+                                                     elem_size, dst3d_ne, dst3d_nb, 3);
+
+    // CANN tensor [ne23, ne1, 2*ne0]: split along CANN dim 2 (last) = 2*ne0.
+    GGML_CANN_CALL_ACLNN_OP(ctx, SwiGlu, acl_src.get(), (int64_t)2, acl_dst.get());
+}
+
+// Fused GeGLU using aclnnGeGluV3: splits input along ne[0] (CANN last dim),
+// activates the LEFT half with GELU, multiplies by right half.
+// approximate: 0=tanh, 1=none(erf). activateLeft=true matches GGML convention.
+// outGelu is a required-but-discard output buffer.
+//
+// Falls back to the generic two-kernel path when src[1] != nullptr (two
+// independent halves) or swapped != 0 (reversed activation order), as
+// aclnnGeGluV3 only handles the single interleaved tensor in standard order.
+void ggml_cann_geglu(ggml_backend_cann_context & ctx, ggml_tensor * dst, int64_t approximate) {
+    auto gelu_fn = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) {
+        GGML_CANN_CALL_ACLNN_OP(ctx, Gelu, acl_src, acl_dst);
+    };
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+    if (dst->src[1] != nullptr || swapped != 0) {
+        ggml_cann_op_unary_gated(gelu_fn, ctx, dst);
+        return;
+    }
+
+    // aclnnGeGluV3 requires the split dim (src->ne[0]) to be even; fall back otherwise.
+    if (dst->src[0]->ne[0] % 2 != 0) {
+        ggml_cann_op_unary_gated(gelu_fn, ctx, dst);
+        return;
+    }
+
+    ggml_tensor * src0 = dst->src[0];
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src0);
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+
+    // Allocate a temporary buffer for the required outGelu output (same shape as dst).
+    // Build contiguous strides since the pool allocation is a fresh buffer.
+    size_t  elem_size    = ggml_element_size(dst);
+    int64_t ne[GGML_MAX_DIMS] = { dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3] };
+    size_t  nb[GGML_MAX_DIMS];
+    nb[0] = elem_size;
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        nb[i] = nb[i - 1] * ne[i - 1];
+    }
+    size_t gelu_out_size = nb[GGML_MAX_DIMS - 1] * ne[GGML_MAX_DIMS - 1];
+    ggml_cann_pool_alloc gelu_out_alloc(ctx.pool(), gelu_out_size);
+
+    acl_tensor_ptr acl_gelu_out = ggml_cann_create_tensor(
+        gelu_out_alloc.get(), ggml_cann_type_mapping(dst->type), elem_size, ne, nb, GGML_MAX_DIMS);
+    // V3 adds activateLeft param; true → Gelu(left)*right, matching GGML convention.
+    // GGML dim 0 → CANN last dim (index GGML_MAX_DIMS-1 = 3 for 4D tensor).
+    GGML_CANN_CALL_ACLNN_OP(ctx, GeGluV3, acl_src.get(), (int64_t)(GGML_MAX_DIMS - 1), approximate, true,
+                             acl_dst.get(), acl_gelu_out.get());
+}
+
 /**
  * @brief Repeats elements of a tensor along each dimension according to the
  * specified repeat array.
@@ -445,28 +554,33 @@ void ggml_cann_l2_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_cann_pool_alloc temp_buffer_allocator(ctx.pool(), n_bytes);
     void *               buffer = temp_buffer_allocator.get();
 
-    int64_t div_ne[] = { 1, src->ne[1], src->ne[2], src->ne[3] };
-    size_t  div_nb[GGML_MAX_DIMS];
-    div_nb[0] = sizeof(float);
+    int64_t norm_ne[] = { 1, src->ne[1], src->ne[2], src->ne[3] };
+    size_t  norm_nb[GGML_MAX_DIMS];
+    norm_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; ++i) {
-        div_nb[i] = div_nb[i - 1] * div_ne[i - 1];
+        norm_nb[i] = norm_nb[i - 1] * norm_ne[i - 1];
     }
-    acl_tensor_ptr acl_div = ggml_cann_create_tensor(buffer, ACL_FLOAT, type_size, div_ne, div_nb, GGML_MAX_DIMS);
+    acl_tensor_ptr acl_norm = ggml_cann_create_tensor(buffer, ACL_FLOAT, sizeof(float), norm_ne, norm_nb, GGML_MAX_DIMS);
 
     std::vector<int64_t> norm_dims  = { 3 };
     acl_int_array_ptr    dims_array = ggml_cann_create_int_array(norm_dims.data(), norm_dims.size());
 
     float          p_value  = 2.0f;
     acl_scalar_ptr p_scalar = ggml_cann_create_scalar(&p_value, aclDataType::ACL_FLOAT);
-    GGML_CANN_CALL_ACLNN_OP(ctx, Norm, acl_src.get(), p_scalar.get(), dims_array.get(), true, acl_div.get());
+    GGML_CANN_CALL_ACLNN_OP(ctx, Norm, acl_src.get(), p_scalar.get(), dims_array.get(), true, acl_norm.get());
 
-    // Clamp norm to at least eps: scale = 1/fmaxf(norm, eps)
-    acl_scalar_ptr acl_min = ggml_cann_create_scalar(&eps, aclDataType::ACL_FLOAT);
-    float          flt_max = FLT_MAX;
-    acl_scalar_ptr acl_max = ggml_cann_create_scalar(&flt_max, aclDataType::ACL_FLOAT);
-    GGML_CANN_CALL_ACLNN_OP(ctx, Clamp, acl_div.get(), acl_min.get(), acl_max.get(), acl_div.get());
+    ggml_cann_pool_alloc clamp_buffer_allocator(ctx.pool());
+    acl_tensor_ptr       acl_clamped;
 
-    GGML_CANN_CALL_ACLNN_OP(ctx, Div, acl_src.get(), acl_div.get(), acl_dst.get());
+    if (eps > 0.0f) {
+        void *         clamp_buf  = clamp_buffer_allocator.alloc(n_bytes);
+        acl_clamped               = ggml_cann_create_tensor(clamp_buf, ACL_FLOAT, sizeof(float), norm_ne, norm_nb, GGML_MAX_DIMS);
+        acl_scalar_ptr eps_scalar = ggml_cann_create_scalar(&eps, aclDataType::ACL_FLOAT);
+        GGML_CANN_CALL_ACLNN_OP(ctx, ClampMin, acl_norm.get(), eps_scalar.get(), acl_clamped.get());
+    }
+
+    aclTensor * acl_div_input = acl_clamped ? acl_clamped.get() : acl_norm.get();
+    GGML_CANN_CALL_ACLNN_OP(ctx, Div, acl_src.get(), acl_div_input, acl_dst.get());
 }
 
 void ggml_cann_cross_entropy_loss(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
@@ -482,56 +596,30 @@ void ggml_cann_cross_entropy_loss(ggml_backend_cann_context & ctx, ggml_tensor *
     logits_nb[1]              = logits_nb[0] * logits_ne[0];
     acl_tensor_ptr acl_logits = ggml_cann_create_tensor(src0->data, ACL_FLOAT, sizeof(float), logits_ne, logits_nb, 2);
 
-    size_t               log_softmax_type_size = sizeof(float);
-    int64_t              log_softmax_n_bytes   = nr * nc * log_softmax_type_size;
-    ggml_cann_pool_alloc log_softmax_allocator(ctx.pool(), log_softmax_n_bytes);
-    void *               log_softmax_buffer = log_softmax_allocator.get();
-
-    int64_t log_softmax_ne[] = { nc, nr };
-    size_t  log_softmax_nb[2];
-    log_softmax_nb[0]              = log_softmax_type_size;
-    log_softmax_nb[1]              = log_softmax_nb[0] * log_softmax_ne[0];
-    acl_tensor_ptr acl_log_softmax = ggml_cann_create_tensor(log_softmax_buffer, ACL_FLOAT, log_softmax_type_size,
-                                                             log_softmax_ne, log_softmax_nb, 2);
-
-    GGML_CANN_CALL_ACLNN_OP(ctx, LogSoftmax, acl_logits.get(), 1, acl_log_softmax.get());
-
     int64_t labels_ne[] = { nc, nr };
     size_t  labels_nb[2];
     labels_nb[0]              = ggml_type_size(src1->type);
     labels_nb[1]              = labels_nb[0] * labels_ne[0];
     acl_tensor_ptr acl_labels = ggml_cann_create_tensor(src1->data, ACL_FLOAT, sizeof(float), labels_ne, labels_nb, 2);
 
-    size_t               mul_type_size = sizeof(float);
-    int64_t              mul_n_bytes   = nr * nc * mul_type_size;
-    ggml_cann_pool_alloc mul_allocator(ctx.pool(), mul_n_bytes);
-    void *               mul_buffer = mul_allocator.get();
+    size_t               loss_per_sample_type_size = sizeof(float);
+    int64_t              loss_per_sample_n_bytes   = nr * loss_per_sample_type_size;
+    ggml_cann_pool_alloc loss_per_sample_allocator(ctx.pool(), loss_per_sample_n_bytes);
+    void *               loss_per_sample_buffer = loss_per_sample_allocator.get();
 
-    int64_t mul_ne[] = { nc, nr };
-    size_t  mul_nb[2];
-    mul_nb[0]                     = mul_type_size;
-    mul_nb[1]                     = mul_nb[0] * mul_ne[0];
-    acl_tensor_ptr acl_mul_result = ggml_cann_create_tensor(mul_buffer, ACL_FLOAT, mul_type_size, mul_ne, mul_nb, 2);
+    int64_t loss_per_sample_ne[] = { nr };
+    size_t  loss_per_sample_nb[1];
+    loss_per_sample_nb[0] = loss_per_sample_type_size;
+    acl_tensor_ptr acl_loss_per_sample = ggml_cann_create_tensor(
+        loss_per_sample_buffer, ACL_FLOAT, loss_per_sample_type_size, loss_per_sample_ne, loss_per_sample_nb, 1);
 
-    GGML_CANN_CALL_ACLNN_OP(ctx, Mul, acl_log_softmax.get(), acl_labels.get(), acl_mul_result.get());
+    size_t               backprop_n_bytes = nr * nc * sizeof(float);
+    ggml_cann_pool_alloc backprop_allocator(ctx.pool(), backprop_n_bytes);
+    void *               backprop_buffer = backprop_allocator.get();
+    acl_tensor_ptr acl_backprop = ggml_cann_create_tensor(backprop_buffer, ACL_FLOAT, sizeof(float), logits_ne, logits_nb, 2);
 
-    size_t               sum_per_sample_type_size = sizeof(float);
-    int64_t              sum_per_sample_n_bytes   = nr * sum_per_sample_type_size;
-    ggml_cann_pool_alloc sum_per_sample_allocator(ctx.pool(), sum_per_sample_n_bytes);
-    void *               sum_per_sample_buffer = sum_per_sample_allocator.get();
-
-    int64_t sum_per_sample_ne[] = { nr };
-    size_t  sum_per_sample_nb[1];
-    sum_per_sample_nb[0]              = sum_per_sample_type_size;
-    acl_tensor_ptr acl_sum_per_sample = ggml_cann_create_tensor(
-        sum_per_sample_buffer, ACL_FLOAT, sum_per_sample_type_size, sum_per_sample_ne, sum_per_sample_nb, 1);
-
-    std::vector<int64_t> sum_dims   = { 1 };
-    acl_int_array_ptr    dims_array = ggml_cann_create_int_array(sum_dims.data(), sum_dims.size());
-    bool                 keep_dims  = false;
-
-    GGML_CANN_CALL_ACLNN_OP(ctx, ReduceSum, acl_mul_result.get(), dims_array.get(), keep_dims, ACL_FLOAT,
-                            acl_sum_per_sample.get());
+    GGML_CANN_CALL_ACLNN_OP(ctx, SoftmaxCrossEntropyWithLogits, acl_logits.get(), acl_labels.get(),
+                            acl_loss_per_sample.get(), acl_backprop.get());
 
     size_t               total_sum_type_size = sizeof(float);
     int64_t              total_sum_n_bytes   = 1 * total_sum_type_size;
@@ -547,11 +635,12 @@ void ggml_cann_cross_entropy_loss(ggml_backend_cann_context & ctx, ggml_tensor *
 
     std::vector<int64_t> total_sum_dims    = { 0 };
     acl_int_array_ptr total_sum_dims_array = ggml_cann_create_int_array(total_sum_dims.data(), total_sum_dims.size());
+    bool              keep_dims            = false;
 
-    GGML_CANN_CALL_ACLNN_OP(ctx, ReduceSum, acl_sum_per_sample.get(), total_sum_dims_array.get(), keep_dims, ACL_FLOAT,
+    GGML_CANN_CALL_ACLNN_OP(ctx, ReduceSum, acl_loss_per_sample.get(), total_sum_dims_array.get(), keep_dims, ACL_FLOAT,
                             acl_total_sum.get());
 
-    float          value        = -1.0f / static_cast<float>(nr);
+    float          value        = 1.0f / static_cast<float>(nr);
     acl_scalar_ptr scale_factor = ggml_cann_create_scalar(&value, aclDataType::ACL_FLOAT);
     acl_tensor_ptr acl_dst =
         ggml_cann_create_tensor(dst->data, ACL_FLOAT, sizeof(float), total_sum_ne, total_sum_nb, 1);
@@ -589,6 +678,33 @@ void ggml_cann_group_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
                             acl_mean_out.get(), acl_rstd_out.get());
 }
 
+void ggml_cann_set(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src1 = dst->src[1];
+
+    size_t nb1     = ((int32_t *) dst->op_params)[0];
+    size_t nb2     = ((int32_t *) dst->op_params)[1];
+    size_t nb3     = ((int32_t *) dst->op_params)[2];
+    size_t offset  = ((int32_t *) dst->op_params)[3];
+    bool   inplace = (bool) ((int32_t *) dst->op_params)[4];
+
+    size_t param_nb[] = { ggml_element_size(src0), nb1, nb2, nb3 };
+
+    // Create a view of dst at the target offset with src1's dimensions
+    acl_tensor_ptr acl_dst  = ggml_cann_create_tensor(dst, src1->ne, param_nb, GGML_MAX_DIMS, ACL_FORMAT_ND, offset);
+    acl_tensor_ptr acl_src1 = ggml_cann_create_tensor(src1);
+
+    if (!inplace) {
+        // First copy src0 to dst entirely
+        size_t cpy_size = ggml_nbytes(dst);
+        ACL_CHECK(
+            aclrtMemcpyAsync(dst->data, cpy_size, src0->data, cpy_size, ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+    }
+
+    // Copy src1 into the target region of dst
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceCopy, acl_dst.get(), acl_src1.get());
+}
+
 void ggml_cann_acc(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_tensor * src0 = dst->src[0];
     ggml_tensor * src1 = dst->src[1];
@@ -652,6 +768,113 @@ void ggml_cann_sum(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     aclnn_reduce_sum(ctx, dst, reduce_dims, 4);
 }
 
+void ggml_cann_cumsum(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src = dst->src[0];
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src);
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+    // GGML cumsum operates along dim 0 (innermost / ne[0]).
+    // ggml_cann_create_tensor reverses dimensions to [ne3,ne2,ne1,ne0],
+    // so GGML dim 0 maps to CANN dim 3 (the last dim of the 4-D tensor).
+    GGML_CANN_CALL_ACLNN_OP(ctx, Cumsum, acl_src.get(), (int64_t)3,
+                            ggml_cann_type_mapping(dst->type), acl_dst.get());
+}
+
+void ggml_cann_solve_tri(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];  // A: [N, N, B2, B3] lower triangular
+    ggml_tensor * src1 = dst->src[1];  // B: [K, N, B2, B3]
+
+    acl_tensor_ptr acl_a = ggml_cann_create_tensor(src0);
+    acl_tensor_ptr acl_b = ggml_cann_create_tensor(src1);
+    acl_tensor_ptr acl_x = ggml_cann_create_tensor(dst);
+
+    // mOut: triangular copy of A (required output), same shape as A.
+    const size_t a_bytes = ggml_nbytes(src0);
+    ggml_cann_pool_alloc m_alloc(ctx.pool(), a_bytes);
+    acl_tensor_ptr acl_m = ggml_cann_create_tensor(
+        m_alloc.get(), ggml_cann_type_mapping(src0->type),
+        ggml_type_size(src0->type), src0->ne, src0->nb, GGML_MAX_DIMS);
+
+    // Solve AX = B: upper=false (lower tri), transpose=false, unitriangular=false.
+    GGML_CANN_CALL_ACLNN_OP(ctx, TriangularSolve,
+        acl_b.get(), acl_a.get(), false, false, false,
+        acl_x.get(), acl_m.get());
+}
+
+void ggml_cann_diag(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src = dst->src[0];
+
+    GGML_ASSERT(src->ne[1] == 1);
+
+    const int64_t N       = src->ne[0];
+    const int64_t n_batch = src->ne[2] * src->ne[3];
+    const size_t  nb_f32  = sizeof(float);
+
+    // Fill dst with zeros.
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+    {
+        float          zero = 0.0f;
+        acl_scalar_ptr acl_zero = ggml_cann_create_scalar(&zero, ACL_FLOAT);
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceFillScalar, acl_dst.get(), acl_zero.get());
+    }
+
+    // Copy src vector onto the diagonal of dst via strided views.
+    // src viewed as [N, n_batch], contiguous strides.
+    int64_t ne_vec[2]      = { N, n_batch };
+    size_t  nb_src_vec[2]  = { nb_f32, N * nb_f32 };
+    // dst diagonal view: stride (N+1)*4 steps along the diagonal.
+    size_t  nb_dst_diag[2] = { (N + 1) * nb_f32, N * N * nb_f32 };
+
+    acl_tensor_ptr acl_src_vec  = ggml_cann_create_tensor(src->data, ACL_FLOAT, nb_f32, ne_vec, nb_src_vec, 2);
+    acl_tensor_ptr acl_dst_diag = ggml_cann_create_tensor(dst->data, ACL_FLOAT, nb_f32, ne_vec, nb_dst_diag, 2);
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceCopy, acl_dst_diag.get(), acl_src_vec.get());
+}
+
+void ggml_cann_fill(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    float c = ggml_get_op_params_f32(dst, 0);
+
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+    acl_scalar_ptr acl_c   = ggml_cann_create_scalar(&c, ACL_FLOAT);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceFillScalar, acl_dst.get(), acl_c.get());
+}
+
+void ggml_cann_tri(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src = dst->src[0];
+
+    const int64_t S       = src->ne[0];
+    const int64_t n_batch = src->ne[2] * src->ne[3];
+    const size_t  nb_f32  = sizeof(float);
+
+    int64_t ne3d[3] = { S, S, n_batch };
+    size_t  nb3d[3] = { nb_f32, S * nb_f32, S * S * nb_f32 };
+
+    const ggml_tri_type ttype = (ggml_tri_type) ggml_get_op_params_i32(dst, 0);
+
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src->data, ACL_FLOAT, nb_f32, ne3d, nb3d, 3);
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst->data, ACL_FLOAT, nb_f32, ne3d, nb3d, 3);
+
+    switch (ttype) {
+        case GGML_TRI_TYPE_LOWER:
+            // Tril(-1): preserve row > col (strict lower), zero upper + diagonal.
+            GGML_CANN_CALL_ACLNN_OP(ctx, Tril, acl_src.get(), (int64_t)-1, acl_dst.get());
+            break;
+        case GGML_TRI_TYPE_UPPER_DIAG:
+            // Triu(0): preserve row <= col (upper + diagonal), zero strict lower.
+            GGML_CANN_CALL_ACLNN_OP(ctx, Triu, acl_src.get(), (int64_t)0, acl_dst.get());
+            break;
+        case GGML_TRI_TYPE_UPPER:
+            // Triu(1): preserve row < col (strict upper), zero lower + diagonal.
+            GGML_CANN_CALL_ACLNN_OP(ctx, Triu, acl_src.get(), (int64_t)1, acl_dst.get());
+            break;
+        case GGML_TRI_TYPE_LOWER_DIAG:
+            // Tril(0): preserve row >= col (lower + diagonal), zero strict upper.
+            GGML_CANN_CALL_ACLNN_OP(ctx, Tril, acl_src.get(), (int64_t)0, acl_dst.get());
+            break;
+        default:
+            GGML_ABORT("unsupported tri type");
+    }
+}
+
 void ggml_cann_upsample_nearest2d(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_tensor *  src     = dst->src[0];
     acl_tensor_ptr acl_src = ggml_cann_create_tensor(src, nullptr, nullptr, 0, ACL_FORMAT_NCHW);
@@ -1695,152 +1918,90 @@ void ggml_cann_softmax(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     aclnn_softmax(ctx, softmax_tensor.get(), 3, acl_dst.get());
 }
 
-/**
- * @brief Performs index select operation on a 4D tensor using the CANN backend.
- *
- * This function applies the `IndexSelect` operation along a specific dimension
- * of the source tensor (`src_buffer`) using the indices from the index tensor (`index`).
- * It iterates over the last two dimensions of the source tensor, creates the corresponding
- * CANN tensors for the source, index, and output slices, and executes the `IndexSelect`
- * operation for each slice.
- *
- * @param ctx The context for CANN backend operations.
- * @param src_buffer The source buffer containing the 4D input tensor data.
- * @param src_ne The dimensions of the source tensor.
- * @param src_nb The strides (byte offsets) of the source tensor.
- * @param dst_buffer The destination buffer where the output tensor data will be written.
- * @param dst_ne The dimensions of the destination tensor.
- * @param dst_nb The strides (byte offsets) of the destination tensor.
- * @param index The index tensor specifying the indices to select from the source tensor.
- * @param type The data type of the source and destination tensors.
- */
-static void aclnn_index_select_4d(ggml_backend_cann_context & ctx,
-                                  void *                      src_buffer,
-                                  int64_t *                   src_ne,
-                                  size_t *                    src_nb,
-                                  void *                      dst_buffer,
-                                  int64_t *                   dst_ne,
-                                  size_t *                    dst_nb,
-                                  ggml_tensor *               index,
-                                  ggml_type                   type) {
-    for (int64_t i = 0; i < src_ne[3]; i++) {
-        for (int64_t j = 0; j < src_ne[2]; j++) {
-            // src
-            acl_tensor_ptr acl_src_tensor =
-                ggml_cann_create_tensor((char *) src_buffer + i * src_nb[3] + j * src_nb[2],
-                                        ggml_cann_type_mapping(type), ggml_type_size(type), src_ne, src_nb, 2);
-
-            // index
-            acl_tensor_ptr acl_index = ggml_cann_create_tensor(
-                (char *) index->data + (i % index->ne[2]) * index->nb[2] + (j % index->ne[1]) * index->nb[1],
-                ggml_cann_type_mapping(index->type), ggml_element_size(index), index->ne, index->nb, 1);
-
-            // out
-            acl_tensor_ptr acl_out =
-                ggml_cann_create_tensor((char *) dst_buffer + i * dst_nb[3] + j * dst_nb[2],
-                                        ggml_cann_type_mapping(type), ggml_type_size(type), dst_ne, dst_nb, 2);
-            GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, acl_src_tensor.get(), 0, acl_index.get(), acl_out.get());
-        }
-    }
-}
-
-/**
- * @brief Performs inplace index copy operation on a 4D tensor using the CANN backend.
- *
- * This function applies the `IndexCopy` operation along a specific dimension of the
- * destination tensor (`dst_buffer`) by copying elements from the source tensor (`src_buffer`)
- * to positions specified by the index tensor (`index`).
- * It iterates over the last two dimensions of the tensors, creates the corresponding
- * CANN tensors for source, index, and destination slices, and performs the index copy
- * operation for each slice.
- *
- * @param ctx The context for CANN backend operations.
- * @param src_buffer The source buffer containing the 4D input tensor data to be copied.
- * @param src_ne The dimensions of the source tensor.
- * @param src_nb The strides (byte offsets) of the source tensor.
- * @param dst_buffer The destination buffer where values will be copied to.
- * @param dst_ne The dimensions of the destination tensor.
- * @param dst_nb The strides (byte offsets) of the destination tensor.
- * @param index The index tensor specifying target positions in the destination tensor.
- * @param type The data type of the source and destination tensors.
- */
-static void aclnn_index_copy_4d(ggml_backend_cann_context & ctx,
-                                void *                      src_buffer,
-                                int64_t *                   src_ne,
-                                size_t *                    src_nb,
-                                void *                      dst_buffer,
-                                int64_t *                   dst_ne,
-                                size_t *                    dst_nb,
-                                ggml_tensor *               index,
-                                ggml_type                   type) {
-    for (int64_t i = 0; i < src_ne[3]; i++) {
-        for (int64_t j = 0; j < src_ne[2]; j++) {
-            // src
-            acl_tensor_ptr acl_src_tensor =
-                ggml_cann_create_tensor((char *) src_buffer + i * src_nb[3] + j * src_nb[2],
-                                        ggml_cann_type_mapping(type), ggml_type_size(type), src_ne, src_nb, 2);
-
-            // index
-            acl_tensor_ptr acl_index = ggml_cann_create_tensor(
-                (char *) index->data + (i % index->ne[2]) * index->nb[2] + (j % index->ne[1]) * index->nb[1],
-                ggml_cann_type_mapping(index->type), ggml_element_size(index), index->ne, index->nb, 1);
-
-            // out
-            acl_tensor_ptr acl_out =
-                ggml_cann_create_tensor((char *) dst_buffer + i * dst_nb[3] + j * dst_nb[2],
-                                        ggml_cann_type_mapping(type), ggml_type_size(type), dst_ne, dst_nb, 2);
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceIndexCopy, acl_out.get(), 0, acl_index.get(), acl_src_tensor.get());
-        }
-    }
-}
 
 void ggml_cann_get_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * src0 = dst->src[0];  // src
+    ggml_tensor * src0 = dst->src[0];  // weight
     ggml_tensor * src1 = dst->src[1];  // index
 
     GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16
                 || dst->type == GGML_TYPE_BF16);
 
+    // n_idx: number of row indices per (i2, i3) batch slice.
+    // ggml guarantees: src0->ne[2] == src1->ne[1], src0->ne[3] == src1->ne[2], src1->ne[3] == 1.
+    const int64_t n_idx = src1->ne[0];
+
+    // Gather all (i2, i3) batch slices from src into dst.
+    // ggml_cann_create_tensor reverses dims, so ACL sees [ne1, ne0].
+    // GatherV2 with dim=0 gathers along ACL dim-0 == ggml ne[1] (the vocabulary / row axis).
+    // nb: the 4 strides of the source buffer (nb[0..1] for the 2D slice shape,
+    //     nb[2..3] for computing per-batch-slice base pointer offsets).
+    auto gather_batched = [&](void * src_base, aclDataType acl_type, size_t type_size,
+                              const size_t * nb) {
+        int64_t src_ne[2]  = { src0->ne[0], src0->ne[1] };
+        size_t  src_nb_2d[2] = { nb[0], nb[1] };
+        int64_t dst_ne[2]  = { src0->ne[0], n_idx };
+        size_t  dst_nb_2d[2] = { dst->nb[0], dst->nb[1] };
+        int64_t idx_ne[1]  = { n_idx };
+        size_t  idx_nb[1]  = { (size_t)ggml_element_size(src1) };
+
+        for (int64_t i3 = 0; i3 < src0->ne[3]; i3++) {
+            for (int64_t i2 = 0; i2 < src0->ne[2]; i2++) {
+                acl_tensor_ptr acl_src = ggml_cann_create_tensor(
+                    (char *)src_base + i3 * nb[3] + i2 * nb[2],
+                    acl_type, type_size, src_ne, src_nb_2d, 2);
+                acl_tensor_ptr acl_idx = ggml_cann_create_tensor(
+                    (char *)src1->data + i3 * src1->nb[2] + i2 * src1->nb[1],
+                    ggml_cann_type_mapping(src1->type), (size_t)ggml_element_size(src1),
+                    idx_ne, idx_nb, 1);
+                acl_tensor_ptr acl_dst = ggml_cann_create_tensor(
+                    (char *)dst->data + i3 * dst->nb[3] + i2 * dst->nb[2],
+                    acl_type, type_size, dst_ne, dst_nb_2d, 2);
+                GGML_CANN_CALL_ACLNN_OP(ctx, GatherV2, acl_src.get(), 0, acl_idx.get(), acl_dst.get());
+            }
+        }
+    };
+
     switch (src0->type) {
         case GGML_TYPE_BF16:
         case GGML_TYPE_F16:
         case GGML_TYPE_F32:
             if (src0->type == dst->type) {
-                aclnn_index_select_4d(ctx, src0->data, src0->ne, src0->nb, dst->data, dst->ne, dst->nb, src1,
-                                      dst->type);
+                gather_batched(src0->data,
+                               ggml_cann_type_mapping(src0->type), ggml_type_size(src0->type),
+                               src0->nb);
             } else {
-                acl_tensor_ptr       acl_src0 = ggml_cann_create_tensor(src0);
-                ggml_cann_pool_alloc src_buffer_allocator(ctx.pool(), ggml_nelements(src0) * ggml_element_size(dst));
-                void *               src_trans_buffer = src_buffer_allocator.get();
-                size_t               src_trans_nb[GGML_MAX_DIMS];
-                src_trans_nb[0] = dst->nb[0];
+                // Cast src0 to dst type, then gather.
+                ggml_cann_pool_alloc src_cast_allocator(ctx.pool(),
+                                                        ggml_nelements(src0) * ggml_element_size(dst));
+                size_t src_cast_nb[GGML_MAX_DIMS];
+                src_cast_nb[0] = ggml_type_size(dst->type);
                 for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                    src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
+                    src_cast_nb[i] = src_cast_nb[i - 1] * src0->ne[i - 1];
                 }
-                acl_tensor_ptr src_trans_tensor =
-                    ggml_cann_create_tensor(src_trans_buffer, ggml_cann_type_mapping(dst->type),
-                                            ggml_type_size(dst->type), src0->ne, src_trans_nb, GGML_MAX_DIMS);
-                aclnn_cast(ctx, acl_src0.get(), src_trans_tensor.get(), ggml_cann_type_mapping(dst->type));
-                aclnn_index_select_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb, dst->data, dst->ne, dst->nb, src1,
-                                      dst->type);
+                acl_tensor_ptr acl_src0     = ggml_cann_create_tensor(src0);
+                acl_tensor_ptr acl_src_cast = ggml_cann_create_tensor(
+                    src_cast_allocator.get(), ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                    src0->ne, src_cast_nb, GGML_MAX_DIMS);
+                aclnn_cast(ctx, acl_src0.get(), acl_src_cast.get(), ggml_cann_type_mapping(dst->type));
+
+                gather_batched(src_cast_allocator.get(),
+                               ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                               src_cast_nb);
             }
             break;
         case GGML_TYPE_Q8_0:
             {
-                // add 1 dim for bcast mul.
+                // Dequantize Q8_0 to dst type, then gather.
                 size_t  weight_nb[GGML_MAX_DIMS + 1], scale_nb[GGML_MAX_DIMS + 1], dequant_nb[GGML_MAX_DIMS + 1];
                 int64_t weight_ne[GGML_MAX_DIMS + 1], scale_ne[GGML_MAX_DIMS + 1], *dequant_ne;
-                int64_t scale_offset = 0;
-                // [3,4,5,64] -> [3,4,5,2,32]
-                weight_ne[0]         = QK8_0;
-                weight_ne[1]         = src0->ne[0] / QK8_0;
-                weight_nb[0]         = sizeof(int8_t);
-                weight_nb[1]         = weight_nb[0] * weight_ne[0];
+                weight_ne[0] = QK8_0;
+                weight_ne[1] = src0->ne[0] / QK8_0;
+                weight_nb[0] = sizeof(int8_t);
+                weight_nb[1] = weight_nb[0] * weight_ne[0];
                 for (int i = 2; i < GGML_MAX_DIMS + 1; i++) {
                     weight_ne[i] = src0->ne[i - 1];
                     weight_nb[i] = weight_nb[i - 1] * weight_ne[i - 1];
                 }
-                // [3,4,5,64] -> [3,4,5,2,1]
                 scale_ne[0] = 1;
                 scale_ne[1] = src0->ne[0] / QK8_0;
                 scale_nb[0] = sizeof(uint16_t);
@@ -1849,31 +2010,33 @@ void ggml_cann_get_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
                     scale_ne[i] = src0->ne[i - 1];
                     scale_nb[i] = scale_nb[i - 1] * scale_ne[i - 1];
                 }
-                // [3,4,5,64] -> [3,4,5,2,32]
                 dequant_ne    = weight_ne;
                 dequant_nb[0] = ggml_type_size(dst->type);
                 for (int i = 1; i < GGML_MAX_DIMS + 1; i++) {
                     dequant_nb[i] = dequant_nb[i - 1] * dequant_ne[i - 1];
                 }
-                scale_offset = ggml_nelements(src0) * sizeof(int8_t);
-                ggml_cann_pool_alloc dequant_buffer_allocator(ctx.pool(),
-                                                              ggml_nelements(src0) * ggml_type_size(dst->type));
-                acl_tensor_ptr       acl_weight_tensor = ggml_cann_create_tensor(src0->data, ACL_INT8, sizeof(int8_t),
-                                                                                 weight_ne, weight_nb, GGML_MAX_DIMS + 1);
-                acl_tensor_ptr       acl_scale_tensor =
-                    ggml_cann_create_tensor(src0->data, ACL_FLOAT16, sizeof(uint16_t), scale_ne, scale_nb,
-                                            GGML_MAX_DIMS + 1, ACL_FORMAT_ND, scale_offset);
-                acl_tensor_ptr dequant_tensor =
-                    ggml_cann_create_tensor(dequant_buffer_allocator.get(), ggml_cann_type_mapping(dst->type),
-                                            ggml_type_size(dst->type), dequant_ne, dequant_nb, GGML_MAX_DIMS + 1);
-                aclnn_mul(ctx, acl_weight_tensor.get(), acl_scale_tensor.get(), dequant_tensor.get());
-                dequant_nb[0] = ggml_type_size(dst->type);
+                const int64_t scale_offset = ggml_nelements(src0) * sizeof(int8_t);
+                ggml_cann_pool_alloc dequant_allocator(ctx.pool(),
+                                                       ggml_nelements(src0) * ggml_type_size(dst->type));
+                acl_tensor_ptr acl_weight = ggml_cann_create_tensor(src0->data, ACL_INT8, sizeof(int8_t),
+                                                                     weight_ne, weight_nb, GGML_MAX_DIMS + 1);
+                acl_tensor_ptr acl_scale  = ggml_cann_create_tensor(
+                    src0->data, ACL_FLOAT16, sizeof(uint16_t), scale_ne, scale_nb,
+                    GGML_MAX_DIMS + 1, ACL_FORMAT_ND, scale_offset);
+                acl_tensor_ptr acl_dequant = ggml_cann_create_tensor(
+                    dequant_allocator.get(), ggml_cann_type_mapping(dst->type),
+                    ggml_type_size(dst->type), dequant_ne, dequant_nb, GGML_MAX_DIMS + 1);
+                aclnn_mul(ctx, acl_weight.get(), acl_scale.get(), acl_dequant.get());
+
+                // Reinterpret dequant buffer as 4D [src0->ne] with contiguous strides.
                 dequant_ne    = src0->ne;
+                dequant_nb[0] = ggml_type_size(dst->type);
                 for (int i = 1; i < GGML_MAX_DIMS; i++) {
                     dequant_nb[i] = dequant_nb[i - 1] * src0->ne[i - 1];
                 }
-                aclnn_index_select_4d(ctx, dequant_buffer_allocator.get(), dequant_ne, dequant_nb, dst->data, dst->ne,
-                                      dst->nb, src1, dst->type);
+                gather_batched(dequant_allocator.get(),
+                               ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                               dequant_nb);
                 break;
             }
         default:
@@ -1883,31 +2046,70 @@ void ggml_cann_get_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
 }
 
 void ggml_cann_set_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * src0 = dst->src[0];  // src
-    ggml_tensor * src1 = dst->src[1];  // index
+    ggml_tensor * src0 = dst->src[0];  // source values
+    ggml_tensor * src1 = dst->src[1];  // row indices
+
+    // n_idx: number of source rows to scatter per batch slice.
+    // ggml guarantees: src0->ne[1] == src1->ne[0].
+    const int64_t n_idx = src1->ne[0];
+
+    // Copy n_idx rows of src [ne0, n_idx] into dst [ne0, ne1] at positions given by a 1D index.
+    // ggml_cann_create_tensor reverses dims, so ACL sees [ne1, ne0] for dst.
+    // InplaceIndexCopy with dim=0 copies along ACL dim-0 == ggml ne[1] (the row axis).
+    // src_nb: the 4 strides of the source buffer (nb[0..1] for the 2D slice shape,
+    //         nb[2..3] for computing per-batch-slice base pointer offsets).
+    auto scatter_batched = [&](void * src_base, aclDataType acl_type, size_t type_size,
+                               const size_t * src_nb) {
+        int64_t d_ne[2]    = { dst->ne[0], dst->ne[1] };
+        size_t  d_nb[2]    = { dst->nb[0], dst->nb[1] };
+        int64_t s_ne[2]    = { dst->ne[0], n_idx };
+        size_t  s_nb_2d[2] = { src_nb[0], src_nb[1] };
+        int64_t i_ne[1]    = { n_idx };
+        size_t  i_nb[1]    = { (size_t)ggml_element_size(src1) };
+
+        for (int64_t i3 = 0; i3 < dst->ne[3]; i3++) {
+            for (int64_t i2 = 0; i2 < dst->ne[2]; i2++) {
+                acl_tensor_ptr acl_dst = ggml_cann_create_tensor(
+                    (char *)dst->data + i3 * dst->nb[3] + i2 * dst->nb[2],
+                    acl_type, type_size, d_ne, d_nb, 2);
+                acl_tensor_ptr acl_idx = ggml_cann_create_tensor(
+                    (char *)src1->data + (i3 % src1->ne[2]) * src1->nb[2] + (i2 % src1->ne[1]) * src1->nb[1],
+                    ggml_cann_type_mapping(src1->type), (size_t)ggml_element_size(src1),
+                    i_ne, i_nb, 1);
+                acl_tensor_ptr acl_src = ggml_cann_create_tensor(
+                    (char *)src_base + i3 * src_nb[3] + i2 * src_nb[2],
+                    acl_type, type_size, s_ne, s_nb_2d, 2);
+                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceIndexCopy, acl_dst.get(), 0, acl_idx.get(), acl_src.get());
+            }
+        }
+    };
 
     switch (dst->type) {
         case GGML_TYPE_F32:
-            {
-                aclnn_index_copy_4d(ctx, src0->data, src0->ne, src0->nb, dst->data, dst->ne, dst->nb, src1, dst->type);
-                break;
-            }
+            scatter_batched(src0->data,
+                            ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                            src0->nb);
+            break;
         case GGML_TYPE_F16:
         case GGML_TYPE_BF16:
             {
-                acl_tensor_ptr       acl_src0 = ggml_cann_create_tensor(src0);
-                ggml_cann_pool_alloc src_buffer_allocator(ctx.pool(), ggml_nelements(src0) * sizeof(uint16_t));
-                void *               src_trans_buffer = src_buffer_allocator.get();
-                size_t               src_trans_nb[GGML_MAX_DIMS];
-                src_trans_nb[0] = sizeof(uint16_t);
+                // Cast src0 (F32) to dst type first.
+                ggml_cann_pool_alloc src_cast_allocator(ctx.pool(),
+                                                        ggml_nelements(src0) * ggml_type_size(dst->type));
+                size_t src_cast_nb[GGML_MAX_DIMS];
+                src_cast_nb[0] = ggml_type_size(dst->type);
                 for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                    src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
+                    src_cast_nb[i] = src_cast_nb[i - 1] * src0->ne[i - 1];
                 }
-                acl_tensor_ptr src_trans_tensor = ggml_cann_create_tensor(
-                    src_trans_buffer, ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type), src0->ne, src_trans_nb, GGML_MAX_DIMS);
-                aclnn_cast(ctx, acl_src0.get(), src_trans_tensor.get(), ggml_cann_type_mapping(dst->type));
-                aclnn_index_copy_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb, dst->data, dst->ne, dst->nb, src1,
-                                    dst->type);
+                acl_tensor_ptr acl_src0     = ggml_cann_create_tensor(src0);
+                acl_tensor_ptr acl_src_cast = ggml_cann_create_tensor(
+                    src_cast_allocator.get(), ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                    src0->ne, src_cast_nb, GGML_MAX_DIMS);
+                aclnn_cast(ctx, acl_src0.get(), acl_src_cast.get(), ggml_cann_type_mapping(dst->type));
+
+                scatter_batched(src_cast_allocator.get(),
+                                ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                                src_cast_nb);
                 break;
             }
         default:
@@ -3268,29 +3470,50 @@ void ggml_cann_pad_reflect_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst
     int64_t           paddingsArray[2] = { opts[0], opts[1] };
     acl_int_array_ptr paddings         = ggml_cann_create_int_array(paddingsArray, 2);
 
-    for (int64_t i = 0; i < src0->ne[3]; i++) {
-        acl_tensor_ptr acl_src =
-            ggml_cann_create_tensor((char *) src0->data + i * src0->ne[3], ggml_cann_type_mapping(src0->type),
-                                    ggml_element_size(src0), src0->ne, src0->nb, 3);
+    // Collapsing ne[2]*ne[3] into a single batch dimension requires that dim3
+    // is contiguous with respect to dim2 in both src and dst.
+    GGML_ASSERT(src0->nb[3] == src0->nb[2] * src0->ne[2]);
+    GGML_ASSERT(dst->nb[3]  == dst->nb[2]  * dst->ne[2]);
 
-        acl_tensor_ptr acl_dst =
-            ggml_cann_create_tensor((char *) dst->data + i * src0->ne[3], ggml_cann_type_mapping(dst->type),
-                                    ggml_element_size(dst), dst->ne, dst->nb, 3);
+    int64_t src_ne_3d[3] = { src0->ne[0], src0->ne[1], src0->ne[2] * src0->ne[3] };
+    int64_t dst_ne_3d[3] = { dst->ne[0],  dst->ne[1],  dst->ne[2]  * dst->ne[3]  };
 
-        GGML_CANN_CALL_ACLNN_OP(ctx, ReflectionPad1d, acl_src.get(), paddings.get(), acl_dst.get());
-    }
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src0->data, ggml_cann_type_mapping(src0->type),
+                                                     ggml_element_size(src0), src_ne_3d, src0->nb, 3);
+
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst->data, ggml_cann_type_mapping(dst->type),
+                                                     ggml_element_size(dst), dst_ne_3d, dst->nb, 3);
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, ReflectionPad1d, acl_src.get(), paddings.get(), acl_dst.get());
 }
 
 void ggml_cann_count_equal(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_tensor * src0 = dst->src[0];
     ggml_tensor * src1 = dst->src[1];
 
+    // Write element-wise equality (0 or 1) into a temporary buffer to avoid
+    // modifying src0 in-place.  Use the same type as src0 so ReduceSum can
+    // consume it directly without a type cast.
+    ggml_cann_pool_alloc eq_alloc(ctx.pool(), ggml_nelements(src0) * ggml_element_size(src0));
+    size_t eq_nb[GGML_MAX_DIMS];
+    eq_nb[0] = ggml_element_size(src0);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        eq_nb[i] = eq_nb[i - 1] * src0->ne[i - 1];
+    }
+    acl_tensor_ptr acl_eq = ggml_cann_create_tensor(
+        eq_alloc.get(), ggml_cann_type_mapping(src0->type), ggml_element_size(src0),
+        src0->ne, eq_nb, GGML_MAX_DIMS);
+
     acl_tensor_ptr acl_self  = ggml_cann_create_tensor(src0);
     acl_tensor_ptr acl_other = ggml_cann_create_tensor(src1);
+    GGML_CANN_CALL_ACLNN_OP(ctx, EqTensor, acl_self.get(), acl_other.get(), acl_eq.get());
 
-    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceEqTensor, acl_self.get(), acl_other.get());
-
-    ggml_cann_sum(ctx, dst);
+    // Sum the 0/1 values into dst.
+    acl_tensor_ptr    acl_dst    = ggml_cann_create_tensor(dst);
+    int64_t           dims[4]    = { 0, 1, 2, 3 };
+    acl_int_array_ptr dims_arr   = ggml_cann_create_int_array(dims, 4);
+    GGML_CANN_CALL_ACLNN_OP(ctx, ReduceSum, acl_eq.get(), dims_arr.get(), true,
+                            ggml_cann_type_mapping(dst->type), acl_dst.get());
 }
 
 void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
@@ -3306,6 +3529,27 @@ void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     GGML_CANN_CALL_ACLNN_OP(ctx, GtScalar, acl_src.get(), alpha.get(), acl_dst.get());
 }
 
+void ggml_cann_softplus(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];
+
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src0);
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+
+    float          beta_val      = 1.0f;
+    float          threshold_val = 20.0f;
+    acl_scalar_ptr beta          = ggml_cann_create_scalar(&beta_val,      ACL_FLOAT);
+    acl_scalar_ptr threshold     = ggml_cann_create_scalar(&threshold_val, ACL_FLOAT);
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, Softplus, acl_src.get(), beta.get(), threshold.get(), acl_dst.get());
+}
+
+void ggml_cann_geglu_quick(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    auto gelu_quick_fn = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) {
+        GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
+    };
+    ggml_cann_op_unary_gated(gelu_quick_fn, ctx, dst);
+}
+
 /**
  * @brief Performs expert-specific matrix multiplication (MoE) with
  * floating-point precision using the CANN backend.
@@ -3892,46 +4136,65 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context & ctx, ggml_tensor * dst
 }
 
 static void ggml_cann_out_prod_fp(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * src0 = dst->src[0];  // weight
-    ggml_tensor * src1 = dst->src[1];  // input
+    ggml_tensor * src0 = dst->src[0];  // weight  [ne00=m, ne01=K, ne02, ne03]
+    ggml_tensor * src1 = dst->src[1];  // input   [ne10=n, ne11=K, ne12, ne13]
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
-    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, acl_dst.get());
+    // dst[i,j] = sum_k src0[i,k] * src1[j,k]  i.e. dst = src0 @ src1^T.
+    //
+    // ggml_cann_create_tensor reverses dimension order, so ACL sees:
+    //   acl_src0 slice:   ggml[m,K]  ->  ACL[K,m]
+    //   acl_src1 slice:   ggml[n,K]  ->  ACL[K,n]
+    //   acl_dst  slice:   ggml[m,n]  ->  ACL[n,m]
+    //
+    // Build a transposed view of src1 by swapping ne[0]/ne[1]:
+    //   src1_t:  ggml[K,n] (swapped strides)  ->  ACL[n,K]
+    //
+    // Matmul(src1_t [n,K], src0 [K,m]) = [n,m] = acl_dst  ✓
+    //
+    // The outer batch loop is kept because src0 may have fewer batch slices than
+    // dst (ne02 <= ne2, ne03 <= ne3): this is a strided-broadcast not supported
+    // by standard CANN Matmul broadcasting.
+
+    const aclDataType src0_acl_type = ggml_cann_type_mapping(src0->type);
+    const aclDataType src1_acl_type = ggml_cann_type_mapping(src1->type);
+    const aclDataType dst_acl_type  = ggml_cann_type_mapping(dst->type);
+    const size_t      src0_type_sz  = ggml_type_size(src0->type);
+    const size_t      src1_type_sz  = ggml_type_size(src1->type);
+    const size_t      dst_type_sz   = ggml_type_size(dst->type);
 
     const int64_t dps2 = ne2 / ne02;
     const int64_t dps3 = ne3 / ne03;
+
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
             const int64_t i02 = i2 / dps2;
             const int64_t i03 = i3 / dps3;
 
-            const int64_t  i12 = i2;
-            const int64_t  i13 = i3;
-            acl_tensor_ptr accumulator =
-                ggml_cann_create_tensor((char *) dst->data + i2 * nb2 + i3 * nb3, ggml_cann_type_mapping(dst->type),
-                                        ggml_type_size(dst->type), dst->ne, dst->nb, 2);
+            // src0 2D slice at [i02, i03]: ggml [m, K] -> ACL [K, m]
+            int64_t src0_ne[2] = { ne00, ne01 };
+            size_t  src0_nb[2] = { nb00, nb01 };
+            acl_tensor_ptr acl_src0_s = ggml_cann_create_tensor(
+                (char *) src0->data + i02 * nb02 + i03 * nb03,
+                src0_acl_type, src0_type_sz, src0_ne, src0_nb, 2);
 
-            // The outer product needs to be accumulated in this dimension.
-            for (int64_t i1 = 0; i1 < ne11; i1++) {
-                acl_tensor_ptr acl_input = ggml_cann_create_tensor(
-                    (char *) src1->data + i1 * nb11 + i12 * nb12 + i13 * nb13, ggml_cann_type_mapping(src0->type),
-                    ggml_type_size(src0->type), src1->ne, src1->nb, 1);
+            // src1 transposed 2D slice at [i2, i3]: swap ne/nb -> ggml[K,n] -> ACL[n,K]
+            int64_t src1_t_ne[2] = { ne11, ne10 };
+            size_t  src1_t_nb[2] = { nb11, nb10 };
+            acl_tensor_ptr acl_src1_t = ggml_cann_create_tensor(
+                (char *) src1->data + i2 * nb12 + i3 * nb13,
+                src1_acl_type, src1_type_sz, src1_t_ne, src1_t_nb, 2);
 
-                acl_tensor_ptr acl_weight = ggml_cann_create_tensor(
-                    (char *) src0->data + i1 * nb01 + i02 * nb02 + i03 * nb03, ggml_cann_type_mapping(src0->type),
-                    ggml_type_size(src0->type), src0->ne, src0->nb, 1);
+            // dst 2D slice at [i2, i3]: ggml [m, n] -> ACL [n, m]
+            int64_t dst_ne[2] = { ne0, ne1 };
+            size_t  dst_nb[2] = { nb0, nb1 };
+            acl_tensor_ptr acl_dst_s = ggml_cann_create_tensor(
+                (char *) dst->data + i2 * nb2 + i3 * nb3,
+                dst_acl_type, dst_type_sz, dst_ne, dst_nb, 2);
 
-                ggml_cann_pool_alloc output_allocator(ctx.pool());
-                void *               output_buffer = output_allocator.alloc(ggml_nbytes(dst));
-                acl_tensor_ptr       acl_out = ggml_cann_create_tensor(output_buffer, ggml_cann_type_mapping(dst->type),
-                                                                       ggml_type_size(dst->type), dst->ne, dst->nb, 2);
-
-                GGML_CANN_CALL_ACLNN_OP(ctx, Ger, acl_input.get(), acl_weight.get(), acl_out.get());
-                float       alpha_value = 1.0f;
-                aclScalar * alpha       = aclCreateScalar(&alpha_value, ACL_FLOAT);
-                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdd, accumulator.get(), acl_out.get(), alpha);
-            }
+            // Matmul(src1_t [n,K], src0 [K,m]) = [n,m] = acl_dst_s  ✓
+            GGML_CANN_CALL_ACLNN_OP(ctx, Matmul,
+                acl_src1_t.get(), acl_src0_s.get(), acl_dst_s.get(), (int8_t) 1);
         }
     }
 }
@@ -4170,3 +4433,4 @@ void ggml_cann_gated_linear_attn(ggml_backend_cann_context & ctx, ggml_tensor *
         }
     }
 }
+

ggml/src/ggml-cann/aclnn_ops.h

@@ -32,6 +32,9 @@
 #include <aclnnop/aclnn_cat.h>
 #include <aclnnop/aclnn_clamp.h>
 #include <aclnnop/aclnn_cos.h>
+#include <aclnnop/aclnn_cumsum.h>
+#include <aclnnop/aclnn_tril.h>
+#include <aclnnop/aclnn_triu.h>
 #include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_gelu.h>
 #include <aclnnop/aclnn_gelu_v2.h>
@@ -47,6 +50,9 @@
 #include <aclnnop/aclnn_sign.h>
 #include <aclnnop/aclnn_silu.h>
 #include <aclnnop/aclnn_sin.h>
+#include <aclnnop/aclnn_softplus.h>
+#include <aclnnop/aclnn_swi_glu.h>
+#include <aclnnop/aclnn_geglu.h>
 #include <aclnnop/aclnn_slice.h>
 #include <aclnnop/aclnn_sqrt.h>
 #include <aclnnop/aclnn_tanh.h>
@@ -69,6 +75,9 @@
  */
 void ggml_cann_repeat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+void ggml_cann_swiglu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_geglu(ggml_backend_cann_context & ctx, ggml_tensor * dst, int64_t approximate);
+
 /**
  * @brief   Applies the Leaky ReLU activation function to a tensor using the CANN
  *          backend.
@@ -325,6 +334,48 @@ void ggml_cann_sum_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 void ggml_cann_sum(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+/**
+ * @brief   Computes the cumulative sum of a ggml tensor along dim 0 using the
+ *          CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_CUMSUM`.
+ */
+void ggml_cann_cumsum(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Computes a triangular mask (tril/triu) of a square ggml tensor
+ *          using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_TRI`.
+ */
+void ggml_cann_tri(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Solves a triangular linear system AX=B using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_SOLVE_TRI`.
+ */
+void ggml_cann_solve_tri(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Creates a diagonal matrix from a vector using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_DIAG`.
+ */
+void ggml_cann_diag(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Fills a tensor with a constant scalar value using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_FILL`.
+ */
+void ggml_cann_fill(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
 /**
  * @brief   Upsamples a ggml tensor using nearest neighbor interpolation using
  *          the CANN backend.
@@ -461,6 +512,9 @@ void ggml_cann_timestep_embedding(ggml_backend_cann_context & ctx, ggml_tensor *
 // @see ggml_cann_dup.
 void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+// @see ggml_cann_acc, but copies src1 into dst instead of adding.
+void ggml_cann_set(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
 /**
  * @brief   Computes the softmax activation with optional masking.
  *
@@ -813,6 +867,8 @@ void ggml_cann_count_equal(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  *            dst->op is expected to be `GGML_OP_STEP`.
  */
 void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_softplus(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_geglu_quick(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Performs the Flash Attention extended operator using the CANN backend.

ggml/src/ggml-cann/ggml-cann.cpp

@@ -1428,6 +1428,22 @@ static bool ggml_backend_cann_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
     return false;
 }
 
+/**
+ * @brief Set a region of a tensor's device memory to a specified value.
+ *
+ * @param buffer The CANN buffer containing the tensor.
+ * @param tensor Pointer to the tensor whose memory will be set.
+ * @param value The value to which each byte in the region will be set.
+ * @param offset Byte offset within the tensor's data to start setting.
+ * @param size Number of bytes to set.
+ */
+static void ggml_backend_cann_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    ggml_backend_cann_buffer_context * ctx = (ggml_backend_cann_buffer_context *) buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+    ACL_CHECK(aclrtMemset((char *) tensor->data + offset, size, value, size));
+}
+
 /**
  * @brief Clear a CANN buffer by setting all its memory to a specified value.
  *
@@ -1454,7 +1470,7 @@ static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
     /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cann_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_cann_buffer_init_tensor,
-    /* .memset_tensor   = */ NULL,
+    /* .memset_tensor   = */ ggml_backend_cann_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_cann_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_cann_buffer_get_tensor,
     /* .set_tensor_2d   = */ NULL,
@@ -1835,6 +1851,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
                 case GGML_UNARY_OP_STEP:
                     ggml_cann_step(ctx, dst);
                     break;
+                case GGML_UNARY_OP_SOFTPLUS:
+                    ggml_cann_softplus(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -1845,20 +1864,16 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
                     GGML_CANN_CALL_OP_UNARY_GATED(Relu);
                     break;
                 case GGML_GLU_OP_GEGLU:
+                    ggml_cann_geglu(ctx, dst, 0);  // approximate=0 → tanh
+                    break;
                 case GGML_GLU_OP_GEGLU_ERF:
-                    // aclnnGelu internally uses the erf-based approximation.
-                    GGML_CANN_CALL_OP_UNARY_GATED(Gelu);
+                    ggml_cann_geglu(ctx, dst, 1);  // approximate=1 → erf
                     break;
                 case GGML_GLU_OP_SWIGLU:
-                    GGML_CANN_CALL_OP_UNARY_GATED(Silu);
+                    ggml_cann_swiglu(ctx, dst);
                     break;
                 case GGML_GLU_OP_GEGLU_QUICK:
-                    {
-                        auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) {
-                            GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
-                        };
-                        ggml_cann_op_unary_gated(lambda, ctx, dst);
-                    }
+                    ggml_cann_geglu_quick(ctx, dst);
                     break;
                 default:
                     return false;
@@ -1920,6 +1935,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
         case GGML_OP_CPY:
             ggml_cann_cpy(ctx, dst);
             break;
+        case GGML_OP_SET:
+            ggml_cann_set(ctx, dst);
+            break;
         case GGML_OP_CONT:
             ggml_cann_dup(ctx, dst);
             break;
@@ -1989,6 +2007,21 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
         case GGML_OP_SSM_CONV:
             ggml_cann_ssm_conv(ctx, dst);
             break;
+        case GGML_OP_CUMSUM:
+            ggml_cann_cumsum(ctx, dst);
+            break;
+        case GGML_OP_TRI:
+            ggml_cann_tri(ctx, dst);
+            break;
+        case GGML_OP_FILL:
+            ggml_cann_fill(ctx, dst);
+            break;
+        case GGML_OP_DIAG:
+            ggml_cann_diag(ctx, dst);
+            break;
+        case GGML_OP_SOLVE_TRI:
+            ggml_cann_solve_tri(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -2324,6 +2357,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
     if (use_cann_graph) {
         // If no matching graph is found, the graph needs to be recaptured.
         graph_capture_required = !cann_ctx->graph_lru_cache.find_and_move_to_front(cgraph);
+
         if (graph_capture_required) {
             // If no matching graph is found, add a new ACL graph.
             ggml_cann_graph * new_graph = ggml_cann_graph::create_from_cgraph(cgraph);
@@ -2382,6 +2416,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                 case GGML_UNARY_OP_SGN:
                 case GGML_UNARY_OP_STEP:
                 case GGML_UNARY_OP_GELU_ERF:
+                case GGML_UNARY_OP_SOFTPLUS:
                     return true;
                 default:
                     return false;
@@ -2572,6 +2607,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
+        case GGML_OP_SET:
         case GGML_OP_GROUP_NORM:
             return true;
         case GGML_OP_PAD:
@@ -2649,6 +2685,16 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
             }
         case GGML_OP_SSM_CONV:
             return true;
+        case GGML_OP_CUMSUM:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_TRI:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_FILL:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_DIAG:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_SOLVE_TRI:
+            return op->src[0]->type == GGML_TYPE_F32;
         default:
             return false;
     }
@@ -2700,8 +2746,8 @@ static const ggml_backend_i ggml_backend_cann_interface = {
     /* .free                    = */ ggml_backend_cann_free,
     /* .set_tensor_async        = */ ggml_backend_cann_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_cann_get_tensor_async,
-    /* .get_tensor_2d_async     = */ NULL,
     /* .set_tensor_2d_async     = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ ggml_backend_cann_cpy_tensor_async,
     /* .synchronize             = */ ggml_backend_cann_synchronize,
     /* .graph_plan_create       = */ NULL,

ggml/src/ggml-cpu/CMakeLists.txt

@@ -485,6 +485,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             if (GGML_RV_ZIHINTPAUSE)
                 string(APPEND MARCH_STR "_zihintpause")
             endif()
+            if (GGML_CPU_RISCV64_SPACEMIT)
+                # `xsmtvdotii' is only required for GCC >= 15.
+                if (CMAKE_C_COMPILER_ID STREQUAL "GNU" AND
+                    CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 15)
+                    string(APPEND MARCH_STR "_xsmtvdotii")
+                endif()
+            endif()
 
             list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
         else()

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -2005,12 +2005,12 @@ void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const v
     const int lda = KB * sizeof(TA);
     //const int ldb = KB * sizeof(TB);
 
-    static thread_local packed_B_t Tile0[TILE_N * TILE_K];
-    static thread_local packed_B_t Tile1[TILE_N * TILE_K];
-    static thread_local int8_t Tile23[TILE_M * TILE_K];
+    alignas(64) static thread_local packed_B_t Tile0[TILE_N * TILE_K];
+    alignas(64) static thread_local packed_B_t Tile1[TILE_N * TILE_K];
+    alignas(64) static thread_local int8_t Tile23[TILE_M * TILE_K];
 
-    static thread_local int32_t TileC0[TILE_M * TILE_N * 4];
-    static thread_local int32_t TileC1[TILE_M * TILE_N * 4];
+    alignas(64) static thread_local int32_t TileC0[TILE_M * TILE_N * 4];
+    alignas(64) static thread_local int32_t TileC1[TILE_M * TILE_N * 4];
 
     // double buffering C to interleave avx512 and amx
     int32_t * C_cur = TileC0;
@@ -2187,21 +2187,21 @@ void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const v
     const int m1 = std::max(M - TILE_M, 0);
     //const int lda = KB * sizeof(TA);
 
-    static thread_local int8_t Tile0[TILE_N * TILE_K];
-    static thread_local int8_t Tile1[TILE_N * TILE_K];
-    static thread_local int8_t Tile23[TILE_M * TILE_K];
+    alignas(64) static thread_local int8_t Tile0[TILE_N * TILE_K];
+    alignas(64) static thread_local int8_t Tile1[TILE_N * TILE_K];
+    alignas(64) static thread_local int8_t Tile23[TILE_M * TILE_K];
 
     // mat mul result for each group
-    static thread_local int32_t Tile4[TILE_M * TILE_N];
-    static thread_local int32_t Tile5[TILE_M * TILE_N];
-    static thread_local int32_t Tile6[TILE_M * TILE_N];
-    static thread_local int32_t Tile7[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Tile4[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Tile5[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Tile6[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Tile7[TILE_M * TILE_N];
 
     // sum of each QK_K block, contains 8 groups, int32
-    static thread_local int32_t Sumi4[TILE_M * TILE_N];
-    static thread_local int32_t Sumi5[TILE_M * TILE_N];
-    static thread_local int32_t Sumi6[TILE_M * TILE_N];
-    static thread_local int32_t Sumi7[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Sumi4[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Sumi5[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Sumi6[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Sumi7[TILE_M * TILE_N];
 
     const int k_group_size = std::is_same<TB, block_q6_K>::value ? 16 : 32;
     for (int i = 0; i < KB; ++i) {

ggml/src/ggml-cpu/arch-fallback.h

@@ -83,7 +83,6 @@
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // quants.c
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
-#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4

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

@@ -151,8 +151,6 @@ void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const block_q1_0 * GGML_RESTRICT x = vx;
     const block_q8_0 * GGML_RESTRICT y = vy;
 
-    float sumf = 0.0f;
-
 #if defined(__ARM_NEON)
     float32x4_t sumv = vdupq_n_f32(0.0f);
 
@@ -212,31 +210,13 @@ void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
     }
 
-    sumf = vaddvq_f32(sumv);
+    *s = vaddvq_f32(sumv);
 #else
-    // Scalar fallback
-    for (int i = 0; i < nb; i++) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
-
-        // Process 4 Q8_0 blocks
-        for (int k = 0; k < 4; k++) {
-            const float d1 = GGML_FP16_TO_FP32(y[i*4 + k].d);
-
-            int sumi = 0;
-            for (int j = 0; j < QK8_0; j++) {
-                const int bit_index = k * QK8_0 + j;
-                const int byte_index = bit_index / 8;
-                const int bit_offset = bit_index % 8;
-
-                const int xi = ((x[i].qs[byte_index] >> bit_offset) & 1) ? 1 : -1;
-                sumi += xi * y[i*4 + k].qs[j];
-            }
-            sumf += d0 * d1 * sumi;
-        }
-    }
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q1_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
-    *s = sumf;
 }
 
 

ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -5023,6 +5023,71 @@ void ggml_gemm_q8_0_4x8_q8_0(int                        n,
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
+#if defined(__aarch64__) && defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (svcntb() * 8 == 256) {
+        const block_q8_0x4 * b_ptr_base = (const block_q8_0x4 *) vx;
+
+        static const uint32_t idx_arr[8] = {0, 1, 4, 5, 2, 3, 6, 7};
+        svuint32_t idx = svld1(svptrue_b32(), idx_arr);
+        static const uint32_t idx_arr1[8] = {0, 1, 2, 3, 1, 2, 3, 0};
+        svuint32_t idx_sc1 = svld1(svptrue_b32(), idx_arr1);
+        static const uint32_t idx_arr2[8] = {0, 1, 2, 3, 0, 1, 2, 3};
+        svuint32_t idx_sc2 = svld1(svptrue_b32(), idx_arr2);
+
+        for (int y = 0; y < nr; y += 4) {
+            const block_q8_0x4 * a_ptr_base = (const block_q8_0x4 *) vy + (y / 4) * nb;
+
+            for (int x = 0; x < nc; x += ncols_interleaved) {
+                const block_q8_0x4 * b_ptr = b_ptr_base + (x / 4) * nb;
+                const block_q8_0x4 * a_ptr = a_ptr_base;
+
+                svfloat32_t acc_f32_01 = svdup_f32(0);
+                svfloat32_t acc_f32_23 = svdup_f32(0);
+
+                for (int b = 0; b < nb; b++) {
+
+                    svint32_t acc_01 = svdup_s32(0);
+                    svint32_t acc_23 = svdup_s32(0);
+
+                    // Process 4 chunks of 8 positions each
+                    for (int chunk = 0; chunk < 4; chunk++) {
+                        svint8_t s_a01 = svld1rq_s8(svptrue_b8(), a_ptr->qs + chunk * 32);
+                        svint8_t s_a23 = svld1rq_s8(svptrue_b8(), a_ptr->qs + chunk * 32 + 16);
+                        svint8_t s_b0123 = svld1_s8(svptrue_b8(), b_ptr->qs + chunk * 32);
+
+                        acc_01 = svmmla_s32(acc_01, s_a01, s_b0123);
+                        acc_23 = svmmla_s32(acc_23, s_a23, s_b0123);
+                    }
+
+                    // Reorder outputs from 2×2 tiles to row-major
+                    // acc[01] = [r0c0, r0c1, r1c0, r1c1, r0c2, r0c3, r1c2, r1c3]
+                    // acc[23] = [r2c0, r2c1, r3c0, r3c1, r2c2, r2c3, r3c2, r3c3]
+
+                    svint32_t row01 = svtbl_s32(acc_01, idx);
+                    svint32_t row23 = svtbl_s32(acc_23, idx);
+
+                    svfloat16_t temp1 = svld1_f16(svptrue_pat_b16(SV_VL4), (const __fp16 *) a_ptr->d);
+                    svfloat16_t temp2 = svld1_f16(svptrue_pat_b16(SV_VL4), (const __fp16 *) b_ptr->d);
+                    svfloat32_t sv_a_d = svtbl_f32(svcvt_f32_f16_x(svptrue_b32(), svzip1_f16(temp1, temp1)), idx_sc1);
+                    svfloat32_t sv_b_d = svtbl_f32(svcvt_f32_f16_x(svptrue_b32(), svzip1_f16(temp2, temp2)), idx_sc2);
+
+                    acc_f32_01 = svmla_f32_x(svptrue_b32(), acc_f32_01, svcvt_f32_s32_x(svptrue_b32(), row01), svmul_lane_f32(sv_b_d, sv_a_d, 0));
+                    acc_f32_23 = svmla_f32_x(svptrue_b32(), acc_f32_23, svcvt_f32_s32_x(svptrue_b32(), row23), svmul_lane_f32(sv_b_d, sv_a_d, 2));
+                    a_ptr++;
+                    b_ptr++;
+                }
+
+                svbool_t pg4 = svptrue_pat_b32(SV_VL4);
+                svst1_f32(pg4, s + (y+0) * bs + x, acc_f32_01);
+                svst1_f32(pg4, s + (y+1) * bs + x, svext_f32(acc_f32_01, acc_f32_01, 4));
+                svst1_f32(pg4, s + (y+2) * bs + x, acc_f32_23);
+                svst1_f32(pg4, s + (y+3) * bs + x, svext_f32(acc_f32_23, acc_f32_23, 4));
+            }
+        }
+        return;
+    }
+#endif  // SVE compile-time end
+
 #if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
     const block_q8_0x4 * b_ptr_base = (const block_q8_0x4 *) vx;
 

ggml/src/ggml-cpu/arch/x86/quants.c

@@ -274,6 +274,18 @@ static inline __m256 quad_mx_delta_float(const uint8_t x0, const float y0, const
 }
 #endif
 #elif defined(__SSSE3__)
+static inline __m128i bytes_from_bits_16(const uint8_t * x) {
+    uint16_t x16;
+    memcpy(&x16, x, sizeof(uint16_t));
+
+    const __m128i shuf_mask = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+    __m128i bytes = _mm_shuffle_epi8(_mm_set1_epi16((short) x16), shuf_mask);
+    const __m128i bit_mask = _mm_set_epi64x(0x7fbfdfeff7fbfdfe, 0x7fbfdfeff7fbfdfe);
+    bytes = _mm_or_si128(bytes, bit_mask);
+
+    return _mm_cmpeq_epi8(bytes, _mm_set1_epi64x(-1));
+}
+
 // horizontally add 4x4 floats
 static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
     __m128 res_0 =_mm_hadd_ps(a, b);
@@ -540,6 +552,152 @@ static inline __m128i get_scale_shuffle(int i) {
 }
 #endif
 
+void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK1_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q1_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__)
+    const __m256i ones_8 = _mm256_set1_epi8(1);
+    const __m256i ones_16 = _mm256_set1_epi16(1);
+    const __m256i byte_shuf = _mm256_setr_epi8(
+            0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+            2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3);
+    const __m256i bit_masks = _mm256_setr_epi8(
+            1, 2, 4, 8, 16, 32, 64, (char) -128, 1, 2, 4, 8, 16, 32, 64, (char) -128,
+            1, 2, 4, 8, 16, 32, 64, (char) -128, 1, 2, 4, 8, 16, 32, 64, (char) -128);
+    const __m256i zero = _mm256_setzero_si256();
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const uint32_t * GGML_RESTRICT qs32 = (const uint32_t *) x[ib].qs;
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+
+        __m256 acc_block;
+        {
+            const __m256i qy = _mm256_loadu_si256((const __m256i *) y_ptr[0].qs);
+            const __m256i sm = _mm256_cmpeq_epi8(
+                    _mm256_and_si256(_mm256_shuffle_epi8(_mm256_set1_epi32((int) qs32[0]), byte_shuf), bit_masks), zero);
+            const __m256i sy = _mm256_sub_epi8(_mm256_xor_si256(qy, sm), sm);
+            const __m256i s32 = _mm256_madd_epi16(_mm256_maddubs_epi16(ones_8, sy), ones_16);
+            acc_block = _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[0].d)), _mm256_cvtepi32_ps(s32));
+        }
+        for (int K = 1; K < 4; ++K) {
+            const __m256i qy = _mm256_loadu_si256((const __m256i *) y_ptr[K].qs);
+            const __m256i sm = _mm256_cmpeq_epi8(
+                    _mm256_and_si256(_mm256_shuffle_epi8(_mm256_set1_epi32((int) qs32[K]), byte_shuf), bit_masks), zero);
+            const __m256i sy = _mm256_sub_epi8(_mm256_xor_si256(qy, sm), sm);
+            const __m256i s32 = _mm256_madd_epi16(_mm256_maddubs_epi16(ones_8, sy), ones_16);
+            acc_block = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[K].d)), _mm256_cvtepi32_ps(s32), acc_block);
+        }
+        acc = _mm256_fmadd_ps(_mm256_set1_ps(d0), acc_block, acc);
+    }
+
+    *s = hsum_float_8(acc);
+#elif defined(__AVX__)
+    const __m128i ones_8 = _mm_set1_epi8(1);
+    const __m128i ones_16 = _mm_set1_epi16(1);
+    const __m128i zero = _mm_setzero_si128();
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+        __m256 acc_block;
+        {
+            const __m256i bit_mask = bytes_from_bits_32(&x[ib].qs[0]);
+            const __m128i bit_mask_0 = _mm256_castsi256_si128(bit_mask);
+            const __m128i bit_mask_1 = _mm256_extractf128_si256(bit_mask, 1);
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[0].qs[0]);
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[0].qs[16]);
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero);
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero);
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0);
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1);
+            const __m128i sum16_0 = _mm_maddubs_epi16(ones_8, sy_0);
+            const __m128i sum16_1 = _mm_maddubs_epi16(ones_8, sy_1);
+            const __m128i sum32_0 = _mm_madd_epi16(sum16_0, ones_16);
+            const __m128i sum32_1 = _mm_madd_epi16(sum16_1, ones_16);
+            const __m256 q = _mm256_cvtepi32_ps(MM256_SET_M128I(sum32_1, sum32_0));
+            acc_block = _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[0].d)), q);
+        }
+        for(int K = 1; K < 4; ++K) {
+            const __m256i bit_mask = bytes_from_bits_32(&x[ib].qs[(K) * 4]);
+            const __m128i bit_mask_0 = _mm256_castsi256_si128(bit_mask);
+            const __m128i bit_mask_1 = _mm256_extractf128_si256(bit_mask, 1);
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[(K)].qs[0]);
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[(K)].qs[16]);
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero);
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero);
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0);
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1);
+            const __m128i sum16_0 = _mm_maddubs_epi16(ones_8, sy_0);
+            const __m128i sum16_1 = _mm_maddubs_epi16(ones_8, sy_1);
+            const __m128i sum32_0 = _mm_madd_epi16(sum16_0, ones_16);
+            const __m128i sum32_1 = _mm_madd_epi16(sum16_1, ones_16);
+            const __m256 q = _mm256_cvtepi32_ps(MM256_SET_M128I(sum32_1, sum32_0));
+            acc_block = _mm256_add_ps(acc_block, _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[(K)].d)), q));
+        }
+#undef Q1_AVX_BLOCK
+
+        acc = _mm256_add_ps(acc, _mm256_mul_ps(_mm256_set1_ps(d0), acc_block));
+    }
+
+    *s = hsum_float_8(acc);
+#elif defined(__SSSE3__)
+    const __m128i ones_8 = _mm_set1_epi8(1);
+    const __m128i ones_16 = _mm_set1_epi16(1);
+    const __m128i zero = _mm_setzero_si128();
+    __m128 acc_0 = _mm_setzero_ps();
+    __m128 acc_1 = _mm_setzero_ps();
+    __m128 acc_2 = _mm_setzero_ps();
+    __m128 acc_3 = _mm_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const __m128 d0 = _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+
+#define Q1_SSSE3_BLOCK(QS_OFF, Y_IDX, ACC) \
+        { \
+            const __m128i bit_mask_0 = bytes_from_bits_16(&x[ib].qs[(QS_OFF) + 0]); \
+            const __m128i bit_mask_1 = bytes_from_bits_16(&x[ib].qs[(QS_OFF) + 2]); \
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[(Y_IDX)].qs[0]); \
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[(Y_IDX)].qs[16]); \
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero); \
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero); \
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0); \
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1); \
+            const __m128i sum_0 = _mm_madd_epi16(_mm_maddubs_epi16(ones_8, sy_0), ones_16); \
+            const __m128i sum_1 = _mm_madd_epi16(_mm_maddubs_epi16(ones_8, sy_1), ones_16); \
+            const __m128 q = _mm_cvtepi32_ps(_mm_add_epi32(sum_0, sum_1)); \
+            (ACC) = _mm_add_ps((ACC), _mm_mul_ps(_mm_mul_ps(d0, _mm_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[(Y_IDX)].d))), q)); \
+        }
+        Q1_SSSE3_BLOCK(0,  0, acc_0)
+        Q1_SSSE3_BLOCK(4,  1, acc_1)
+        Q1_SSSE3_BLOCK(8,  2, acc_2)
+        Q1_SSSE3_BLOCK(12, 3, acc_3)
+#undef Q1_SSSE3_BLOCK
+    }
+
+    *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q1_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -2142,9 +2300,8 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 #if defined __AVX2__
 
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-    const __m256i m2 = _mm256_set1_epi8(3);
-    const __m256i m32s = _mm256_set1_epi8(32);
+    const __m256i m3 = _mm256_set1_epi8(3);
+    const __m256i m15 = _mm256_set1_epi8(15);
 
     __m256 acc = _mm256_setzero_ps();
 
@@ -2156,53 +2313,45 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
+        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
         const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+        const __m256i scales_16 = _mm256_cvtepi8_epi16(scales);
+        const __m256i q8sclsub = _mm256_slli_epi32(_mm256_madd_epi16(q8sums, scales_16), 5);
 
         __m256i sumi = _mm256_setzero_si256();
 
         int is = 0;
 
         for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
-            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
-            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
-            is += 4;
-
             const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
             const __m256i q4bits2 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
             const __m256i q4bitsH = _mm256_loadu_si256((const __m256i*)qh); qh += 32;
 
-            const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m2), 4);
-            const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 2), m2), 4);
-            const __m256i q4h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 4), m2), 4);
-            const __m256i q4h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 6), m2), 4);
+            const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m3), 4);
+            const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, _mm256_set1_epi8(12)), 2);
+            const __m256i q4h_2 = _mm256_and_si256(q4bitsH, _mm256_set1_epi8(48));
+            const __m256i q4h_3 = _mm256_srli_epi16(_mm256_and_si256(q4bitsH, _mm256_set1_epi8(-64)), 2);
 
-            const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
-            const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m4), q4h_1);
-            const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_2);
-            const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m4), q4h_3);
+            const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m15), q4h_0);
+            const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m15), q4h_1);
+            const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m15), q4h_2);
+            const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m15), q4h_3);
 
             const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
             const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
             const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
             const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
 
-            __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
-            __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
-            __m256i q8s_2 = _mm256_maddubs_epi16(m32s, q8_2);
-            __m256i q8s_3 = _mm256_maddubs_epi16(m32s, q8_3);
-
             __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
             __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
             __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
             __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
 
-            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
+            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
+            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
+            is += 4;
 
             p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
             p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
@@ -2214,6 +2363,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         }
 
+        sumi = _mm256_sub_epi32(sumi, q8sclsub);
         acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
     }
 

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -195,8 +195,8 @@ static const struct ggml_backend_i ggml_backend_cpu_i = {
     /* .free                    = */ ggml_backend_cpu_free,
     /* .set_tensor_async        = */ NULL,
     /* .get_tensor_async        = */ NULL,
-    /* .get_tensor_2d_async     = */ NULL,
     /* .set_tensor_2d_async     = */ NULL,
+    /* .get_tensor_2d_async     = */ NULL,
     /* .cpy_tensor_async        = */ NULL,
     /* .synchronize             = */ NULL,
     /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,

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

@@ -2321,6 +2321,9 @@ class tinyBLAS_Q0_PPC {
     }
 
     void matmul(int64_t m, int64_t n) {
+    #if defined(_AIX) || defined(__BIG_ENDIAN__)
+        mnpack(0, m, 0, n);
+    #else
         const int64_t mc = 64;
         const int64_t kc = 64;
         int64_t nc = 64;
@@ -2334,7 +2337,6 @@ class tinyBLAS_Q0_PPC {
         } else {
             n_aligned = (n / 64) * 64;
         }
-
         if (n_aligned > 0) {
             if (n_aligned % 64 == 0)      nc = 64;
             else if (n_aligned == n)      nc = n;
@@ -2352,6 +2354,7 @@ class tinyBLAS_Q0_PPC {
         } else {
             mnpack(0, m, 0, n);
         }
+    #endif
     }
 
   private:
@@ -3191,12 +3194,16 @@ class tinyBLAS_PPC {
     }
 
     void matmul(int64_t m, int64_t n) {
+    #if defined(_AIX) || defined(__BIG_ENDIAN__)
+        mnpack(0, m, 0, n);
+    #else
         int64_t mc = 256; int64_t nc = 256; int64_t kc = 256;
         if (m % mc == 0 && n % nc == 0 && k % kc == 0) {
             matmul_tiled(m, n, mc, nc, kc);
         } else {
             mnpack(0, m, 0, n);
         }
+    #endif
     }
 
   private:

ggml/src/ggml-cpu/quants.c

@@ -137,22 +137,28 @@ void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
     float sumf = 0.0;
 
     for (int i = 0; i < nb; i++) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[i].d);
 
         float sumi = 0.0f;
 
         for (int k = 0; k < 4; k++) {
-            const float d1 = GGML_FP16_TO_FP32(y[i*4 + k].d);
-
+            const block_q8_0 * GGML_RESTRICT yb = &y[i * 4 + k];
+            const float d1 = GGML_CPU_FP16_TO_FP32(yb->d);
             int sumi_block = 0;
 
-            for (int j = 0; j < QK8_0; j++) {
-                const int bit_index = k * QK8_0 + j;
-                const int byte_index = bit_index / 8;
-                const int bit_offset = bit_index % 8;
+            const uint8_t * GGML_RESTRICT bits = &x[i].qs[k * 4];
+            const int8_t  * GGML_RESTRICT qy   = yb->qs;
 
-                const int xi = ((x[i].qs[byte_index] >> bit_offset) & 1) ? 1 : -1;
-                sumi_block += xi * y[i*4 + k].qs[j];
+            for (int b = 0; b < 4; ++b, qy += 8) {
+                const unsigned mask = bits[b];
+                sumi_block += ((mask & 0x01) ? qy[0] : -qy[0])
+                           +  ((mask & 0x02) ? qy[1] : -qy[1])
+                           +  ((mask & 0x04) ? qy[2] : -qy[2])
+                           +  ((mask & 0x08) ? qy[3] : -qy[3])
+                           +  ((mask & 0x10) ? qy[4] : -qy[4])
+                           +  ((mask & 0x20) ? qy[5] : -qy[5])
+                           +  ((mask & 0x40) ? qy[6] : -qy[6])
+                           +  ((mask & 0x80) ? qy[7] : -qy[7]);
             }
 
             sumi += d1 * sumi_block;

ggml/src/ggml-cpu/vec.h

@@ -1036,12 +1036,12 @@ inline static float ggml_gelu_quick_f32(float x) {
     return x*(1.0f/(1.0f+expf(GELU_QUICK_COEF*x)));
 }
 
-//inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
-//    const uint16_t * i16 = (const uint16_t *) x;
-//    for (int i = 0; i < n; ++i) {
-//        y[i] = ggml_table_gelu_quick_f16[i16[i]];
-//    }
-//}
+inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    const uint16_t * i16 = (const uint16_t *) x;
+    for (int i = 0; i < n; ++i) {
+        y[i] = ggml_table_gelu_quick_f16[i16[i]];
+    }
+}
 
 #ifdef GGML_GELU_QUICK_FP16
 inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
@@ -1060,13 +1060,6 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float *
 }
 #endif
 
-inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
-    for (int i = 0; i < n; ++i) {
-        float v = GGML_CPU_FP16_TO_FP32(x[i]);
-        y[i] = GGML_CPU_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v))));
-    }
-}
-
 // Sigmoid Linear Unit (SiLU) function
 inline static float ggml_silu_f32(float x) {
     return x/(1.0f + expf(-x));

ggml/src/ggml-cuda/common.cuh

@@ -269,10 +269,6 @@ static const char * cu_get_error_str(CUresult err) {
 #define FLASH_ATTN_AVAILABLE
 #endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
-#if defined(TURING_MMA_AVAILABLE)
-#define LDMATRIX_TRANS_AVAILABLE
-#endif // defined(TURING_MMA_AVAILABLE)
-
 static bool fp16_available(const int cc) {
     return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
         (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
@@ -834,6 +830,18 @@ static __device__ __forceinline__ float ggml_cuda_ue4m3_to_fp32(uint8_t x) {
 #endif // defined(GGML_USE_HIP) && defined(CDNA3) && defined(FP8_AVAILABLE) && HIP_VERSION >= 60200000
 }
 
+static __device__ __forceinline__ uint8_t ggml_cuda_fp32_to_ue4m3(float x) {
+#if defined(BLACKWELL_MMA_AVAILABLE) // This is used for NVFP4 subblock scale quantizations only
+    if (!(x > 0.0f)) {
+        return 0;
+    }
+    const __nv_fp8_e4m3 xf(x);
+    return xf.__x;
+#else
+     NO_DEVICE_CODE; // Used only for NVFP4 Scales for Activations, only for Blackwell
+#endif // defined(BLACKWELL_MMA_AVAILABLE)
+}
+
 __device__ __forceinline__ uint8_t ggml_cuda_float_to_fp4_e2m1(float x, float e) {
     const uint8_t sign_bit = (x < 0.0f) << 3;
     float         ax       = fabsf(x) * e;

ggml/src/ggml-cuda/concat.cu

@@ -1,96 +1,79 @@
 #include "concat.cuh"
 
 // contiguous kernels
-static __global__ void concat_f32_dim0(const float * x, const float * y, float * dst, const int ne0, const int ne00) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
+template <int dim>
+static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE) concat_f32_cont(const float * x,
+                                                                                 const float * y,
+                                                                                 float *       dst,
+                                                                                 int64_t       ne00,
+                                                                                 int64_t       ne01,
+                                                                                 int64_t       ne02,
+                                                                                 int64_t       ne0,
+                                                                                 int64_t       ne1,
+                                                                                 int64_t       ne2) {
+    static_assert(dim >= 0 && dim <= 2, "dim must be in [0, 2]");
 
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
+    const int64_t n = ne0 * ne1 * ne2;
 
-    if (nidx < ne00) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne00 +
-            blockIdx.z * ne00 * gridDim.y;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            (nidx - ne00) +
-            blockIdx.y * (ne0 - ne00) +
-            blockIdx.z * (ne0 - ne00) * gridDim.y;
-        dst[offset_dst] = y[offset_src];
+    for (int64_t i = (int64_t) blockIdx.x * blockDim.x + threadIdx.x; i < n; i += (int64_t) blockDim.x * gridDim.x) {
+        if constexpr (dim == 0) {
+            const int64_t row = i / ne0;
+            const int64_t i0  = i - row * ne0;
+
+            if (i0 < ne00) {
+                dst[i] = x[row * ne00 + i0];
+            } else {
+                dst[i] = y[row * (ne0 - ne00) + (i0 - ne00)];
+            }
+        } else if constexpr (dim == 1) {
+            const int64_t dst_plane  = ne0 * ne1;
+            const int64_t src0_plane = ne0 * ne01;
+            const int64_t src1_plane = dst_plane - src0_plane;
+            const int64_t i2         = i / dst_plane;
+            const int64_t i01        = i - i2 * dst_plane;
+
+            if (i01 < src0_plane) {
+                dst[i] = x[i2 * src0_plane + i01];
+            } else {
+                dst[i] = y[i2 * src1_plane + (i01 - src0_plane)];
+            }
+        } else {
+            const int64_t src0_size = ne0 * ne1 * ne02;
+
+            if (i < src0_size) {
+                dst[i] = x[i];
+            } else {
+                dst[i] = y[i - src0_size];
+            }
+        }
     }
 }
 
-static __global__ void concat_f32_dim1(const float * x, const float * y, float * dst, const int ne0, const int ne01) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
+static void concat_f32_cuda(const float * x,
+                            const float * y,
+                            float *       dst,
+                            int64_t       ne00,
+                            int64_t       ne01,
+                            int64_t       ne02,
+                            int64_t       ne0,
+                            int64_t       ne1,
+                            int64_t       ne2,
+                            int           dim,
+                            cudaStream_t  stream) {
+    const int64_t n          = ne0 * ne1 * ne2;
+    const int     num_blocks = (n + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
 
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-
-    if (blockIdx.y < (unsigned)ne01) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            blockIdx.z * ne0 * ne01;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            nidx +
-            (blockIdx.y - ne01) * ne0 +
-            blockIdx.z * ne0 * (gridDim.y - ne01);
-        dst[offset_dst] = y[offset_src];
-    }
-}
-
-static __global__ void concat_f32_dim2(const float * x, const float * y, float * dst, const int ne0, const int ne02) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-
-    if (blockIdx.z < (unsigned)ne02) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            blockIdx.z * ne0 * gridDim.y;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            (blockIdx.z - ne02) * ne0 *  gridDim.y;
-        dst[offset_dst] = y[offset_src];
-    }
-}
-
-static void concat_f32_cuda(const float * x, const float * y, float * dst, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2, int dim, cudaStream_t stream) {
-    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, ne1, ne2);
     if (dim == 0) {
-        concat_f32_dim0<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne00);
+        concat_f32_cont<0>
+            <<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
         return;
     }
     if (dim == 1) {
-        concat_f32_dim1<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne01);
+        concat_f32_cont<1>
+            <<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
         return;
     }
-    concat_f32_dim2<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
+    concat_f32_cont<2><<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
 }
 
 // non-contiguous kernel (slow)

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -66,6 +66,9 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  32, 128, 128, 128, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  32, 128, 128, 128, 2, true);
 
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(320, 256, 32, 128, 2,  32, 128, 128, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(320, 256, 64, 256, 1,  32, 128, 128, 128, 1, false);
+
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512,  8,  64, 4,  32, 256, 256, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 16,  64, 4,  32, 256, 256, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 32, 128, 2,  32, 128, 128, 128, 1, false);
@@ -85,6 +88,9 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
 
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(320, 256, 32, 128, 2,  32, 128, 128, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(320, 256, 64, 256, 1,  32, 128, 128, 128, 1, false);
+
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512,  8,  64, 4,  32,  96,  64, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 32, 128, 2,  32, 128, 128, 128, 1, false);
@@ -118,6 +124,9 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
 
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(320, 256, 32, 128, 2,  64, 160, 128,  64, 2, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(320, 256, 64, 128, 2,  64, 160, 128,  64, 2, false);
+
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 16,  64, 4,  32, 128, 128, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 32, 128, 2,  32, 128, 128, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 64, 256, 1,  32, 128, 128, 128, 1, false);
@@ -305,12 +314,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int D2, const int stride_KV, const int i_sup) {
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     // K/V data is loaded with decreasing granularity for D for better memory bandwidth.
-    // The minimum granularity with cp.async is 16 bytes, with synchronous data loading it's 4 bytes.
+    // The minimum granularity is 16 bytes.
+    constexpr int h2_per_chunk = 16/sizeof(half2);
+    const int chunks_per_row = D2 / h2_per_chunk;
     if constexpr (use_cp_async) {
+        static_assert(warp_size == 32, "bad warp_size");
         static_assert(!oob_check, "OOB check not compatible with cp_async");
         constexpr int preload = 64;
-        constexpr int h2_per_chunk = 16/sizeof(half2);
-        const int chunks_per_row = D2 / h2_per_chunk;
 
         const unsigned int tile_KV_32 = ggml_cuda_cvta_generic_to_shared(tile_KV);
 
@@ -348,11 +358,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         // 6: max  1*16= 16 bytes,   8 half
         ggml_cuda_unroll<6>{}(load);
     } else {
-        // TODO use ggml_cuda_memcpy_1
+        const half2 zero[4] = {{0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}};
         auto load = [&] __device__ (const int n) {
-            const int stride_k = warp_size >> n;
-            const int k0_start = stride_k == warp_size ? 0 : D2 - D2 % (2*stride_k);
-            const int k0_stop  =                             D2 - D2 % (1*stride_k);
+            const int stride_k = 32 >> n;
+            const int k0_start = stride_k == 32 ? 0 : chunks_per_row - chunks_per_row % (2*stride_k);
+            const int k0_stop  =                      chunks_per_row - chunks_per_row % (1*stride_k);
             const int stride_i = warp_size / stride_k;
 
             if (k0_start == k0_stop) {
@@ -371,15 +381,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                     const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
-                    tile_KV[i*stride_tile + k] = !oob_check || i < i_sup ? KV[i*stride_KV + k] : make_half2(0.0f, 0.0f);
+                    ggml_cuda_memcpy_1<16>(tile_KV + i*stride_tile + k*4,
+                        !oob_check || i < i_sup ? KV + i*stride_KV + k*h2_per_chunk : zero);
                 }
             }
         };
-        // 1: max 32* 4=128 bytes,  64 half
-        // 2: max 16* 4= 64 bytes,  32 half
-        // 3: max  8* 4= 32 bytes,  16 half
-        // 4: max  4* 4= 16 bytes,   8 half
-        ggml_cuda_unroll<4>{}(load);
+        // 1: max 32*16=512 bytes, 256 half
+        // 2: max 16*16=256 bytes, 128 half
+        // 3: max  8*16=128 bytes,  64 half
+        // 4: max  4*16= 64 bytes,  32 half
+        // 5: max  2*16= 32 bytes,  16 half
+        // 6: max  1*16= 16 bytes,   8 half
+        ggml_cuda_unroll<6>{}(load);
     }
 }
 
@@ -862,11 +875,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     }
 
 
-#if defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
-    T_A_VKQ A_identity;
-    make_identity_mat(A_identity);
-#endif // defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
-
     // Calculate VKQ tile, need to use logical rather than physical elements for i0 due to transposition of V:
 #pragma unroll
     for (int i0_start = 0; i0_start < DV; i0_start += 2*nbatch_V2) {
@@ -897,29 +905,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 const int k0 = k00 + (threadIdx.y % np)*T_A_VKQ::J;
 
                 T_A_VKQ A; // Transposed in SRAM but not in registers, gets transposed on load.
-#if defined(LDMATRIX_TRANS_AVAILABLE)
                 load_ldmatrix_trans(A, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
-#elif defined(AMD_MFMA_AVAILABLE)
-                // MFMA A register layout: A_mat[i=lane%16][k=4*(lane/16)+reg].
-                // Normal load gives A_mat[seq][dv] but we need A_mat[dv][seq] = V^T.
-                // Load with transposed addressing: 4 strided half loads.
-                {
-                    const half2 * xs0 = tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2;
-                    const half * xs0_h = (const half *) xs0;
-                    const int stride_h = stride_tile_V * 2; // stride in half units
-                    half * A_h = (half *) A.x;
-#pragma unroll
-                    for (int l = 0; l < 4; ++l) {
-                        A_h[l] = xs0_h[(4*(threadIdx.x / 16) + l) * stride_h + threadIdx.x % 16];
-                    }
-                }
-#else
-                // TODO: Try to transpose tile_V when loading gmem to smem.
-                // Use mma to transpose T_A_VKQ for RDNA.
-                T_A_VKQ A_trans;
-                load_ldmatrix(A_trans, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
-                mma(A, A_trans, A_identity);
-#endif // defined(LDMATRIX_TRANS_AVAILABLE)
                 if constexpr (T_B_KQ::I == 8) {
                     mma(VKQ_C[i_VKQ_0/i0_stride], A, B[k00/(np*T_A_VKQ::J)]);
                 } else {
@@ -1240,7 +1226,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         float KQ_max_scale[cols_per_thread];
 #pragma unroll
         for (int col = 0; col < cols_per_thread; ++col) {
-            const int jc = cols_per_warp == 8 ? T_C_KQ::get_j(col) : T_C_KQ::get_i(2*col);
+            const int jc = (threadIdx.y/np)*cols_per_warp + (cols_per_warp == 8 ? T_C_KQ::get_j(col) : T_C_KQ::get_i(2*col));
             const float sink = sinks_f[jc % ncols2];
 
             const float KQ_max_new = fmaxf(KQ_max[col], sink);
@@ -1848,6 +1834,10 @@ extern DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);
 
+// Mistral Small 4 (DKQ=320, DV=256), GQA=32-only build:
+extern DECL_FATTN_MMA_F16_CASE(320, 256,  1, 32);
+extern DECL_FATTN_MMA_F16_CASE(320, 256,  2, 32);
+
 // For GLM 4.7 Flash
 extern DECL_FATTN_MMA_F16_CASE(576, 512,  4,  4);
 extern DECL_FATTN_MMA_F16_CASE(576, 512,  8,  4);

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

@@ -38,6 +38,10 @@ void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_cuda_flash_attn_ext_tile_case<256, 256>(ctx, dst);
         } break;
+        case 320: {
+            GGML_ASSERT(V->ne[0] == 256);
+            ggml_cuda_flash_attn_ext_tile_case<320, 256>(ctx, dst);
+        } break;
         case 512: {
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_cuda_flash_attn_ext_tile_case<512, 512>(ctx, dst);

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

@@ -68,6 +68,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 16, 256, 2,  64,  64)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2,  64,  64)
@@ -128,6 +130,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 16, 256, 2,  32,  64)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  32,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  32,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2,  32,  64)
@@ -195,6 +199,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 32, 512, 1, 128,  64)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2,  64,  64)
@@ -264,6 +270,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 32, 256, 2, 128,  64)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 4,  64,  64)
@@ -1116,7 +1124,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
     constexpr size_t nbytes_shared = 0;
 
 #ifdef GGML_USE_HIP
-    if constexpr (DV <= 128) {
+    if constexpr (DKQ <= 128) {
         if (Q->ne[1] > 32/ncols2) {
             constexpr int cols_per_block = 64;
             const int nwarps    = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
@@ -1130,7 +1138,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
 #endif // GGML_USE_HIP
 
 #ifndef GGML_USE_HIP
-    if constexpr (DV <= 256)
+    if constexpr (DKQ <= 256)
 #endif // GGML_USE_HIP
     {
         if (Q->ne[1] > 16/ncols2) {
@@ -1144,14 +1152,16 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
         }
     }
 
-    if (Q->ne[1] > 8/ncols2) {
-        constexpr int cols_per_block = 16;
-        const int nwarps    = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
-        const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
-        fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
-        launch_fattn<DV, cols_per_block/ncols2, ncols2>
-            (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, warp_size);
-        return;
+    if constexpr (ncols2 <= 16) {
+        if (Q->ne[1] > 8/ncols2) {
+            constexpr int cols_per_block = 16;
+            const int nwarps    = ggml_cuda_fattn_tile_get_nthreads (DKQ, DV, cols_per_block, cc) / warp_size;
+            const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
+            fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
+            launch_fattn<DV, cols_per_block/ncols2, ncols2>
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, warp_size);
+            return;
+        }
     }
 
     if constexpr (ncols2 <= 8) {
@@ -1210,6 +1220,25 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
     const int gqa_limit = nvidia && gqa_ratio <= 4 && DV <= 256 ? 16 : INT_MAX;
     const bool use_gqa_opt = mask && max_bias == 0.0f && Q->ne[1] <= gqa_limit && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
+    if constexpr (DKQ == 320) {
+        // This branch is only used for Mistral Small 4 which has a GQA ratio of 32.
+        // On AMD, simply use that GQA ratio with 32 columns / block since we always have enough SRAM.
+        // On NVIDIA however, the tile kernel is only used for GPUs that can't use the mma kernel (Pascal and older).
+        // Therefore, use a GQA ratio of 16 with 16 columns / block to stay below 48 kiB of SRAM / block.
+#ifdef GGML_USE_HIP
+        if (use_gqa_opt && gqa_ratio % 32 == 0) {
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 32, use_logit_softcap>(ctx, dst);
+            return;
+        }
+#else
+        if (use_gqa_opt && gqa_ratio % 16 == 0) {
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
+            return;
+        }
+#endif // GGML_USE_HIP
+        GGML_ABORT("flash-attn tile (320/256): expected GQA ratio multiple of 32");
+    }
+
     if constexpr (DKQ == 576) {
         if (use_gqa_opt && gqa_ratio % 16 == 0) {
             launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
@@ -1221,7 +1250,7 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
         }
     }
 
-    if constexpr (DKQ <= 512) {
+    if constexpr (DKQ <= 512 && DKQ != 320) {
         if (use_gqa_opt && gqa_ratio % 8 == 0) {
             launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
             return;
@@ -1275,5 +1304,6 @@ extern DECL_FATTN_TILE_CASE( 96,  96);
 extern DECL_FATTN_TILE_CASE(112, 112);
 extern DECL_FATTN_TILE_CASE(128, 128);
 extern DECL_FATTN_TILE_CASE(256, 256);
+extern DECL_FATTN_TILE_CASE(320, 256);
 extern DECL_FATTN_TILE_CASE(512, 512);
 extern DECL_FATTN_TILE_CASE(576, 512);

ggml/src/ggml-cuda/fattn.cu

@@ -143,6 +143,22 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
             GGML_ASSERT(V->ne[0] == 256);
             ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
             break;
+        case 320:
+            // For Mistral Small 4, go straight to the ncols1 switch (ncols2=32-only build).
+            GGML_ASSERT(V->ne[0] == 256);
+            {
+                float max_bias = 0.0f;
+                memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+
+                const bool use_gqa_opt = mask && max_bias == 0.0f;
+                GGML_ASSERT(use_gqa_opt);
+                GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
+                const int gqa_ratio = Q->ne[2] / K->ne[2];
+                GGML_ASSERT(gqa_ratio % 32 == 0);
+
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<320, 256, 32>(ctx, dst);
+            }
+            break;
         case 512:
             GGML_ASSERT(V->ne[0] == 512);
             ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<512, 512>(ctx, dst);
@@ -352,6 +368,14 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
                 return BEST_FATTN_KERNEL_NONE;
             }
             break;
+        case 320:
+            if (V->ne[0] != 256 || !gqa_opt_applies) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            if (gqa_ratio % 32 != 0) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            break;
         case 512:
             if (V->ne[0] != K->ne[0]) {
                 return BEST_FATTN_KERNEL_NONE;

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

@@ -368,15 +368,21 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
     }
 
     ~ggml_cuda_pool_leg() {
+        clear_pool();
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    void clear_pool() {
         ggml_cuda_set_device(device);
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cuda_buffer & b = buffer_pool[i];
             if (b.ptr != nullptr) {
                 CUDA_CHECK(cudaFree(b.ptr));
                 pool_size -= b.size;
+                b.ptr  = nullptr;
+                b.size = 0;
             }
         }
-        GGML_ASSERT(pool_size == 0);
     }
 
     void * alloc(size_t size, size_t * actual_size) override {
@@ -421,7 +427,20 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
         size_t look_ahead_size = (size_t) (1.05 * size);
         look_ahead_size = 256 * ((look_ahead_size + 255)/256);
         ggml_cuda_set_device(device);
-        CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
+        cudaError_t err = ggml_cuda_device_malloc(&ptr, look_ahead_size, device);
+        if (err == cudaErrorMemoryAllocation) {
+            (void)cudaGetLastError();
+            const size_t cached_bytes = pool_size;
+            GGML_LOG_DEBUG(GGML_CUDA_NAME " pool[%d]: alloc of %.2f MiB failed, flushing %.2f MiB of cached buffers and retrying\n",
+                           device, look_ahead_size/1024.0/1024.0, cached_bytes/1024.0/1024.0);
+            CUDA_CHECK(cudaDeviceSynchronize());
+            clear_pool();
+            err = ggml_cuda_device_malloc(&ptr, look_ahead_size, device);
+            if (err == cudaSuccess) {
+                GGML_LOG_DEBUG(GGML_CUDA_NAME " pool[%d]: retry succeeded\n", device);
+            }
+        }
+        CUDA_CHECK(err);
         *actual_size = look_ahead_size;
         pool_size += look_ahead_size;
 #ifdef DEBUG_CUDA_MALLOC
@@ -1203,6 +1222,13 @@ static bool ggml_backend_cuda_comm_allreduce_tensor(void * comm_ctx_v, struct gg
     // For small tensors, simply reduce them as FP32.
     // The following heuristic for how "small" a tensor should be is based on RTX 4090s connected via 16x PCIe 4.0.
     if ((n_backends <= 2 && ne < 32768) || (n_backends == 3 && ne < 131072) || (n_backends >= 4 && ne < 262144)) {
+        for (size_t i = 0; i < n_backends; ++i) {
+            if ((tensors[i]->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+                ggml_cuda_set_device(cuda_ctx->device);
+                CUDA_CHECK(cudaMemsetAsync(tensors[i]->data, 0, ggml_nbytes(tensors[i]), cuda_ctx->stream()));
+            }
+        }
         NCCL_CHECK(ncclGroupStart());
         for (size_t i = 0; i < n_backends; ++i) {
             ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
@@ -1224,7 +1250,11 @@ static bool ggml_backend_cuda_comm_allreduce_tensor(void * comm_ctx_v, struct gg
         tmp[i].alloc(ne);
 
         ggml_cuda_set_device(cuda_ctx->device);
-        to_bf16(tensors[i]->data, tmp[i].get(), ne, cuda_ctx->stream());
+        if (tensors[i]->flags & GGML_TENSOR_FLAG_COMPUTE) {
+            to_bf16(tensors[i]->data, tmp[i].get(), ne, cuda_ctx->stream());
+        } else {
+            CUDA_CHECK(cudaMemsetAsync(tmp[i].get(), 0, ne * sizeof(nv_bfloat16), cuda_ctx->stream()));
+        }
         CUDA_CHECK(cudaGetLastError());
     }
 
@@ -3526,6 +3556,9 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
      && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_SILU) {
         const ggml_tensor * ssm_conv = cgraph->nodes[node_idx];
         const ggml_tensor * silu     = cgraph->nodes[node_idx+1];
+        if (ggml_get_unary_op(silu) != unary_ops.begin()[0]) {
+            return false;
+        }
 
         if (ssm_conv->type != GGML_TYPE_F32 || silu->type != GGML_TYPE_F32) {
             return false;
@@ -3534,6 +3567,31 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
         return true;
     }
 
+    if (ops.size() == 3 && ops.begin()[0] == GGML_OP_SSM_CONV && ops.begin()[1] == GGML_OP_ADD
+     && ops.begin()[2] == GGML_OP_UNARY && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_SILU) {
+        const ggml_tensor * ssm_conv = cgraph->nodes[node_idx];
+        const ggml_tensor * add      = cgraph->nodes[node_idx+1];
+        const ggml_tensor * silu     = cgraph->nodes[node_idx+2];
+        if (ggml_get_unary_op(silu) != unary_ops.begin()[0]) {
+            return false;
+        }
+
+        if (ssm_conv->type != GGML_TYPE_F32 || add->type != GGML_TYPE_F32 || silu->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        // ADD must consume ssm_conv's output and broadcast a 1-D channel-wise bias.
+        const ggml_tensor * bias = (add->src[0] == ssm_conv) ? add->src[1] : add->src[0];
+        if (bias->type != GGML_TYPE_F32 || !ggml_is_contiguous(bias)) {
+            return false;
+        }
+        if (ggml_nelements(bias) != ssm_conv->ne[0] || bias->ne[0] != ssm_conv->ne[0]) {
+            return false;
+        }
+
+        return true;
+    }
+
     if (ops.size() == 2 && ops.begin()[0] == GGML_OP_UNARY && ops.begin()[1] == GGML_OP_MUL
      && unary_ops.size() == 1 && (unary_ops.begin()[0] == GGML_UNARY_OP_SILU || unary_ops.begin()[0] == GGML_UNARY_OP_SIGMOID || unary_ops.begin()[0] == GGML_UNARY_OP_SOFTPLUS)) {
         const ggml_tensor * unary = cgraph->nodes[node_idx];
@@ -3562,6 +3620,30 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
         return true;
     }
 
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_UNARY && ops.begin()[1] == GGML_OP_SQR
+     && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_RELU) {
+        const ggml_tensor * unary = cgraph->nodes[node_idx];
+        const ggml_tensor * sqr   = cgraph->nodes[node_idx+1];
+
+        if (ggml_get_unary_op(unary) != GGML_UNARY_OP_RELU) {
+            return false;
+        }
+
+        if (unary->type != GGML_TYPE_F32 && unary->type != GGML_TYPE_F16) {
+            return false;
+        }
+
+        if (unary->type != sqr->type) {
+            return false;
+        }
+
+        if (!ggml_is_contiguous(unary->src[0])) {
+            return false;
+        }
+
+        return true;
+    }
+
     if (ops.size() == 3 && ops.begin()[0] == GGML_OP_SCALE && ops.begin()[1] == GGML_OP_UNARY && ops.begin()[2] == GGML_OP_SCALE
      && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_TANH) {
         const ggml_tensor *scale  = cgraph->nodes[node_idx];
@@ -3586,6 +3668,362 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
     return false;
 }
 
+// try and fuse nodes and return the number of nodes to skip
+static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, int i) {
+
+    static bool disable_fusion = getenv("GGML_CUDA_DISABLE_FUSION") != nullptr && std::atoi(getenv("GGML_CUDA_DISABLE_FUSION"));
+    if (disable_fusion) {
+        return 0;
+    }
+
+    ggml_tensor * node = cgraph->nodes[i];
+
+    //topk-moe
+    if (cgraph->nodes[i]->op == GGML_OP_UNARY || cgraph->nodes[i]->op == GGML_OP_SOFT_MAX ||
+            cgraph->nodes[i]->op == GGML_OP_ARGSORT) {
+        ggml_cuda_topk_moe_args args;
+        const bool              can_fuse = ggml_cuda_topk_moe_fusion(cgraph, i, args);
+        std::vector<ggml_op>    ops;
+
+        if (can_fuse) {
+            const ggml_tensor * logits  = node->src[0];
+            ggml_tensor *       weights = nullptr;
+            ggml_tensor *       ids     = nullptr;
+            const ggml_tensor * bias    = nullptr;
+            const ggml_tensor * clamp   = nullptr;
+            const ggml_tensor * scale   = nullptr;
+
+            if (!args.delayed_softmax) {
+                ggml_op gating_op = args.sigmoid ? GGML_OP_UNARY : GGML_OP_SOFT_MAX;
+                int     out_nodes[2];  // nodes which can't be elided
+
+                if (args.prob_bias) {
+                    bias = cgraph->nodes[i + 2]->src[1];
+                    ops.insert(ops.end(), { gating_op, GGML_OP_RESHAPE, GGML_OP_ADD, GGML_OP_ARGSORT, GGML_OP_VIEW,
+                                            GGML_OP_GET_ROWS });
+                    out_nodes[0] = i + 4;
+                    ids          = cgraph->nodes[i + 4];
+                } else {
+                    ops.insert(ops.end(),
+                               { gating_op, GGML_OP_RESHAPE, GGML_OP_ARGSORT, GGML_OP_VIEW, GGML_OP_GET_ROWS });
+                    out_nodes[0] = i + 3;
+                    ids          = cgraph->nodes[i + 3];
+                }
+
+                if (args.norm) {
+                    ops.insert(ops.end(),
+                               { GGML_OP_RESHAPE, GGML_OP_SUM_ROWS, GGML_OP_CLAMP, GGML_OP_DIV, GGML_OP_RESHAPE });
+                    clamp = cgraph->nodes[i + ops.size() - 3];
+                }
+                if (args.scale) {
+                    ops.insert(ops.end(), { GGML_OP_SCALE });
+                    scale = cgraph->nodes[i + ops.size() - 1];
+                }
+
+                weights      = cgraph->nodes[i + ops.size() - 1];
+                out_nodes[1] = i + ops.size() - 1;
+
+                if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) &&
+                        ggml_cuda_should_use_topk_moe(node, logits, weights, ids) &&
+                        ggml_cuda_check_fusion_memory_ranges(cgraph, i, ops.size(), out_nodes, 2, /*is_topk_moe=*/true)) {
+                    ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args);
+                    return ops.size() - 1;
+                }
+            } else if (!args.norm && !args.prob_bias) {
+                //special case gpt-oss, no norm, no bias.
+                ops.insert(ops.end(), { GGML_OP_ARGSORT, GGML_OP_VIEW, GGML_OP_GET_ROWS, GGML_OP_RESHAPE,
+                                        GGML_OP_SOFT_MAX, GGML_OP_RESHAPE });
+                weights                     = cgraph->nodes[i + 5];
+                ids                         = cgraph->nodes[i + 1];
+                const ggml_tensor * softmax = cgraph->nodes[i + 4];
+
+                int out_nodes[2] = { i + 1, i + 5 };
+                if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) &&
+                        ggml_cuda_should_use_topk_moe(softmax, logits, weights, ids) &&
+                        ggml_cuda_check_fusion_memory_ranges(cgraph, i, ops.size(), out_nodes, 2, /*is_topk_moe=*/true)) {
+                    ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args);
+                    return ops.size() - 1;
+                }
+            }
+        }
+    }
+
+    //RoPE + view + set-rows
+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, {})) {
+        ggml_tensor * rope     = cgraph->nodes[i];
+        ggml_tensor * set_rows = cgraph->nodes[i + 2];
+
+        ggml_cuda_op_rope_fused(*cuda_ctx, rope, set_rows);
+        return 2;
+    }
+
+    // multi-(add or mul)
+    if (node->op == GGML_OP_ADD || node->op == GGML_OP_MUL) {
+        int     n_fuse = 0;
+        ggml_op ops[8];
+        std::fill(ops, ops + 8, node->op);
+
+        for (; n_fuse <= 6; ++n_fuse) {
+            if (!ggml_can_fuse(cgraph, i + n_fuse, ops + n_fuse, 2)) {
+                break;
+            }
+            if (cgraph->nodes[i + n_fuse] != cgraph->nodes[i + n_fuse + 1]->src[0]) {
+                break;
+            }
+            if (!ggml_are_same_layout(cgraph->nodes[i + n_fuse]->src[1], cgraph->nodes[i + n_fuse + 1]->src[1])) {
+                break;
+            }
+        }
+
+        n_fuse++;
+
+        if (n_fuse > 1) {
+            ggml_tensor fused_node;
+            memcpy(&fused_node, node, sizeof(ggml_tensor));
+            for (int j = 0; j < n_fuse - 1; ++j) {
+                fused_node.src[j + 2] = cgraph->nodes[i + j + 1]->src[1];
+            }
+            fused_node.data = cgraph->nodes[i + n_fuse - 1]->data;
+            if (node->op == GGML_OP_ADD) {
+                ggml_cuda_op_fused_add(*cuda_ctx, &fused_node, n_fuse);
+            } else {
+                ggml_cuda_op_fused_mul(*cuda_ctx, &fused_node, n_fuse);
+            }
+            return n_fuse - 1;
+        }
+    }
+
+    bool fused_mul_mat_vec = false;
+    int  fused_node_count  = 0;
+
+    // gate + glu + up
+    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+        if (ggml_cuda_can_fuse(cgraph, i, { op, bias_op, op, bias_op, GGML_OP_GLU }, {})) {
+            ggml_tensor * glu         = cgraph->nodes[i + 4];
+            ggml_tensor * gate_bias_n = glu->src[0];
+            ggml_tensor * up_bias_n   = glu->src[1];
+
+            //we don't assume the order for {gate, up}. Instead infer it from the bias tensor
+            ggml_tensor * gate_n = nullptr;
+            ggml_tensor * up_n   = nullptr;
+
+            if (gate_bias_n->src[0] == cgraph->nodes[i] || gate_bias_n->src[1] == cgraph->nodes[i]) {
+                gate_n = cgraph->nodes[i];
+                up_n   = cgraph->nodes[i + 2];
+            } else if (gate_bias_n->src[0] == cgraph->nodes[i + 2] || gate_bias_n->src[1] == cgraph->nodes[i + 2]) {
+                gate_n = cgraph->nodes[i + 2];
+                up_n   = cgraph->nodes[i];
+            } else {
+                continue;
+            }
+
+            auto get_bias_tensor = [](const ggml_tensor * bias_node, const ggml_tensor * mul_node, ggml_op op_bias) {
+                if (op_bias == GGML_OP_ADD) {
+                    if (bias_node->src[0] == mul_node) {
+                        return bias_node->src[1];
+                    }
+                    if (bias_node->src[1] == mul_node) {
+                        return bias_node->src[0];
+                    }
+                    return (ggml_tensor *) nullptr;
+                }
+                GGML_ASSERT(op_bias == GGML_OP_ADD_ID);
+                GGML_ASSERT(bias_node->src[0] == mul_node);
+                return bias_node->src[1];
+            };
+
+            ggml_tensor * up_bias_tensor   = get_bias_tensor(up_bias_n, up_n, bias_op);
+            ggml_tensor * gate_bias_tensor = get_bias_tensor(gate_bias_n, gate_n, bias_op);
+
+            if (!up_bias_tensor || !gate_bias_tensor) {
+                continue;
+            }
+
+            // we don't support repeating adds
+            if (bias_op == GGML_OP_ADD && (!ggml_are_same_shape(gate_bias_n->src[0], gate_bias_n->src[1]) ||
+                                           !ggml_are_same_shape(up_bias_n->src[0], up_bias_n->src[1]))) {
+                continue;
+            }
+
+            const ggml_tensor * src0 = up_n->src[0];
+            const ggml_tensor * src1 = up_n->src[1];
+            const ggml_tensor * ids  = up_n->src[2];
+
+            if (ggml_cuda_should_fuse_mul_mat_vec_f(up_n)) {
+                ggml_cuda_mm_fusion_args_host fusion_data{};
+                fusion_data.gate      = gate_n->src[0];
+                fusion_data.x_bias    = up_bias_tensor;
+                fusion_data.gate_bias = gate_bias_tensor;
+                fusion_data.glu_op    = ggml_get_glu_op(glu);
+
+                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                fused_mul_mat_vec = true;
+                fused_node_count  = 5;
+                break;
+            }
+
+            if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
+                ggml_cuda_mm_fusion_args_host fusion_data{};
+                fusion_data.gate      = gate_n->src[0];
+                fusion_data.x_bias    = up_bias_tensor;
+                fusion_data.gate_bias = gate_bias_tensor;
+                fusion_data.glu_op    = ggml_get_glu_op(glu);
+
+                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                fused_mul_mat_vec = true;
+                fused_node_count  = 5;
+                break;
+            }
+        } else if (ggml_cuda_can_fuse(cgraph, i, { op, op, GGML_OP_GLU }, {})) {
+            ggml_tensor * glu  = cgraph->nodes[i + 2];
+            ggml_tensor * gate = glu->src[0];
+            ggml_tensor * up   = glu->src[1];
+
+            bool ok = (gate == cgraph->nodes[i] && up == cgraph->nodes[i + 1]) ||
+                      (gate == cgraph->nodes[i + 1] && up == cgraph->nodes[i]);
+
+            if (!ok) {
+                continue;
+            }
+
+            const ggml_tensor * src0 = up->src[0];
+            const ggml_tensor * src1 = up->src[1];
+            const ggml_tensor * ids  = up->src[2];
+
+            if (ggml_cuda_should_fuse_mul_mat_vec_f(up)) {
+                ggml_cuda_mm_fusion_args_host fusion_data{};
+                fusion_data.gate   = gate->src[0];
+                fusion_data.glu_op = ggml_get_glu_op(glu);
+
+                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                fused_mul_mat_vec = true;
+                fused_node_count  = 3;
+                break;
+            }
+
+            if (ggml_cuda_should_fuse_mul_mat_vec_q(up)) {
+                ggml_cuda_mm_fusion_args_host fusion_data{};
+                fusion_data.gate   = gate->src[0];
+                fusion_data.glu_op = ggml_get_glu_op(glu);
+
+                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
+                fused_mul_mat_vec = true;
+                fused_node_count  = 3;
+                break;
+            }
+        }
+    }
+
+    if (fused_mul_mat_vec) {
+        return fused_node_count - 1;
+    }
+
+    fused_mul_mat_vec = false;
+    fused_node_count  = 0;
+
+    // gate + add + glu + up + add
+    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
+        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
+
+        if (!ggml_can_fuse(cgraph, i, { op, bias_op })) {
+            continue;
+        }
+
+        ggml_tensor * mm_node   = cgraph->nodes[i];
+        ggml_tensor * bias_node = cgraph->nodes[i + 1];
+
+        ggml_tensor * bias_tensor = nullptr;
+        if (bias_op == GGML_OP_ADD) {
+            if (bias_node->src[0] == mm_node) {
+                bias_tensor = bias_node->src[1];
+            } else if (bias_node->src[1] == mm_node) {
+                bias_tensor = bias_node->src[0];
+            } else {
+                continue;
+            }
+        } else {
+            if (bias_node->src[0] != mm_node) {
+                continue;
+            }
+            bias_tensor = bias_node->src[1];
+        }
+
+        const ggml_tensor * src0 = mm_node->src[0];
+        const ggml_tensor * src1 = mm_node->src[1];
+        const ggml_tensor * ids  = mm_node->src[2];
+
+        if (bias_op == GGML_OP_ADD_ID && bias_node->src[2] != ids) {
+            continue;
+        }
+
+        if (bias_op == GGML_OP_ADD && !ggml_are_same_shape(bias_node->src[0], bias_node->src[1])) {
+            continue;
+        }
+
+        ggml_cuda_mm_fusion_args_host fusion_data{};
+        fusion_data.x_bias = bias_tensor;
+
+        if (ggml_cuda_should_fuse_mul_mat_vec_f(mm_node)) {
+            ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
+            fused_mul_mat_vec = true;
+            fused_node_count  = 2;
+            break;
+        }
+
+        if (ggml_cuda_should_fuse_mul_mat_vec_q(mm_node)) {
+            ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
+            fused_mul_mat_vec = true;
+            fused_node_count  = 2;
+            break;
+        }
+    }
+
+    if (fused_mul_mat_vec) {
+        return fused_node_count - 1;
+    }
+
+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ADD }, {})) {
+        ggml_cuda_op_rms_norm_fused_add(*cuda_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2]);
+        return 2;
+    }
+
+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL }, {})) {
+        ggml_cuda_op_rms_norm_fused(*cuda_ctx, node, cgraph->nodes[i + 1]);
+        return 1;
+    }
+
+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SSM_CONV, GGML_OP_ADD, GGML_OP_UNARY }, { GGML_UNARY_OP_SILU })) {
+        ggml_cuda_op_ssm_conv(*cuda_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2]);
+        return 2;
+    }
+
+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SSM_CONV, GGML_OP_UNARY }, { GGML_UNARY_OP_SILU })) {
+        ggml_cuda_op_ssm_conv(*cuda_ctx, node, /*bias_add_node=*/ nullptr, cgraph->nodes[i + 1]);
+        return 1;
+    }
+
+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SILU }) ||
+        ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SIGMOID }) ||
+        ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SOFTPLUS })) {
+        ggml_cuda_op_unary_mul(*cuda_ctx, node, cgraph->nodes[i + 1]);
+        return 1;
+    }
+
+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_SQR }, { GGML_UNARY_OP_RELU })) {
+        ggml_cuda_op_relu_sqr(*cuda_ctx, node, cgraph->nodes[i + 1]);
+        return 1;
+    }
+
+    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SCALE, GGML_OP_UNARY, GGML_OP_SCALE }, { GGML_UNARY_OP_TANH })) {
+        ggml_cuda_op_softcap(*cuda_ctx, cgraph->nodes[i + 2], node);
+        return 2;
+    }
+
+    return 0;
+}
+
 static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, const bool use_cuda_graph, const bool cuda_graph_update_required, const void * graph_key) {
     bool graph_evaluated_or_captured = false;
 
@@ -3732,349 +4170,11 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                     continue;
                 }
 
-                // start of fusion operations
-                static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
-                if (!disable_fusion) {
-                    ggml_cuda_topk_moe_args args;
+                int nodes_to_skip = ggml_cuda_try_fuse(cuda_ctx, cgraph, i);
 
-                    if (cgraph->nodes[i]->op == GGML_OP_UNARY || cgraph->nodes[i]->op == GGML_OP_SOFT_MAX ||
-                        cgraph->nodes[i]->op == GGML_OP_ARGSORT) {
-                        const bool can_fuse = ggml_cuda_topk_moe_fusion(cgraph, i, args);
-
-                        std::vector<ggml_op> ops;
-
-                        if (can_fuse) {
-                            const ggml_tensor * logits  = node->src[0];
-                            ggml_tensor *       weights = nullptr;
-                            ggml_tensor *       ids     = nullptr;
-                            const ggml_tensor * bias    = nullptr;
-                            const ggml_tensor * clamp   = nullptr;
-                            const ggml_tensor * scale   = nullptr;
-
-                            if (!args.delayed_softmax) {
-                                ggml_op gating_op = args.sigmoid ? GGML_OP_UNARY : GGML_OP_SOFT_MAX;
-                                int     out_nodes[2];  // nodes which can't be elided
-
-                                if (args.prob_bias) {
-                                    bias = cgraph->nodes[i + 2]->src[1];
-                                    ops.insert(ops.end(), { gating_op, GGML_OP_RESHAPE, GGML_OP_ADD, GGML_OP_ARGSORT,
-                                                            GGML_OP_VIEW, GGML_OP_GET_ROWS });
-                                    out_nodes[0] = i + 4;
-                                    ids          = cgraph->nodes[i + 4];
-                                } else {
-                                    ops.insert(ops.end(), { gating_op, GGML_OP_RESHAPE, GGML_OP_ARGSORT, GGML_OP_VIEW,
-                                                            GGML_OP_GET_ROWS });
-                                    out_nodes[0] = i + 3;
-                                    ids          = cgraph->nodes[i + 3];
-                                }
-
-                                if (args.norm) {
-                                    ops.insert(ops.end(), { GGML_OP_RESHAPE, GGML_OP_SUM_ROWS, GGML_OP_CLAMP,
-                                                            GGML_OP_DIV, GGML_OP_RESHAPE });
-                                    clamp = cgraph->nodes[i + ops.size() - 3];
-                                }
-                                if (args.scale) {
-                                    ops.insert(ops.end(), { GGML_OP_SCALE });
-                                    scale = cgraph->nodes[i + ops.size() - 1];
-                                }
-
-                                weights      = cgraph->nodes[i + ops.size() - 1];
-                                out_nodes[1] = i + ops.size() - 1;
-
-                                if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) &&
-                                    ggml_cuda_should_use_topk_moe(node, logits, weights, ids) &&
-                                    ggml_cuda_check_fusion_memory_ranges(cgraph, i, ops.size(), out_nodes, 2, /*is_topk_moe=*/ true)) {
-                                    ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args);
-                                    i += ops.size() - 1;
-                                    continue;
-                                }
-                            } else if (!args.norm && !args.prob_bias) {
-                                //special case gpt-oss, no norm, no bias.
-                                ops.insert(ops.end(), { GGML_OP_ARGSORT, GGML_OP_VIEW, GGML_OP_GET_ROWS,
-                                                        GGML_OP_RESHAPE, GGML_OP_SOFT_MAX, GGML_OP_RESHAPE });
-                                weights                     = cgraph->nodes[i + 5];
-                                ids                         = cgraph->nodes[i + 1];
-                                const ggml_tensor * softmax = cgraph->nodes[i + 4];
-
-                                int out_nodes[2] = { i + 1, i + 5 };
-                                if (ggml_can_fuse_subgraph(cgraph, i, ops.size(), ops.data(), out_nodes, 2) &&
-                                    ggml_cuda_should_use_topk_moe(softmax, logits, weights, ids) &&
-                                    ggml_cuda_check_fusion_memory_ranges(cgraph, i, ops.size(), out_nodes, 2, /*is_topk_moe=*/ true)) {
-                                    ggml_cuda_op_topk_moe(*cuda_ctx, logits, weights, ids, clamp, scale, bias, args);
-                                    i += ops.size() - 1;
-                                    continue;
-                                }
-                            }
-                        }
-                    }
-
-                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }, {})) {
-                        ggml_tensor * rope = cgraph->nodes[i];
-                        ggml_tensor * set_rows = cgraph->nodes[i + 2];
-
-                        ggml_cuda_op_rope_fused(*cuda_ctx, rope, set_rows);
-                        i += 2;
-                        continue;
-                    }
-
-                    if (node->op == GGML_OP_ADD || node->op == GGML_OP_MUL) {
-                        int n_fuse = 0;
-                        ggml_op ops[8];
-                        std::fill(ops, ops + 8, node->op);
-
-                        for (; n_fuse <= 6; ++n_fuse){
-                            if (!ggml_can_fuse(cgraph, i + n_fuse, ops + n_fuse, 2)) {
-                                break;
-                            }
-                            if (cgraph->nodes[i + n_fuse] != cgraph->nodes[i + n_fuse + 1]->src[0]) {
-                                break;
-                            }
-                            if (!ggml_are_same_layout(cgraph->nodes[i + n_fuse]->src[1], cgraph->nodes[i + n_fuse + 1]->src[1])) {
-                                break;
-                            }
-                        }
-
-                        n_fuse++;
-
-                        if (n_fuse > 1) {
-                            ggml_tensor fused_node;
-                            memcpy(&fused_node, node, sizeof(ggml_tensor));
-                            for (int j = 0; j < n_fuse - 1; ++j) {
-                                fused_node.src[j + 2] = cgraph->nodes[i + j + 1]->src[1];
-                            }
-                            fused_node.data = cgraph->nodes[i + n_fuse - 1]->data;
-                            if (node->op == GGML_OP_ADD) {
-                                ggml_cuda_op_fused_add(*cuda_ctx, &fused_node, n_fuse);
-                            } else {
-                                ggml_cuda_op_fused_mul(*cuda_ctx, &fused_node, n_fuse);
-                            }
-                            i += n_fuse - 1;
-
-                            continue;
-                        }
-                    }
-
-                    bool fused_mul_mat_vec = false;
-                    int fused_node_count = 0;
-
-                    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
-                        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
-
-                        if (ggml_cuda_can_fuse(cgraph, i, { op, bias_op, op, bias_op, GGML_OP_GLU }, {})) {
-                            ggml_tensor * glu         = cgraph->nodes[i + 4];
-                            ggml_tensor * gate_bias_n = glu->src[0];
-                            ggml_tensor * up_bias_n   = glu->src[1];
-
-                            //we don't assume the order for {gate, up}. Instead infer it from the bias tensor
-                            ggml_tensor * gate_n      = nullptr;
-                            ggml_tensor * up_n        = nullptr;
-
-                            if (gate_bias_n->src[0] == cgraph->nodes[i] || gate_bias_n->src[1] == cgraph->nodes[i]) {
-                                gate_n = cgraph->nodes[i];
-                                up_n   = cgraph->nodes[i + 2];
-                            } else if (gate_bias_n->src[0] == cgraph->nodes[i + 2] || gate_bias_n->src[1] == cgraph->nodes[i + 2]) {
-                                gate_n = cgraph->nodes[i + 2];
-                                up_n   = cgraph->nodes[i];
-                            } else {
-                                continue;
-                            }
-
-                            auto get_bias_tensor = [](const ggml_tensor * bias_node, const ggml_tensor * mul_node, ggml_op op_bias) {
-                                if (op_bias == GGML_OP_ADD) {
-                                    if (bias_node->src[0] == mul_node) {
-                                        return bias_node->src[1];
-                                    }
-                                    if (bias_node->src[1] == mul_node) {
-                                        return bias_node->src[0];
-                                    }
-                                    return (ggml_tensor *) nullptr;
-                                }
-                                GGML_ASSERT(op_bias == GGML_OP_ADD_ID);
-                                GGML_ASSERT(bias_node->src[0] == mul_node);
-                                return bias_node->src[1];
-                            };
-
-                            ggml_tensor * up_bias_tensor   = get_bias_tensor(up_bias_n, up_n, bias_op);
-                            ggml_tensor * gate_bias_tensor = get_bias_tensor(gate_bias_n, gate_n, bias_op);
-
-                            if (!up_bias_tensor || !gate_bias_tensor) {
-                                continue;
-                            }
-
-                            // we don't support repeating adds
-                            if (bias_op == GGML_OP_ADD &&
-                                (!ggml_are_same_shape(gate_bias_n->src[0], gate_bias_n->src[1]) ||
-                                 !ggml_are_same_shape(up_bias_n->src[0], up_bias_n->src[1]))) {
-                                continue;
-                            }
-
-                            const ggml_tensor * src0 = up_n->src[0];
-                            const ggml_tensor * src1 = up_n->src[1];
-                            const ggml_tensor * ids  = up_n->src[2];
-
-                            if (ggml_cuda_should_fuse_mul_mat_vec_f(up_n)) {
-                                ggml_cuda_mm_fusion_args_host fusion_data{};
-                                fusion_data.gate      = gate_n->src[0];
-                                fusion_data.x_bias    = up_bias_tensor;
-                                fusion_data.gate_bias = gate_bias_tensor;
-                                fusion_data.glu_op    = ggml_get_glu_op(glu);
-
-                                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
-                                fused_mul_mat_vec = true;
-                                fused_node_count = 5;
-                                break;
-                            }
-
-                            if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
-                                ggml_cuda_mm_fusion_args_host fusion_data{};
-                                fusion_data.gate      = gate_n->src[0];
-                                fusion_data.x_bias    = up_bias_tensor;
-                                fusion_data.gate_bias = gate_bias_tensor;
-                                fusion_data.glu_op    = ggml_get_glu_op(glu);
-
-                                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
-                                fused_mul_mat_vec = true;
-                                fused_node_count = 5;
-                                break;
-                            }
-                        } else if (ggml_cuda_can_fuse(cgraph, i, { op, op, GGML_OP_GLU }, {})) {
-                            ggml_tensor * glu  = cgraph->nodes[i + 2];
-                            ggml_tensor * gate = glu->src[0];
-                            ggml_tensor * up   = glu->src[1];
-
-                            bool ok = (gate == cgraph->nodes[i] && up == cgraph->nodes[i + 1])
-                                || (gate == cgraph->nodes[i + 1] && up == cgraph->nodes[i]);
-
-                            if (!ok) continue;
-
-                            const ggml_tensor * src0 = up->src[0];
-                            const ggml_tensor * src1 = up->src[1];
-                            const ggml_tensor * ids  = up->src[2];
-
-                            if (ggml_cuda_should_fuse_mul_mat_vec_f(up)) {
-                                ggml_cuda_mm_fusion_args_host fusion_data{};
-                                fusion_data.gate   = gate->src[0];
-                                fusion_data.glu_op = ggml_get_glu_op(glu);
-
-                                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
-                                fused_mul_mat_vec = true;
-                                fused_node_count = 3;
-                                break;
-                            }
-
-                            if (ggml_cuda_should_fuse_mul_mat_vec_q(up)) {
-                                ggml_cuda_mm_fusion_args_host fusion_data{};
-                                fusion_data.gate   = gate->src[0];
-                                fusion_data.glu_op = ggml_get_glu_op(glu);
-
-                                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
-                                fused_mul_mat_vec = true;
-                                fused_node_count = 3;
-                                break;
-                            }
-                        }
-                    }
-
-                    if (fused_mul_mat_vec) {
-                        i += fused_node_count - 1;
-                        continue;
-                    }
-
-                    fused_mul_mat_vec = false;
-                    fused_node_count = 0;
-
-                    for (ggml_op op : { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID }) {
-                        const ggml_op bias_op = op == GGML_OP_MUL_MAT ? GGML_OP_ADD : GGML_OP_ADD_ID;
-
-                        if (!ggml_can_fuse(cgraph, i, { op, bias_op })) {
-                            continue;
-                        }
-
-                        ggml_tensor * mm_node   = cgraph->nodes[i];
-                        ggml_tensor * bias_node = cgraph->nodes[i + 1];
-
-                        ggml_tensor * bias_tensor = nullptr;
-                        if (bias_op == GGML_OP_ADD) {
-                            if (bias_node->src[0] == mm_node) {
-                                bias_tensor = bias_node->src[1];
-                            } else if (bias_node->src[1] == mm_node) {
-                                bias_tensor = bias_node->src[0];
-                            } else {
-                                continue;
-                            }
-                        } else {
-                            if (bias_node->src[0] != mm_node) {
-                                continue;
-                            }
-                            bias_tensor = bias_node->src[1];
-                        }
-
-                        const ggml_tensor * src0 = mm_node->src[0];
-                        const ggml_tensor * src1 = mm_node->src[1];
-                        const ggml_tensor * ids  = mm_node->src[2];
-
-                        if (bias_op == GGML_OP_ADD_ID && bias_node->src[2] != ids) {
-                            continue;
-                        }
-
-                        if (bias_op == GGML_OP_ADD && !ggml_are_same_shape(bias_node->src[0], bias_node->src[1])) {
-                            continue;
-                        }
-
-                        ggml_cuda_mm_fusion_args_host fusion_data{};
-                        fusion_data.x_bias = bias_tensor;
-
-                        if (ggml_cuda_should_fuse_mul_mat_vec_f(mm_node)) {
-                            ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
-                            fused_mul_mat_vec = true;
-                            fused_node_count = 2;
-                            break;
-                        }
-
-                        if (ggml_cuda_should_fuse_mul_mat_vec_q(mm_node)) {
-                            ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, bias_node, &fusion_data);
-                            fused_mul_mat_vec = true;
-                            fused_node_count = 2;
-                            break;
-                        }
-                    }
-
-                    if (fused_mul_mat_vec) {
-                        i += fused_node_count - 1;
-                        continue;
-                    }
-
-                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL, GGML_OP_ADD}, {})) {
-                        ggml_cuda_op_rms_norm_fused_add(*cuda_ctx, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
-                        i += 2;
-                        continue;
-                    }
-
-                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL}, {})) {
-                        ggml_cuda_op_rms_norm_fused(*cuda_ctx, node, cgraph->nodes[i+1]);
-                        i++;
-                        continue;
-                    }
-
-                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SSM_CONV, GGML_OP_UNARY }, { GGML_UNARY_OP_SILU })) {
-                        ggml_cuda_op_ssm_conv(*cuda_ctx, node, cgraph->nodes[i+1]);
-                        i++;
-                        continue;
-                    }
-
-                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SILU }) ||
-                        ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SIGMOID }) ||
-                        ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_MUL }, { GGML_UNARY_OP_SOFTPLUS })) {
-                        ggml_cuda_op_unary_mul(*cuda_ctx, node, cgraph->nodes[i+1]);
-                        i++;
-                        continue;
-                    }
-
-                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SCALE, GGML_OP_UNARY, GGML_OP_SCALE }, { GGML_UNARY_OP_TANH })) {
-                        i += 2;
-                        ggml_cuda_op_softcap(*cuda_ctx, cgraph->nodes[i], node);
-                        continue;
-                    }
+                if (nodes_to_skip != 0) {
+                    i += nodes_to_skip;
+                    continue;
                 }
 #ifndef NDEBUG
                 assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
@@ -4488,8 +4588,8 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
     /* .free                    = */ ggml_backend_cuda_free,
     /* .set_tensor_async        = */ ggml_backend_cuda_set_tensor_async,
     /* .get_tensor_async        = */ ggml_backend_cuda_get_tensor_async,
-    /* .get_tensor_2d_async     = */ ggml_backend_cuda_set_tensor_2d_async,
-    /* .set_tensor_2d_async     = */ ggml_backend_cuda_get_tensor_2d_async,
+    /* .set_tensor_2d_async     = */ ggml_backend_cuda_set_tensor_2d_async,
+    /* .get_tensor_2d_async     = */ ggml_backend_cuda_get_tensor_2d_async,
     /* .cpy_tensor_async        = */ ggml_backend_cuda_cpy_tensor_async,
     /* .synchronize             = */ ggml_backend_cuda_synchronize,
     /* .graph_plan_create       = */ NULL,

ggml/src/ggml-cuda/mma.cuh

@@ -86,17 +86,12 @@ namespace ggml_cuda_mma {
     //   - (I_MAJOR, I_MAJOR_MIRRORED) -> I_MAJOR
     //   - (I_MAJOR, J_MAJOR_MIRRORED) -> I_MAJOR
 
-    static constexpr bool is_i_major(const data_layout dl) {
-        return dl == DATA_LAYOUT_I_MAJOR ||
-               dl == DATA_LAYOUT_I_MAJOR_MIRRORED;
-    }
-
     static constexpr __device__ data_layout get_input_data_layout() {
-#if defined(RDNA3) || __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
         return DATA_LAYOUT_I_MAJOR_MIRRORED;
 #else
         return DATA_LAYOUT_I_MAJOR;
-#endif // defined(RDNA3) || __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <int I_, int J_, typename T, data_layout ds_=DATA_LAYOUT_I_MAJOR>
@@ -113,7 +108,6 @@ namespace ggml_cuda_mma {
         T x[ne] = {0};
 
         static constexpr __device__ bool supported() {
-            if (I == 64 && J ==  2) return true;
             if (I == 16 && J ==  8) return true;
             if (I == 32 && J ==  4) return true;
             if (I == 16 && J == 16) return true;
@@ -122,7 +116,7 @@ namespace ggml_cuda_mma {
         }
 
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
+            if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 16;
             } else if constexpr (I == 16 && J == 8) {
                 return threadIdx.x % 16;
@@ -139,8 +133,8 @@ namespace ggml_cuda_mma {
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
-            if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
-                return (2 * ((threadIdx.x / 16) % 2) + l);
+            if constexpr (I == 16 && J == 4) {
+                return threadIdx.x / 16;
             } else if constexpr (I == 16 && J == 8) {
                 return 2 * (threadIdx.x / 16) + l;
             } else if constexpr (I == 32 && J == 4) {
@@ -154,7 +148,7 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#elif defined(VOLTA_MMA_AVAILABLE)
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
 
@@ -283,7 +277,7 @@ namespace ggml_cuda_mma {
         static constexpr int         J  = J_;
         static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
 
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
@@ -407,7 +401,7 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     };
 
     template <int I_, int J_>
@@ -701,57 +695,12 @@ namespace ggml_cuda_mma {
     }
 #endif // defined(TURING_MMA_AVAILABLE)
 
-    static __device__ __forceinline__ void make_identity_mat(tile<16, 8, half2> & t) {
-#if defined(RDNA4)
-        const int row = t.get_i(0);
-        const int left_right = t.get_j(0) / 4;
-        const int up_down = row / 8;
-        const int idx = row % 8;
-        reinterpret_cast<half*>(t.x)[idx] = left_right == up_down ? 1.0f : 0.0f;
-#else
-        GGML_UNUSED_VARS(t);
-        NO_DEVICE_CODE;
-#endif // defined(RDNA4)
-    }
-
     template <int I, int J, typename T, data_layout dl>
     static __device__ __forceinline__ void load_generic(tile<I, J, T, dl> & t, const T * __restrict__ xs0, const int stride) {
-#if defined(AMD_MFMA_AVAILABLE)
-        if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
-#pragma unroll
-            for (int l = 0; l < t.ne; ++l) {
-                t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
-            }
-        } else {
-            ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
-        }
-#elif defined(AMD_WMMA_AVAILABLE)
-        // All wmma layout has contiguous data when i-major.
-        if constexpr (is_i_major(dl)) {
-            // the data must be aligned to 16 bytes when bigger than ggml_cuda_get_max_cpy_bytes()
-            constexpr int aligned_copy_bytes = ggml_cuda_get_max_cpy_bytes();
-            if constexpr (sizeof(t.x) > aligned_copy_bytes) {
-                static_assert(sizeof(t.x) % aligned_copy_bytes == 0, "bad type size");
-                constexpr int aligned_copy_count = sizeof(t.x)/aligned_copy_bytes;
-#pragma unroll
-                for (int i = 0; i < aligned_copy_count; ++i) {
-                    ggml_cuda_memcpy_1<aligned_copy_bytes>(t.x + t.ne/aligned_copy_count*i, xs0 + t.get_i(0) * stride + t.get_j(t.ne/aligned_copy_count*i));
-                }
-            } else {
-                ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
-            }
-        } else {
-#pragma unroll
-            for (int l = 0; l < t.ne; ++l) {
-                t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
-            }
-        }
-#else
 #pragma unroll
         for (int l = 0; l < t.ne; ++l) {
             t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
         }
-#endif // defined(AMD_MFMA_AVAILABLE)
     }
 
     template <typename T>
@@ -764,26 +713,37 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
 #else
-        load_generic(t, xs0, stride);
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
 
-    template <typename T>
+    template <typename T, data_layout dl>
     static __device__ __forceinline__ void load_ldmatrix(
-            tile<16, 4, T> & t, const T * __restrict__ xs0, const int stride) {
+            tile<16, 4, T, dl> & t, const T * __restrict__ xs0, const int stride) {
 #ifdef TURING_MMA_AVAILABLE
         int * xi = (int *) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride;
         asm volatile("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];"
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
+#elif defined(AMD_WMMA_AVAILABLE)
+#ifdef RDNA3
+        static_assert(dl == DATA_LAYOUT_I_MAJOR_MIRRORED, "bad data layout");
+        static_assert(sizeof(t.x) == 16, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<8>(t.x + 2, xs0 + t.get_i(0)*stride + 2);
+#else
+        static_assert(dl == DATA_LAYOUT_I_MAJOR, "bad data layout");
+        static_assert(sizeof(t.x) == 8, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
+#endif // RDNA3
+#elif defined(AMD_MFMA_AVAILABLE)
+        static_assert(sizeof(t.x) == 4, "bad ne");
+        ggml_cuda_memcpy_1<4>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
 #else
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
-#else
-        load_generic(t, xs0, stride);
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -796,19 +756,26 @@ namespace ggml_cuda_mma {
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
             : "l"(xs));
-#else
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
-#if 1
-        // TODO: more generic handling
-        static_assert(sizeof(T) == 4, "bad type size");
+#elif defined(VOLTA_MMA_AVAILABLE)
         ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
         ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 4, xs0 + t.get_i(4)*stride + 4);
+#elif defined(AMD_WMMA_AVAILABLE)
+#ifdef RDNA3
+        static_assert(dl == DATA_LAYOUT_I_MAJOR_MIRRORED, "bad data layout");
+        static_assert(sizeof(t.x) == 32, "bad ne");
+        ggml_cuda_memcpy_1<16>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<16>(t.x + 4, xs0 + t.get_i(0)*stride + 4);
 #else
-        load_generic(t, xs0, stride);
-#endif // 1
+        static_assert(dl == DATA_LAYOUT_I_MAJOR, "bad data layout");
+        static_assert(sizeof(t.x) == 16, "bad ne");
+        ggml_cuda_memcpy_1<16>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
+#endif // RDNA3
+#elif defined(AMD_MFMA_AVAILABLE)
+        static_assert(sizeof(t.x) == 8, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
 #else
-        load_generic(t, xs0, stride);
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -827,23 +794,30 @@ namespace ggml_cuda_mma {
 
     static __device__ __forceinline__ void load_ldmatrix(
             tile<32, 4, half2> & t, const half2 * __restrict__ xs0, const int stride) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         ggml_cuda_memcpy_1<4*sizeof(half2)>(t.x, xs0 + t.get_i(0)*stride);
 #else
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix_trans(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
 #ifdef TURING_MMA_AVAILABLE
-        int * xi = (int * ) t.x;
+        int * xi = (int *) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + (threadIdx.x / t.I) * (t.J / 2);
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.trans.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[2]), "=r"(xi[1]), "=r"(xi[3])
             : "l"(xs));
+#elif defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        half * xh = (half *) t.x;
+#pragma unroll
+        for (int l = 0; l < t.ne; ++l) {
+            xh[2*l + 0] = ((const half *) xs0)[(2*t.get_j(l) + 0)*(2*stride) + t.get_i(l)];
+            xh[2*l + 1] = ((const half *) xs0)[(2*t.get_j(l) + 1)*(2*stride) + t.get_i(l)];
+        }
 #else
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
@@ -1041,25 +1015,35 @@ namespace ggml_cuda_mma {
 #endif // AMD_MFMA_AVAILABLE
     }
 
-    static __device__ __forceinline__ void mma_block_scaled(tile<16, 8, float> &     D,
-                                                            const tile<16, 8, int> & A,
-                                                            const tile<8, 8, int> &  B,
-                                                            uint32_t                 a_scale,
-                                                            uint32_t                 b_scale) {
+    template <ggml_type type>
+    static __device__ __forceinline__ void mma_block_scaled_fp4(tile<16, 8, float> &     D,
+                                                                const tile<16, 8, int> & A,
+                                                                const tile<8, 8, int> &  B,
+                                                                uint32_t                 a_scale,
+                                                                uint32_t                 b_scale) {
 #ifdef BLACKWELL_MMA_AVAILABLE
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         float *     Dxi = (float *) D.x;
 
-        asm volatile(
-            "mma.sync.aligned.kind::mxf4.block_scale.scale_vec::2X.m16n8k64.row.col.f32.e2m1.e2m1.f32.ue8m0 "
-            "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3}, "
-            "%10, {0, 0}, %11, {0, 0};"
-            : "+f"(Dxi[0]), "+f"(Dxi[1]), "+f"(Dxi[2]), "+f"(Dxi[3])
-            : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1]), "r"(a_scale), "r"(b_scale));
+        if constexpr (type == GGML_TYPE_MXFP4) {
+            asm volatile(
+                "mma.sync.aligned.kind::mxf4.block_scale.scale_vec::2X.m16n8k64.row.col.f32.e2m1.e2m1.f32.ue8m0 "
+                "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3}, "
+                "%10, {0, 0}, %11, {0, 0};"
+                : "+f"(Dxi[0]), "+f"(Dxi[1]), "+f"(Dxi[2]), "+f"(Dxi[3])
+                : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1]), "r"(a_scale), "r"(b_scale));
+        } else {
+            asm volatile(
+                "mma.sync.aligned.kind::mxf4nvf4.block_scale.scale_vec::4X.m16n8k64.row.col.f32.e2m1.e2m1.f32.ue4m3 "
+                "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3}, "
+                "%10, {0, 0}, %11, {0, 0};"
+                : "+f"(Dxi[0]), "+f"(Dxi[1]), "+f"(Dxi[2]), "+f"(Dxi[3])
+                : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1]), "r"(a_scale), "r"(b_scale));
+        }
 #else
         GGML_UNUSED_VARS(D, A, B, a_scale, b_scale);
-#endif  // BLACKWELL_MMA_AVAILABLE
+#endif // BLACKWELL_MMA_AVAILABLE
     }
 
     static __device__ __forceinline__ void mma(
@@ -1218,73 +1202,27 @@ namespace ggml_cuda_mma {
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * acc = (int32x4_t *) D.x;
 #if defined(CDNA4) || defined(CDNA3)
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(((int64_t *) A.x)[0],
-                                                       ((int64_t *) B.x)[0],
-                                                       acc[0],
-                                                       0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(((int64_t *) A.x)[0], ((int64_t *) B.x)[0], acc[0], 0, 0, 0);
 #elif defined(CDNA2) || defined(CDNA1)
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0],
-                                                      B.x[0],
-                                                      acc[0],
-                                                      0, 0, 0);
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[1],
-                                                      B.x[1],
-                                                      acc[0],
-                                                      0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[1], B.x[1], acc[0], 0, 0, 0);
 #endif // defined(CDNA4) || defined(CDNA3)
-
 #elif defined(AMD_WMMA_AVAILABLE)
-
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
-
 #if defined(RDNA4)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            true
-        );
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[1],
-            true,
-            b_vec[1],
-            acc[0],
-            true
-        );
-
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], true);
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[1], true, b_vec[1], acc[0], true);
 #elif defined(RDNA3)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * a_vec = (int32x4_t *) A.x;
         int32x4_t * b_vec = (int32x4_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            true
-        );
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[1],
-            true,
-            b_vec[1],
-            acc[0],
-            true
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], true);
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[1], true, b_vec[1], acc[0], true);
 #endif // RDNA4
-
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -1297,19 +1235,10 @@ namespace ggml_cuda_mma {
         using int32x16_t = __attribute__((__vector_size__(16 * sizeof(int)))) int;
         int32x16_t * acc = (int32x16_t *) D.x;
 #if defined(CDNA4) || defined(CDNA3)
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x16_i8(((int64_t *) A.x)[0],
-                                                       ((int64_t *) B.x)[0],
-                                                       acc[0],
-                                                       0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x16_i8(((int64_t *) A.x)[0], ((int64_t *) B.x)[0], acc[0], 0, 0, 0);
 #elif defined(CDNA2) || defined(CDNA1)
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[0],
-                                                     B.x[0],
-                                                     acc[0],
-                                                     0, 0, 0);
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[1],
-                                                     B.x[1],
-                                                     acc[0],
-                                                     0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[1], B.x[1], acc[0], 0, 0, 0);
 #endif // defined(CDNA4) || defined(CDNA3)
 
 #else
@@ -1329,7 +1258,7 @@ namespace ggml_cuda_mma {
 
     static __device__ __forceinline__ void mma(
             tile<32, 8, float> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & B) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -1344,12 +1273,12 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     static __device__ __forceinline__ void mma(
             tile<32, 4, half2> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_J_MAJOR_MIRRORED> & B) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -1364,41 +1293,35 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <data_layout dl_d, data_layout dl_ab>
     static __device__ __forceinline__ void mma(
             tile<16, 16, int, dl_d> & D, const tile<16, 4, int, dl_ab> & A, const tile<16, 4, int, dl_ab> & B) {
-#if defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE)
+        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
+        int32x4_t * acc = (int32x4_t *) D.x;
+#if defined(CDNA4) || defined(CDNA3)
+        const int64_t xA = uint32_t(A.x[0]);
+        const int64_t xB = uint32_t(B.x[0]);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(xA, xB, acc[0], 0, 0, 0);
+#elif defined(CDNA2) || defined(CDNA1)
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+#endif // defined(CDNA4) || defined(CDNA3)
+#elif defined(AMD_WMMA_AVAILABLE)
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
 #if defined(RDNA4)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            false
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], false);
 #elif defined(RDNA3)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * a_vec = (int32x4_t *) A.x;
         int32x4_t * b_vec = (int32x4_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            false
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], false);
 #endif // RDNA4
 #else
         GGML_UNUSED(D);

ggml/src/ggml-cuda/mmq.cu

@@ -122,7 +122,7 @@ void ggml_cuda_mul_mat_q(
                             || GGML_CUDA_CC_IS_CDNA(cc);
 
     // TODO: tighter pool buffer size vs q8 path
-    const bool use_native_mxfp4 = blackwell_mma_available(cc) && src0->type == GGML_TYPE_MXFP4;
+    const bool use_native_fp4 = blackwell_mma_available(cc) && (src0->type == GGML_TYPE_MXFP4 || src0->type == GGML_TYPE_NVFP4);
 
     if (!ids) {
         const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 +
@@ -133,9 +133,9 @@ void ggml_cuda_mul_mat_q(
             const int64_t s11 = src1->nb[1] / ts_src1;
             const int64_t s12 = src1->nb[2] / ts_src1;
             const int64_t s13 = src1->nb[3] / ts_src1;
-            if (use_native_mxfp4) {
+            if (use_native_fp4) {
                 static_assert(sizeof(block_fp4_mmq) == 4 * sizeof(block_q8_1));
-                quantize_mmq_mxfp4_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded,
+                quantize_mmq_fp4_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded,
                                         ne11, ne12, ne13, stream);
 
             } else {
@@ -146,10 +146,8 @@ void ggml_cuda_mul_mat_q(
         }
 
         // Stride depends on quantization format
-        const int64_t s12 = use_native_mxfp4 ?
-                                ne11 * ne10_padded * sizeof(block_fp4_mmq) /
-                                    (8 * QK_MXFP4 * sizeof(int))  // block_fp4_mmq holds 256 values (8 blocks of 32)
-                                :
+        const int64_t s12 = use_native_fp4 ?
+                                ne11 * ne10_padded * sizeof(block_fp4_mmq) / (QK_K * sizeof(int)) :  // block_fp4_mmq holds 256 values
                                 ne11 * ne10_padded * sizeof(block_q8_1) / (QK8_1 * sizeof(int));
         const int64_t s13 = ne12*s12;
 
@@ -198,8 +196,8 @@ void ggml_cuda_mul_mat_q(
         const int64_t s12 = src1->nb[2] / ts_src1;
         const int64_t s13 = src1->nb[3] / ts_src1;
 
-        if (use_native_mxfp4) {
-            quantize_mmq_mxfp4_cuda(src1_d, ids_src1.get(), src1_q8_1.get(), src0->type, ne10, s11, s12, s13,
+        if (use_native_fp4) {
+            quantize_mmq_fp4_cuda(src1_d, ids_src1.get(), src1_q8_1.get(), src0->type, ne10, s11, s12, s13,
                                     ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
         } else {
             quantize_mmq_q8_1_cuda(src1_d, ids_src1.get(), src1_q8_1.get(), src0->type, ne10, s11, s12, s13,
@@ -208,8 +206,9 @@ void ggml_cuda_mul_mat_q(
         CUDA_CHECK(cudaGetLastError());
     }
 
-    const int64_t s12 = use_native_mxfp4 ? ne11 * ne10_padded * sizeof(block_fp4_mmq) / (8 * QK_MXFP4 * sizeof(int)) :
-                                           ne11 * ne10_padded * sizeof(block_q8_1) / (QK8_1 * sizeof(int));
+    static_assert(QK_K == 8 * QK_MXFP4, "QK_K needs to be 8 * QK_MXFP4");
+    const int64_t s12 = use_native_fp4 ? ne11 * ne10_padded * sizeof(block_fp4_mmq) / (QK_K * sizeof(int)) :
+                                         ne11 * ne10_padded * sizeof(block_q8_1) / (QK8_1 * sizeof(int));
     const int64_t s13 = ne12*s12;
 
     // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.

ggml/src/ggml-cuda/mmvq.cu

@@ -115,6 +115,7 @@ static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_pascal_older(gg
         case GGML_TYPE_IQ4_NL:  return 6;
         case GGML_TYPE_IQ4_XS:  return 5;
         case GGML_TYPE_MXFP4:   return 4;
+        case GGML_TYPE_NVFP4:   return 4;
         case GGML_TYPE_Q2_K:    return 4;
         case GGML_TYPE_Q3_K:    return 4;
         case GGML_TYPE_Q4_0:    return 6;
@@ -135,6 +136,7 @@ static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_turing_plus(ggm
         case GGML_TYPE_IQ3_S:   return 6;
         case GGML_TYPE_IQ3_XXS: return 7;
         case GGML_TYPE_MXFP4:   return 7;
+        case GGML_TYPE_NVFP4:   return 8;
         case GGML_TYPE_Q2_K:    return 7;
         case GGML_TYPE_Q3_K:    return 5;
         default:                return MMVQ_MAX_BATCH_SIZE;
@@ -221,6 +223,7 @@ static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna4(ggml_type
         case GGML_TYPE_IQ4_NL:  return 7;
         case GGML_TYPE_IQ4_XS:  return 5;
         case GGML_TYPE_MXFP4:   return 5;
+        case GGML_TYPE_NVFP4:   return 5;
         case GGML_TYPE_Q3_K:    return 4;
         case GGML_TYPE_Q4_0:    return 7;
         case GGML_TYPE_Q4_1:    return 7;

ggml/src/ggml-cuda/quantize.cu

@@ -70,6 +70,102 @@ __device__ __forceinline__ uint8_t compute_e8m0_scale(float amax) {
     return static_cast<uint8_t>(biased);
 }
 
+
+static __global__ void quantize_mmq_nvfp4(
+        const float * __restrict__ x, const int32_t * __restrict__ ids, void * __restrict__ vy,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2) {
+#if defined(BLACKWELL_MMA_AVAILABLE)
+
+    const int64_t i0_base = ((int64_t) blockDim.x * blockIdx.y + threadIdx.x) * QK_NVFP4_SUB;
+    if (i0_base >= ne0) {
+        return;
+    }
+
+    const int64_t i1 = blockIdx.x;
+    const int64_t i2 = blockIdx.z % ne2;
+    const int64_t i3 = blockIdx.z / ne2;
+    const int64_t i01 = ids ? ids[i1] : i1;
+    const int64_t k_block = i0_base / QK_K;
+    const int64_t blocks_per_col = (ne0 + QK_K - 1) / QK_K;
+    if (k_block >= blocks_per_col) {
+        return;
+    }
+
+    const int64_t ib = blockIdx.z * ((int64_t) blocks_per_col * ne1) + k_block * ne1 + blockIdx.x;
+    block_fp4_mmq * y = (block_fp4_mmq *) vy;
+    block_fp4_mmq * yb = y + ib;
+
+    const int sub = (i0_base % QK_K) / QK_NVFP4_SUB;
+
+    float vals_raw[QK_NVFP4_SUB];
+    float amax_raw = 0.0f;
+    const int64_t base_idx = i3 * s03 + i2 * s02 + i01 * s01;
+#pragma unroll
+    for (int k = 0; k < QK_NVFP4_SUB; k++) {
+        const int64_t i00 = i0_base + k;
+        if (i00 < ne00) {
+            const float v = x[base_idx + i00];
+            vals_raw[k] = v;
+            amax_raw = fmaxf(amax_raw, fabsf(v));
+        } else {
+            vals_raw[k] = 0.0f;
+        }
+    }
+
+    static constexpr int test_offsets[5] = { 0, -1, 1, -2, 2};
+    const int first_fp8_code = (int) ggml_cuda_fp32_to_ue4m3(amax_raw / 6.0f);
+
+    float best_err = FLT_MAX;
+    uint8_t fp8_code = 0;
+    float subblock_scale = 0.0f;
+
+#pragma unroll // Check +/- 2 to find best code to reduce NVFP4 activation loss. Negligible overhead on Blackwell.
+    for (int i = 0; i < 5; i++) {
+        const int test_code = first_fp8_code + test_offsets[i];
+        if (test_code < 0 || test_code > 0x7e) {
+            continue;
+        }
+        const uint8_t code = (uint8_t) test_code;
+        const float test_scale = ggml_cuda_ue4m3_to_fp32(code);
+        const float test_inv_scale = test_scale > 0.0f ? 0.5f / test_scale : 0.0f;
+        float cur_err = 0.0f;
+#pragma unroll
+        for (int k = 0; k < QK_NVFP4_SUB; ++k) {
+            const float v = vals_raw[k];
+            const uint8_t q = ggml_cuda_float_to_fp4_e2m1(v, test_inv_scale);
+            const float err_diff = fabsf(v) - fabsf(kvalues_mxfp4[q & 0x7]) * test_scale;
+            cur_err = fmaf(err_diff, err_diff, cur_err);
+        }
+
+        if (cur_err < best_err) {
+            best_err = cur_err;
+            fp8_code = test_code;
+            subblock_scale = test_scale;
+        }
+    }
+
+    const float inv_scale = subblock_scale > 0.0f ? 0.5f / subblock_scale : 0.0f;
+    uint32_t q0 = 0;
+    uint32_t q1 = 0;
+#pragma unroll // this is faster than the previous __nv_fp4x4_e2m1
+    for (int k = 0; k < QK_NVFP4_SUB / 4; ++k) {
+        q0 |= (uint32_t) ggml_cuda_float_to_fp4_e2m1(vals_raw[k +  0], inv_scale) << (8 * k);
+        q0 |= (uint32_t) ggml_cuda_float_to_fp4_e2m1(vals_raw[k +  8], inv_scale) << (8 * k + 4);
+        q1 |= (uint32_t) ggml_cuda_float_to_fp4_e2m1(vals_raw[k +  4], inv_scale) << (8 * k);
+        q1 |= (uint32_t) ggml_cuda_float_to_fp4_e2m1(vals_raw[k + 12], inv_scale) << (8 * k + 4);
+    }
+
+    uint32_t * yqs = reinterpret_cast<uint32_t *>(yb->qs);
+    yqs[2 * sub + 0] = q0;
+    yqs[2 * sub + 1] = q1;
+    reinterpret_cast<uint8_t *>(yb->d4)[sub] = fp8_code;
+#else
+    NO_DEVICE_CODE; // This is for Blackwell NVFP4 activations only.
+#endif // defined(BLACKWELL_MMA_AVAILABLE)
+
+}
+
 // quantize values in the format mxfp4 is stored which is interleaved nibbles
 // i.e. a block a0-a31 is represented as a0a16,a1a17 ...a15a31
 static __global__ void quantize_mmq_mxfp4(const float * __restrict__ x,
@@ -316,28 +412,32 @@ void quantize_mmq_q8_1_cuda(
     }
 }
 
-void quantize_mmq_mxfp4_cuda(const float *                    x,
-                             const int32_t *                  ids,
-                             void *                           vy,
-                             [[maybe_unused]] const ggml_type type_src0,
-                             const int64_t                    ne00,
-                             const int64_t                    s01,
-                             const int64_t                    s02,
-                             const int64_t                    s03,
-                             const int64_t                    ne0,
-                             const int64_t                    ne1,
-                             const int64_t                    ne2,
-                             const int64_t                    ne3,
-                             cudaStream_t                     stream) {
-    GGML_ASSERT(ne0 % (2 * QK_MXFP4) == 0);
+void quantize_mmq_fp4_cuda(
+        const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
+    GGML_ASSERT(type_src0 == GGML_TYPE_MXFP4 || type_src0 == GGML_TYPE_NVFP4);
+    GGML_ASSERT(ne0 > 0);
 
-    constexpr int nwarps = 8;
-    constexpr int vals_per_warp  = 2 * QK_MXFP4;
-    constexpr int vals_per_block = nwarps * vals_per_warp;
+    if (type_src0 == GGML_TYPE_NVFP4) {
+        GGML_ASSERT(ne00 % QK_NVFP4 == 0);
+        constexpr int nvfp4_block_size = 128;
+        const int64_t block_num_y = (ne0 + QK_NVFP4_SUB * nvfp4_block_size - 1) / (QK_NVFP4_SUB * nvfp4_block_size);
+        const dim3 block_size(nvfp4_block_size, 1, 1);
+        const dim3 num_blocks(ne1, block_num_y, ne2 * ne3);
+        quantize_mmq_nvfp4<<<num_blocks, block_size, 0, stream>>>(
+            x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
+    } else {
+        GGML_ASSERT(ne0 % (2 * QK_MXFP4) == 0);
 
-    const int64_t block_num_y = (ne0 + vals_per_block - 1) / vals_per_block;
-    const dim3    num_blocks(ne1, block_num_y, ne2 * ne3);
-    const dim3    block_size(WARP_SIZE, nwarps, 1);
+        constexpr int nwarps = 8;
+        constexpr int vals_per_warp  = 2 * QK_MXFP4;
+        constexpr int vals_per_block = nwarps * vals_per_warp;
 
-    quantize_mmq_mxfp4<<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
+        const int64_t block_num_y = (ne0 + vals_per_block - 1) / vals_per_block;
+        const dim3    num_blocks(ne1, block_num_y, ne2 * ne3);
+        const dim3    block_size(WARP_SIZE, nwarps, 1);
+
+        quantize_mmq_mxfp4<<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
+    }
 }

ggml/src/ggml-cuda/quantize.cuh

@@ -26,7 +26,7 @@ void quantize_mmq_q8_1_cuda(
         ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
         int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
 
-void quantize_mmq_mxfp4_cuda(const float *   x,
+void quantize_mmq_fp4_cuda(const float *   x,
                              const int32_t * ids,
                              void *          vy,
                              ggml_type       type_src0,

ggml/src/ggml-cuda/ssm-conv.cu

@@ -3,6 +3,7 @@
 
 template <bool apply_silu, size_t split_d_inner, size_t d_conv>
 static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float * __restrict__ src1,
+                                    const float * __restrict__ bias,
                                     const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src1_nb1,
                                     float * __restrict__ dst, const int dst_nb0, const int dst_nb1, const int dst_nb2,
                                     const int64_t n_t) {
@@ -27,6 +28,8 @@ static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float
         w[j] = w_block[tid * stride_w + j];
     }
 
+    float b = bias != nullptr ? bias[bidy * split_d_inner + tid] : 0.0f;
+
     for (int64_t i = 0; i < n_t; i++) {
         float sumf = 0.0f;
 
@@ -42,12 +45,14 @@ static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float
         for (size_t j = 0; j < d_conv; j++) {
             sumf += x[(i + j) % d_conv] * w[j];
         }
+        sumf += b;
         y_block[i * stride_y + tid] = apply_silu ? ggml_cuda_op_silu_single(sumf) : sumf;
     }
 }
 
 template <bool apply_silu, size_t split_d_inner, size_t d_conv, int64_t split_n_t>
 static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0, const float * __restrict__ src1,
+                                               const float * __restrict__ bias,
                                                const int src0_nb0, const int src0_nb1, const int src0_nb2,
                                                const int src1_nb1, float * __restrict__ dst, const int dst_nb0,
                                                const int dst_nb1, const int dst_nb2, const int64_t n_t) {
@@ -97,6 +102,8 @@ static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0,
         w[j] = w_block[tid * stride_w + j];
     }
 
+    float b = bias != nullptr ? bias[bidy * split_d_inner + tid] : 0.0f;
+
     // Compute from shared memory
     for (int64_t i = 0; i < local_n_t; i++) {
         float sumf = 0.0f;
@@ -104,12 +111,13 @@ static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0,
         for (size_t j = 0; j < d_conv; j++) {
             sumf += smem[tid * n_cols + i + j] * w[j];
         }
+        sumf += b;
         y_block[i * stride_y + tid] = apply_silu ? ggml_cuda_op_silu_single(sumf) : sumf;
     }
 }
 
 template <bool apply_silu>
-static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int src0_nb0, const int src0_nb1,
+static void ssm_conv_f32_cuda(const float * src0, const float * src1, const float * bias, const int src0_nb0, const int src0_nb1,
                               const int src0_nb2, const int src1_nb1, float * dst, const int dst_nb0, const int dst_nb1,
                               const int dst_nb2, const int64_t nc, const int64_t nr, const int64_t n_t,
                               const int64_t n_s, cudaStream_t stream) {
@@ -120,14 +128,14 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
         constexpr int kNC = decltype(NC)::value;
         if (n_t <= 32) {
             const dim3 blocks(n_s, (nr + threads - 1) / threads, 1);
-            ssm_conv_f32<apply_silu, threads, kNC><<<blocks, threads, 0, stream>>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
+            ssm_conv_f32<apply_silu, threads, kNC><<<blocks, threads, 0, stream>>>(src0, src1, bias, src0_nb0, src0_nb1, src0_nb2, src1_nb1,
                                                                        dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         } else {
             const int64_t split_n_t = 32;
             dim3          blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t);
             const size_t  smem_size = threads * (kNC - 1 + split_n_t) * sizeof(float);
             ssm_conv_long_token_f32<apply_silu, threads, kNC, split_n_t><<<blocks, threads, smem_size, stream>>>(
-                src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
+                src0, src1, bias, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t);
         }
     };
 
@@ -140,11 +148,18 @@ static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int
     }
 }
 
-void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * silu_dst) {
+void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * bias_add_node, ggml_tensor * silu_dst) {
     const struct ggml_tensor * src0 = dst->src[0];  // conv_x
     const struct ggml_tensor * src1 = dst->src[1];  // conv1d.weight
+    const bool fuse_bias = bias_add_node != nullptr;
     const bool fuse_silu = silu_dst != nullptr;
 
+    // bias always comes with silu.
+    GGML_ASSERT(!fuse_bias || fuse_silu);
+
+    // The bias (when fused) is the non-conv operand of the ADD node.
+    const struct ggml_tensor * bias = fuse_bias ? (bias_add_node->src[0] == dst ? bias_add_node->src[1] : bias_add_node->src[0]) : nullptr;
+
     // When fusing, write to silu_dst (the node downstream references).
     const struct ggml_tensor * out = fuse_silu ? silu_dst : dst;
 
@@ -160,16 +175,23 @@ void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, g
 
     const float * src0_d = (const float *) src0->data;
     const float * src1_d = (const float *) src1->data;
+    const float * bias_d = fuse_bias ? (const float *) bias->data : nullptr;
     float *       dst_d  = (float *) out->data;
     cudaStream_t  stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(out->type == GGML_TYPE_F32);
+    if (fuse_bias) {
+        GGML_ASSERT(bias->type == GGML_TYPE_F32);
+        GGML_ASSERT(ggml_is_contiguous(bias));
+        GGML_ASSERT(ggml_nelements(bias) == nr);
+    }
+
     if (fuse_silu) {
-        ssm_conv_f32_cuda<true>(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, out->nb[0], out->nb[1],
+        ssm_conv_f32_cuda<true>(src0_d, src1_d, bias_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, out->nb[0], out->nb[1],
                           out->nb[2], nc, nr, n_t, n_s, stream);
     } else {
-        ssm_conv_f32_cuda<false>(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, out->nb[0], out->nb[1],
+        ssm_conv_f32_cuda<false>(src0_d, src1_d, bias_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, out->nb[0], out->nb[1],
                           out->nb[2], nc, nr, n_t, n_s, stream);
     }
 }

ggml/src/ggml-cuda/ssm-conv.cuh

@@ -1,3 +1,3 @@
 #include "common.cuh"
 
-void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * silu_dst = nullptr);
+void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst, ggml_tensor * bias_add_node = nullptr, ggml_tensor * silu_dst = nullptr);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_32.cu

@@ -2,4 +2,5 @@
 
 #include "../fattn-mma-f16.cuh"
 
+DECL_FATTN_MMA_F16_CASE(320, 256, 1, 32);
 DECL_FATTN_MMA_F16_CASE(576, 512, 1, 32);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_32.cu

@@ -2,4 +2,5 @@
 
 #include "../fattn-mma-f16.cuh"
 
+DECL_FATTN_MMA_F16_CASE(320, 256, 2, 32);
 DECL_FATTN_MMA_F16_CASE(576, 512, 2, 32);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq320-dv256.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(320, 256);

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

@@ -3,7 +3,7 @@
 from glob import glob
 import os
 
-HEAD_SIZES_KQ = [40, 64, 72, 80, 96, 112, 128, 256, 512, 576]
+HEAD_SIZES_KQ = [40, 64, 72, 80, 96, 112, 128, 256, 320, 512, 576]
 
 TYPES_KV = ["GGML_TYPE_F16", "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0", "GGML_TYPE_BF16"]
 
@@ -62,7 +62,7 @@ for filename in glob("*.cu"):
     os.remove(filename)
 
 for head_size_kq in HEAD_SIZES_KQ:
-    head_size_v = head_size_kq if head_size_kq != 576 else 512
+    head_size_v = 256 if head_size_kq == 320 else (head_size_kq if head_size_kq != 576 else 512)
     with open(f"fattn-tile-instance-dkq{head_size_kq}-dv{head_size_v}.cu", "w") as f:
         f.write(SOURCE_FATTN_TILE.format(head_size_kq=head_size_kq, head_size_v=head_size_v))
 
@@ -84,13 +84,16 @@ for ncols in [8, 16, 32, 64]:
                     continue
                 if head_size_kq == 72:
                     continue
-                if head_size_kq == 512 and ncols2 not in (4, 8):
+                # Skip compilation of unused ncols2 values for niche head sizes:
+                if head_size_kq == 320 and ncols2 != 32: # Mistral Small 4
                     continue
-                if head_size_kq != 576 and ncols2 in (16, 32):
+                if head_size_kq == 512 and ncols2 not in (4, 8): # Gemma 4
                     continue
-                if head_size_kq == 576 and ncols2 not in (4, 16, 32):
+                if head_size_kq == 576 and ncols2 not in (4, 16, 32): # Deepseek, GLM 4.7 Flash
                     continue
-                head_size_v = head_size_kq if head_size_kq != 576 else 512
+                if head_size_kq not in (320, 576) and ncols2 in (16, 32):
+                    continue
+                head_size_v = 256 if head_size_kq == 320 else (head_size_kq if head_size_kq != 576 else 512)
                 f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size_kq=head_size_kq, head_size_v=head_size_v))
 
 for type in TYPES_MMQ: