vulkan: improve mul_mat_vec_iq1_s speed (#17874)

* q6_k faster mul mat

* 8 values

* fix comment

* switch to two at a time

* start ci for .glsl files

.devops/cann.Dockerfile

@@ -4,7 +4,7 @@
 
 # Define the CANN base image for easier version updates later
 ARG CHIP_TYPE=910b
-ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.3.rc1.alpha001-${CHIP_TYPE}-openeuler22.03-py3.11
+ARG CANN_BASE_IMAGE=quay.io/ascend/cann:8.3.rc2-${CHIP_TYPE}-openeuler24.03-py3.11
 
 # ==============================================================================
 # BUILD STAGE
@@ -111,7 +111,7 @@ ENTRYPOINT ["/app/tools.sh"]
 # ==============================================================================
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 ENTRYPOINT [ "/app/llama-cli" ]
 

.devops/cpu.Dockerfile

@@ -68,7 +68,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 

.devops/cuda.Dockerfile

@@ -74,7 +74,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 

.devops/intel.Dockerfile

@@ -73,7 +73,7 @@ ENTRYPOINT ["/app/tools.sh"]
 FROM base AS light
 
 COPY --from=build /app/lib/ /app
-COPY --from=build /app/full/llama-cli /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 

.devops/musa.Dockerfile

@@ -81,7 +81,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 

.devops/rocm.Dockerfile

@@ -94,7 +94,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 

.devops/s390x.Dockerfile

@@ -105,7 +105,7 @@ WORKDIR /llama.cpp/bin
 
 # Copy llama.cpp binaries and libraries
 COPY --from=collector /llama.cpp/bin/*.so /llama.cpp/bin
-COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin
+COPY --from=collector /llama.cpp/bin/llama-cli /llama.cpp/bin/llama-completion /llama.cpp/bin
 
 ENTRYPOINT [ "/llama.cpp/bin/llama-cli" ]
 

.devops/tools.sh

@@ -13,6 +13,8 @@ elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
     exec ./llama-quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
     exec ./llama-cli "$@"
+elif [[ "$arg1" == '--run-legacy' || "$arg1" == '-l' ]]; then
+    exec ./llama-completion "$@"
 elif [[ "$arg1" == '--bench' || "$arg1" == '-b' ]]; then
     exec ./llama-bench "$@"
 elif [[ "$arg1" == '--perplexity' || "$arg1" == '-p' ]]; then
@@ -32,8 +34,10 @@ elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "
-    echo "  --run (-r): Run a model previously converted into ggml"
-    echo "              ex: -m /models/7B/ggml-model-q4_0.bin -p \"Building a website can be done in 10 simple steps:\" -n 512"
+    echo "  --run (-r): Run a model (chat) previously converted into ggml"
+    echo "              ex: -m /models/7B/ggml-model-q4_0.bin"
+    echo "  --run-legacy (-l): Run a model (legacy completion) previously converted into ggml"
+    echo "              ex: -m /models/7B/ggml-model-q4_0.bin -no-cnv -p \"Building a website can be done in 10 simple steps:\" -n 512"
     echo "  --bench (-b): Benchmark the performance of the inference for various parameters."
     echo "              ex: -m model.gguf"
     echo "  --perplexity (-p): Measure the perplexity of a model over a given text."

.devops/vulkan.Dockerfile

@@ -68,7 +68,7 @@ ENTRYPOINT ["/app/tools.sh"]
 ### Light, CLI only
 FROM base AS light
 
-COPY --from=build /app/full/llama-cli /app
+COPY --from=build /app/full/llama-cli /app/full/llama-completion /app
 
 WORKDIR /app
 

.github/workflows/build.yml

@@ -20,7 +20,8 @@ on:
       '**/*.swift',
       '**/*.m',
       '**/*.metal',
-      '**/*.comp'
+      '**/*.comp',
+      '**/*.glsl'
     ]
 
   pull_request:
@@ -40,7 +41,8 @@ on:
       '**/*.swift',
       '**/*.m',
       '**/*.metal',
-      '**/*.comp'
+      '**/*.comp',
+      '**/*.glsl'
     ]
 
 concurrency:
@@ -243,7 +245,7 @@ jobs:
           echo "Fetch llama2c model"
           wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
           ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
-          ./bin/llama-cli -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
       - name: Test llama2c (s390x)
         id: llama2c_test_s390x
@@ -252,7 +254,7 @@ jobs:
           cd build
           echo "Fetch llama2c big-endian model"
           wget https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K-be.gguf
-          ./bin/llama-cli -m stories260K-be.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+          ./bin/llama-completion -m stories260K-be.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
   ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
@@ -1400,25 +1402,54 @@ jobs:
         chip_type: ['910b', '310p']
         build: ['Release']
     runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
-    container: ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc1.alpha001-910b-openeuler22.03-py3.11' || '8.2.rc1-310p-openeuler22.03-py3.11' }}
     steps:
       - name: Checkout
         uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
 
-      - name: Dependencies
+      - name: Free up disk space
+        uses: ggml-org/free-disk-space@v1.3.1
+        with:
+          tool-cache: true
+
+      - name: Set container image
+        id: cann-image
         run: |
-          yum update -y
-          yum install -y git gcc gcc-c++ make cmake libcurl-devel
+          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc2-910b-openeuler24.03-py3.11' || '8.3.rc2-310p-openeuler24.03-py3.11' }}"
+          echo "image=${image}" >> "${GITHUB_OUTPUT}"
+
+      - name: Pull container image
+        run: docker pull "${{ steps.cann-image.outputs.image }}"
 
       - name: Build
+        env:
+          BUILD_TYPE: ${{ matrix.build }}
+          SOC_TYPE: ascend${{ matrix.chip_type }}
         run: |
-          export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
+          HOST_UID=$(id -u)
+          HOST_GID=$(id -g)
 
-          cmake -S . -B build \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build }} \
-              -DGGML_CANN=on \
-              -DSOC_TYPE=ascend${{ matrix.chip_type }}
-          cmake --build build -j $(nproc)
+          docker run --rm \
+            -v "${PWD}:/workspace" \
+            -w /workspace \
+            -e SOC_TYPE=${SOC_TYPE} \
+            -e BUILD_TYPE=${BUILD_TYPE} \
+            "${{ steps.cann-image.outputs.image }}" \
+            bash -lc '
+              set -e
+              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake libcurl-devel
+              yum clean all && rm -rf /var/cache/yum
+              git config --global --add safe.directory "/workspace"
+              export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
+              cmake -S . -B build \
+                  -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+                  -DGGML_CANN=on \
+                  -DSOC_TYPE=${SOC_TYPE}
+              cmake --build build -j $(nproc)
+
+              chown -R '"${HOST_UID}"':'"${HOST_GID}"' /workspace/build
+            '
 
 # TODO: simplify the following workflows using a matrix
 # TODO: run lighter CI on PRs and the full CI only on master (if needed)
@@ -1770,7 +1801,7 @@ jobs:
           echo "Fetch llama2c model"
           wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
           ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
-          ./bin/llama-cli -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
+          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
   ubuntu-cmake-sanitizer-riscv64-native:
     runs-on: RISCV64

.github/workflows/release.yml

@@ -731,6 +731,78 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-xcframework.tar.gz
           name: llama-${{ steps.tag.outputs.name }}-xcframework.tar.gz
 
+
+  openEuler-cann:
+    strategy:
+      matrix:
+        arch: [x86, aarch64]
+        chip_type: ['910b', '310p']
+        build: ['Release']
+    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Free up disk space
+        uses: ggml-org/free-disk-space@v1.3.1
+        with:
+          tool-cache: true
+
+      - name: Set container image
+        id: cann-image
+        run: |
+          image="ascendai/cann:${{ matrix.chip_type == '910b' &&  '8.3.rc2-910b-openeuler24.03-py3.11' || '8.3.rc2-310p-openeuler24.03-py3.11' }}"
+          echo "image=${image}" >> "${GITHUB_OUTPUT}"
+
+      - name: Pull container image
+        run: docker pull "${{ steps.cann-image.outputs.image }}"
+
+      - name: Build
+        env:
+          BUILD_TYPE: ${{ matrix.build }}
+          SOC_TYPE: ascend${{ matrix.chip_type }}
+        run: |
+          HOST_UID=$(id -u)
+          HOST_GID=$(id -g)
+
+          docker run --rm \
+            -v "${PWD}:/workspace" \
+            -w /workspace \
+            -e SOC_TYPE=${SOC_TYPE} \
+            -e BUILD_TYPE=${BUILD_TYPE} \
+            "${{ steps.cann-image.outputs.image }}" \
+            bash -lc '
+              set -e
+              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake libcurl-devel
+              yum clean all && rm -rf /var/cache/yum
+              git config --global --add safe.directory "/workspace"
+              export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
+              cmake -S . -B build \
+                  -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+                  -DGGML_CANN=on \
+                  -DSOC_TYPE=${SOC_TYPE}
+              cmake --build build -j $(nproc)
+
+              chown -R '"${HOST_UID}"':'"${HOST_GID}"' /workspace/build
+            '
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+
+      - name: Upload artifacts (tar)
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}.tar.gz
+          name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}.tar.gz
+
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
@@ -752,6 +824,7 @@ jobs:
       - macOS-arm64
       - macOS-x64
       - ios-xcode-build
+      - openEuler-cann
 
     steps:
       - name: Clone
@@ -844,6 +917,12 @@ jobs:
             - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip)
             - [Windows x64 (HIP)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-hip-radeon-x64.zip)
 
+            **openEuler:**
+            - [openEuler x86 (310p)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-310p-openEuler-x86.tar.gz)
+            - [openEuler x86 (910b)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-910b-openEuler-x86.tar.gz)
+            - [openEuler aarch64 (310p)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-310p-openEuler-aarch64.tar.gz)
+            - [openEuler aarch64 (910b)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-910b-openEuler-aarch64.tar.gz)
+
       - name: Upload release
         id: upload_release
         uses: actions/github-script@v3

.gitignore

@@ -54,6 +54,7 @@
 /out/
 /tmp/
 /autogen-*.md
+/common/build-info.cpp
 
 # Deprecated
 

CONTRIBUTING.md

@@ -15,6 +15,7 @@ The project differentiates between 3 levels of contributors:
     - If you modified the `ggml` source, run the `test-backend-ops` tool to check whether different backend implementations of the `ggml` operators produce consistent results (this requires access to at least two different `ggml` backends)
     - If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops`
 - Create separate PRs for each feature or fix. Avoid combining unrelated changes in a single PR
+- When adding support for a new model or feature, focus on **CPU support only** in the initial PR unless you have a good reason not to. Add support for other backends like CUDA in follow-up PRs
 - Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
 - If your PR becomes stale, rebase it on top of latest `master` to get maintainers attention
 - Maintainers will rely on your insights and approval when making a final decision to approve and merge a PR

README.md

@@ -347,19 +347,6 @@ To learn more about model quantization, [read this documentation](tools/quantize
 
     </details>
 
-- <details>
-    <summary>Run simple text completion</summary>
-
-    To disable conversation mode explicitly, use `-no-cnv`
-
-    ```bash
-    llama-cli -m model.gguf -p "I believe the meaning of life is" -n 128 -no-cnv
-
-    # I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey.
-    ```
-
-    </details>
-
 - <details>
     <summary>Constrain the output with a custom grammar</summary>
 

ci/run.sh

@@ -398,18 +398,18 @@ function gg_run_qwen3_0_6b {
     ./bin/llama-quantize ${model_bf16} ${model_q5_k} q5_k $(nproc)
     ./bin/llama-quantize ${model_bf16} ${model_q6_k} q6_k $(nproc)
 
-    (time ./bin/llama-cli -no-cnv --model ${model_f16}  -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli -no-cnv --model ${model_bf16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q8_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q2_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q3_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q4_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q5_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli -no-cnv --model ${model_q6_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-completion -no-cnv --model ${model_f16}  -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-completion -no-cnv --model ${model_bf16} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-bf16.log
+    (time ./bin/llama-completion -no-cnv --model ${model_q8_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-completion -no-cnv --model ${model_q4_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-completion -no-cnv --model ${model_q4_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-completion -no-cnv --model ${model_q5_0} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-completion -no-cnv --model ${model_q5_1} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-completion -no-cnv --model ${model_q2_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-completion -no-cnv --model ${model_q3_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-completion -no-cnv --model ${model_q4_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-completion -no-cnv --model ${model_q5_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-completion -no-cnv --model ${model_q6_k} -ngl 99 -c 1024 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
     (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     if [ -z ${GG_BUILD_NO_BF16} ]; then

common/CMakeLists.txt

@@ -73,6 +73,8 @@ add_library(${TARGET} STATIC
     ngram-cache.h
     peg-parser.cpp
     peg-parser.h
+    preset.cpp
+    preset.h
     regex-partial.cpp
     regex-partial.h
     sampling.cpp

common/arg.h

@@ -3,8 +3,10 @@
 #include "common.h"
 
 #include <set>
+#include <map>
 #include <string>
 #include <vector>
+#include <cstring>
 
 //
 // CLI argument parsing
@@ -14,6 +16,7 @@ struct common_arg {
     std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
     std::set<enum llama_example> excludes = {};
     std::vector<const char *> args;
+    std::vector<const char *> args_neg;  // for negated args like --no-xxx
     const char * value_hint   = nullptr; // help text or example for arg value
     const char * value_hint_2 = nullptr; // for second arg value
     const char * env          = nullptr;
@@ -23,6 +26,9 @@ struct common_arg {
     void (*handler_string) (common_params & params, const std::string &) = nullptr;
     void (*handler_str_str)(common_params & params, const std::string &, const std::string &) = nullptr;
     void (*handler_int)    (common_params & params, int) = nullptr;
+    void (*handler_bool)   (common_params & params, bool) = nullptr;
+
+    common_arg() = default;
 
     common_arg(
         const std::initializer_list<const char *> & args,
@@ -44,6 +50,13 @@ struct common_arg {
         void (*handler)(common_params & params)
     ) : args(args), help(help), handler_void(handler) {}
 
+    common_arg(
+        const std::initializer_list<const char *> & args,
+        const std::initializer_list<const char *> & args_neg,
+        const std::string & help,
+        void (*handler)(common_params & params, bool)
+    ) : args(args), args_neg(args_neg), help(help), handler_bool(handler) {}
+
     // support 2 values for arg
     common_arg(
         const std::initializer_list<const char *> & args,
@@ -61,9 +74,33 @@ struct common_arg {
     bool is_exclude(enum llama_example ex);
     bool get_value_from_env(std::string & output) const;
     bool has_value_from_env() const;
-    std::string to_string();
+    std::string to_string() const;
+
+    // for using as key in std::map
+    bool operator<(const common_arg& other) const {
+        if (args.empty() || other.args.empty()) {
+            return false;
+        }
+        return strcmp(args[0], other.args[0]) < 0;
+    }
+    bool operator==(const common_arg& other) const {
+        if (args.empty() || other.args.empty()) {
+            return false;
+        }
+        return strcmp(args[0], other.args[0]) == 0;
+    }
+
+    // get all args and env vars (including negated args/env)
+    std::vector<std::string> get_args() const;
+    std::vector<std::string> get_env() const;
 };
 
+namespace common_arg_utils {
+    bool is_truthy(const std::string & value);
+    bool is_falsey(const std::string & value);
+    bool is_autoy(const std::string & value);
+}
+
 struct common_params_context {
     enum llama_example ex = LLAMA_EXAMPLE_COMMON;
     common_params & params;
@@ -76,7 +113,11 @@ struct common_params_context {
 // if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
 bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
 
-// function to be used by test-arg-parser
+// parse input arguments from CLI into a map
+// TODO: support repeated args in the future
+bool common_params_to_map(int argc, char ** argv, llama_example ex, std::map<common_arg, std::string> & out_map);
+
+// initialize argument parser context - used by test-arg-parser and preset
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
 
 struct common_remote_params {

common/chat.cpp

@@ -1,5 +1,6 @@
 #include "chat.h"
 #include "chat-parser.h"
+#include "chat-peg-parser.h"
 #include "common.h"
 #include "json-partial.h"
 #include "json-schema-to-grammar.h"
@@ -150,6 +151,7 @@ struct templates_params {
     common_chat_tool_choice tool_choice;
     json json_schema;
     bool parallel_tool_calls;
+    common_reasoning_format reasoning_format;
     bool stream;
     std::string grammar;
     bool add_generation_prompt = true;
@@ -589,6 +591,16 @@ common_chat_templates_ptr common_chat_templates_init(
             "{%- if false %}");
     }
 
+    // TODO @aldehir : this is a temporary fix, pending Minja changes
+    // Ref: https://github.com/ggml-org/llama.cpp/pull/17713#issuecomment-3631342664
+    if (default_template_src.find("[TOOL_CALLS]") != std::string::npos
+            // search for the error message and patch it
+            && default_template_src.find("if (message['content'] is none or") != std::string::npos) {
+        string_replace_all(default_template_src,
+            "{%- if (message['content'] is none or message['content'] == '' or message['content']|length == 0) and (message['tool_calls'] is not defined or message['tool_calls'] is none or message['tool_calls']|length == 0) %}",
+            "{%- if false %}");
+    }
+
     std::string token_bos = bos_token_override;
     std::string token_eos = eos_token_override;
     bool add_bos = false;
@@ -987,6 +999,118 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
     return data;
 }
 
+static common_chat_params common_chat_params_init_ministral_3(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+
+    // Build up messages to follow the format: https://huggingface.co/mistralai/Ministral-3-14B-Reasoning-2512/blob/main/chat_template.jinja
+    auto adjusted_messages = json::array();
+    for (const auto & msg : inputs.messages) {
+        auto role = msg.value("role", "");
+        if (role != "system" && role != "assistant") {
+            // Only adjust system and assistant messages. Interestingly, the system message may contain thinking.
+            adjusted_messages.push_back(msg);
+            continue;
+        }
+
+        auto content = json::array();
+
+        // If message contains `reasoning_content`, add it as a block of type `thinking`
+        if (msg.contains("reasoning_content") && msg.at("reasoning_content").is_string()) {
+            content.push_back({
+                {"type", "thinking"},
+                {"thinking", msg.at("reasoning_content").get<std::string>()},
+            });
+        }
+
+        // If message contains `content`, add it as a block of type `text`
+        if (msg.contains("content")) {
+            if (msg.at("content").is_string()) {
+                content.push_back({
+                    {"type", "text"},
+                    {"text", msg.at("content").get<std::string>()},
+                });
+            } else if (msg.at("content").is_array()) {
+                auto blocks = msg.at("content");
+                content.insert(content.end(), blocks.begin(), blocks.end());
+            }
+        }
+
+        auto adjusted = msg;
+        adjusted["content"] = content;
+        adjusted.erase("reasoning_content");
+        adjusted_messages.push_back(adjusted);
+    }
+
+    auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar = true;
+
+    data.prompt = apply(tmpl, inputs, /* messages_override = */ adjusted_messages);
+    data.format = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.preserved_tokens = {
+        "[THINK]",
+        "[/THINK]",
+        "[TOOL_CALLS]",
+        "[ARGS]",
+    };
+
+    auto parser = build_chat_peg_native_parser([&](common_chat_peg_native_builder & p) {
+        auto reasoning = extract_reasoning ? p.optional("[THINK]" + p.reasoning(p.until("[/THINK]")) + "[/THINK]") : p.eps();
+
+        // Response format parser
+        if (inputs.json_schema.is_object() && !inputs.json_schema.empty()) {
+            // Ministral wants to emit json surrounded by code fences
+            return reasoning << "```json" << p.content(p.schema(p.json(), "response-format", inputs.json_schema)) << "```";
+        }
+
+        // Tool call parser
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            auto tool_choice = p.choice();
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                const auto & schema = function.at("parameters");
+
+                tool_choice |= p.rule("tool-" + name,
+                    p.tool_open(p.tool_name(p.literal(name)) + "[ARGS]")
+                    + p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema))
+                );
+            });
+
+            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+            auto max_calls = inputs.parallel_tool_calls ? -1 : 1;
+            auto tool_calls = p.trigger_rule("tool-call", p.repeat("[TOOL_CALLS]" + tool_choice, min_calls, max_calls));
+
+            return reasoning << p.content(p.until("[TOOL_CALLS]")) << tool_calls;
+        }
+
+        // Content only parser
+        include_grammar = false;
+        return reasoning << p.content(p.rest());
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto schema = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            {COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "[TOOL_CALLS]"}
+        };
+    }
+
+    return data;
+}
+
 static common_chat_params common_chat_params_init_magistral(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
     data.prompt = apply(tmpl, inputs);
@@ -2341,6 +2465,7 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.messages = common_chat_msgs_to_json_oaicompat<json>(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
     params.add_generation_prompt = inputs.add_generation_prompt;
     params.tool_choice = inputs.tool_choice;
+    params.reasoning_format = inputs.reasoning_format;
     params.enable_thinking = inputs.enable_thinking;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
@@ -2504,6 +2629,13 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_llama_3_x(tmpl, params, allow_python_tag_builtin_tools);
     }
 
+    // Ministral/Mistral Large 3
+    if (src.find("[SYSTEM_PROMPT]") != std::string::npos &&
+        src.find("[TOOL_CALLS]") != std::string::npos &&
+        src.find("[ARGS]") != std::string::npos) {
+        return common_chat_params_init_ministral_3(tmpl, params);
+    }
+
     if (src.find("[THINK]") != std::string::npos && src.find("[/THINK]") != std::string::npos) {
         return common_chat_params_init_magistral(tmpl, params);
     }

common/common.h

@@ -82,7 +82,8 @@ int32_t cpu_get_num_math();
 enum llama_example {
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
-    LLAMA_EXAMPLE_MAIN,
+    LLAMA_EXAMPLE_COMPLETION,
+    LLAMA_EXAMPLE_CLI,
     LLAMA_EXAMPLE_EMBEDDING,
     LLAMA_EXAMPLE_PERPLEXITY,
     LLAMA_EXAMPLE_RETRIEVAL,
@@ -406,6 +407,7 @@ struct common_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool no_perf           = false; // disable performance metrics
+    bool show_timings      = true;  // show timing information on CLI
     bool ctx_shift         = false; // context shift on infinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     bool kv_unified        = false; // enable unified KV cache
@@ -462,7 +464,7 @@ struct common_params {
     std::string public_path   = "";                                                                         // NOLINT
     std::string api_prefix    = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
-    bool use_jinja = false;                                                                                 // NOLINT
+    bool use_jinja = true;                                                                                  // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int reasoning_budget = -1;
@@ -482,9 +484,10 @@ struct common_params {
     bool endpoint_metrics = false;
 
     // router server configs
-    std::string models_dir = ""; // directory containing models for the router server
-    int models_max = 4;          // maximum number of models to load simultaneously
-    bool models_autoload = true; // automatically load models when requested via the router server
+    std::string models_dir    = ""; // directory containing models for the router server
+    std::string models_preset = ""; // directory containing model presets for the router server
+    int models_max = 4;             // maximum number of models to load simultaneously
+    bool models_autoload = true;    // automatically load models when requested via the router server
 
     bool log_json = false;
 

common/console.cpp

@@ -1,6 +1,16 @@
 #include "console.h"
+#include "log.h"
 #include <vector>
 #include <iostream>
+#include <cassert>
+#include <cstddef>
+#include <cctype>
+#include <cwctype>
+#include <cstdint>
+#include <condition_variable>
+#include <mutex>
+#include <thread>
+#include <stdarg.h>
 
 #if defined(_WIN32)
 #define WIN32_LEAN_AND_MEAN
@@ -30,26 +40,44 @@
 #define ANSI_COLOR_BLUE    "\x1b[34m"
 #define ANSI_COLOR_MAGENTA "\x1b[35m"
 #define ANSI_COLOR_CYAN    "\x1b[36m"
+#define ANSI_COLOR_GRAY    "\x1b[90m"
 #define ANSI_COLOR_RESET   "\x1b[0m"
 #define ANSI_BOLD          "\x1b[1m"
 
 namespace console {
 
+#if defined (_WIN32)
+    namespace {
+        // Use private-use unicode values to represent special keys that are not reported
+        // as characters (e.g. arrows on Windows). These values should never clash with
+        // real input and let the rest of the code handle navigation uniformly.
+        static constexpr char32_t KEY_ARROW_LEFT       = 0xE000;
+        static constexpr char32_t KEY_ARROW_RIGHT      = 0xE001;
+        static constexpr char32_t KEY_ARROW_UP         = 0xE002;
+        static constexpr char32_t KEY_ARROW_DOWN       = 0xE003;
+        static constexpr char32_t KEY_HOME             = 0xE004;
+        static constexpr char32_t KEY_END              = 0xE005;
+        static constexpr char32_t KEY_CTRL_ARROW_LEFT  = 0xE006;
+        static constexpr char32_t KEY_CTRL_ARROW_RIGHT = 0xE007;
+        static constexpr char32_t KEY_DELETE           = 0xE008;
+    }
+
     //
     // Console state
     //
+#endif
 
-    static bool      advanced_display = false;
-    static bool      simple_io        = true;
-    static display_t current_display  = reset;
+    static bool         advanced_display = false;
+    static bool         simple_io        = true;
+    static display_type current_display  = DISPLAY_TYPE_RESET;
 
-    static FILE*     out              = stdout;
+    static FILE*        out              = stdout;
 
 #if defined (_WIN32)
-    static void*     hConsole;
+    static void*        hConsole;
 #else
-    static FILE*     tty              = nullptr;
-    static termios   initial_state;
+    static FILE*        tty              = nullptr;
+    static termios      initial_state;
 #endif
 
     //
@@ -120,7 +148,7 @@ namespace console {
 
     void cleanup() {
         // Reset console display
-        set_display(reset);
+        set_display(DISPLAY_TYPE_RESET);
 
 #if !defined(_WIN32)
         // Restore settings on POSIX systems
@@ -140,20 +168,26 @@ namespace console {
     //
 
     // Keep track of current display and only emit ANSI code if it changes
-    void set_display(display_t display) {
+    void set_display(display_type display) {
         if (advanced_display && current_display != display) {
-            fflush(stdout);
+            common_log_flush(common_log_main());
             switch(display) {
-                case reset:
+                case DISPLAY_TYPE_RESET:
                     fprintf(out, ANSI_COLOR_RESET);
                     break;
-                case prompt:
+                case DISPLAY_TYPE_INFO:
+                    fprintf(out, ANSI_COLOR_MAGENTA);
+                    break;
+                case DISPLAY_TYPE_PROMPT:
                     fprintf(out, ANSI_COLOR_YELLOW);
                     break;
-                case user_input:
+                case DISPLAY_TYPE_REASONING:
+                    fprintf(out, ANSI_COLOR_GRAY);
+                    break;
+                case DISPLAY_TYPE_USER_INPUT:
                     fprintf(out, ANSI_BOLD ANSI_COLOR_GREEN);
                     break;
-                case error:
+                case DISPLAY_TYPE_ERROR:
                     fprintf(out, ANSI_BOLD ANSI_COLOR_RED);
             }
             current_display = display;
@@ -176,7 +210,18 @@ namespace console {
             if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
                 wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
                 if (wc == 0) {
-                    continue;
+                    const DWORD ctrl_mask = LEFT_CTRL_PRESSED | RIGHT_CTRL_PRESSED;
+                    const bool ctrl_pressed = (record.Event.KeyEvent.dwControlKeyState & ctrl_mask) != 0;
+                    switch (record.Event.KeyEvent.wVirtualKeyCode) {
+                        case VK_LEFT:   return ctrl_pressed ? KEY_CTRL_ARROW_LEFT  : KEY_ARROW_LEFT;
+                        case VK_RIGHT:  return ctrl_pressed ? KEY_CTRL_ARROW_RIGHT : KEY_ARROW_RIGHT;
+                        case VK_UP:     return KEY_ARROW_UP;
+                        case VK_DOWN:   return KEY_ARROW_DOWN;
+                        case VK_HOME:   return KEY_HOME;
+                        case VK_END:    return KEY_END;
+                        case VK_DELETE: return KEY_DELETE;
+                        default:        continue;
+                    }
                 }
 
                 if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
@@ -315,6 +360,52 @@ namespace console {
 #endif
     }
 
+    static char32_t decode_utf8(const std::string & input, size_t pos, size_t & advance) {
+        unsigned char c = static_cast<unsigned char>(input[pos]);
+        if ((c & 0x80u) == 0u) {
+            advance = 1;
+            return c;
+        }
+        if ((c & 0xE0u) == 0xC0u && pos + 1 < input.size()) {
+            unsigned char c1 = static_cast<unsigned char>(input[pos + 1]);
+            if ((c1 & 0xC0u) != 0x80u) {
+                advance = 1;
+                return 0xFFFD;
+            }
+            advance = 2;
+            return ((c & 0x1Fu) << 6) | (static_cast<unsigned char>(input[pos + 1]) & 0x3Fu);
+        }
+        if ((c & 0xF0u) == 0xE0u && pos + 2 < input.size()) {
+            unsigned char c1 = static_cast<unsigned char>(input[pos + 1]);
+            unsigned char c2 = static_cast<unsigned char>(input[pos + 2]);
+            if ((c1 & 0xC0u) != 0x80u || (c2 & 0xC0u) != 0x80u) {
+                advance = 1;
+                return 0xFFFD;
+            }
+            advance = 3;
+            return ((c & 0x0Fu) << 12) |
+                   ((static_cast<unsigned char>(input[pos + 1]) & 0x3Fu) << 6) |
+                   (static_cast<unsigned char>(input[pos + 2]) & 0x3Fu);
+        }
+        if ((c & 0xF8u) == 0xF0u && pos + 3 < input.size()) {
+            unsigned char c1 = static_cast<unsigned char>(input[pos + 1]);
+            unsigned char c2 = static_cast<unsigned char>(input[pos + 2]);
+            unsigned char c3 = static_cast<unsigned char>(input[pos + 3]);
+            if ((c1 & 0xC0u) != 0x80u || (c2 & 0xC0u) != 0x80u || (c3 & 0xC0u) != 0x80u) {
+                advance = 1;
+                return 0xFFFD;
+            }
+            advance = 4;
+            return ((c & 0x07u) << 18) |
+                   ((static_cast<unsigned char>(input[pos + 1]) & 0x3Fu) << 12) |
+                   ((static_cast<unsigned char>(input[pos + 2]) & 0x3Fu) << 6) |
+                   (static_cast<unsigned char>(input[pos + 3]) & 0x3Fu);
+        }
+
+        advance = 1;
+        return 0xFFFD; // replacement character for invalid input
+    }
+
     static void append_utf8(char32_t ch, std::string & out) {
         if (ch <= 0x7F) {
             out.push_back(static_cast<unsigned char>(ch));
@@ -336,22 +427,319 @@ namespace console {
     }
 
     // Helper function to remove the last UTF-8 character from a string
-    static void pop_back_utf8_char(std::string & line) {
-        if (line.empty()) {
+    static size_t prev_utf8_char_pos(const std::string & line, size_t pos) {
+        if (pos == 0) return 0;
+        pos--;
+        while (pos > 0 && (line[pos] & 0xC0) == 0x80) {
+            pos--;
+        }
+        return pos;
+    }
+
+    static size_t next_utf8_char_pos(const std::string & line, size_t pos) {
+        if (pos >= line.length()) return line.length();
+        pos++;
+        while (pos < line.length() && (line[pos] & 0xC0) == 0x80) {
+            pos++;
+        }
+        return pos;
+    }
+
+    static void move_cursor(int delta);
+    static void move_word_left(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line);
+    static void move_word_right(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line);
+    static void move_to_line_start(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths);
+    static void move_to_line_end(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line);
+
+    static void delete_at_cursor(std::string & line, std::vector<int> & widths, size_t & char_pos, size_t & byte_pos) {
+        if (char_pos >= widths.size()) {
             return;
         }
 
-        size_t pos = line.length() - 1;
+        size_t next_pos = next_utf8_char_pos(line, byte_pos);
+        int w = widths[char_pos];
+        size_t char_len = next_pos - byte_pos;
 
-        // Find the start of the last UTF-8 character (checking up to 4 bytes back)
-        for (size_t i = 0; i < 3 && pos > 0; ++i, --pos) {
-            if ((line[pos] & 0xC0) != 0x80) {
-                break; // Found the start of the character
-            }
+        line.erase(byte_pos, char_len);
+        widths.erase(widths.begin() + char_pos);
+
+        size_t p = byte_pos;
+        int tail_width = 0;
+        for (size_t i = char_pos; i < widths.size(); ++i) {
+            size_t following = next_utf8_char_pos(line, p);
+            put_codepoint(line.c_str() + p, following - p, widths[i]);
+            tail_width += widths[i];
+            p = following;
         }
-        line.erase(pos);
+
+        for (int i = 0; i < w; ++i) {
+            fputc(' ', out);
+        }
+
+        move_cursor(-(tail_width + w));
     }
 
+    static void clear_current_line(const std::vector<int> & widths) {
+        int total_width = 0;
+        for (int w : widths) {
+            total_width += (w > 0 ? w : 1);
+        }
+
+        if (total_width > 0) {
+            std::string spaces(total_width, ' ');
+            fwrite(spaces.c_str(), 1, total_width, out);
+            move_cursor(-total_width);
+        }
+    }
+
+    static void set_line_contents(std::string new_line, std::string & line, std::vector<int> & widths, size_t & char_pos,
+                                  size_t & byte_pos) {
+        move_to_line_start(char_pos, byte_pos, widths);
+        clear_current_line(widths);
+
+        line = std::move(new_line);
+        widths.clear();
+        byte_pos = 0;
+        char_pos = 0;
+
+        size_t idx = 0;
+        while (idx < line.size()) {
+            size_t advance = 0;
+            char32_t cp = decode_utf8(line, idx, advance);
+            int expected_width = estimateWidth(cp);
+            int real_width = put_codepoint(line.c_str() + idx, advance, expected_width);
+            if (real_width < 0) real_width = 0;
+            widths.push_back(real_width);
+            idx += advance;
+            ++char_pos;
+            byte_pos = idx;
+        }
+    }
+
+    static void move_to_line_start(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths) {
+        int back_width = 0;
+        for (size_t i = 0; i < char_pos; ++i) {
+            back_width += widths[i];
+        }
+        move_cursor(-back_width);
+        char_pos = 0;
+        byte_pos = 0;
+    }
+
+    static void move_to_line_end(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line) {
+        int forward_width = 0;
+        for (size_t i = char_pos; i < widths.size(); ++i) {
+            forward_width += widths[i];
+        }
+        move_cursor(forward_width);
+        char_pos = widths.size();
+        byte_pos = line.length();
+    }
+
+    static bool has_ctrl_modifier(const std::string & params) {
+        size_t start = 0;
+        while (start < params.size()) {
+            size_t end = params.find(';', start);
+            size_t len = (end == std::string::npos) ? params.size() - start : end - start;
+            if (len > 0) {
+                int value = 0;
+                for (size_t i = 0; i < len; ++i) {
+                    char ch = params[start + i];
+                    if (!std::isdigit(static_cast<unsigned char>(ch))) {
+                        value = -1;
+                        break;
+                    }
+                    value = value * 10 + (ch - '0');
+                }
+                if (value == 5) {
+                    return true;
+                }
+            }
+
+            if (end == std::string::npos) {
+                break;
+            }
+            start = end + 1;
+        }
+        return false;
+    }
+
+    static bool is_space_codepoint(char32_t cp) {
+        return std::iswspace(static_cast<wint_t>(cp)) != 0;
+    }
+
+    static void move_word_left(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line) {
+        if (char_pos == 0) {
+            return;
+        }
+
+        size_t new_char_pos = char_pos;
+        size_t new_byte_pos = byte_pos;
+        int move_width = 0;
+
+        while (new_char_pos > 0) {
+            size_t prev_byte = prev_utf8_char_pos(line, new_byte_pos);
+            size_t advance = 0;
+            char32_t cp = decode_utf8(line, prev_byte, advance);
+            if (!is_space_codepoint(cp)) {
+                break;
+            }
+            move_width += widths[new_char_pos - 1];
+            new_char_pos--;
+            new_byte_pos = prev_byte;
+        }
+
+        while (new_char_pos > 0) {
+            size_t prev_byte = prev_utf8_char_pos(line, new_byte_pos);
+            size_t advance = 0;
+            char32_t cp = decode_utf8(line, prev_byte, advance);
+            if (is_space_codepoint(cp)) {
+                break;
+            }
+            move_width += widths[new_char_pos - 1];
+            new_char_pos--;
+            new_byte_pos = prev_byte;
+        }
+
+        move_cursor(-move_width);
+        char_pos = new_char_pos;
+        byte_pos = new_byte_pos;
+    }
+
+    static void move_word_right(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line) {
+        if (char_pos >= widths.size()) {
+            return;
+        }
+
+        size_t new_char_pos = char_pos;
+        size_t new_byte_pos = byte_pos;
+        int move_width = 0;
+
+        while (new_char_pos < widths.size()) {
+            size_t advance = 0;
+            char32_t cp = decode_utf8(line, new_byte_pos, advance);
+            if (!is_space_codepoint(cp)) {
+                break;
+            }
+            move_width += widths[new_char_pos];
+            new_char_pos++;
+            new_byte_pos += advance;
+        }
+
+        while (new_char_pos < widths.size()) {
+            size_t advance = 0;
+            char32_t cp = decode_utf8(line, new_byte_pos, advance);
+            if (is_space_codepoint(cp)) {
+                break;
+            }
+            move_width += widths[new_char_pos];
+            new_char_pos++;
+            new_byte_pos += advance;
+        }
+
+        while (new_char_pos < widths.size()) {
+            size_t advance = 0;
+            char32_t cp = decode_utf8(line, new_byte_pos, advance);
+            if (!is_space_codepoint(cp)) {
+                break;
+            }
+            move_width += widths[new_char_pos];
+            new_char_pos++;
+            new_byte_pos += advance;
+        }
+
+        move_cursor(move_width);
+        char_pos = new_char_pos;
+        byte_pos = new_byte_pos;
+    }
+
+    static void move_cursor(int delta) {
+        if (delta == 0) return;
+#if defined(_WIN32)
+        if (hConsole != NULL) {
+            CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
+            GetConsoleScreenBufferInfo(hConsole, &bufferInfo);
+            COORD newCursorPosition = bufferInfo.dwCursorPosition;
+            int width = bufferInfo.dwSize.X;
+            int newX = newCursorPosition.X + delta;
+            int newY = newCursorPosition.Y;
+
+            while (newX >= width) {
+                newX -= width;
+                newY++;
+            }
+            while (newX < 0) {
+                newX += width;
+                newY--;
+            }
+
+            newCursorPosition.X = newX;
+            newCursorPosition.Y = newY;
+            SetConsoleCursorPosition(hConsole, newCursorPosition);
+        }
+#else
+        if (delta < 0) {
+            for (int i = 0; i < -delta; i++) fprintf(out, "\b");
+        } else {
+            for (int i = 0; i < delta; i++) fprintf(out, "\033[C");
+        }
+#endif
+    }
+
+    struct history_t {
+        std::vector<std::string> entries;
+        size_t viewing_idx = SIZE_MAX;
+        std::string backup_line; // current line before viewing history
+        void add(const std::string & line) {
+            if (line.empty()) {
+                return;
+            }
+            // avoid duplicates with the last entry
+            if (entries.empty() || entries.back() != line) {
+                entries.push_back(line);
+            }
+            // also clear viewing state
+            end_viewing();
+        }
+        bool prev(std::string & cur_line) {
+            if (entries.empty()) {
+                return false;
+            }
+            if (viewing_idx == SIZE_MAX) {
+                return false;
+            }
+            if (viewing_idx > 0) {
+                viewing_idx--;
+            }
+            cur_line = entries[viewing_idx];
+            return true;
+        }
+        bool next(std::string & cur_line) {
+            if (entries.empty() || viewing_idx == SIZE_MAX) {
+                return false;
+            }
+            viewing_idx++;
+            if (viewing_idx >= entries.size()) {
+                cur_line = backup_line;
+                end_viewing();
+            } else {
+                cur_line = entries[viewing_idx];
+            }
+            return true;
+        }
+        void begin_viewing(const std::string & line) {
+            backup_line = line;
+            viewing_idx = entries.size();
+        }
+        void end_viewing() {
+            viewing_idx = SIZE_MAX;
+            backup_line.clear();
+        }
+        bool is_viewing() const {
+            return viewing_idx != SIZE_MAX;
+        }
+    } history;
+
     static bool readline_advanced(std::string & line, bool multiline_input) {
         if (out != stdout) {
             fflush(stdout);
@@ -362,8 +750,33 @@ namespace console {
         bool is_special_char = false;
         bool end_of_stream = false;
 
+        size_t byte_pos = 0; // current byte index
+        size_t char_pos = 0; // current character index (one char can be multiple bytes)
+
         char32_t input_char;
         while (true) {
+            assert(char_pos <= byte_pos);
+            assert(char_pos <= widths.size());
+            auto history_prev = [&]() {
+                if (!history.is_viewing()) {
+                    history.begin_viewing(line);
+                }
+                std::string new_line;
+                if (!history.prev(new_line)) {
+                    return;
+                }
+                set_line_contents(new_line, line, widths, char_pos, byte_pos);
+            };
+            auto history_next = [&]() {
+                if (history.is_viewing()) {
+                    std::string new_line;
+                    if (!history.next(new_line)) {
+                        return;
+                    }
+                    set_line_contents(new_line, line, widths, char_pos, byte_pos);
+                }
+            };
+
             fflush(out); // Ensure all output is displayed before waiting for input
             input_char = getchar32();
 
@@ -371,20 +784,83 @@ namespace console {
                 break;
             }
 
-            if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
+            if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D */) {
                 end_of_stream = true;
                 break;
             }
 
             if (is_special_char) {
-                set_display(user_input);
                 replace_last(line.back());
                 is_special_char = false;
             }
 
             if (input_char == '\033') { // Escape sequence
                 char32_t code = getchar32();
-                if (code == '[' || code == 0x1B) {
+                if (code == '[') {
+                    std::string params;
+                    while (true) {
+                        code = getchar32();
+                        if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~' || code == (char32_t) WEOF) {
+                            break;
+                        }
+                        params.push_back(static_cast<char>(code));
+                    }
+
+                    const bool ctrl_modifier = has_ctrl_modifier(params);
+
+                    if (code == 'D') { // left
+                        if (ctrl_modifier) {
+                            move_word_left(char_pos, byte_pos, widths, line);
+                        } else if (char_pos > 0) {
+                            int w = widths[char_pos - 1];
+                            move_cursor(-w);
+                            char_pos--;
+                            byte_pos = prev_utf8_char_pos(line, byte_pos);
+                        }
+                    } else if (code == 'C') { // right
+                        if (ctrl_modifier) {
+                            move_word_right(char_pos, byte_pos, widths, line);
+                        } else if (char_pos < widths.size()) {
+                            int w = widths[char_pos];
+                            move_cursor(w);
+                            char_pos++;
+                            byte_pos = next_utf8_char_pos(line, byte_pos);
+                        }
+                    } else if (code == 'H') { // home
+                        move_to_line_start(char_pos, byte_pos, widths);
+                    } else if (code == 'F') { // end
+                        move_to_line_end(char_pos, byte_pos, widths, line);
+                    } else if (code == 'A' || code == 'B') {
+                        // up/down
+                        if (code == 'A') {
+                            history_prev();
+                            is_special_char = false;
+                        } else if (code == 'B') {
+                            history_next();
+                            is_special_char = false;
+                        }
+                    } else if ((code == '~' || (code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z')) && !params.empty()) {
+                        std::string digits;
+                        for (char ch : params) {
+                            if (ch == ';') {
+                                break;
+                            }
+                            if (std::isdigit(static_cast<unsigned char>(ch))) {
+                                digits.push_back(ch);
+                            }
+                        }
+
+                        if (code == '~') {
+                            if (digits == "1" || digits == "7") { // home
+                                move_to_line_start(char_pos, byte_pos, widths);
+                            } else if (digits == "4" || digits == "8") { // end
+                                move_to_line_end(char_pos, byte_pos, widths, line);
+                            } else if (digits == "3") { // delete
+                                delete_at_cursor(line, widths, char_pos, byte_pos);
+                            }
+                        }
+                    }
+                } else if (code == 0x1B) {
                     // Discard the rest of the escape sequence
                     while ((code = getchar32()) != (char32_t) WEOF) {
                         if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
@@ -392,32 +868,110 @@ namespace console {
                         }
                     }
                 }
+#if defined(_WIN32)
+            } else if (input_char == KEY_ARROW_LEFT) {
+                if (char_pos > 0) {
+                    int w = widths[char_pos - 1];
+                    move_cursor(-w);
+                    char_pos--;
+                    byte_pos = prev_utf8_char_pos(line, byte_pos);
+                }
+            } else if (input_char == KEY_ARROW_RIGHT) {
+                if (char_pos < widths.size()) {
+                    int w = widths[char_pos];
+                    move_cursor(w);
+                    char_pos++;
+                    byte_pos = next_utf8_char_pos(line, byte_pos);
+                }
+            } else if (input_char == KEY_CTRL_ARROW_LEFT) {
+                move_word_left(char_pos, byte_pos, widths, line);
+            } else if (input_char == KEY_CTRL_ARROW_RIGHT) {
+                move_word_right(char_pos, byte_pos, widths, line);
+            } else if (input_char == KEY_HOME) {
+                move_to_line_start(char_pos, byte_pos, widths);
+            } else if (input_char == KEY_END) {
+                move_to_line_end(char_pos, byte_pos, widths, line);
+            } else if (input_char == KEY_DELETE) {
+                delete_at_cursor(line, widths, char_pos, byte_pos);
+            } else if (input_char == KEY_ARROW_UP || input_char == KEY_ARROW_DOWN) {
+                if (input_char == KEY_ARROW_UP) {
+                    history_prev();
+                    is_special_char = false;
+                } else if (input_char == KEY_ARROW_DOWN) {
+                    history_next();
+                    is_special_char = false;
+                }
+#endif
             } else if (input_char == 0x08 || input_char == 0x7F) { // Backspace
-                if (!widths.empty()) {
-                    int count;
-                    do {
-                        count = widths.back();
-                        widths.pop_back();
-                        // Move cursor back, print space, and move cursor back again
-                        for (int i = 0; i < count; i++) {
-                            replace_last(' ');
-                            pop_cursor();
-                        }
-                        pop_back_utf8_char(line);
-                    } while (count == 0 && !widths.empty());
+                if (char_pos > 0) {
+                    int w = widths[char_pos - 1];
+                    move_cursor(-w);
+                    char_pos--;
+                    size_t prev_pos = prev_utf8_char_pos(line, byte_pos);
+                    size_t char_len = byte_pos - prev_pos;
+                    byte_pos = prev_pos;
+
+                    // remove the character
+                    line.erase(byte_pos, char_len);
+                    widths.erase(widths.begin() + char_pos);
+
+                    // redraw tail
+                    size_t p = byte_pos;
+                    int tail_width = 0;
+                    for (size_t i = char_pos; i < widths.size(); ++i) {
+                        size_t next_p = next_utf8_char_pos(line, p);
+                        put_codepoint(line.c_str() + p, next_p - p, widths[i]);
+                        tail_width += widths[i];
+                        p = next_p;
+                    }
+
+                    // clear display
+                    for (int i = 0; i < w; ++i) {
+                        fputc(' ', out);
+                    }
+                    move_cursor(-(tail_width + w));
                 }
             } else {
-                int offset = line.length();
-                append_utf8(input_char, line);
-                int width = put_codepoint(line.c_str() + offset, line.length() - offset, estimateWidth(input_char));
-                if (width < 0) {
-                    width = 0;
+                // insert character
+                std::string new_char_str;
+                append_utf8(input_char, new_char_str);
+                int w = estimateWidth(input_char);
+
+                if (char_pos == widths.size()) {
+                    // insert at the end
+                    line += new_char_str;
+                    int real_w = put_codepoint(new_char_str.c_str(), new_char_str.length(), w);
+                    if (real_w < 0) real_w = 0;
+                    widths.push_back(real_w);
+                    byte_pos += new_char_str.length();
+                    char_pos++;
+                } else {
+                    // insert in middle
+                    line.insert(byte_pos, new_char_str);
+
+                    int real_w = put_codepoint(new_char_str.c_str(), new_char_str.length(), w);
+                    if (real_w < 0) real_w = 0;
+
+                    widths.insert(widths.begin() + char_pos, real_w);
+
+                    // print the tail
+                    size_t p = byte_pos + new_char_str.length();
+                    int tail_width = 0;
+                    for (size_t i = char_pos + 1; i < widths.size(); ++i) {
+                        size_t next_p = next_utf8_char_pos(line, p);
+                        put_codepoint(line.c_str() + p, next_p - p, widths[i]);
+                        tail_width += widths[i];
+                        p = next_p;
+                    }
+
+                    move_cursor(-tail_width);
+
+                    byte_pos += new_char_str.length();
+                    char_pos++;
                 }
-                widths.push_back(width);
             }
 
             if (!line.empty() && (line.back() == '\\' || line.back() == '/')) {
-                set_display(prompt);
                 replace_last(line.back());
                 is_special_char = true;
             }
@@ -451,6 +1005,15 @@ namespace console {
             }
         }
 
+        if (!end_of_stream && !line.empty()) {
+            // remove the trailing newline for history storage
+            if (!line.empty() && line.back() == '\n') {
+                line.pop_back();
+            }
+            // TODO: maybe support multiline history entries?
+            history.add(line);
+        }
+
         fflush(out);
         return has_more;
     }
@@ -493,12 +1056,82 @@ namespace console {
     }
 
     bool readline(std::string & line, bool multiline_input) {
-        set_display(user_input);
-
         if (simple_io) {
             return readline_simple(line, multiline_input);
         }
         return readline_advanced(line, multiline_input);
     }
 
+    namespace spinner {
+        static const char LOADING_CHARS[] = {'|', '/', '-', '\\'};
+        static std::condition_variable cv_stop;
+        static std::thread th;
+        static size_t frame = 0; // only modified by one thread
+        static bool running = false;
+        static std::mutex mtx;
+        static auto wait_time = std::chrono::milliseconds(100);
+        static void draw_next_frame() {
+            // don't need lock because only one thread modifies running
+            frame = (frame + 1) % sizeof(LOADING_CHARS);
+            replace_last(LOADING_CHARS[frame]);
+            fflush(out);
+        }
+        void start() {
+            std::unique_lock<std::mutex> lock(mtx);
+            if (simple_io || running) {
+                return;
+            }
+            common_log_flush(common_log_main());
+            fprintf(out, "%c", LOADING_CHARS[0]);
+            fflush(out);
+            frame = 1;
+            running = true;
+            th = std::thread([]() {
+                std::unique_lock<std::mutex> lock(mtx);
+                while (true) {
+                    if (cv_stop.wait_for(lock, wait_time, []{ return !running; })) {
+                        break;
+                    }
+                    draw_next_frame();
+                }
+            });
+        }
+        void stop() {
+            {
+                std::unique_lock<std::mutex> lock(mtx);
+                if (simple_io || !running) {
+                    return;
+                }
+                running = false;
+                cv_stop.notify_all();
+            }
+            if (th.joinable()) {
+                th.join();
+            }
+            replace_last(' ');
+            pop_cursor();
+            fflush(out);
+        }
+    }
+
+    void log(const char * fmt, ...) {
+        va_list args;
+        va_start(args, fmt);
+        vfprintf(out, fmt, args);
+        va_end(args);
+    }
+
+    void error(const char * fmt, ...) {
+        va_list args;
+        va_start(args, fmt);
+        display_type cur = current_display;
+        set_display(DISPLAY_TYPE_ERROR);
+        vfprintf(out, fmt, args);
+        set_display(cur); // restore previous color
+        va_end(args);
+    }
+
+    void flush() {
+        fflush(out);
+    }
 }

common/console.h

@@ -2,18 +2,40 @@
 
 #pragma once
 
+#include "common.h"
+
 #include <string>
 
-namespace console {
-    enum display_t {
-        reset = 0,
-        prompt,
-        user_input,
-        error
-    };
+enum display_type {
+    DISPLAY_TYPE_RESET = 0,
+    DISPLAY_TYPE_INFO,
+    DISPLAY_TYPE_PROMPT,
+    DISPLAY_TYPE_REASONING,
+    DISPLAY_TYPE_USER_INPUT,
+    DISPLAY_TYPE_ERROR
+};
 
+namespace console {
     void init(bool use_simple_io, bool use_advanced_display);
     void cleanup();
-    void set_display(display_t display);
+    void set_display(display_type display);
     bool readline(std::string & line, bool multiline_input);
+
+    namespace spinner {
+        void start();
+        void stop();
+    }
+
+    // note: the logging API below output directly to stdout
+    // it can negatively impact performance if used on inference thread
+    // only use in in a dedicated CLI thread
+    // for logging in inference thread, use log.h instead
+
+    LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
+    void log(const char * fmt, ...);
+
+    LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
+    void error(const char * fmt, ...);
+
+    void flush();
 }

common/download.cpp

@@ -12,6 +12,8 @@
 #include <filesystem>
 #include <fstream>
 #include <future>
+#include <map>
+#include <mutex>
 #include <regex>
 #include <string>
 #include <thread>
@@ -472,36 +474,79 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
 
 #elif defined(LLAMA_USE_HTTPLIB)
 
-static bool is_output_a_tty() {
+class ProgressBar {
+    static inline std::mutex mutex;
+    static inline std::map<const ProgressBar *, int> lines;
+    static inline int max_line = 0;
+
+    static void cleanup(const ProgressBar * line) {
+        lines.erase(line);
+        if (lines.empty()) {
+            max_line = 0;
+        }
+    }
+
+    static bool is_output_a_tty() {
 #if defined(_WIN32)
-    return _isatty(_fileno(stdout));
+        return _isatty(_fileno(stdout));
 #else
-    return isatty(1);
+        return isatty(1);
 #endif
-}
-
-static void print_progress(size_t current, size_t total) {
-    if (!is_output_a_tty()) {
-        return;
     }
 
-    if (!total) {
-        return;
+public:
+    ProgressBar() = default;
+
+    ~ProgressBar() {
+        std::lock_guard<std::mutex> lock(mutex);
+        cleanup(this);
     }
 
-    size_t width = 50;
-    size_t pct = (100 * current) / total;
-    size_t pos = (width * current) / total;
+    void update(size_t current, size_t total) {
+        if (!is_output_a_tty()) {
+            return;
+        }
 
-    std::cout << "["
-              << std::string(pos, '=')
-              << (pos < width ? ">" : "")
-              << std::string(width - pos, ' ')
-              << "] " << std::setw(3) << pct << "%  ("
-              << current / (1024 * 1024) << " MB / "
-              << total / (1024 * 1024) << " MB)\r";
-    std::cout.flush();
-}
+        if (!total) {
+            return;
+        }
+
+        std::lock_guard<std::mutex> lock(mutex);
+
+        if (lines.find(this) == lines.end()) {
+            lines[this] = max_line++;
+            std::cout << "\n";
+        }
+        int lines_up = max_line - lines[this];
+
+        size_t width = 50;
+        size_t pct = (100 * current) / total;
+        size_t pos = (width * current) / total;
+
+        std::cout << "\033[s";
+
+        if (lines_up > 0) {
+            std::cout << "\033[" << lines_up << "A";
+        }
+        std::cout << "\033[2K\r["
+            << std::string(pos, '=')
+            << (pos < width ? ">" : "")
+            << std::string(width - pos, ' ')
+            << "] " << std::setw(3) << pct << "%  ("
+            << current / (1024 * 1024) << " MB / "
+            << total / (1024 * 1024) << " MB) "
+            << "\033[u";
+
+        std::cout.flush();
+
+        if (current == total) {
+             cleanup(this);
+        }
+    }
+
+    ProgressBar(const ProgressBar &) = delete;
+    ProgressBar & operator=(const ProgressBar &) = delete;
+};
 
 static bool common_pull_file(httplib::Client & cli,
                              const std::string & resolve_path,
@@ -523,6 +568,7 @@ static bool common_pull_file(httplib::Client & cli,
     const char * func = __func__; // avoid __func__ inside a lambda
     size_t downloaded = existing_size;
     size_t progress_step = 0;
+    ProgressBar bar;
 
     auto res = cli.Get(resolve_path, headers,
         [&](const httplib::Response &response) {
@@ -554,7 +600,7 @@ static bool common_pull_file(httplib::Client & cli,
             progress_step += len;
 
             if (progress_step >= total_size / 1000 || downloaded == total_size) {
-                print_progress(downloaded, total_size);
+                bar.update(downloaded, total_size);
                 progress_step = 0;
             }
             return true;
@@ -562,8 +608,6 @@ static bool common_pull_file(httplib::Client & cli,
         nullptr
     );
 
-    std::cout << "\n";
-
     if (!res) {
         LOG_ERR("%s: error during download. Status: %d\n", __func__, res ? res->status : -1);
         return false;

common/log.cpp

@@ -420,6 +420,11 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
 
+void common_log_flush(struct common_log * log) {
+    log->pause();
+    log->resume();
+}
+
 static int common_get_verbosity(enum ggml_log_level level) {
     switch (level) {
         case GGML_LOG_LEVEL_DEBUG: return LOG_LEVEL_DEBUG;

common/log.h

@@ -84,6 +84,7 @@ void common_log_set_file      (struct common_log * log, const char * file); // n
 void common_log_set_colors    (struct common_log * log, log_colors colors); // not thread-safe
 void common_log_set_prefix    (struct common_log * log, bool prefix);       // whether to output prefix to each log
 void common_log_set_timestamps(struct common_log * log, bool timestamps);   // whether to output timestamps in the prefix
+void common_log_flush         (struct common_log * log);                    // flush all pending log messages
 
 // helper macros for logging
 // use these to avoid computing log arguments if the verbosity of the log is higher than the threshold

common/preset.cpp

@@ -0,0 +1,186 @@
+#include "arg.h"
+#include "preset.h"
+#include "peg-parser.h"
+#include "log.h"
+
+#include <fstream>
+#include <sstream>
+#include <filesystem>
+
+static std::string rm_leading_dashes(const std::string & str) {
+    size_t pos = 0;
+    while (pos < str.size() && str[pos] == '-') {
+        ++pos;
+    }
+    return str.substr(pos);
+}
+
+std::vector<std::string> common_preset::to_args() const {
+    std::vector<std::string> args;
+
+    for (const auto & [opt, value] : options) {
+        args.push_back(opt.args.back()); // use the last arg as the main arg
+        if (opt.value_hint == nullptr && opt.value_hint_2 == nullptr) {
+            // flag option, no value
+            if (common_arg_utils::is_falsey(value)) {
+                // use negative arg if available
+                if (!opt.args_neg.empty()) {
+                    args.back() = opt.args_neg.back();
+                } else {
+                    // otherwise, skip the flag
+                    // TODO: maybe throw an error instead?
+                    args.pop_back();
+                }
+            }
+        }
+        if (opt.value_hint != nullptr) {
+            // single value
+            args.push_back(value);
+        }
+        if (opt.value_hint != nullptr && opt.value_hint_2 != nullptr) {
+            throw std::runtime_error(string_format(
+                "common_preset::to_args(): option '%s' has two values, which is not supported yet",
+                opt.args.back()
+            ));
+        }
+    }
+
+    return args;
+}
+
+std::string common_preset::to_ini() const {
+    std::ostringstream ss;
+
+    ss << "[" << name << "]\n";
+    for (const auto & [opt, value] : options) {
+        auto espaced_value = value;
+        string_replace_all(espaced_value, "\n", "\\\n");
+        ss << rm_leading_dashes(opt.args.back()) << " = ";
+        ss << espaced_value << "\n";
+    }
+    ss << "\n";
+
+    return ss.str();
+}
+
+static std::map<std::string, std::map<std::string, std::string>> parse_ini_from_file(const std::string & path) {
+    std::map<std::string, std::map<std::string, std::string>> parsed;
+
+    if (!std::filesystem::exists(path)) {
+        throw std::runtime_error("preset file does not exist: " + path);
+    }
+
+    std::ifstream file(path);
+    if (!file.good()) {
+        throw std::runtime_error("failed to open server preset file: " + path);
+    }
+
+    std::string contents((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+
+    static const auto parser = build_peg_parser([](auto & p) {
+        // newline ::= "\r\n" / "\n" / "\r"
+        auto newline = p.rule("newline", p.literal("\r\n") | p.literal("\n") | p.literal("\r"));
+
+        // ws ::= [ \t]*
+        auto ws = p.rule("ws", p.chars("[ \t]", 0, -1));
+
+        // comment ::= [;#] (!newline .)*
+        auto comment = p.rule("comment", p.chars("[;#]", 1, 1) + p.zero_or_more(p.negate(newline) + p.any()));
+
+        // eol ::= ws comment? (newline / EOF)
+        auto eol = p.rule("eol", ws + p.optional(comment) + (newline | p.end()));
+
+        // ident ::= [a-zA-Z_] [a-zA-Z0-9_.-]*
+        auto ident = p.rule("ident", p.chars("[a-zA-Z_]", 1, 1) + p.chars("[a-zA-Z0-9_.-]", 0, -1));
+
+        // value ::= (!eol-start .)*
+        auto eol_start = p.rule("eol-start", ws + (p.chars("[;#]", 1, 1) | newline | p.end()));
+        auto value = p.rule("value", p.zero_or_more(p.negate(eol_start) + p.any()));
+
+        // header-line ::= "[" ws ident ws "]" eol
+        auto header_line = p.rule("header-line", "[" + ws + p.tag("section-name", p.chars("[^]]")) + ws + "]" + eol);
+
+        // kv-line ::= ident ws "=" ws value eol
+        auto kv_line = p.rule("kv-line", p.tag("key", ident) + ws + "=" + ws + p.tag("value", value) + eol);
+
+        // comment-line ::= ws comment (newline / EOF)
+        auto comment_line = p.rule("comment-line", ws + comment + (newline | p.end()));
+
+        // blank-line ::= ws (newline / EOF)
+        auto blank_line = p.rule("blank-line", ws + (newline | p.end()));
+
+        // line ::= header-line / kv-line / comment-line / blank-line
+        auto line = p.rule("line", header_line | kv_line | comment_line | blank_line);
+
+        // ini ::= line* EOF
+        auto ini = p.rule("ini", p.zero_or_more(line) + p.end());
+
+        return ini;
+    });
+
+    common_peg_parse_context ctx(contents);
+    const auto result = parser.parse(ctx);
+    if (!result.success()) {
+        throw std::runtime_error("failed to parse server config file: " + path);
+    }
+
+    std::string current_section = COMMON_PRESET_DEFAULT_NAME;
+    std::string current_key;
+
+    ctx.ast.visit(result, [&](const auto & node) {
+        if (node.tag == "section-name") {
+            const std::string section = std::string(node.text);
+            current_section = section;
+            parsed[current_section] = {};
+        } else if (node.tag == "key") {
+            const std::string key = std::string(node.text);
+            current_key = key;
+        } else if (node.tag == "value" && !current_key.empty() && !current_section.empty()) {
+            parsed[current_section][current_key] = std::string(node.text);
+            current_key.clear();
+        }
+    });
+
+    return parsed;
+}
+
+static std::map<std::string, common_arg> get_map_key_opt(common_params_context & ctx_params) {
+    std::map<std::string, common_arg> mapping;
+    for (const auto & opt : ctx_params.options) {
+        for (const auto & env : opt.get_env()) {
+            mapping[env] = opt;
+        }
+        for (const auto & arg : opt.get_args()) {
+            mapping[rm_leading_dashes(arg)] = opt;
+        }
+    }
+    return mapping;
+}
+
+common_presets common_presets_load(const std::string & path, common_params_context & ctx_params) {
+    common_presets out;
+    auto key_to_opt = get_map_key_opt(ctx_params);
+    auto ini_data = parse_ini_from_file(path);
+
+    for (auto section : ini_data) {
+        common_preset preset;
+        if (section.first.empty()) {
+            preset.name = COMMON_PRESET_DEFAULT_NAME;
+        } else {
+            preset.name = section.first;
+        }
+        LOG_DBG("loading preset: %s\n", preset.name.c_str());
+        for (const auto & [key, value] : section.second) {
+            LOG_DBG("option: %s = %s\n", key.c_str(), value.c_str());
+            if (key_to_opt.find(key) != key_to_opt.end()) {
+                preset.options[key_to_opt[key]] = value;
+                LOG_DBG("accepted option: %s = %s\n", key.c_str(), value.c_str());
+            } else {
+                // TODO: maybe warn about unknown key?
+            }
+        }
+        out[preset.name] = preset;
+    }
+
+    return out;
+}

common/preset.h

@@ -0,0 +1,32 @@
+#pragma once
+
+#include "common.h"
+#include "arg.h"
+
+#include <string>
+#include <vector>
+#include <map>
+
+//
+// INI preset parser and writer
+//
+
+constexpr const char * COMMON_PRESET_DEFAULT_NAME = "default";
+
+struct common_preset {
+    std::string name;
+    // TODO: support repeated args in the future
+    std::map<common_arg, std::string> options;
+
+    // convert preset to CLI argument list
+    std::vector<std::string> to_args() const;
+
+    // convert preset to INI format string
+    std::string to_ini() const;
+
+    // TODO: maybe implement to_env() if needed
+};
+
+// interface for multiple presets in one file
+using common_presets = std::map<std::string, common_preset>;
+common_presets common_presets_load(const std::string & path, common_params_context & ctx_params);

convert_hf_to_gguf.py

@@ -383,6 +383,17 @@ class ModelBase:
                         s = self.model_tensors[name]
                         self.model_tensors[weight_name] = lambda w=w, s=s, bs=block_size: dequant_simple(w(), s(), bs)
                         tensors_to_remove.append(name)
+                    if name.endswith(".activation_scale"):  # unused
+                        tensors_to_remove.append(name)
+                    # mistral format
+                    if name.endswith(".qscale_weight"):
+                        weight_name = name.removesuffix("qscale_weight") + "weight"
+                        w = self.model_tensors[weight_name]
+                        s = self.model_tensors[name]
+                        self.model_tensors[weight_name] = lambda w=w, s=s, bs=block_size: dequant_simple(w(), s(), bs)
+                        tensors_to_remove.append(name)
+                    if name.endswith(".qscale_act"):
+                        tensors_to_remove.append(name)
             elif quant_method == "gptq":
                 for name in self.model_tensors.keys():
                     if name.endswith(".qweight"):
@@ -2854,13 +2865,10 @@ class Mistral3Model(LlamaModel):
             self.gguf_writer.add_attn_temperature_scale(rope_params["llama_4_scaling_beta"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
-        # TODO: probably not worth supporting quantized weight, as official BF16 is also available
-        if name.endswith("weight_scale_inv"):
-            raise ValueError("This is a quantized weight, please use BF16 weight instead")
-
         name = name.replace("language_model.", "")
         if "multi_modal_projector" in name or "vision_tower" in name:
             return []
+
         return super().modify_tensors(data_torch, name, bid)
 
 
@@ -5825,9 +5833,11 @@ class Gemma3Model(TextModel):
     norm_shift = 1.0  # Gemma3RMSNorm adds 1.0 to the norm value
 
     def set_vocab(self):
-        self._set_vocab_sentencepiece()
-
-        self.gguf_writer.add_add_space_prefix(False)
+        if (self.dir_model / "tokenizer.model").is_file():
+            self._set_vocab_sentencepiece()
+            self.gguf_writer.add_add_space_prefix(False)
+        else:
+            self._set_vocab_gpt2()
 
     def set_gguf_parameters(self):
         hparams = self.hparams
@@ -5845,13 +5855,24 @@ class Gemma3Model(TextModel):
         self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1_000_000.0)) # for global layers
         # attn_logit_softcapping is removed in Gemma3
         assert hparams.get("attn_logit_softcapping") is None
-        self.gguf_writer.add_sliding_window(hparams["sliding_window"])
+        if (final_logit_softcap := hparams.get("final_logit_softcapping")):
+            self.gguf_writer.add_final_logit_softcapping(final_logit_softcap)
+        if hparams.get("sliding_window_pattern") != 1:
+            self.gguf_writer.add_sliding_window(hparams["sliding_window"])
         self.gguf_writer.add_head_count_kv(hparams.get("num_key_value_heads", 4))
         if hparams.get("rope_scaling") is not None:
-            assert hparams["rope_scaling"]["rope_type"] == "linear"
-            # important: this rope_scaling is only applied for global layers, and not used by 1B model
-            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
-            self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
+            rope_scaling = hparams["rope_scaling"]
+            if rope_scaling["rope_type"] == "linear":
+                # important: this rope_scaling is only applied for global layers, and not used by 1B model
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            elif rope_scaling["rope_type"] == "yarn":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+                self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+                self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+                self.gguf_writer.add_rope_scaling_yarn_ext_factor(rope_scaling["extrapolation_factor"])
+                self.gguf_writer.add_rope_scaling_yarn_beta_fast(rope_scaling["beta_fast"])
+                self.gguf_writer.add_rope_scaling_yarn_beta_slow(rope_scaling["beta_slow"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -5865,8 +5886,10 @@ class Gemma3Model(TextModel):
 
         # remove OOV (out-of-vocabulary) rows in token_embd
         if "embed_tokens.weight" in name:
-            vocab = self._create_vocab_sentencepiece()
-            tokens = vocab[0]
+            if (self.dir_model / "tokenizer.model").is_file():
+                tokens = self._create_vocab_sentencepiece()[0]
+            else:
+                tokens = self.get_vocab_base()[0]
             data_torch = data_torch[:len(tokens)]
 
         # ref code in Gemma3RMSNorm
@@ -7263,6 +7286,10 @@ class DeepseekV2Model(TextModel):
             self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
             self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
             self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+
+            # [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+            # note: for legacy reasons, this is not consistent with the other usages of self.gguf_writer.add_rope_scaling_yarn_log_mul
+            # ref https://github.com/ggml-org/llama.cpp/pull/17945
             self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1 * rope_scaling["mscale_all_dim"])
 
     _experts: list[dict[str, Tensor]] | None = None
@@ -9883,6 +9910,18 @@ class MistralModel(LlamaModel):
             self.gguf_writer.add_architecture()
             self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
+    def dequant_model(self):
+        # transform quantization config into HF format
+        quant_config = self.hparams.get("quantization")
+        if quant_config is not None:
+            assert quant_config["qformat_weight"] == "fp8_e4m3"
+            self.hparams["quantization_config"] = {
+                "activation_scheme": "static",
+                "quant_method": "fp8",
+                "weight_block_size": None,
+            }
+        return super().dequant_model()
+
     @staticmethod
     def get_community_chat_template(vocab: MistralVocab, templates_dir: Path, is_mistral_format: bool):
         assert TokenizerVersion is not None and Tekkenizer is not None and SentencePieceTokenizer is not None, _mistral_import_error_msg
@@ -10006,6 +10045,10 @@ class MistralMoeModel(DeepseekV2Model):
         MistralModel.set_mistral_config(self.gguf_writer, self.hparams)
         yarn_params = self.hparams["yarn"]
         self.gguf_writer.add_attn_temperature_length(yarn_params["original_max_position_embeddings"])
+
+        # [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+        # note: for legacy reasons, this is not consistent with the other usages of self.gguf_writer.add_rope_scaling_yarn_log_mul
+        # ref https://github.com/ggml-org/llama.cpp/pull/17945
         self.gguf_writer.add_rope_scaling_yarn_log_mul(0.1) # mscale_all_dim * 0.1
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):

docs/docker.md

@@ -56,7 +56,7 @@ docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:light -m /model
 or with a server image:
 
 ```bash
-docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggml-org/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
+docker run -v /path/to/models:/models -p 8080:8080 ghcr.io/ggml-org/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8080 --host 0.0.0.0 -n 512
 ```
 
 ## Docker With CUDA
@@ -91,7 +91,7 @@ After building locally, Usage is similar to the non-CUDA examples, but you'll ne
 ```bash
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
-docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8080 --host 0.0.0.0 -n 512 --n-gpu-layers 1
 ```
 
 ## Docker With MUSA
@@ -125,5 +125,5 @@ After building locally, Usage is similar to the non-MUSA examples, but you'll ne
 ```bash
 docker run -v /path/to/models:/models local/llama.cpp:full-musa --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
 docker run -v /path/to/models:/models local/llama.cpp:light-musa -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
-docker run -v /path/to/models:/models local/llama.cpp:server-musa -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
+docker run -v /path/to/models:/models local/llama.cpp:server-musa -m /models/7B/ggml-model-q4_0.gguf --port 8080 --host 0.0.0.0 -n 512 --n-gpu-layers 1
 ```

docs/ops.md

@@ -16,12 +16,12 @@ Legend:
 |-----------|------|------|------|------|------|------|------|------|------|------|------|
 |                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -36,15 +36,16 @@ Legend:
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                           CUMSUM | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                             DIAG | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
 |                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                             FILL | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                             FILL | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -64,8 +65,8 @@ Legend:
 |                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
-|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -74,7 +75,7 @@ Legend:
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -83,7 +84,7 @@ Legend:
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
@@ -102,21 +103,21 @@ Legend:
 |                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
-|                        SOLVE_TRI | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
-|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              SUM | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ |
+|                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            TOP_K | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
-|                              TRI | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
-|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
+|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ |
 |                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |

docs/ops/CPU.csv

@@ -4964,6 +4964,7 @@
 "CPU","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1","support","1","yes","CPU"
 "CPU","CONV_TRANSPOSE_2D","ne_input=[3,2,3,1],ne_kernel=[2,2,1,3],stride=1","support","1","yes","CPU"
 "CPU","CONV_TRANSPOSE_2D","ne_input=[10,10,9,1],ne_kernel=[3,3,1,9],stride=2","support","1","yes","CPU"
+"CPU","CONV_TRANSPOSE_2D","ne_input=[129,63,35,1],ne_kernel=[3,3,48,35],stride=1","support","1","yes","CPU"
 "CPU","COUNT_EQUAL","type=f32,ne=[4,500,1,1]","support","1","yes","CPU"
 "CPU","COUNT_EQUAL","type=f32,ne=[4,5000,1,1]","support","1","yes","CPU"
 "CPU","ARGMAX","type=f32,ne=[32,1,1,1]","support","1","yes","CPU"
@@ -5419,17 +5420,45 @@
 "CPU","CPY","type_src=f16,type_dst=f16,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","CPU"
 "CPU","CPY","type_src=f32,type_dst=f32,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","CPU"
 "CPU","CPY","type_src=bf16,type_dst=bf16,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","CPU"
+"CPU","CPY","type_src=i32,type_dst=i32,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","CPU"
+"CPU","CPY","type_src=i32,type_dst=i32,ne=[256,1,4,1],permute_src=[1,2,0,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","CPU"
 "CPU","CPY","type_src=f32,type_dst=f32,ne=[256,1,4,1],permute_src=[1,2,0,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","CPU"
-"CPU","CONT","type=f32,ne=[10,10,10,1]","support","1","yes","CPU"
-"CPU","CONT","type=f32,ne=[2,1,1,1]","support","1","yes","CPU"
-"CPU","CONT","type=f32,ne=[2,1,3,5]","support","1","yes","CPU"
-"CPU","CONT","type=f32,ne=[2,3,5,7]","support","1","yes","CPU"
-"CPU","CONT","type=f16,ne=[2,1,1,1]","support","1","yes","CPU"
-"CPU","CONT","type=f16,ne=[2,1,3,5]","support","1","yes","CPU"
-"CPU","CONT","type=f16,ne=[2,3,5,7]","support","1","yes","CPU"
-"CPU","CONT","type=bf16,ne=[2,1,1,1]","support","1","yes","CPU"
-"CPU","CONT","type=bf16,ne=[2,1,3,5]","support","1","yes","CPU"
-"CPU","CONT","type=bf16,ne=[2,3,5,7]","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[2,1,1,1],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[2,1,3,5],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[2,3,5,7],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[1,4,4,1],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[1,8,17,1],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[10,10,10,1],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[2,1,1,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[2,1,3,5],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[2,3,5,7],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[1,4,4,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[1,8,17,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f32,ne=[10,10,10,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[2,1,1,1],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[2,1,3,5],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[2,3,5,7],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[1,4,4,1],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[1,8,17,1],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[10,10,10,1],use_view_slice=1","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[2,1,1,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[2,1,3,5],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[2,3,5,7],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[1,4,4,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[1,8,17,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=i32,ne=[10,10,10,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f16,ne=[2,1,1,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f16,ne=[2,1,3,5],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f16,ne=[2,3,5,7],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f16,ne=[1,4,4,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f16,ne=[1,8,17,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=f16,ne=[10,10,10,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=bf16,ne=[2,1,1,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=bf16,ne=[2,1,3,5],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=bf16,ne=[2,3,5,7],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=bf16,ne=[1,4,4,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=bf16,ne=[1,8,17,1],use_view_slice=0","support","1","yes","CPU"
+"CPU","CONT","type=bf16,ne=[10,10,10,1],use_view_slice=0","support","1","yes","CPU"
 "CPU","ADD","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","1","yes","CPU"
 "CPU","SUB","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","1","yes","CPU"
 "CPU","MUL","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","1","yes","CPU"
@@ -5655,6 +5684,7 @@
 "CPU","MUL","type=f32,ne=[64,262144,1,1],nr=[1,1,1,1],nf=1","support","1","yes","CPU"
 "CPU","DIV","type=f32,ne=[64,262144,1,1],nr=[1,1,1,1],nf=1","support","1","yes","CPU"
 "CPU","ADD1","type=f32,ne=[10,5,4,3]","support","1","yes","CPU"
+"CPU","ADD1","type=f32,ne=[1024,1024,1,1]","support","1","yes","CPU"
 "CPU","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=0.000000,inplace=0","support","1","yes","CPU"
 "CPU","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=1.000000,inplace=0","support","1","yes","CPU"
 "CPU","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=1.000000,inplace=1","support","1","yes","CPU"
@@ -8644,9 +8674,13 @@
 "CPU","CLAMP","type=f16,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","CPU"
 "CPU","LEAKY_RELU","type=f16,ne_a=[7,1,5,3],negative_slope=0.100000","support","1","yes","CPU"
 "CPU","FLOOR","type=f16,ne=[7,1,5,3]","support","1","yes","CPU"
+"CPU","FLOOR","type=f16,ne=[1024,1024,1,1]","support","1","yes","CPU"
 "CPU","CEIL","type=f16,ne=[7,1,5,3]","support","1","yes","CPU"
+"CPU","CEIL","type=f16,ne=[1024,1024,1,1]","support","1","yes","CPU"
 "CPU","ROUND","type=f16,ne=[7,1,5,3]","support","1","yes","CPU"
+"CPU","ROUND","type=f16,ne=[1024,1024,1,1]","support","1","yes","CPU"
 "CPU","TRUNC","type=f16,ne=[7,1,5,3]","support","1","yes","CPU"
+"CPU","TRUNC","type=f16,ne=[1024,1024,1,1]","support","1","yes","CPU"
 "CPU","SQR","type=f32,ne=[10,5,4,3]","support","1","yes","CPU"
 "CPU","SQRT","type=f32,ne=[10,3,3,2]","support","1","yes","CPU"
 "CPU","LOG","type=f32,ne=[10,5,4,3]","support","1","yes","CPU"
@@ -8666,9 +8700,13 @@
 "CPU","CLAMP","type=f32,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","CPU"
 "CPU","LEAKY_RELU","type=f32,ne_a=[7,1,5,3],negative_slope=0.100000","support","1","yes","CPU"
 "CPU","FLOOR","type=f32,ne=[7,1,5,3]","support","1","yes","CPU"
+"CPU","FLOOR","type=f32,ne=[1024,1024,1,1]","support","1","yes","CPU"
 "CPU","CEIL","type=f32,ne=[7,1,5,3]","support","1","yes","CPU"
+"CPU","CEIL","type=f32,ne=[1024,1024,1,1]","support","1","yes","CPU"
 "CPU","ROUND","type=f32,ne=[7,1,5,3]","support","1","yes","CPU"
+"CPU","ROUND","type=f32,ne=[1024,1024,1,1]","support","1","yes","CPU"
 "CPU","TRUNC","type=f32,ne=[7,1,5,3]","support","1","yes","CPU"
+"CPU","TRUNC","type=f32,ne=[1024,1024,1,1]","support","1","yes","CPU"
 "CPU","DIAG_MASK_INF","type=f32,ne=[10,10,1,1],n_past=5","support","1","yes","CPU"
 "CPU","DIAG_MASK_INF","type=f32,ne=[10,10,3,1],n_past=5","support","1","yes","CPU"
 "CPU","DIAG_MASK_INF","type=f32,ne=[10,10,3,2],n_past=5","support","1","yes","CPU"
@@ -9411,18 +9449,405 @@
 "CPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","1","yes","CPU"
 "CPU","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","1","yes","CPU"
 "CPU","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[3,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[4,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[7,1,1,1],order=0","support","1","yes","CPU"
 "CPU","ARGSORT","type=f32,ne=[8,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[15,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[16,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[31,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[32,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[63,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[64,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[127,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[128,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[255,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[256,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[511,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[512,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1023,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1024,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[2047,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[2048,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[4095,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[4096,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[8191,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[8192,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[16383,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[16384,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[32767,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[32768,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[65535,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[65536,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[131071,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[131072,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[262143,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[262144,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[524287,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[524288,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1048575,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1048576,1,1,1],order=0","support","1","yes","CPU"
 "CPU","ARGSORT","type=f32,ne=[16,10,10,10],order=0","support","1","yes","CPU"
 "CPU","ARGSORT","type=f32,ne=[60,10,10,10],order=0","support","1","yes","CPU"
-"CPU","ARGSORT","type=f32,ne=[1024,1,1,1],order=0","support","1","yes","CPU"
-"CPU","ARGSORT","type=f32,ne=[16384,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1023,2,1,3],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1024,2,1,3],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1025,2,1,3],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[2047,2,1,3],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[2048,2,1,3],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[2049,2,1,3],order=0","support","1","yes","CPU"
 "CPU","ARGSORT","type=f32,ne=[2,8,8192,1],order=0","support","1","yes","CPU"
-"CPU","ARGSORT","type=f32,ne=[8,1,1,1],order=1","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[3,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[4,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[7,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[8,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[15,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[16,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[31,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[32,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[63,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[64,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[127,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[128,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[255,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[256,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[511,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[512,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1023,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1024,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[2047,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[2048,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[4095,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[4096,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[8191,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[8192,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[16383,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[16384,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[32767,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[32768,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[65535,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[65536,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[131071,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[131072,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[262143,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[262144,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[524287,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[524288,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1048575,1,1,1],order=0","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1048576,1,1,1],order=0","support","1","yes","CPU"
 "CPU","ARGSORT","type=f32,ne=[16,10,10,10],order=1","support","1","yes","CPU"
 "CPU","ARGSORT","type=f32,ne=[60,10,10,10],order=1","support","1","yes","CPU"
-"CPU","ARGSORT","type=f32,ne=[1024,1,1,1],order=1","support","1","yes","CPU"
-"CPU","ARGSORT","type=f32,ne=[16384,1,1,1],order=1","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1023,2,1,3],order=1","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1024,2,1,3],order=1","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[1025,2,1,3],order=1","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[2047,2,1,3],order=1","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[2048,2,1,3],order=1","support","1","yes","CPU"
+"CPU","ARGSORT","type=f32,ne=[2049,2,1,3],order=1","support","1","yes","CPU"
 "CPU","ARGSORT","type=f32,ne=[2,8,8192,1],order=1","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[12,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[13,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[13,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[15,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[15,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[15,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[19,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[19,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[19,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[19,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[27,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[27,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[27,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[27,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[27,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[43,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[43,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[43,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[43,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[43,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[64,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[75,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[64,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[75,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[64,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[75,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[64,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[75,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[64,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[75,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[128,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[139,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[128,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[139,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[128,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[139,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[128,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[139,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[128,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[139,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[128,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[139,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[256,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[267,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[256,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[267,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[256,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[267,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[256,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[267,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[256,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[267,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[256,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[267,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[512,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[523,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[512,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[523,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[512,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[523,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[512,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[523,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[512,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[523,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[512,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[523,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[512,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[523,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1035,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1035,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1035,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1035,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1035,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1035,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1035,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,1,1,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1035,1,2,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2059,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2059,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2059,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2059,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2059,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2059,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2059,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,1,1,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2059,1,2,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4096,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4107,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4096,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4107,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4096,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4107,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4096,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4107,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4096,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4107,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4096,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4107,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4096,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4107,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4096,1,1,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[4107,1,2,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8192,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8203,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8192,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8203,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8192,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8203,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8192,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8203,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8192,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8203,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8192,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8203,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8192,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8203,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8192,1,1,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[8203,1,2,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16395,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16395,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16395,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16395,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16395,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16395,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16395,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16395,1,2,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16395,1,2,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32768,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32779,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32768,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32779,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32768,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32779,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32768,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32779,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32768,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32779,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32768,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32779,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32768,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32779,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32768,1,1,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32779,1,2,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32768,1,1,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[32779,1,2,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65536,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65547,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65536,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65547,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65536,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65547,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65536,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65547,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65536,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65547,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65536,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65547,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65536,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65547,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65536,1,1,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65547,1,2,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65536,1,1,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[65547,1,2,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131072,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131083,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131072,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131083,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131072,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131083,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131072,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131083,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131072,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131083,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131072,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131083,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131072,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131083,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131072,1,1,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131083,1,2,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131072,1,1,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[131083,1,2,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262144,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262155,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262144,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262155,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262144,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262155,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262144,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262155,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262144,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262155,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262144,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262155,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262144,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262155,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262144,1,1,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262155,1,2,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262144,1,1,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[262155,1,2,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524288,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524299,1,2,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524288,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524299,1,2,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524288,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524299,1,2,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524288,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524299,1,2,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524288,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524299,1,2,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524288,1,1,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524299,1,2,1],k=100,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524288,1,1,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524299,1,2,1],k=500,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524288,1,1,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524299,1,2,1],k=1023,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524288,1,1,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[524299,1,2,1],k=9999,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16,10,10,10],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[60,10,10,10],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1023,2,1,3],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,2,1,3],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1025,2,1,3],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2047,2,1,3],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,2,1,3],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2049,2,1,3],k=1,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16,10,10,10],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[60,10,10,10],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1023,2,1,3],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,2,1,3],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1025,2,1,3],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2047,2,1,3],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,2,1,3],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2049,2,1,3],k=2,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16,10,10,10],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[60,10,10,10],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1023,2,1,3],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,2,1,3],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1025,2,1,3],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2047,2,1,3],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,2,1,3],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2049,2,1,3],k=3,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16,10,10,10],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[60,10,10,10],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1023,2,1,3],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,2,1,3],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1025,2,1,3],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2047,2,1,3],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,2,1,3],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2049,2,1,3],k=7,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16,10,10,10],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[60,10,10,10],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1023,2,1,3],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1024,2,1,3],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[1025,2,1,3],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2047,2,1,3],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2048,2,1,3],k=15,ties=0","support","1","yes","CPU"
+"CPU","TOP_K","type=f32,ne=[2049,2,1,3],k=15,ties=0","support","1","yes","CPU"
 "CPU","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=0","support","1","yes","CPU"
 "CPU","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","1","yes","CPU"
 "CPU","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest,flags=none","support","1","yes","CPU"
@@ -9435,6 +9860,10 @@
 "CPU","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=1","support","1","yes","CPU"
 "CPU","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic,flags=none","support","1","yes","CPU"
 "CPU","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic,flags=none","support","1","yes","CPU"
+"CPU","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=0","support","1","yes","CPU"
+"CPU","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=1","support","1","yes","CPU"
+"CPU","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=none","support","1","yes","CPU"
+"CPU","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear,flags=none","support","1","yes","CPU"
 "CPU","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=align_corners","support","1","yes","CPU"
 "CPU","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear,flags=align_corners","support","1","yes","CPU"
 "CPU","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear,flags=align_corners","support","1","yes","CPU"
@@ -9463,15 +9892,30 @@
 "CPU","GROUP_NORM","type=f32,ne=[64,64,320,1],num_groups=32,eps=0.000001","support","1","yes","CPU"
 "CPU","GROUP_NORM","type=f32,ne=[9,9,1280,1],num_groups=32,eps=0.000001","support","1","yes","CPU"
 "CPU","ACC","type=f32,ne_a=[256,17,1,1],ne_b=[256,16,1,1]","support","1","yes","CPU"
-"CPU","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1","support","1","yes","CPU"
-"CPU","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1,circular=0","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[33,17,2,1],pad_0=4,pad_1=3,circular=1","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0,circular=0","support","1","yes","CPU"
 "CPU","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","1","yes","CPU"
 "CPU","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9","support","1","yes","CPU"
 "CPU","ROLL","shift0=3,shift1=-2,shift3=1,shift4=-1","support","1","yes","CPU"
 "CPU","ARANGE","type=f32,start=0.000000,stop=10.000000,step=1.000000","support","1","yes","CPU"
+"CPU","ARANGE","type=f32,start=0.000000,stop=1048576.000000,step=1.000000","support","1","yes","CPU"
 "CPU","TIMESTEP_EMBEDDING","type=f32,ne_a=[2,1,1,1],dim=320,max_period=10000","support","1","yes","CPU"
 "CPU","LEAKY_RELU","type=f32,ne_a=[10,5,4,3],negative_slope=0.100000","support","1","yes","CPU"
 "CPU","CUMSUM","type=f32,ne=[10,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[127,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[128,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[128,128,4,4]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[255,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[256,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[511,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[512,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[1023,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[1024,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[2047,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[2048,5,4,3]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[242004,1,1,1]","support","1","yes","CPU"
+"CPU","CUMSUM","type=f32,ne=[375960,1,1,1]","support","1","yes","CPU"
 "CPU","XIELU","type=f32,ne=[10,5,4,3]","support","1","yes","CPU"
 "CPU","TRI","type=f32,ne=[10,10,4,3],tri_type=3","support","1","yes","CPU"
 "CPU","TRI","type=f32,ne=[10,10,4,3],tri_type=2","support","1","yes","CPU"
@@ -9480,6 +9924,10 @@
 "CPU","FILL","type=f32,ne=[10,10,4,3],c=0.000000","support","1","yes","CPU"
 "CPU","FILL","type=f32,ne=[303,207,11,3],c=2.000000","support","1","yes","CPU"
 "CPU","FILL","type=f32,ne=[800,600,4,4],c=-152.000000","support","1","yes","CPU"
+"CPU","FILL","type=f32,ne=[2048,512,2,2],c=3.500000","support","1","yes","CPU"
+"CPU","DIAG","type=f32,ne=[10,1,4,3]","support","1","yes","CPU"
+"CPU","DIAG","type=f32,ne=[79,1,19,13]","support","1","yes","CPU"
+"CPU","DIAG","type=f32,ne=[256,1,8,16]","support","1","yes","CPU"
 "CPU","SOLVE_TRI","type=f32,ne_lhs=[10,10,4,3],ne_rhs=[3,10,4,3]","support","1","yes","CPU"
 "CPU","SOLVE_TRI","type=f32,ne_lhs=[11,11,1,1],ne_rhs=[5,11,1,1]","support","1","yes","CPU"
 "CPU","SOLVE_TRI","type=f32,ne_lhs=[17,17,2,4],ne_rhs=[9,17,2,4]","support","1","yes","CPU"
@@ -9487,10 +9935,16 @@
 "CPU","SOLVE_TRI","type=f32,ne_lhs=[42,42,5,2],ne_rhs=[10,42,5,2]","support","1","yes","CPU"
 "CPU","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[10,64,2,2]","support","1","yes","CPU"
 "CPU","SOLVE_TRI","type=f32,ne_lhs=[100,100,4,4],ne_rhs=[41,100,4,4]","support","1","yes","CPU"
-"CPU","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0","support","1","yes","CPU"
-"CPU","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0","support","1","yes","CPU"
-"CPU","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1","support","1","yes","CPU"
-"CPU","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1","support","1","yes","CPU"
+"CPU","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,4],ne_rhs=[31,128,4,4]","support","1","yes","CPU"
+"CPU","SOLVE_TRI","type=f32,ne_lhs=[64,64,4,4],ne_rhs=[300,64,4,4]","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0,circular=0","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0,circular=0","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0,circular=1","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0,circular=1","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1,circular=0","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1,circular=0","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1,circular=1","support","1","yes","CPU"
+"CPU","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1,circular=1","support","1","yes","CPU"
 "CPU","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","1","yes","CPU"
 "CPU","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","1","yes","CPU"
 "CPU","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","1","yes","CPU"

docs/ops/CUDA.csv

@@ -4964,6 +4964,7 @@
 "CUDA0","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1","support","1","yes","CUDA"
 "CUDA0","CONV_TRANSPOSE_2D","ne_input=[3,2,3,1],ne_kernel=[2,2,1,3],stride=1","support","1","yes","CUDA"
 "CUDA0","CONV_TRANSPOSE_2D","ne_input=[10,10,9,1],ne_kernel=[3,3,1,9],stride=2","support","1","yes","CUDA"
+"CUDA0","CONV_TRANSPOSE_2D","ne_input=[129,63,35,1],ne_kernel=[3,3,48,35],stride=1","support","1","yes","CUDA"
 "CUDA0","COUNT_EQUAL","type=f32,ne=[4,500,1,1]","support","1","yes","CUDA"
 "CUDA0","COUNT_EQUAL","type=f32,ne=[4,5000,1,1]","support","1","yes","CUDA"
 "CUDA0","ARGMAX","type=f32,ne=[32,1,1,1]","support","1","yes","CUDA"
@@ -5419,17 +5420,45 @@
 "CUDA0","CPY","type_src=f16,type_dst=f16,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","CUDA"
 "CUDA0","CPY","type_src=f32,type_dst=f32,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","CUDA"
 "CUDA0","CPY","type_src=bf16,type_dst=bf16,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","CUDA"
+"CUDA0","CPY","type_src=i32,type_dst=i32,ne=[256,4,1,1],permute_src=[0,0,0,0],permute_dst=[0,0,0,0],_src_transpose=1","support","1","yes","CUDA"
+"CUDA0","CPY","type_src=i32,type_dst=i32,ne=[256,1,4,1],permute_src=[1,2,0,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","CUDA"
 "CUDA0","CPY","type_src=f32,type_dst=f32,ne=[256,1,4,1],permute_src=[1,2,0,3],permute_dst=[0,0,0,0],_src_transpose=0","support","1","yes","CUDA"
-"CUDA0","CONT","type=f32,ne=[10,10,10,1]","support","1","yes","CUDA"
-"CUDA0","CONT","type=f32,ne=[2,1,1,1]","support","1","yes","CUDA"
-"CUDA0","CONT","type=f32,ne=[2,1,3,5]","support","1","yes","CUDA"
-"CUDA0","CONT","type=f32,ne=[2,3,5,7]","support","1","yes","CUDA"
-"CUDA0","CONT","type=f16,ne=[2,1,1,1]","support","1","yes","CUDA"
-"CUDA0","CONT","type=f16,ne=[2,1,3,5]","support","1","yes","CUDA"
-"CUDA0","CONT","type=f16,ne=[2,3,5,7]","support","1","yes","CUDA"
-"CUDA0","CONT","type=bf16,ne=[2,1,1,1]","support","1","yes","CUDA"
-"CUDA0","CONT","type=bf16,ne=[2,1,3,5]","support","1","yes","CUDA"
-"CUDA0","CONT","type=bf16,ne=[2,3,5,7]","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[2,1,1,1],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[2,1,3,5],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[2,3,5,7],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[1,4,4,1],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[1,8,17,1],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[10,10,10,1],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[2,1,1,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[2,1,3,5],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[2,3,5,7],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[1,4,4,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[1,8,17,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f32,ne=[10,10,10,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[2,1,1,1],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[2,1,3,5],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[2,3,5,7],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[1,4,4,1],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[1,8,17,1],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[10,10,10,1],use_view_slice=1","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[2,1,1,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[2,1,3,5],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[2,3,5,7],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[1,4,4,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[1,8,17,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=i32,ne=[10,10,10,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f16,ne=[2,1,1,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f16,ne=[2,1,3,5],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f16,ne=[2,3,5,7],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f16,ne=[1,4,4,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f16,ne=[1,8,17,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=f16,ne=[10,10,10,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=bf16,ne=[2,1,1,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=bf16,ne=[2,1,3,5],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=bf16,ne=[2,3,5,7],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=bf16,ne=[1,4,4,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=bf16,ne=[1,8,17,1],use_view_slice=0","support","1","yes","CUDA"
+"CUDA0","CONT","type=bf16,ne=[10,10,10,1],use_view_slice=0","support","1","yes","CUDA"
 "CUDA0","ADD","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","1","yes","CUDA"
 "CUDA0","SUB","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","1","yes","CUDA"
 "CUDA0","MUL","type=f16,ne=[1,1,8,1],nr=[1,1,1,1],nf=1","support","1","yes","CUDA"
@@ -5655,6 +5684,7 @@
 "CUDA0","MUL","type=f32,ne=[64,262144,1,1],nr=[1,1,1,1],nf=1","support","1","yes","CUDA"
 "CUDA0","DIV","type=f32,ne=[64,262144,1,1],nr=[1,1,1,1],nf=1","support","1","yes","CUDA"
 "CUDA0","ADD1","type=f32,ne=[10,5,4,3]","support","1","yes","CUDA"
+"CUDA0","ADD1","type=f32,ne=[1024,1024,1,1]","support","1","yes","CUDA"
 "CUDA0","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=0.000000,inplace=0","support","1","yes","CUDA"
 "CUDA0","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=1.000000,inplace=0","support","1","yes","CUDA"
 "CUDA0","SCALE","type=f32,ne=[10,10,10,10],scale=2.000000,bias=1.000000,inplace=1","support","1","yes","CUDA"
@@ -8644,9 +8674,13 @@
 "CUDA0","CLAMP","type=f16,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","CUDA"
 "CUDA0","LEAKY_RELU","type=f16,ne_a=[7,1,5,3],negative_slope=0.100000","support","1","yes","CUDA"
 "CUDA0","FLOOR","type=f16,ne=[7,1,5,3]","support","1","yes","CUDA"
+"CUDA0","FLOOR","type=f16,ne=[1024,1024,1,1]","support","1","yes","CUDA"
 "CUDA0","CEIL","type=f16,ne=[7,1,5,3]","support","1","yes","CUDA"
+"CUDA0","CEIL","type=f16,ne=[1024,1024,1,1]","support","1","yes","CUDA"
 "CUDA0","ROUND","type=f16,ne=[7,1,5,3]","support","1","yes","CUDA"
+"CUDA0","ROUND","type=f16,ne=[1024,1024,1,1]","support","1","yes","CUDA"
 "CUDA0","TRUNC","type=f16,ne=[7,1,5,3]","support","1","yes","CUDA"
+"CUDA0","TRUNC","type=f16,ne=[1024,1024,1,1]","support","1","yes","CUDA"
 "CUDA0","SQR","type=f32,ne=[10,5,4,3]","support","1","yes","CUDA"
 "CUDA0","SQRT","type=f32,ne=[10,3,3,2]","support","1","yes","CUDA"
 "CUDA0","LOG","type=f32,ne=[10,5,4,3]","support","1","yes","CUDA"
@@ -8666,9 +8700,13 @@
 "CUDA0","CLAMP","type=f32,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","CUDA"
 "CUDA0","LEAKY_RELU","type=f32,ne_a=[7,1,5,3],negative_slope=0.100000","support","1","yes","CUDA"
 "CUDA0","FLOOR","type=f32,ne=[7,1,5,3]","support","1","yes","CUDA"
+"CUDA0","FLOOR","type=f32,ne=[1024,1024,1,1]","support","1","yes","CUDA"
 "CUDA0","CEIL","type=f32,ne=[7,1,5,3]","support","1","yes","CUDA"
+"CUDA0","CEIL","type=f32,ne=[1024,1024,1,1]","support","1","yes","CUDA"
 "CUDA0","ROUND","type=f32,ne=[7,1,5,3]","support","1","yes","CUDA"
+"CUDA0","ROUND","type=f32,ne=[1024,1024,1,1]","support","1","yes","CUDA"
 "CUDA0","TRUNC","type=f32,ne=[7,1,5,3]","support","1","yes","CUDA"
+"CUDA0","TRUNC","type=f32,ne=[1024,1024,1,1]","support","1","yes","CUDA"
 "CUDA0","DIAG_MASK_INF","type=f32,ne=[10,10,1,1],n_past=5","support","1","yes","CUDA"
 "CUDA0","DIAG_MASK_INF","type=f32,ne=[10,10,3,1],n_past=5","support","1","yes","CUDA"
 "CUDA0","DIAG_MASK_INF","type=f32,ne=[10,10,3,2],n_past=5","support","1","yes","CUDA"
@@ -9411,18 +9449,405 @@
 "CUDA0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=2,v=3","support","0","no","CUDA"
 "CUDA0","CONCAT","type=f32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","1","yes","CUDA"
 "CUDA0","CONCAT","type=i32,ne_a=[11,12,13,14],ne_b_d=7,dim=3,v=3","support","0","no","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[3,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[4,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[7,1,1,1],order=0","support","1","yes","CUDA"
 "CUDA0","ARGSORT","type=f32,ne=[8,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[15,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[16,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[31,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[32,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[63,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[64,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[127,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[128,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[255,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[256,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[511,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[512,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1023,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1024,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[2047,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[2048,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[4095,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[4096,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[8191,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[8192,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[16383,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[16384,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[32767,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[32768,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[65535,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[65536,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[131071,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[131072,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[262143,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[262144,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[524287,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[524288,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1048575,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1048576,1,1,1],order=0","support","1","yes","CUDA"
 "CUDA0","ARGSORT","type=f32,ne=[16,10,10,10],order=0","support","1","yes","CUDA"
 "CUDA0","ARGSORT","type=f32,ne=[60,10,10,10],order=0","support","1","yes","CUDA"
-"CUDA0","ARGSORT","type=f32,ne=[1024,1,1,1],order=0","support","1","yes","CUDA"
-"CUDA0","ARGSORT","type=f32,ne=[16384,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1023,2,1,3],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1024,2,1,3],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1025,2,1,3],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[2047,2,1,3],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[2048,2,1,3],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[2049,2,1,3],order=0","support","1","yes","CUDA"
 "CUDA0","ARGSORT","type=f32,ne=[2,8,8192,1],order=0","support","1","yes","CUDA"
-"CUDA0","ARGSORT","type=f32,ne=[8,1,1,1],order=1","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[3,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[4,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[7,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[8,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[15,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[16,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[31,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[32,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[63,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[64,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[127,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[128,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[255,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[256,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[511,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[512,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1023,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1024,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[2047,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[2048,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[4095,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[4096,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[8191,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[8192,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[16383,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[16384,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[32767,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[32768,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[65535,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[65536,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[131071,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[131072,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[262143,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[262144,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[524287,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[524288,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1048575,1,1,1],order=0","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1048576,1,1,1],order=0","support","1","yes","CUDA"
 "CUDA0","ARGSORT","type=f32,ne=[16,10,10,10],order=1","support","1","yes","CUDA"
 "CUDA0","ARGSORT","type=f32,ne=[60,10,10,10],order=1","support","1","yes","CUDA"
-"CUDA0","ARGSORT","type=f32,ne=[1024,1,1,1],order=1","support","1","yes","CUDA"
-"CUDA0","ARGSORT","type=f32,ne=[16384,1,1,1],order=1","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1023,2,1,3],order=1","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1024,2,1,3],order=1","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[1025,2,1,3],order=1","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[2047,2,1,3],order=1","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[2048,2,1,3],order=1","support","1","yes","CUDA"
+"CUDA0","ARGSORT","type=f32,ne=[2049,2,1,3],order=1","support","1","yes","CUDA"
 "CUDA0","ARGSORT","type=f32,ne=[2,8,8192,1],order=1","support","1","yes","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[12,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[13,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[13,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[15,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[15,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[15,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[19,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[19,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[19,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[19,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[27,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[27,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[27,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[27,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[27,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[43,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[43,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[43,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[43,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[43,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[64,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[75,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[64,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[75,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[64,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[75,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[64,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[75,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[64,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[75,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[128,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[139,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[128,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[139,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[128,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[139,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[128,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[139,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[128,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[139,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[128,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[139,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[256,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[267,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[256,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[267,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[256,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[267,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[256,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[267,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[256,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[267,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[256,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[267,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[512,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[523,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[512,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[523,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[512,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[523,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[512,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[523,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[512,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[523,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[512,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[523,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[512,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[523,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1035,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1035,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1035,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1035,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1035,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1035,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1035,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,1,1,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1035,1,2,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2059,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2059,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2059,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2059,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2059,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2059,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2059,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,1,1,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2059,1,2,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4096,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4107,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4096,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4107,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4096,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4107,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4096,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4107,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4096,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4107,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4096,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4107,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4096,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4107,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4096,1,1,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[4107,1,2,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8192,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8203,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8192,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8203,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8192,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8203,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8192,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8203,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8192,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8203,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8192,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8203,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8192,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8203,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8192,1,1,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[8203,1,2,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16395,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16395,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16395,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16395,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16395,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16395,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16395,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16395,1,2,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16395,1,2,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32768,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32779,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32768,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32779,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32768,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32779,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32768,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32779,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32768,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32779,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32768,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32779,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32768,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32779,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32768,1,1,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32779,1,2,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32768,1,1,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[32779,1,2,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65536,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65547,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65536,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65547,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65536,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65547,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65536,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65547,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65536,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65547,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65536,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65547,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65536,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65547,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65536,1,1,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65547,1,2,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65536,1,1,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[65547,1,2,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131072,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131083,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131072,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131083,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131072,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131083,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131072,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131083,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131072,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131083,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131072,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131083,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131072,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131083,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131072,1,1,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131083,1,2,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131072,1,1,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[131083,1,2,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262144,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262155,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262144,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262155,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262144,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262155,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262144,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262155,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262144,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262155,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262144,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262155,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262144,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262155,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262144,1,1,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262155,1,2,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262144,1,1,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[262155,1,2,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524288,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524299,1,2,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524288,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524299,1,2,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524288,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524299,1,2,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524288,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524299,1,2,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524288,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524299,1,2,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524288,1,1,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524299,1,2,1],k=100,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524288,1,1,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524299,1,2,1],k=500,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524288,1,1,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524299,1,2,1],k=1023,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524288,1,1,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[524299,1,2,1],k=9999,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16,10,10,10],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[60,10,10,10],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1023,2,1,3],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,2,1,3],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1025,2,1,3],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2047,2,1,3],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,2,1,3],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2049,2,1,3],k=1,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16,10,10,10],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[60,10,10,10],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1023,2,1,3],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,2,1,3],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1025,2,1,3],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2047,2,1,3],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,2,1,3],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2049,2,1,3],k=2,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16,10,10,10],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[60,10,10,10],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1023,2,1,3],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,2,1,3],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1025,2,1,3],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2047,2,1,3],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,2,1,3],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2049,2,1,3],k=3,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16,10,10,10],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[60,10,10,10],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1023,2,1,3],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,2,1,3],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1025,2,1,3],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2047,2,1,3],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,2,1,3],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2049,2,1,3],k=7,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16,10,10,10],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[60,10,10,10],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1023,2,1,3],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1024,2,1,3],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[1025,2,1,3],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2047,2,1,3],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2048,2,1,3],k=15,ties=0","support","0","no","CUDA"
+"CUDA0","TOP_K","type=f32,ne=[2049,2,1,3],k=15,ties=0","support","0","no","CUDA"
 "CUDA0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=0","support","1","yes","CUDA"
 "CUDA0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","1","yes","CUDA"
 "CUDA0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest,flags=none","support","1","yes","CUDA"
@@ -9435,6 +9860,10 @@
 "CUDA0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=1","support","1","yes","CUDA"
 "CUDA0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic,flags=none","support","1","yes","CUDA"
 "CUDA0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic,flags=none","support","1","yes","CUDA"
+"CUDA0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=0","support","1","yes","CUDA"
+"CUDA0","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=1","support","1","yes","CUDA"
+"CUDA0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=none","support","1","yes","CUDA"
+"CUDA0","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear,flags=none","support","1","yes","CUDA"
 "CUDA0","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=align_corners","support","1","yes","CUDA"
 "CUDA0","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear,flags=align_corners","support","1","yes","CUDA"
 "CUDA0","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear,flags=align_corners","support","1","yes","CUDA"
@@ -9463,34 +9892,59 @@
 "CUDA0","GROUP_NORM","type=f32,ne=[64,64,320,1],num_groups=32,eps=0.000001","support","1","yes","CUDA"
 "CUDA0","GROUP_NORM","type=f32,ne=[9,9,1280,1],num_groups=32,eps=0.000001","support","1","yes","CUDA"
 "CUDA0","ACC","type=f32,ne_a=[256,17,1,1],ne_b=[256,16,1,1]","support","1","yes","CUDA"
-"CUDA0","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1","support","1","yes","CUDA"
-"CUDA0","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0","support","1","yes","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1,circular=0","support","1","yes","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[33,17,2,1],pad_0=4,pad_1=3,circular=1","support","1","yes","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0,circular=0","support","1","yes","CUDA"
 "CUDA0","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","1","yes","CUDA"
 "CUDA0","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9","support","1","yes","CUDA"
 "CUDA0","ROLL","shift0=3,shift1=-2,shift3=1,shift4=-1","support","1","yes","CUDA"
 "CUDA0","ARANGE","type=f32,start=0.000000,stop=10.000000,step=1.000000","support","1","yes","CUDA"
+"CUDA0","ARANGE","type=f32,start=0.000000,stop=1048576.000000,step=1.000000","support","1","yes","CUDA"
 "CUDA0","TIMESTEP_EMBEDDING","type=f32,ne_a=[2,1,1,1],dim=320,max_period=10000","support","1","yes","CUDA"
 "CUDA0","LEAKY_RELU","type=f32,ne_a=[10,5,4,3],negative_slope=0.100000","support","1","yes","CUDA"
-"CUDA0","CUMSUM","type=f32,ne=[10,5,4,3]","support","0","no","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[10,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[127,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[128,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[128,128,4,4]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[255,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[256,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[511,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[512,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[1023,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[1024,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[2047,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[2048,5,4,3]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[242004,1,1,1]","support","1","yes","CUDA"
+"CUDA0","CUMSUM","type=f32,ne=[375960,1,1,1]","support","1","yes","CUDA"
 "CUDA0","XIELU","type=f32,ne=[10,5,4,3]","support","0","no","CUDA"
-"CUDA0","TRI","type=f32,ne=[10,10,4,3],tri_type=3","support","0","no","CUDA"
-"CUDA0","TRI","type=f32,ne=[10,10,4,3],tri_type=2","support","0","no","CUDA"
-"CUDA0","TRI","type=f32,ne=[10,10,4,3],tri_type=1","support","0","no","CUDA"
-"CUDA0","TRI","type=f32,ne=[10,10,4,3],tri_type=0","support","0","no","CUDA"
-"CUDA0","FILL","type=f32,ne=[10,10,4,3],c=0.000000","support","0","no","CUDA"
-"CUDA0","FILL","type=f32,ne=[303,207,11,3],c=2.000000","support","0","no","CUDA"
-"CUDA0","FILL","type=f32,ne=[800,600,4,4],c=-152.000000","support","0","no","CUDA"
-"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[10,10,4,3],ne_rhs=[3,10,4,3]","support","0","no","CUDA"
-"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[11,11,1,1],ne_rhs=[5,11,1,1]","support","0","no","CUDA"
-"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[17,17,2,4],ne_rhs=[9,17,2,4]","support","0","no","CUDA"
-"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[30,30,7,1],ne_rhs=[8,30,7,1]","support","0","no","CUDA"
-"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[42,42,5,2],ne_rhs=[10,42,5,2]","support","0","no","CUDA"
-"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[10,64,2,2]","support","0","no","CUDA"
+"CUDA0","TRI","type=f32,ne=[10,10,4,3],tri_type=3","support","1","yes","CUDA"
+"CUDA0","TRI","type=f32,ne=[10,10,4,3],tri_type=2","support","1","yes","CUDA"
+"CUDA0","TRI","type=f32,ne=[10,10,4,3],tri_type=1","support","1","yes","CUDA"
+"CUDA0","TRI","type=f32,ne=[10,10,4,3],tri_type=0","support","1","yes","CUDA"
+"CUDA0","FILL","type=f32,ne=[10,10,4,3],c=0.000000","support","1","yes","CUDA"
+"CUDA0","FILL","type=f32,ne=[303,207,11,3],c=2.000000","support","1","yes","CUDA"
+"CUDA0","FILL","type=f32,ne=[800,600,4,4],c=-152.000000","support","1","yes","CUDA"
+"CUDA0","FILL","type=f32,ne=[2048,512,2,2],c=3.500000","support","1","yes","CUDA"
+"CUDA0","DIAG","type=f32,ne=[10,1,4,3]","support","1","yes","CUDA"
+"CUDA0","DIAG","type=f32,ne=[79,1,19,13]","support","1","yes","CUDA"
+"CUDA0","DIAG","type=f32,ne=[256,1,8,16]","support","1","yes","CUDA"
+"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[10,10,4,3],ne_rhs=[3,10,4,3]","support","1","yes","CUDA"
+"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[11,11,1,1],ne_rhs=[5,11,1,1]","support","1","yes","CUDA"
+"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[17,17,2,4],ne_rhs=[9,17,2,4]","support","1","yes","CUDA"
+"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[30,30,7,1],ne_rhs=[8,30,7,1]","support","1","yes","CUDA"
+"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[42,42,5,2],ne_rhs=[10,42,5,2]","support","1","yes","CUDA"
+"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[10,64,2,2]","support","1","yes","CUDA"
 "CUDA0","SOLVE_TRI","type=f32,ne_lhs=[100,100,4,4],ne_rhs=[41,100,4,4]","support","0","no","CUDA"
-"CUDA0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0","support","1","yes","CUDA"
-"CUDA0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0","support","1","yes","CUDA"
-"CUDA0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1","support","0","no","CUDA"
-"CUDA0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1","support","0","no","CUDA"
+"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,4],ne_rhs=[31,128,4,4]","support","0","no","CUDA"
+"CUDA0","SOLVE_TRI","type=f32,ne_lhs=[64,64,4,4],ne_rhs=[300,64,4,4]","support","0","no","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0,circular=0","support","1","yes","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0,circular=0","support","1","yes","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0,circular=1","support","1","yes","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0,circular=1","support","1","yes","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1,circular=0","support","0","no","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1,circular=0","support","0","no","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1,circular=1","support","0","no","CUDA"
+"CUDA0","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1,circular=1","support","0","no","CUDA"
 "CUDA0","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","1","yes","CUDA"
 "CUDA0","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","1","yes","CUDA"
 "CUDA0","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","CUDA"

examples/gen-docs/gen-docs.cpp

@@ -14,12 +14,13 @@ static void write_table_header(std::ofstream & file) {
 static void write_table_entry(std::ofstream & file, const common_arg & opt) {
     file << "| `";
     // args
-    for (const auto & arg : opt.args) {
-    if (arg == opt.args.front()) {
+    auto all_args = opt.get_args();
+    for (const auto & arg : all_args) {
+    if (arg == all_args.front()) {
             file << arg;
-            if (opt.args.size() > 1) file << ", ";
+            if (all_args.size() > 1) file << ", ";
         } else {
-            file << arg << (arg != opt.args.back() ? ", " : "");
+            file << arg << (arg != all_args.back() ? ", " : "");
         }
     }
     // value hint
@@ -76,7 +77,7 @@ static void export_md(std::string fname, llama_example ex) {
 }
 
 int main(int, char **) {
-    export_md("autogen-main.md", LLAMA_EXAMPLE_MAIN);
+    export_md("autogen-main.md", LLAMA_EXAMPLE_COMPLETION);
     export_md("autogen-server.md", LLAMA_EXAMPLE_SERVER);
 
     return 0;

examples/model-conversion/logits.cpp

@@ -144,7 +144,7 @@ int main(int argc, char ** argv) {
             return 1;
         }
         std::string s(buf, n);
-        printf("%s", s.c_str());
+        printf("%s (%d)", s.c_str(), id);
     }
     printf("\n");
 

examples/model-conversion/scripts/causal/compare-logits.py

@@ -1,10 +1,13 @@
 #!/usr/bin/env python3
 
-import numpy as np
 import sys
-import os
+import numpy as np
 from pathlib import Path
 
+# Add utils directory to path for direct script execution
+sys.path.insert(0, str(Path(__file__).parent.parent / "utils"))
+from common import get_model_name_from_env_path  # type: ignore[import-not-found]
+
 def quick_logits_check(pytorch_file, llamacpp_file):
     """Lightweight sanity check before NMSE"""
 
@@ -32,27 +35,16 @@ def quick_logits_check(pytorch_file, llamacpp_file):
     print(f"Top 10 llama.cpp logits: {llamacpp_logits[llamacpp_top10]}")
     print(f"Max absolute difference: {max_diff:.4f}")
 
-    if max_diff > 1.0:
-        print(f"❌ NOK: Large differences detected - max diff: {max_diff:.4f}")
-        return False
-
     return True
 
 def main():
-    model_path = os.getenv('MODEL_PATH')
-    if not model_path:
-        print("Error: MODEL_PATH environment variable not set")
-        sys.exit(1)
-
-    if not os.path.exists(model_path):
-        print(f"Error: Model file not found: {model_path}")
-        sys.exit(1)
-
-    model_name = os.path.basename(model_path)
+    model_name = get_model_name_from_env_path('MODEL_PATH')
     data_dir = Path("data")
-
     pytorch_file = data_dir / f"pytorch-{model_name}.bin"
-    llamacpp_file = data_dir / f"llamacpp-{model_name}.bin"
+
+    llamacpp_model_name = get_model_name_from_env_path('CONVERTED_MODEL')
+    print(f"Using converted model: {llamacpp_model_name}")
+    llamacpp_file = data_dir / f"llamacpp-{llamacpp_model_name}.bin"
 
     if not pytorch_file.exists():
         print(f"Error: PyTorch logits file not found: {pytorch_file}")

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

@@ -200,7 +200,7 @@ with torch.no_grad():
     logits = outputs.logits
 
     # Extract logits for the last token (next token prediction)
-    last_logits = logits[0, -1, :].cpu().numpy()
+    last_logits = logits[0, -1, :].float().cpu().numpy()
 
     print(f"Logits shape: {logits.shape}")
     print(f"Last token logits shape: {last_logits.shape}")

examples/model-conversion/scripts/utils/check-nmse.py

@@ -5,6 +5,7 @@ import sys
 import os
 import argparse
 from pathlib import Path
+from common import get_model_name_from_env_path  # type: ignore[import-not-found]
 
 def calculate_nmse(reference, test):
     mse = np.mean((test - reference) ** 2)
@@ -67,11 +68,13 @@ def main():
     parser.add_argument('-m', '--model-path', required=True,  help='Path to the model directory')
     args = parser.parse_args()
 
-    model_name = os.path.basename(args.model_path)
+    model_name = get_model_name_from_env_path('MODEL_PATH')
     data_dir = Path("data")
 
     pytorch_file = data_dir / f"pytorch-{model_name}.bin"
-    llamacpp_file = data_dir / f"llamacpp-{model_name}.bin"
+
+    llamacpp_model_name = get_model_name_from_env_path('CONVERTED_MODEL')
+    llamacpp_file = data_dir / f"llamacpp-{llamacpp_model_name}.bin"
 
     print(f"Model name: {model_name}")
     print(f"PyTorch logits file: {pytorch_file}")

examples/model-conversion/scripts/utils/common.py

@@ -0,0 +1,20 @@
+#!/usr/bin/env python3
+
+import os
+import sys
+
+def get_model_name_from_env_path(env_path_name):
+    model_path = os.getenv(env_path_name)
+    if not model_path:
+        print(f"Error: {env_path_name} environment variable not set")
+        sys.exit(1)
+
+    if not os.path.exists(model_path):
+        print(f"Error: Model file not found: {model_path}")
+        sys.exit(1)
+
+    name = os.path.basename(os.path.normpath(model_path))
+    if name.endswith(".gguf"):
+        name = name[:-5]
+
+    return name

examples/speculative-simple/speculative-simple.cpp

@@ -255,6 +255,8 @@ int main(int argc, char ** argv) {
     LOG_INF("target:\n\n");
     common_perf_print(ctx_tgt, smpl);
 
+    llama_batch_free(batch_tgt);
+
     common_sampler_free(smpl);
     common_speculative_free(spec);
 

ggml/CMakeLists.txt

@@ -54,6 +54,10 @@ if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     # TODO
 else()
     set(GGML_STANDALONE OFF)
+
+    if (NOT CMAKE_RUNTIME_OUTPUT_DIRECTORY)
+        set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+    endif()
 endif()
 
 if (EMSCRIPTEN)

ggml/include/ggml-cpu.h

@@ -99,6 +99,7 @@ extern "C" {
     GGML_BACKEND_API int ggml_cpu_has_sme        (void);
     // other
     GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
+    GGML_BACKEND_API int ggml_cpu_get_rvv_vlen   (void);  // risc-v vector length in bytes
     GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
     GGML_BACKEND_API int ggml_cpu_has_vxe        (void);
     GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);

ggml/include/ggml.h

@@ -2305,13 +2305,11 @@ extern "C" {
             float                 stop,
             float                 step);
 
-#define GGML_KQ_MASK_PAD 1
-
-    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
-    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
-    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
-    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
-    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
+    // q:    [n_embd_k, n_batch, n_head,    ne3 ]
+    // k:    [n_embd_k, n_kv,    n_head_kv, ne3 ]
+    // v:    [n_embd_v, n_kv,    n_head_kv, ne3 ] !! not transposed !!
+    // mask: [n_kv,     n_batch, ne32,      ne33]
+    // res:  [n_embd_v, n_head,  n_batch,   ne3 ] !! permuted !!
     //
     // broadcast:
     //   n_head % n_head_kv == 0

ggml/src/ggml-alloc.c

@@ -25,6 +25,7 @@ static bool ggml_is_view(const struct ggml_tensor * t) {
 // ops that return true for this function must not use restrict pointers for their backend implementations
 bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {
+        case GGML_OP_FILL:
         case GGML_OP_SCALE:
         case GGML_OP_DIAG_MASK_ZERO:
         case GGML_OP_DIAG_MASK_INF:
@@ -311,16 +312,9 @@ static struct buffer_address ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * al
 }
 
 // this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, size_t size, const struct ggml_tensor * tensor) {
+static void ggml_dyn_tallocr_free_bytes(struct ggml_dyn_tallocr * alloc, struct buffer_address addr, size_t size) {
     size = aligned_offset(NULL, size, alloc->alignment);
 
-    AT_PRINTF("%s: freeing %s at {chunk=%d, offset=%zu} (%zu bytes) - n_free_blocks = %d\n",
-        __func__, tensor->name, addr.chunk, addr.offset, size, alloc->chunks[addr.chunk]->n_free_blocks);
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    remove_allocated_tensor(alloc, addr, tensor);
-#endif
-
     struct tallocr_chunk * chunk = alloc->chunks[addr.chunk];
 
     // see if we can merge with an existing block
@@ -356,8 +350,6 @@ static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, struct
     }
     // otherwise, add a new block
     ggml_dyn_tallocr_insert_block(chunk, addr.offset, size);
-
-    GGML_UNUSED(tensor);
 }
 
 static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) {
@@ -615,13 +607,17 @@ static void ggml_gallocr_free_extra_space(ggml_gallocr_t galloc, struct ggml_ten
 
     GGML_ASSERT(parent_size >= node_size);
 
+    // note: we want after the freeing the chunks to continue to be aligned
+    struct ggml_dyn_tallocr * p_alloc = galloc->buf_tallocs[p_hn->buffer_id];
+    parent_size = aligned_offset(NULL, parent_size, p_alloc->alignment);
+    node_size = aligned_offset(NULL, node_size, p_alloc->alignment);
+
     if (parent_size > node_size) {
-        struct ggml_dyn_tallocr * p_alloc = galloc->buf_tallocs[p_hn->buffer_id];
         struct buffer_address p_addr = p_hn->addr;
         p_addr.offset += node_size;
         size_t extra_size = parent_size - node_size;
         AT_PRINTF("freeing extra %zu bytes from parent %s for %s\n", extra_size, parent->name, node->name);
-        ggml_dyn_tallocr_free_tensor(p_alloc, p_addr, extra_size, parent);
+        ggml_dyn_tallocr_free_bytes(p_alloc, p_addr, extra_size);
     }
 }
 
@@ -705,7 +701,14 @@ static void ggml_gallocr_free_node(ggml_gallocr_t galloc, struct ggml_tensor * n
     struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
     ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
     size_t size = ggml_backend_buft_get_alloc_size(buft, node);
-    ggml_dyn_tallocr_free_tensor(alloc, hn->addr, size, node);
+
+    AT_PRINTF("%s: freeing %s at {chunk=%d, offset=%zu} (%zu bytes) - n_free_blocks = %d\n",
+        __func__, node->name, hn->addr.chunk, hn->addr.offset, size, alloc->chunks[hn->addr.chunk]->n_free_blocks);
+#ifdef GGML_ALLOCATOR_DEBUG
+    remove_allocated_tensor(alloc, hn->addr, node);
+#endif
+
+    ggml_dyn_tallocr_free_bytes(alloc, hn->addr, size);
     hn->allocated = false;
 }
 

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -2251,12 +2251,12 @@ static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
                                   int                         sections[4],
                                   bool                        mrope_used,
                                   bool                        is_imrope,
-                                  bool                        indep_sects) {
-    ggml_tensor * src0 = dst->src[0];  // input
+                                  bool                        indep_sects,
+                                  int64_t                     rope_dims) {
     ggml_tensor * src1 = dst->src[1];  // position
     ggml_tensor * src2 = dst->src[2];  // freq_factors
 
-    int64_t theta_scale_length = src0->ne[0] / 2;
+    int64_t theta_scale_length = rope_dims / 2;
     int64_t position_length    = dst->ne[2];
 
     // TODO: check theta_scale_length and position_length.
@@ -2331,18 +2331,17 @@ static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
         ACL_CHECK(aclrtMemcpyAsync(ctx.rope_cache.theta_scale_cache, theta_scale_length * sizeof(float),
                                    ctx.rope_cache.theta_scale_exp_host, theta_scale_length * sizeof(float),
                                    ACL_MEMCPY_HOST_TO_DEVICE, ctx.stream()));
-
-        acl_theta_scale_tensor = ggml_cann_create_tensor(ctx.rope_cache.theta_scale_cache, ACL_FLOAT, sizeof(float),
-                                                         theta_scale_ne, theta_scale_nb, 1);
     }
+    acl_theta_scale_tensor = ggml_cann_create_tensor(ctx.rope_cache.theta_scale_cache, ACL_FLOAT, sizeof(float),
+                                                     theta_scale_ne, theta_scale_nb, 1);
 
     // Step1.2: prepare rope_yarn_ramp, if this part updated, should update theta_scale_tensor.
+    // TODO: acl_yarn_ramp_tensor use rope cache.
     bool                 yarn_ramp_tensor_updated = false;
     ggml_cann_pool_alloc yarn_ramp_allocator(ctx.pool());
     acl_tensor_ptr       acl_yarn_ramp_tensor;
-    if (ext_factor != 0 &&
-        // TODO: check more parameter.
-        (ctx.rope_cache.theta_scale_length != theta_scale_length || ctx.rope_cache.freq_scale != freq_scale)) {
+    if (ext_factor != 0 && (theta_scale_updated || ctx.rope_cache.theta_scale_length != theta_scale_length ||
+                            ctx.rope_cache.freq_scale != freq_scale)) {
         yarn_ramp_tensor_updated = true;
 
         // -rope_yarn_ramp
@@ -2590,7 +2589,7 @@ static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
         aclnn_muls(ctx, acl_cos_tensor.get(), attn_factor, nullptr, true);
     }
 
-    int64_t sin_reshape_ne[4] = { src0->ne[0], 1, dst->ne[2], 1 };
+    int64_t sin_reshape_ne[4] = { rope_dims, 1, dst->ne[2], 1 };
     size_t  sin_reshape_nb[GGML_MAX_DIMS];
     sin_reshape_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
@@ -2645,7 +2644,7 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
 
     // param
     float     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-    int sections[4];
+    int       sections[4];
     // const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
     const int mode       = ((int32_t *) dst->op_params)[2];
@@ -2654,44 +2653,60 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
-    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
-    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
-    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
-    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
-    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
-    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
-    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
+    memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
+    memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
+    memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
+    memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
+    memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
+    memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections, (int32_t *) dst->op_params + 11, sizeof(int) * 4);
 
-    // TODO: n_dims <= ne0
-    GGML_ASSERT(n_dims == ne0);
     GGML_ASSERT(n_dims % 2 == 0);
+    GGML_ASSERT(n_dims <= ne00);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
     float corr_dims[2];
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
-    const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
-    const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, note: also true for vision (24 & 8 == true) and for imrope
-    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+    bool       is_neox    = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_imrope  = mode == GGML_ROPE_TYPE_IMROPE;  // qwen3vl apply interleaved mrope
+    // mrope_used means the GGML_ROPE_TYPE_MROPE bit is set.
+    // Note: this bit is also set for imrope and some vision modes,
+    // so mrope_used does NOT exclusively indicate pure mrope.
+    const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision  = mode == GGML_ROPE_TYPE_VISION;
 
     if (mrope_used) {
         GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
     }
 
     if (is_vision) {
-        GGML_ASSERT(n_dims == ne0/2);
+        GGML_ASSERT(n_dims == ne0 / 2);
     }
 
     if (is_imrope || mrope_used) {
         is_neox = true;
     }
 
-    // init ctx.rope_cos/rope_sin cache
-    aclnn_rope_cache_init(ctx, dst, corr_dims, ext_factor, theta_scale, freq_scale, attn_factor, is_neox, sections, mrope_used, is_imrope, is_vision);
+    int64_t rope_dims = n_dims;
 
-    int64_t sin_reshape_ne[4] = { ne00, 1, ne02, 1 };
+    //Our current RotaryPositionEmbedding does not support the VISION mode,
+    //but essentially it only modifies theta_base in mrope,
+    //then repeats it at the end in the same way as is_neox.
+    //In fact, RoPE is still applied across all dimensions.
+    if (is_vision) {
+        rope_dims = src0->ne[0];
+    }
+    int64_t tail_dims = ne00 - rope_dims;
+    bool    has_tail  = tail_dims > 0;
+
+    // init ctx.rope_cos/rope_sin cache
+    aclnn_rope_cache_init(ctx, dst, corr_dims, ext_factor, theta_scale, freq_scale, attn_factor, is_neox, sections,
+                          mrope_used, is_imrope, is_vision, rope_dims);
+
+    // Cache is generated with ne00 dimensions, so we use ne00 for reshape
+    int64_t sin_reshape_ne[4] = { rope_dims, 1, ne02, 1 };
     size_t  sin_reshape_nb[GGML_MAX_DIMS];
     sin_reshape_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
@@ -2704,7 +2719,6 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
 
     acl_tensor_ptr acl_src = ggml_cann_create_tensor(src0);
     acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
-
 #ifdef ASCEND_310P
     // Special ROPE operation for 310P
 
@@ -2844,46 +2858,124 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     }
     return;
 #endif
-
     int64_t acl_mode = is_neox ? 0 : 1;
 
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                GGML_CANN_CALL_ACLNN_OP(ctx, RotaryPositionEmbedding, acl_src.get(), acl_cos_reshape_tensor.get(),
-                                        acl_sin_reshape_tensor.get(), acl_mode, acl_dst.get());
-                break;
-            }
-        case GGML_TYPE_F16:
-            {
-                ggml_cann_pool_alloc src_trans_allocator(ctx.pool(), ggml_nelements(src0) * sizeof(float));
-                void *               src_trans_buffer = src_trans_allocator.get();
-                ggml_cann_pool_alloc dst_trans_allocator(ctx.pool(), ggml_nelements(dst) * sizeof(float));
-                void *               dst_trans_buffer = dst_trans_allocator.get();
+    // Pre-define head and tail dimensions for reuse
+    int64_t head_ne[GGML_MAX_DIMS] = { rope_dims, ne01, ne02, ne03 };
+    int64_t tail_ne[GGML_MAX_DIMS] = { tail_dims, ne01, ne02, ne03 };
 
-                size_t src_trans_nb[GGML_MAX_DIMS];
-                src_trans_nb[0] = sizeof(float);
-                for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                    src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
-                }
+    // Step 1: Prepare trans tensors for F16 type conversion to F32 if needed
+    bool                 src_dst_need_trans = false;
+    ggml_cann_pool_alloc src_trans_allocator(ctx.pool());
+    ggml_cann_pool_alloc dst_trans_allocator(ctx.pool());
+    acl_tensor_ptr       acl_src_trans_tensor;
+    acl_tensor_ptr       acl_dst_trans_tensor;
+    void *               src_trans_buffer = nullptr;
+    void *               dst_trans_buffer = nullptr;
+    size_t               src_dst_trans_nb[GGML_MAX_DIMS];
+    if (src0->type == GGML_TYPE_F16) {
+        src_dst_need_trans = true;
+        src_trans_buffer   = src_trans_allocator.alloc(ggml_nelements(src0) * sizeof(float));
+        dst_trans_buffer   = dst_trans_allocator.alloc(ggml_nelements(dst) * sizeof(float));
 
-                acl_tensor_ptr acl_src_trans_tensor = ggml_cann_create_tensor(
-                    src_trans_buffer, ACL_FLOAT, sizeof(float), src0->ne, src_trans_nb, GGML_MAX_DIMS);
-                acl_tensor_ptr acl_dst_trans_tensor = ggml_cann_create_tensor(
-                    dst_trans_buffer, ACL_FLOAT, sizeof(float), dst->ne, src_trans_nb, GGML_MAX_DIMS);
+        src_dst_trans_nb[0] = sizeof(float);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            src_dst_trans_nb[i] = src_dst_trans_nb[i - 1] * src0->ne[i - 1];
+        }
+        acl_src_trans_tensor = ggml_cann_create_tensor(src_trans_buffer, ACL_FLOAT, sizeof(float), src0->ne,
+                                                       src_dst_trans_nb, GGML_MAX_DIMS);
+        acl_dst_trans_tensor = ggml_cann_create_tensor(dst_trans_buffer, ACL_FLOAT, sizeof(float), dst->ne,
+                                                       src_dst_trans_nb, GGML_MAX_DIMS);
+        aclnn_cast(ctx, acl_src.get(), acl_src_trans_tensor.get(), ACL_FLOAT);
+    }
 
-                aclnn_cast(ctx, acl_src.get(), acl_src_trans_tensor.get(), ACL_FLOAT);
+    // Step 2: Prepare head tensors for tail splitting if needed
+    acl_tensor_ptr acl_src_head;
+    acl_tensor_ptr acl_dst_head;
+    if (has_tail) {
+        // Create head views for RotaryPositionEmbedding (only first rope_dims dimensions)
+        // RotaryPositionEmbedding requires contiguous dst tensor, so we use a temporary buffer
+        if (src_dst_need_trans) {
+            // Use F32 trans tensor strides
+            acl_src_head = ggml_cann_create_tensor((char *) src_trans_buffer, ACL_FLOAT, sizeof(float), head_ne,
+                                                   src_dst_trans_nb, GGML_MAX_DIMS);
+        } else {
+            // Use original F32 tensor strides
+            acl_src_head = ggml_cann_create_tensor((char *) src0->data, ACL_FLOAT, sizeof(float), head_ne, src0->nb,
+                                                   GGML_MAX_DIMS);
+        }
 
-                GGML_CANN_CALL_ACLNN_OP(ctx, RotaryPositionEmbedding, acl_src_trans_tensor.get(),
-                                        acl_cos_reshape_tensor.get(), acl_sin_reshape_tensor.get(), acl_mode,
-                                        acl_dst_trans_tensor.get());
+        int64_t              head_elements = rope_dims * ne01 * ne02 * ne03;
+        ggml_cann_pool_alloc dst_head_contiguous_allocator(ctx.pool(), head_elements * sizeof(float));
+        void *               dst_head_contiguous_buffer = dst_head_contiguous_allocator.get();
 
-                aclnn_cast(ctx, acl_dst_trans_tensor.get(), acl_dst.get(), ACL_FLOAT16);
-                break;
-            }
-        default:
-            GGML_ABORT("Unsupported tensor type for GGML_OP_ROPE");
-            break;
+        size_t head_contiguous_nb[GGML_MAX_DIMS];
+        head_contiguous_nb[0] = sizeof(float);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            head_contiguous_nb[i] = head_contiguous_nb[i - 1] * head_ne[i - 1];
+        }
+        acl_dst_head = ggml_cann_create_tensor(dst_head_contiguous_buffer, ACL_FLOAT, sizeof(float), head_ne,
+                                               head_contiguous_nb, GGML_MAX_DIMS);
+    }
+
+    // Step 3: Execute RotaryPositionEmbedding
+    if (has_tail) {
+        // Rotate only the head portion (first rope_dims dimensions)
+        GGML_CANN_CALL_ACLNN_OP(ctx, RotaryPositionEmbedding, acl_src_head.get(), acl_cos_reshape_tensor.get(),
+                                acl_sin_reshape_tensor.get(), acl_mode, acl_dst_head.get());
+
+        // Copy head result from contiguous buffer back to destination tensor
+        if (src_dst_need_trans) {
+            acl_tensor_ptr acl_dst_head_target = ggml_cann_create_tensor(
+                (char *) dst_trans_buffer, ACL_FLOAT, sizeof(float), head_ne, src_dst_trans_nb, GGML_MAX_DIMS);
+            cann_copy(ctx, acl_dst_head.get(), acl_dst_head_target.get());
+        } else {
+            acl_tensor_ptr acl_dst_head_target =
+                ggml_cann_create_tensor((char *) dst->data, ACL_FLOAT, sizeof(float), head_ne, dst->nb, GGML_MAX_DIMS);
+            cann_copy(ctx, acl_dst_head.get(), acl_dst_head_target.get());
+        }
+    } else if (src_dst_need_trans) {
+        // Rotate full tensor (no tail), using trans tensors
+        GGML_CANN_CALL_ACLNN_OP(ctx, RotaryPositionEmbedding, acl_src_trans_tensor.get(), acl_cos_reshape_tensor.get(),
+                                acl_sin_reshape_tensor.get(), acl_mode, acl_dst_trans_tensor.get());
+    } else {
+        // Rotate full tensor (no tail), using original tensors
+        GGML_CANN_CALL_ACLNN_OP(ctx, RotaryPositionEmbedding, acl_src.get(), acl_cos_reshape_tensor.get(),
+                                acl_sin_reshape_tensor.get(), acl_mode, acl_dst.get());
+    }
+
+    // Step 4: Copy unrotated tail portion from source to destination
+    if (has_tail) {
+        size_t src_tail_offset;
+        size_t dst_tail_offset;
+
+        auto copy_tail_device = [&](void * src_ptr, void * dst_ptr, aclDataType dtype, size_t elem_size,
+                                    size_t * nb_src_arr, size_t * nb_dst_arr) {
+            acl_tensor_ptr acl_src_tail =
+                ggml_cann_create_tensor(src_ptr, dtype, elem_size, tail_ne, nb_src_arr, GGML_MAX_DIMS);
+            acl_tensor_ptr acl_dst_tail =
+                ggml_cann_create_tensor(dst_ptr, dtype, elem_size, tail_ne, nb_dst_arr, GGML_MAX_DIMS);
+            cann_copy(ctx, acl_src_tail.get(), acl_dst_tail.get());
+        };
+
+        if (src_dst_need_trans) {
+            // Use F32 trans tensor strides and offsets
+            src_tail_offset = rope_dims * src_dst_trans_nb[0];
+            dst_tail_offset = rope_dims * src_dst_trans_nb[0];
+            copy_tail_device((char *) src_trans_buffer + src_tail_offset, (char *) dst_trans_buffer + dst_tail_offset,
+                             ACL_FLOAT, sizeof(float), src_dst_trans_nb, src_dst_trans_nb);
+        } else {
+            // Use original tensor strides and offsets
+            src_tail_offset = rope_dims * nb00;
+            dst_tail_offset = rope_dims * nb0;
+            copy_tail_device((char *) src0->data + src_tail_offset, (char *) dst->data + dst_tail_offset,
+                             ggml_cann_type_mapping(dst->type), ggml_element_size(dst), src0->nb, dst->nb);
+        }
+    }
+
+    // Step 5: Cast back to F16 if needed
+    if (src_dst_need_trans) {
+        aclnn_cast(ctx, acl_dst_trans_tensor.get(), acl_dst.get(), ACL_FLOAT16);
     }
 }
 

ggml/src/ggml-cann/common.h

@@ -315,7 +315,7 @@ struct ggml_cann_rope_cache {
         if (theta_scale_exp_host) {
             free(theta_scale_exp_host);
         }
-        if(position_select_index_host) {
+        if (position_select_index_host) {
             free(position_select_index_host);
         }
     }
@@ -340,7 +340,7 @@ struct ggml_cann_rope_cache {
 
     void set(int64_t theta_scale_length,
              int64_t position_length,
-             float    ext_factor,
+             float   ext_factor,
              float   theta_scale,
              float   freq_scale,
              float   attn_factor,

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

@@ -2308,7 +2308,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
 
     bool cann_graph_update_required = false;
 #ifdef USE_ACL_GRAPH
-    bool use_cann_graph             = true;
+    bool use_cann_graph = true;
 
     static bool prefill_use_graph = parse_bool(get_env("GGML_CANN_PREFILL_USE_GRAPH").value_or(""));
     if (!prefill_use_graph) {
@@ -2338,7 +2338,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
         }
     }
 #else
-    bool use_cann_graph             = false;
+    bool use_cann_graph = false;
 #endif  // USE_ACL_GRAPH
     evaluate_and_capture_cann_graph(cann_ctx, cgraph, use_cann_graph, cann_graph_update_required);
 
@@ -2474,16 +2474,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
             }
         case GGML_OP_ROPE:
             {
-                // TODO: with ops-test v == 1
-                // TODO: n_dims <= ne0
-                if (op->src[0]->ne[0] != op->op_params[1]) {
-                    return false;
-                }
-
                 if (op->src[0]->ne[0] > 896) {
                     return false;
                 }
 #ifdef ASCEND_310P
+                // TODO: Support rope_dim < ne00(dim)
+                if (op->src[0]->ne[0] != op->op_params[1]) {
+                    return false;
+                }
                 if (!ggml_is_contiguous(op->src[0])) {
                     return false;
                 }
@@ -2550,6 +2548,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
         case GGML_OP_GROUP_NORM:
+            return true;
         case GGML_OP_PAD:
             // TODO: add circular padding support for cann, see https://github.com/ggml-org/llama.cpp/pull/16985
             return ggml_get_op_params_i32(op, 8) == 0;

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

@@ -24,6 +24,7 @@
 
 #define UNUSED GGML_UNUSED
 
+#if defined(__aarch64__) && defined(__ARM_NEON) && (defined(__ARM_FEATURE_MATMUL_INT8) || defined(__ARM_FEATURE_DOTPROD))
 static inline void decode_q4_Kx8_scales_mins(const uint8_t * scales_in,
                                              int16x8_t *     out_mins,
                                              int8_t *        out_scales) {
@@ -46,6 +47,7 @@ static inline void decode_q4_Kx8_scales_mins(const uint8_t * scales_in,
     scales_u32[1] = (sm[2] & kmask2) | (((sm[0] >> 6) & kmask3) << 4);
     memcpy(out_scales, scales_u32, 8);
 }
+#endif
 
 void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);

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

@@ -81,6 +81,11 @@ struct ggml_arm_arch_features_type {
 } ggml_arm_arch_features = { 0 };
 #endif
 
+#if defined(__riscv)
+struct ggml_riscv_arch_features_type {
+    int rvv_vlen;
+} ggml_riscv_arch_features = { 0 };
+#endif
 
 #if defined(_WIN32)
 
@@ -187,6 +192,9 @@ typedef void * thread_ret_t;
 
 typedef pthread_t ggml_thread_t;
 
+#define GGML_THREADPOOL_N_THREADS_MASK (0xffffU)
+#define GGML_THREADPOOL_N_THREADS_BITS (16)
+
 #if defined(__APPLE__)
 #include <unistd.h>
 #include <mach/mach.h>
@@ -449,7 +457,7 @@ struct ggml_threadpool {
     struct ggml_cplan  * cplan;
 
     // synchronization primitives
-    atomic_int n_graph;       // incremented when there is work to be done (i.e each graph)
+    atomic_int n_graph;       // updated when there is work to be done (i.e each graph) holds graph and active thread counts.
     atomic_int GGML_CACHE_ALIGN n_barrier;
     atomic_int GGML_CACHE_ALIGN n_barrier_passed;
     atomic_int GGML_CACHE_ALIGN current_chunk; // currently processing chunk during Mat_Mul, shared between all the threads.
@@ -457,12 +465,10 @@ struct ggml_threadpool {
     // these are atomic as an annotation for thread-sanitizer
     atomic_bool stop;         // Used for stopping the threadpool altogether
     atomic_bool pause;        // Used for pausing the threadpool or individual threads
-    atomic_int abort;         // Used for aborting processing of a graph
+    atomic_int  abort;        // Used for aborting processing of a graph
 
     struct ggml_compute_state * workers;   // per thread state
-    int          n_threads_max; // number of threads in the pool
-    atomic_int   n_threads_cur; // number of threads used in the current graph
-
+    int          n_threads;   // Number of threads in the pool
     int32_t      prio;        // Scheduling priority
     uint32_t     poll;        // Polling level (0 - no polling)
 
@@ -539,7 +545,7 @@ struct ggml_state {
 static struct ggml_state g_state = {0};
 
 void ggml_barrier(struct ggml_threadpool * tp) {
-    int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
+    int n_threads = atomic_load_explicit(&tp->n_graph, memory_order_relaxed) & GGML_THREADPOOL_N_THREADS_MASK;
     if (n_threads == 1) {
         return;
     }
@@ -556,7 +562,7 @@ void ggml_barrier(struct ggml_threadpool * tp) {
         // last thread
         atomic_store_explicit(&tp->n_barrier, 0, memory_order_relaxed);
 
-        // exit barrier (fill seq-cst fence)
+        // exit barrier (full seq-cst fence)
         atomic_fetch_add_explicit(&tp->n_barrier_passed, 1, memory_order_seq_cst);
         return;
     }
@@ -702,6 +708,15 @@ static void ggml_init_arm_arch_features(void) {}
 #endif
 #endif // __ARM_ARCH
 
+#if defined(__riscv) && defined(__riscv_v_intrinsic)
+#include <riscv_vector.h>
+static void ggml_init_riscv_arch_features(void) {
+    ggml_riscv_arch_features.rvv_vlen = __riscv_vlenb();
+}
+#else
+static void ggml_init_riscv_arch_features(void) {}
+#endif
+
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
     GGML_ASSERT(!ggml_get_no_alloc(ctx));
 
@@ -2628,7 +2643,7 @@ static void ggml_thread_cpumask_next(const bool * global_mask, bool * local_mask
 void ggml_threadpool_free(struct ggml_threadpool* threadpool) {
     if (!threadpool) return;
 
-    const int n_threads = threadpool->n_threads_max;
+    const int n_threads = threadpool->n_threads;
 
 #ifndef GGML_USE_OPENMP
     struct ggml_compute_state* workers = threadpool->workers;
@@ -2704,7 +2719,7 @@ struct ggml_cplan ggml_graph_plan(
         //GGML_PRINT_DEBUG("Threadpool is not specified. Will create a disposable threadpool : n_threads %d\n", n_threads);
     }
     if (n_threads <= 0) {
-        n_threads = threadpool ? threadpool->n_threads_max : GGML_DEFAULT_N_THREADS;
+        n_threads = threadpool ? threadpool->n_threads : GGML_DEFAULT_N_THREADS;
     }
 
 #if defined(__EMSCRIPTEN__) && !defined(__EMSCRIPTEN_PTHREADS__)
@@ -2912,12 +2927,14 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
 
     struct ggml_compute_params params = {
         /*.ith       =*/ state->ith,
-        /*.nth       =*/ atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed),
+        /*.nth       =*/ atomic_load_explicit(&tp->n_graph, memory_order_relaxed) & GGML_THREADPOOL_N_THREADS_MASK,
         /*.wsize     =*/ cplan->work_size,
         /*.wdata     =*/ cplan->work_data,
         /*.threadpool=*/ tp,
     };
 
+    GGML_PRINT_DEBUG("thread #%d compute-start cplan %p last-graph %d \n", state->ith, cplan, state->last_graph);
+
     for (int node_n = 0; node_n < cgraph->n_nodes && atomic_load_explicit(&tp->abort, memory_order_relaxed) != node_n; node_n++) {
         struct ggml_tensor * node = cgraph->nodes[node_n];
 
@@ -2939,6 +2956,8 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
         }
     }
 
+    GGML_PRINT_DEBUG("thread #%d compute-done cplan %p last-graph %d \n", state->ith, cplan, state->last_graph);
+
     ggml_barrier(state->threadpool);
 
     return 0;
@@ -2946,27 +2965,23 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
 
 #ifndef GGML_USE_OPENMP
 
-// check if thread is active
-static inline bool ggml_graph_compute_thread_active(struct ggml_compute_state * state) {
-    struct ggml_threadpool * threadpool = state->threadpool;
-    int n_threads = atomic_load_explicit(&threadpool->n_threads_cur, memory_order_relaxed);
-    return (state->ith < n_threads);
-}
-
 // check if thread is ready to proceed (exit from polling or sleeping)
+// returns true if loops should exit, sets state->pending to indicate new work
 static inline bool ggml_graph_compute_thread_ready(struct ggml_compute_state * state) {
     struct ggml_threadpool * threadpool = state->threadpool;
 
     if (state->pending || threadpool->stop || threadpool->pause) { return true; }
 
     // check for new graph/work
-    int new_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed);
-    if (new_graph != state->last_graph) {
-        state->pending    = ggml_graph_compute_thread_active(state);
-        state->last_graph = new_graph;
+    int n_graph   = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed);
+    int n_threads = n_graph & GGML_THREADPOOL_N_THREADS_MASK;
+    if (n_graph != state->last_graph) {
+        state->pending    = (state->ith < n_threads);
+        state->last_graph = n_graph;
+        return true;
     }
 
-    return state->pending;
+    return false;
 }
 
 // sync thread state after polling
@@ -2983,11 +2998,6 @@ static inline void ggml_graph_compute_thread_sync(struct ggml_compute_state * st
 static inline bool ggml_graph_compute_poll_for_work(struct ggml_compute_state * state) {
     struct ggml_threadpool * threadpool = state->threadpool;
 
-    // Skip polling for unused threads
-    if (!ggml_graph_compute_thread_active(state)) {
-        return state->pending;
-    }
-
     // This seems to make 0 ... 100 a decent range for polling level across modern processors.
     // Perhaps, we can adjust it dynamically based on load and things.
     const uint64_t n_rounds = 1024UL * 128 * threadpool->poll;
@@ -3049,7 +3059,6 @@ static thread_ret_t ggml_graph_compute_secondary_thread(void* data) {
         ggml_graph_compute_check_for_work(state);
         if (state->pending) {
             state->pending = false;
-
             ggml_graph_compute_thread(state);
         }
     }
@@ -3064,14 +3073,15 @@ static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int
 
     ggml_mutex_lock(&threadpool->mutex);
 
-    GGML_PRINT_DEBUG("threadpool: n_threads_cur %d n_threads %d\n", threadpool->n_threads_cur, n_threads);
+    // Update the number of active threads and the graph count
+    int n_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed) >> GGML_THREADPOOL_N_THREADS_BITS;
+    n_graph = ((n_graph + 1) << GGML_THREADPOOL_N_THREADS_BITS) | (n_threads & GGML_THREADPOOL_N_THREADS_MASK);
 
-    // Update the number of active threads
-    atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
+    GGML_PRINT_DEBUG("compute-kickoff: n_threads %d n_graph %d\n", n_threads, n_graph);
 
     // Indicate the graph is ready to be processed
     // We need the full seq-cst fence here because of the polling threads (used in thread_sync)
-    atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_seq_cst);
+    atomic_store_explicit(&threadpool->n_graph, n_graph, memory_order_seq_cst);
 
     if (threadpool->pause) {
        // Update main thread prio and affinity to match the threadpool settings
@@ -3109,8 +3119,7 @@ static struct ggml_threadpool * ggml_threadpool_new_impl(
         threadpool->pause            = tpp->paused;
         threadpool->abort            = -1;
         threadpool->workers          = NULL;
-        threadpool->n_threads_max    = tpp->n_threads;
-        threadpool->n_threads_cur    = tpp->n_threads;
+        threadpool->n_threads        = tpp->n_threads;
         threadpool->poll             = tpp->poll;
         threadpool->prio             = tpp->prio;
         threadpool->ec               = GGML_STATUS_SUCCESS;
@@ -3205,7 +3214,7 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
             {
                 // update the number of threads from the actual number of threads that we got from OpenMP
                 n_threads = omp_get_num_threads();
-                atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
+                atomic_store_explicit(&threadpool->n_graph, n_threads, memory_order_relaxed);
             }
 
             // Apply thread CPU mask and priority
@@ -3218,13 +3227,13 @@ enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cpl
             ggml_graph_compute_thread(&threadpool->workers[ith]);
         }
     } else {
-        atomic_store_explicit(&threadpool->n_threads_cur, 1, memory_order_relaxed);
+        atomic_store_explicit(&threadpool->n_graph, 1, memory_order_relaxed);
         ggml_graph_compute_thread(&threadpool->workers[0]);
     }
 #else
-    if (n_threads > threadpool->n_threads_max) {
-        GGML_LOG_WARN("cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads_max);
-        n_threads = threadpool->n_threads_max;
+    if (n_threads > threadpool->n_threads) {
+        GGML_LOG_WARN("cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads);
+        n_threads = threadpool->n_threads;
     }
 
     // Kick all threads to start the new graph
@@ -3464,6 +3473,14 @@ int ggml_cpu_has_riscv_v(void) {
 #endif
 }
 
+int ggml_cpu_get_rvv_vlen(void) {
+#if defined(__riscv) && defined(__riscv_v_intrinsic)
+    return ggml_riscv_arch_features.rvv_vlen;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_f16c(void) {
 #if defined(__F16C__)
     return 1;
@@ -3630,6 +3647,10 @@ void ggml_cpu_init(void) {
         ggml_init_arm_arch_features();
 #endif
 
+#if defined(__riscv)
+        ggml_init_riscv_arch_features();
+#endif
+
         is_first_call = false;
     }
 

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

@@ -583,6 +583,10 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_riscv_v()) {
             features.push_back({ "RISCV_V", "1" });
         }
+        if (ggml_cpu_get_rvv_vlen() > 0) {
+            static std::string rvv_vlen = std::to_string(ggml_cpu_get_rvv_vlen());
+            features.push_back({ "RVV_VLEN", rvv_vlen.c_str() });
+        }
         if (ggml_cpu_has_vsx()) {
             features.push_back({ "VSX", "1" });
         }

ggml/src/ggml-cpu/repack.cpp

@@ -2169,7 +2169,8 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0> iq4_nl_8x8_q8_0;
 
     if (cur->type == GGML_TYPE_Q4_0) {
-        if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
+        if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)
+            || (ggml_cpu_has_riscv_v() && (ggml_cpu_get_rvv_vlen() >= QK4_0))) {
             if (cur->ne[1] % 8 == 0) {
                 return &q4_0_8x8_q8_0;
             }

ggml/src/ggml-cuda/common.cuh

@@ -67,19 +67,22 @@
 #define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000
 #define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
 #define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
+#define GGML_CUDA_CC_RDNA3_5    (GGML_CUDA_CC_OFFSET_AMD + 0x1150) // AI 370, AI Max 395 laptops.
 #define GGML_CUDA_CC_RDNA4      (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000
 
-#define GGML_CUDA_CC_IS_AMD(cc)   (cc >= GGML_CUDA_CC_OFFSET_AMD)
-#define GGML_CUDA_CC_IS_RDNA(cc)  (cc >= GGML_CUDA_CC_RDNA1)
-#define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
-#define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
-#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4)
-#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4)
-#define GGML_CUDA_CC_IS_GCN(cc)   (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1)
-#define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
-#define GGML_CUDA_CC_IS_CDNA1(cc) (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)
-#define GGML_CUDA_CC_IS_CDNA2(cc) (cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)
-#define GGML_CUDA_CC_IS_CDNA3(cc) (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_AMD(cc)     (cc >= GGML_CUDA_CC_OFFSET_AMD)
+#define GGML_CUDA_CC_IS_RDNA(cc)    (cc >= GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_RDNA1(cc)   (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
+#define GGML_CUDA_CC_IS_RDNA2(cc)   (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
+#define GGML_CUDA_CC_IS_RDNA3_0(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA3_5)
+#define GGML_CUDA_CC_IS_RDNA3_5(cc) (cc >= GGML_CUDA_CC_RDNA3_5 && cc < GGML_CUDA_CC_RDNA4)
+#define GGML_CUDA_CC_IS_RDNA3(cc)   (GGML_CUDA_CC_IS_RDNA3_0(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc))
+#define GGML_CUDA_CC_IS_RDNA4(cc)   (cc >= GGML_CUDA_CC_RDNA4)
+#define GGML_CUDA_CC_IS_GCN(cc)     (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1)
+#define GGML_CUDA_CC_IS_CDNA(cc)    (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_RDNA1)
+#define GGML_CUDA_CC_IS_CDNA1(cc)   (cc >= GGML_CUDA_CC_CDNA1 && cc < GGML_CUDA_CC_CDNA2)
+#define GGML_CUDA_CC_IS_CDNA2(cc)   (cc >= GGML_CUDA_CC_CDNA2 && cc < GGML_CUDA_CC_CDNA3)
+#define GGML_CUDA_CC_IS_CDNA3(cc)   (cc >= GGML_CUDA_CC_CDNA3 && cc < GGML_CUDA_CC_RDNA1)
 
 // Moore Threads
 #define MUSART_HMASK 40300 // MUSA rc4.3, min. ver. for half2 -> uint mask comparisons

ggml/src/ggml-cuda/diag.cu

@@ -0,0 +1,77 @@
+#include "convert.cuh"
+#include "diag.cuh"
+#include "ggml.h"
+
+template <typename T>
+static __global__ void diag_kernel(T * __restrict__ dst,
+                                   const T * __restrict__ src,
+                                   const int64_t ne0,
+                                   const int64_t ne1,
+                                   const int64_t ne2,
+                                   const int64_t ne3,
+                                   const int64_t total_elements) {
+    const int64_t global_idx = blockIdx.x * blockDim.x + threadIdx.x;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    const int64_t i0 = global_idx % ne0;
+    const int64_t i1 = (global_idx / ne0) % ne1;
+    const int64_t i2 = (global_idx / (ne0 * ne1)) % ne2;
+    const int64_t i3 = global_idx / (ne0 * ne1 * ne2);
+
+    const int64_t dst_idx = ((i3 * ne2 + i2) * ne1 + i1) * ne0 + i0;
+
+    if (i0 == i1) {
+        const int64_t batch_idx = i3 * ne2 + i2;
+        const int64_t src_idx   = batch_idx * ne0 + i0;
+        dst[dst_idx]            = src[src_idx];
+    } else {
+        dst[dst_idx] = ggml_cuda_cast<T>(0);
+    }
+    GGML_UNUSED_VARS(ne3);
+}
+
+void ggml_cuda_op_diag(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    void *       dst_d  = dst->data;
+    const void * src0_d = src0->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    const int64_t ne3 = dst->ne[3];
+
+    GGML_ASSERT(ne00 == ne0);
+    GGML_ASSERT(ne01 == 1);
+    GGML_ASSERT(ne02 == ne2);
+    GGML_ASSERT(ne03 == ne3);
+
+    const int64_t n_elems    = ggml_nelements(dst);
+    const int64_t num_blocks = (n_elems + CUDA_DIAG_BLOCK_SIZE - 1) / CUDA_DIAG_BLOCK_SIZE;
+
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            diag_kernel<<<num_blocks, CUDA_DIAG_BLOCK_SIZE, 0, stream>>>((float *) dst_d, (const float *) src0_d, ne0,
+                                                                         ne1, ne2, ne3, n_elems);
+            break;
+        case GGML_TYPE_F16:
+            diag_kernel<<<num_blocks, CUDA_DIAG_BLOCK_SIZE, 0, stream>>>((half *) dst_d, (const half *) src0_d, ne0,
+                                                                         ne1, ne2, ne3, n_elems);
+            break;
+        default:
+            GGML_ABORT("unsupported type");
+    }
+}

ggml/src/ggml-cuda/diag.cuh

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

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

@@ -642,8 +642,8 @@ static __global__ void flash_attn_stream_k_fixup(
     const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa;
     const int iter_j = (ne01 + (ncols1    - 1)) / ncols1;
 
-    const int kbc0      = (bidx0 + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
-    const int kbc0_stop = (bidx0 + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    const int kbc0      = int64_t(bidx0 + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    const int kbc0_stop = int64_t(bidx0 + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
 
     const bool did_not_have_any_data   = kbc0 == kbc0_stop;
     const bool wrote_beginning_of_tile = kbc0 % iter_k == 0;
@@ -679,7 +679,7 @@ static __global__ void flash_attn_stream_k_fixup(
     int bidx = bidx0 - 1;
     int kbc_stop = kbc0;
     while(true) {
-        const int kbc = bidx*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+        const int kbc = int64_t(bidx)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
         if (kbc == kbc_stop) { // Did not have any data.
             bidx--;
             kbc_stop = kbc;

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

@@ -955,22 +955,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             (K_h2 + int64_t(kb0)*nbatch_fa*stride_K, tile_K, nbatch_K2, stride_K, k_VKQ_sup);
     }
 
-    for (; kb0 < kb0_stop-1; ++kb0) {
-        constexpr bool last_iter = false;
-        constexpr bool oob_check = false;
-        constexpr int  k_VKQ_sup = nbatch_fa;
-        flash_attn_ext_f16_iter
-            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
-             T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
-            (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
-             ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
-             KQ_max, KQ_rowsum, jt, kb0, k_VKQ_sup);
-    }
     // kb0_start is always < kb0_stop so the last iter can be executed unconditionally.
     if constexpr (ncols2 == 1) {
-        if (ne11 % nbatch_fa == 0) {
-            constexpr bool last_iter = true;
-            constexpr bool oob_check = false;
+        constexpr bool oob_check = true;
+        for (; kb0 < kb0_stop-1; ++kb0) {
+            constexpr bool last_iter = false;
             constexpr int  k_VKQ_sup = nbatch_fa;
             flash_attn_ext_f16_iter
                 <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
@@ -978,10 +967,20 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                 (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
                  ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
                  KQ_max, KQ_rowsum, jt, kb0, k_VKQ_sup);
-        } else {
-            constexpr bool last_iter = true;
-            constexpr bool oob_check = true;
-            const     int  k_VKQ_sup = ne11 - kb0*nbatch_fa;
+        }
+        constexpr bool last_iter = true;
+        const     int  k_VKQ_sup = ne11 - kb0*nbatch_fa;
+        flash_attn_ext_f16_iter
+            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
+              T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
+            (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
+             ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
+             KQ_max, KQ_rowsum, jt, kb0, k_VKQ_sup);
+    } else {
+        constexpr bool oob_check = false;
+        for (; kb0 < kb0_stop-1; ++kb0) {
+            constexpr bool last_iter = false;
+            constexpr int  k_VKQ_sup = nbatch_fa;
             flash_attn_ext_f16_iter
                 <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
                  T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
@@ -989,9 +988,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                  ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
                  KQ_max, KQ_rowsum, jt, kb0, k_VKQ_sup);
         }
-    } else {
         constexpr bool last_iter = true;
-        constexpr bool oob_check = false;
         constexpr int  k_VKQ_sup = nbatch_fa;
         flash_attn_ext_f16_iter
             <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter, oob_check,
@@ -1383,8 +1380,8 @@ static __global__ void flash_attn_ext_f16(
     const int iter_j = (ne01.z + (ncols1    - 1)) / ncols1;
 
     // kbc == k block continuous, current index in continuous ijk space.
-    int       kbc      = (blockIdx.x + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
-    const int kbc_stop = (blockIdx.x + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    int       kbc      = int64_t(blockIdx.x + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
+    const int kbc_stop = int64_t(blockIdx.x + 1)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
 
     // If the seams of 2 CUDA blocks fall within an output tile their results need to be combined.
     // For this we need to track both the block that starts the tile (needs_fixup) and the block that finishes the tile (is_fixup).
@@ -1404,7 +1401,7 @@ static __global__ void flash_attn_ext_f16(
         const float2 * Q_f2   = (const float2 *) (Q + nb03*sequence + nb02* head0);
         const half2  * K_h2   = (const half2  *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
         const half   * mask_h = ncols2 == 1 && !mask ? nullptr :
-            (const half  *) (mask + nb33*(sequence % ne33));
+            (const half *) (mask + nb33*(sequence % ne33));
         float2       * dstk   = ((float2 *) dst) + (sequence*ne01.z*ne02 + head0) * (DV/2);
 
         const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio));

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

@@ -564,6 +564,12 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
         for (int i_KQ_0 = 0; i_KQ_0 < nbatch_fa; i_KQ_0 += np*warp_size) {
             const int i_KQ = i_KQ_0 + (threadIdx.y % np)*warp_size + threadIdx.x;
 
+#if defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
+            // Without the v_dot2_f32_f16 instruction there is a higher risk of numerical overflow in the KQ calculation.
+            // Therefore, scale down Q values and apply the inverse scale the FP32 KQ values afterwards again.
+            KQ_acc[i_KQ_0/(np*warp_size)*cpw + jc0] *= 4.0f;
+#endif // defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
+
             if (use_logit_softcap) {
                 KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] = logit_softcap * tanhf(KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0]);
             }
@@ -858,6 +864,11 @@ static __global__ void flash_attn_tile(
 #pragma unroll
                 for (int i1 = 0; i1 < cpy_ne_D; i1 += 2) {
                     tmp_h2[i1/2] = make_half2(tmp_f[i1 + 0], tmp_f[i1 + 1]);
+#if defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
+                    // Without the v_dot2_f32_f16 instruction there is a higher risk of numerical overflow in the KQ calculation.
+                    // Therefore, scale down Q values and apply the inverse scale the FP32 KQ values afterwards again.
+                    tmp_h2[i1/2] *= make_half2(0.25f, 0.25f);
+#endif // defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
                 }
                 ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
                     &Q_tmp[jc*(DKQ/2) + i0/2 + (threadIdx.y % np)*(warp_size*cpy_ne_D/2) + threadIdx.x*(cpy_ne_D/2)],

ggml/src/ggml-cuda/fattn.cu

@@ -36,12 +36,26 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
     const ggml_tensor * KQV  = dst;
     const ggml_tensor * Q    = dst->src[0];
     const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * V    = dst->src[2];
     const ggml_tensor * mask = dst->src[3];
 
     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 && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    // Edge cases like no mask, ALiBi, unpadded K/V, or misaligned addresses for large data transfers
+    //     are put into the template specialization without GQA optimizations.
+    bool use_gqa_opt = mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    for (const ggml_tensor * t : {Q, K, V, mask}) {
+        if (t == nullptr) {
+            continue;
+        }
+        for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
+            if (t->nb[i] % 16 != 0) {
+                use_gqa_opt = false;
+                break;
+            }
+        }
+    }
 
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
     const int gqa_ratio = Q->ne[2] / K->ne[2];

ggml/src/ggml-cuda/fill.cu

@@ -4,7 +4,7 @@
 #define CUDA_FILL_BLOCK_SIZE 256
 
 template <typename T>
-static __global__ void fill_kernel(T * __restrict__ dst, const int64_t k, const T value) {
+static __global__ void fill_kernel(T * dst, const int64_t k, const T value) {
     const int64_t i = (int64_t)blockDim.x * blockIdx.x + threadIdx.x;
     if (i >= k) {
         return;

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

@@ -20,6 +20,7 @@
 #include "ggml-cuda/cpy.cuh"
 #include "ggml-cuda/cross-entropy-loss.cuh"
 #include "ggml-cuda/diagmask.cuh"
+#include "ggml-cuda/diag.cuh"
 #include "ggml-cuda/fattn.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
@@ -2641,6 +2642,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
                 break;
+        case GGML_OP_DIAG:
+            ggml_cuda_op_diag(ctx, dst);
+            break;
         case GGML_OP_DIAG_MASK_INF:
             ggml_cuda_op_diag_mask_inf(ctx, dst);
             break;
@@ -4309,6 +4313,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_UNARY_OP_EXPM1:
                 case GGML_UNARY_OP_SOFTPLUS:
                 case GGML_UNARY_OP_ELU:
+                case GGML_UNARY_OP_XIELU:
                 case GGML_UNARY_OP_FLOOR:
                 case GGML_UNARY_OP_CEIL:
                 case GGML_UNARY_OP_ROUND:
@@ -4624,9 +4629,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_FILL:
         case GGML_OP_CUMSUM:
         case GGML_OP_TRI:
-            return true;
+        case GGML_OP_DIAG:
         case GGML_OP_SOLVE_TRI:
-            return op->src[0]->ne[0] <= 64 && op->src[1]->ne[0] <= 32;
+            return true;
+
         default:
             return false;
     }

ggml/src/ggml-cuda/mma.cuh

@@ -189,6 +189,9 @@ namespace ggml_cuda_mma {
                 return 8 * (threadIdx.x / 16) + l;
 #elif defined(RDNA3)
                 return 2 * l + (threadIdx.x / 16);
+#else
+                NO_DEVICE_CODE;
+                return -1;
 #endif // defined(RDNA4)
             } else {
                 NO_DEVICE_CODE;
@@ -290,8 +293,12 @@ namespace ggml_cuda_mma {
             }
         }
 #elif defined(AMD_WMMA_AVAILABLE)
-
+#if defined(RDNA3)
+        // RDNA3 has duplicated data as input.
+        static constexpr int ne = I * J / 32 * 2;
+#else
         static constexpr int ne = I * J / 32;
+#endif // defined(RDNA3)
         half2 x[ne] = {{0.0f, 0.0f}};
 
         static constexpr __device__ bool supported() {
@@ -310,7 +317,14 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 16 && J == 8) {
+#if defined(RDNA4)
                 return 4 * (threadIdx.x / 16) + l;
+#elif defined(RDNA3)
+                return l;
+#else
+                NO_DEVICE_CODE;
+                return -1;
+#endif // defined(RDNA4)
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -366,11 +380,16 @@ namespace ggml_cuda_mma {
         static constexpr int         I  = I_;
         static constexpr int         J  = J_;
         static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
-        static constexpr int         ne = I * J / WARP_SIZE;
-
-        nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
 
 #if defined(AMD_WMMA_AVAILABLE)
+#if defined(RDNA3)
+        // RDNA3 has duplicated data as input.
+        static constexpr int ne = I * J / 32 * 2;
+#else
+        static constexpr int ne = I * J / 32;
+#endif // defined(RDNA3)
+        nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
+
         static constexpr __device__ bool supported() {
             if (I == 16 && J == 8) return true;
             return false;
@@ -387,13 +406,23 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 16 && J == 8) {
+#if defined(RDNA4)
                 return 4 * (threadIdx.x / 16) + l;
+#elif defined(RDNA3)
+                return l;
+#else
+                NO_DEVICE_CODE;
+                return -1;
+#endif // defined(RDNA4)
             } else {
                 NO_DEVICE_CODE;
                 return -1;
             }
         }
 #else
+        static constexpr int ne = I * J / WARP_SIZE;
+        nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
+
         static constexpr __device__ bool supported() {
             if (I ==  8 && J ==  8) return true;
             if (I == 16 && J ==  4) return true;
@@ -546,8 +575,14 @@ namespace ggml_cuda_mma {
         }
 #elif defined(AMD_WMMA_AVAILABLE)
         if constexpr (std::is_same_v<T, half2> || std::is_same_v<T, nv_bfloat162>) {
-            ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
-
+#if defined(RDNA4)
+                ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
+#elif defined(RDNA3)
+                ggml_cuda_memcpy_1<sizeof(t.x)/2>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
+                ggml_cuda_memcpy_1<sizeof(t.x)/2>(t.x + t.ne/2, xs0 + t.get_i(0) * stride + t.get_j(t.ne/2));
+#else
+                NO_DEVICE_CODE;
+#endif // defined(RDNA4)
         } else if constexpr (std::is_same_v<T, int>) {
             if constexpr (I == 16 && J == 4) {
                 int64_t * xi = (int64_t *) t.x;
@@ -888,6 +923,16 @@ namespace ggml_cuda_mma {
         const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
         const halfx8_t& b_frag = reinterpret_cast<const halfx8_t&>(B.x[0]);
         acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12(a_frag, b_frag, acc_frag);
+#elif defined(RDNA3)
+        using halfx16_t = __attribute__((ext_vector_type(16))) _Float16;
+        using floatx8_t = __attribute__((ext_vector_type(8))) float;
+        floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
+        const halfx16_t& a_frag = reinterpret_cast<const halfx16_t&>(A.x[0]);
+        const halfx16_t& b_frag = reinterpret_cast<const halfx16_t&>(B.x[0]);
+        acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(a_frag, b_frag, acc_frag);
+#else
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
 #endif // RDNA4
 #else
         GGML_UNUSED_VARS(D, A, B);
@@ -905,6 +950,16 @@ namespace ggml_cuda_mma {
         const bf16x8_t& a_frag = reinterpret_cast<const bf16x8_t&>(A.x[0]);
         const bf16x8_t& b_frag = reinterpret_cast<const bf16x8_t&>(B.x[0]);
         acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(a_frag, b_frag, acc_frag);
+#elif defined(RDNA3)
+        using bf16x16_t = __attribute__((ext_vector_type(16))) __bf16;
+        using floatx8_t = __attribute__((ext_vector_type(8))) float;
+        floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
+        const bf16x16_t& a_frag = reinterpret_cast<const bf16x16_t&>(A.x[0]);
+        const bf16x16_t& b_frag = reinterpret_cast<const bf16x16_t&>(B.x[0]);
+        acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32(a_frag, b_frag, acc_frag);
+#else
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
 #endif // RDNA4
 #else
         GGML_UNUSED_VARS(D, A, B);

ggml/src/ggml-cuda/mmf.cu

@@ -151,7 +151,9 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
             return false;
         }
     } else {
-        if (src1_ncols > 16) {
+        if (GGML_CUDA_CC_IS_RDNA3_0(cc) && src1_ncols > 8) {
+            return false;
+        } else if (src1_ncols > 16) {
             return false;
         }
     }
@@ -160,9 +162,9 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
         case GGML_TYPE_F32:
             return ampere_mma_available(cc);
         case GGML_TYPE_F16:
-            return volta_mma_available(cc) || turing_mma_available(cc) || (amd_wmma_available(cc) && GGML_CUDA_CC_IS_RDNA4(cc));
+            return volta_mma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc);
         case GGML_TYPE_BF16:
-            return ampere_mma_available(cc) || (amd_wmma_available(cc) && GGML_CUDA_CC_IS_RDNA4(cc));
+            return ampere_mma_available(cc) || amd_wmma_available(cc);
         default:
             return false;
     }

ggml/src/ggml-cuda/mmvf.cu

@@ -765,7 +765,10 @@ bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, const int64_t * src0
                 return ne11 <= 8;
             } else if (GGML_CUDA_CC_IS_AMD(cc)) {
                 if (fp16_mma_hardware_available(cc)) {
-                    if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+                    if (GGML_CUDA_CC_IS_RDNA3(cc)) {
+                        return ne11 <= 3;
+                    }
+                    if (GGML_CUDA_CC_IS_RDNA4(cc)) {
                         return ne11 <= 5;
                     }
                     return ne11 <= 2;

ggml/src/ggml-cuda/solve_tri.cu

@@ -3,6 +3,80 @@
 #include "solve_tri.cuh"
 
 #define MAX_N_FAST 64
+#define MAX_K_FAST 32
+
+static __global__ void get_batch_pointers(const float *  A,
+                                          float *        X,
+                                          const float ** A_ptrs,
+                                          float **       X_ptrs,
+                                          int64_t        ne02,
+                                          int64_t        total_batches,
+                                          size_t         s02,
+                                          size_t         s03,
+                                          size_t         s2,
+                                          size_t         s3) {
+    const int idx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (idx >= total_batches) {
+        return;
+    }
+
+    const int64_t i3 = idx / ne02;
+    const int64_t i2 = idx % ne02;
+
+    A_ptrs[idx] = A + i3 * s03 + i2 * s02;
+    X_ptrs[idx] = X + i3 * s3 + i2 * s2;
+}
+
+static void solve_tri_f32_cublas(ggml_backend_cuda_context & ctx,
+                                 const float *               A,
+                                 const float *               B,
+                                 float *                     X,
+                                 int                         n,
+                                 int                         k,
+                                 int64_t                     ne02,
+                                 int64_t                     ne03,
+                                 size_t                      s02,
+                                 size_t                      s03,
+                                 size_t                      s12,
+                                 size_t                      s13,
+                                 size_t                      s2,
+                                 size_t                      s3,
+                                 cudaStream_t                stream) {
+    const float   alpha         = 1.0f;
+    const int64_t total_batches = ne02 * ne03;
+    if (total_batches == 0) {
+        return;
+    }
+
+    // Bulk copy B -> X (contiguous tensors)
+    if (X != B) {
+        const int64_t total_elements_BX = n * k * total_batches;
+        CUDA_CHECK(cudaMemcpyAsync(X, B, total_elements_BX * sizeof(float), cudaMemcpyDeviceToDevice, stream));
+    }
+
+    const int id = ggml_cuda_get_device();
+
+    ggml_cuda_pool_alloc<const float *> A_ptrs_alloc(ctx.pool(id), total_batches);
+    ggml_cuda_pool_alloc<float *>       X_ptrs_alloc(ctx.pool(id), total_batches);
+
+    const float ** A_ptrs_dev = A_ptrs_alloc.get();
+    float **       X_ptrs_dev = X_ptrs_alloc.get();
+
+    get_batch_pointers<<<(total_batches + 255) / 256, 256, 0, stream>>>(A, X, A_ptrs_dev, X_ptrs_dev, ne02,
+                                                                        total_batches, s02, s03, s2, s3);
+
+    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
+
+    // Yes, this is necessary, without this we get RMSE errors
+    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(id), CUBLAS_DEFAULT_MATH));
+    CUBLAS_CHECK(cublasStrsmBatched(ctx.cublas_handle(id), CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N,
+                                    CUBLAS_DIAG_NON_UNIT, k, n, &alpha, A_ptrs_dev, n, X_ptrs_dev, k, total_batches));
+
+    // revert to standard mode from common.cuh
+    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(id), CUBLAS_TF32_TENSOR_OP_MATH));
+
+    GGML_UNUSED_VARS(s12, s13);
+}
 
 // ======================
 // Fast Kernel (n <= 64, k <= 32) - Warp-based parallel reduction
@@ -63,7 +137,7 @@ static __global__ void solve_tri_f32_fast(const float * __restrict__ A,
     float x_low  = (lane < n) ? B_batch[lane * k + col_idx] : 0.0f;
     float x_high = (WARP_SIZE + lane < n) ? B_batch[(WARP_SIZE + lane) * k + col_idx] : 0.0f;
 
-    const int half = WARP_SIZE;
+    const int half      = WARP_SIZE;
     const int nrows_low = (n < half) ? n : half;
 
 #pragma unroll
@@ -81,8 +155,8 @@ static __global__ void solve_tri_f32_fast(const float * __restrict__ A,
 
 #pragma unroll
     for (int row = half; row < n; ++row) {
-        float sum = sA[row * n + lane] * x_low;
-        const int j = half + lane;
+        float     sum = sA[row * n + lane] * x_low;
+        const int j   = half + lane;
         if (j < row) {
             sum += sA[row * n + j] * x_high;
         }
@@ -97,7 +171,7 @@ static __global__ void solve_tri_f32_fast(const float * __restrict__ A,
     for (int rr = 0; rr < 2; ++rr) {
         const int row = rr * WARP_SIZE + lane;
         if (row < n) {
-            const float val = (row < half) ? x_low : x_high;
+            const float val            = (row < half) ? x_low : x_high;
             X_batch[row * k + col_idx] = val;
         }
     }
@@ -176,20 +250,26 @@ static void solve_tri_f32_cuda(const float * A,
 }
 
 void ggml_cuda_op_solve_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];  // A (triangular n x x matrix)
-    const ggml_tensor * src1 = dst->src[1];  // B (right hand side of n x k equation columns)
+    const ggml_tensor * src0 = dst->src[0];  // A (n×n, lower triangular)
+    const ggml_tensor * src1 = dst->src[1];  // B (n×k)
 
     ggml_is_contiguous(src0);
     ggml_is_contiguous(src1);
 
-    const int64_t n = src0->ne[0];
-    const int64_t k = src1->ne[0];
+    const int64_t n    = src0->ne[0];
+    const int64_t k    = src1->ne[0];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
 
-    GGML_ASSERT(n <= 64);
-    GGML_ASSERT(k <= 32);
-
-    solve_tri_f32_cuda((const float *) src0->data, (const float *) src1->data, (float *) dst->data, n, k, src0->ne[2],
-                       src0->ne[3], src0->nb[2] / sizeof(float), src0->nb[3] / sizeof(float),
-                       src1->nb[2] / sizeof(float), src1->nb[3] / sizeof(float), dst->nb[2] / sizeof(float),
-                       dst->nb[3] / sizeof(float), ctx.stream());
+    if (n <= MAX_N_FAST && k <= MAX_K_FAST) {
+        solve_tri_f32_cuda((const float *) src0->data, (const float *) src1->data, (float *) dst->data, n, k,
+                           src0->ne[2], src0->ne[3], src0->nb[2] / sizeof(float), src0->nb[3] / sizeof(float),
+                           src1->nb[2] / sizeof(float), src1->nb[3] / sizeof(float), dst->nb[2] / sizeof(float),
+                           dst->nb[3] / sizeof(float), ctx.stream());
+    } else {
+        solve_tri_f32_cublas(ctx, (const float *) src0->data, (const float *) src1->data, (float *) dst->data, n, k,
+                             ne02, ne03, src0->nb[2] / sizeof(float), src0->nb[3] / sizeof(float),
+                             src1->nb[2] / sizeof(float), src1->nb[3] / sizeof(float), dst->nb[2] / sizeof(float),
+                             dst->nb[3] / sizeof(float), ctx.stream());
+    }
 }

ggml/src/ggml-cuda/vendors/hip.h

@@ -19,6 +19,9 @@
 #define CUDA_R_16F  HIPBLAS_R_16F
 #define CUDA_R_16BF HIPBLAS_R_16B
 #define CUDA_R_32F  HIPBLAS_R_32F
+#define CUBLAS_SIDE_RIGHT HIPBLAS_SIDE_RIGHT
+#define CUBLAS_FILL_MODE_UPPER HIPBLAS_FILL_MODE_UPPER
+#define CUBLAS_DIAG_NON_UNIT HIPBLAS_DIAG_NON_UNIT
 #define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED hipDeviceAttributeVirtualMemoryManagementSupported
 #define CU_MEM_ALLOC_GRANULARITY_RECOMMENDED hipMemAllocationGranularityRecommended
 #define CU_MEM_ALLOCATION_TYPE_PINNED hipMemAllocationTypePinned
@@ -30,6 +33,7 @@
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
 #define __all_sync(mask, var) __all(var)
 #define __any_sync(mask, var) __any(var)
+#define cublasStrsmBatched hipblasStrsmBatched
 #define cublasCreate hipblasCreate
 #define cublasDestroy hipblasDestroy
 #define cublasGemmEx hipblasGemmEx

ggml/src/ggml-cuda/vendors/musa.h

@@ -12,11 +12,16 @@
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP MUBLAS_GEMM_DEFAULT
 #define CUBLAS_OP_N MUBLAS_OP_N
 #define CUBLAS_OP_T MUBLAS_OP_T
+#define CUBLAS_DEFAULT_MATH MUBLAS_DEFAULT_MATH
+#define CUBLAS_SIDE_RIGHT MUBLAS_SIDE_RIGHT
+#define CUBLAS_FILL_MODE_UPPER MUBLAS_FILL_MODE_UPPER
+#define CUBLAS_DIAG_NON_UNIT MUBLAS_DIAG_NON_UNIT
 #define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS
 #define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_TENSOR_OP_MATH
 #define CUDA_R_16F  MUSA_R_16F
 #define CUDA_R_16BF MUSA_R_16BF
 #define CUDA_R_32F  MUSA_R_32F
+#define cublasStrsmBatched mublasStrsmBatched
 #define cublasComputeType_t cudaDataType_t
 #define cublasCreate mublasCreate
 #define cublasDestroy mublasDestroy

ggml/src/ggml-hexagon/htp/rope-ops.c

@@ -73,15 +73,15 @@ static float rope_yarn_ramp(const float low, const float high, const int i0) {
     return (1 - MIN(1, MAX(0, y)));
 }
 
-static void rope_cache_init(const float   theta_base,
-                            float         freq_scale,
-                            const float * freq_factors,
-                            float *       corr_dims,
-                            uint32_t      ne0,
-                            float         ext_factor,
-                            float         mscale,
-                            float *       cache,
-                            float         theta_scale) {
+static void rope_cache_init(const float    theta_base,
+                            const float    freq_scale,
+                            const float *  freq_factors,
+                            float *        corr_dims,
+                            const uint32_t ne0,
+                            const float    ext_factor,
+                            const float    mscale,
+                            float *        cache,
+                            const float    theta_scale) {
     // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
     float theta = theta_base;
 
@@ -92,18 +92,19 @@ static void rope_cache_init(const float   theta_base,
 
         // Get n-d rotational scaling corrected for extrapolation
         float theta_interp = freq_scale * theta_extrap;
-        float theta2       = theta_interp;
+        float theta_final  = theta_interp;
+        float mscale_final = mscale;
 
         if (ext_factor != 0.0f) {
             float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
-            theta2         = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+            theta_final    = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
 
             // Get n-d magnitude scaling corrected for interpolation
-            mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale);
+            mscale_final *= 1.0f + 0.1f * logf(1.0f / freq_scale);
         }
 
-        cache[i0 + 0] = cosf(theta2) * mscale;
-        cache[i0 + 1] = sinf(theta2) * mscale;
+        cache[i0 + 0] = cosf(theta_final) * mscale_final;
+        cache[i0 + 1] = sinf(theta_final) * mscale_final;
 
         theta *= theta_scale;
     }
@@ -151,9 +152,9 @@ static void init_rope_ctx(struct rope_th_ctx * rope_ctx, struct htp_ops_context
 }
 
 static void hvx_calc_rope_neox_f32(const float * restrict src0,
-                              float * restrict dst,
-                              const int num_elems,
-                              const float * restrict theta_cache) {
+                                   float * restrict dst,
+                                   const int num_elems,
+                                   const float * restrict theta_cache) {
     // for (int i = 0; i < num_elems; i += 2) {
     //const float cos_theta = theta_cache[i + 0];
     //const float sin_theta = theta_cache[i + 1];
@@ -192,7 +193,7 @@ static void hvx_calc_rope_neox_f32(const float * restrict src0,
         HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
         HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);
 
-        *(HVX_Vector *) dst_curr          = Q6_Vsf_equals_Vqf32(v4);
+        *(HVX_Vector *) dst_curr               = Q6_Vsf_equals_Vqf32(v4);
         *(HVX_Vector *) (dst_curr + half_size) = Q6_Vsf_equals_Vqf32(v5);
 
         src0_curr += VLEN;
@@ -259,7 +260,7 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
                          const uint32_t       ir1,
                          int                  nth,
                          int                  ith,
-                         int                  opt_path) {
+                         const int            opt_path) {
     struct htp_ops_context * octx = rope_ctx->octx;
 
     const struct htp_tensor * src0 = &octx->src0;
@@ -267,8 +268,8 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
     const struct htp_tensor * src2 = &octx->src2;
     struct htp_tensor *       dst  = &octx->dst;
 
-    const int32_t mode  = rope_ctx->mode;
-    const bool is_neox  = mode & HTP_ROPE_TYPE_NEOX;
+    const int32_t mode    = rope_ctx->mode;
+    const bool    is_neox = mode & HTP_ROPE_TYPE_NEOX;
 
     htp_rope_preamble;
 
@@ -281,8 +282,9 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
         freq_factors = (const float *) src2->data;
     }
 
-    int ir = 0;
-
+    const uint32_t i1_end       = MIN(ir1, ne1);
+    const int32_t  half_dims    = rope_ctx->n_dims / 2;
+    const size_t   remain_bytes = (ne0 - rope_ctx->n_dims) * sizeof(float);
     for (uint32_t i3 = 0; i3 < ne3; i3++) {      // batch
         for (uint32_t i2 = 0; i2 < ne2; i2++) {  // seq-len
             const int32_t p = pos[i2];
@@ -290,14 +292,7 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
             rope_cache_init(p, rope_ctx->freq_scale, freq_factors, rope_ctx->corr_dims, ne0, rope_ctx->ext_factor,
                             rope_ctx->attn_factor, wp0, rope_ctx->theta_scale);
 
-            for (uint32_t i1 = 0; i1 < ne1; i1++) {  // attn-heads
-                if (ir++ < ir0) {
-                    continue;
-                }
-                if (ir > ir1) {
-                    break;
-                }
-
+            for (uint32_t i1 = ir0; i1 < i1_end; i1++) {  // attn-heads
                 const float * src      = (float *) ((char *) src0->data + i3 * nb03 + i2 * nb02 + i1 * nb01);
                 float *       dst_data = (float *) ((char *) dst->data + i3 * nb3 + i2 * nb2 + i1 * nb1);
 
@@ -310,6 +305,9 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
                     } else {
                         hvx_calc_rope_f32(src_loc, dst_data_loc, rope_ctx->n_dims, wp0);
                     }
+
+                    src_loc += rope_ctx->n_dims;
+                    dst_data_loc += rope_ctx->n_dims;
                 } else {
                     for (uint32_t i0 = 0; i0 < rope_ctx->n_dims; i0 += 2) {
                         const float cos_theta = wp0[i0 + 0];
@@ -317,10 +315,10 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
 
                         if (is_neox) {
                             const float x0 = src_loc[0];
-                            const float x1 = src_loc[rope_ctx->n_dims/2];
+                            const float x1 = src_loc[half_dims];
 
-                            dst_data_loc[0] = x0 * cos_theta - x1 * sin_theta;
-                            dst_data_loc[rope_ctx->n_dims/2] = x0 * sin_theta + x1 * cos_theta;
+                            dst_data_loc[0]         = x0 * cos_theta - x1 * sin_theta;
+                            dst_data_loc[half_dims] = x0 * sin_theta + x1 * cos_theta;
 
                             src_loc += 1;
                             dst_data_loc += 1;
@@ -335,15 +333,13 @@ static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
                             dst_data_loc += 2;
                         }
                     }
+
+                    src_loc += (is_neox ? half_dims : 0);
+                    dst_data_loc += (is_neox ? half_dims : 0);
                 }
 
-                for (uint32_t i0 = rope_ctx->n_dims; i0 < ne0; i0 += 2) {
-                    dst_data_loc[0] = src_loc[0];
-                    dst_data_loc[1] = src_loc[1];
-
-                    src_loc += 2;
-                    dst_data_loc += 2;
-                }
+                // TODO: use simd to speed up the remaining elements copy
+                memcpy(dst_data_loc, src_loc, remain_bytes);
             }
         }
     }

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

@@ -411,6 +411,38 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv(ggml_me
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched(ggml_metal_library_t lib, const ggml_tensor * op, int ssm_conv_bs) {
+    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(op->src[0]));
+    GGML_ASSERT(ggml_is_contiguous(op->src[1]));
+
+    char base[256];
+    char name[256];
+
+    const char * suffix = "";
+    if (op->src[1]->ne[0] % 4 == 0) {
+        suffix = "_4";
+    }
+
+    snprintf(base, 256, "kernel_ssm_conv_%s_%s_batched%s", ggml_type_name(op->src[0]->type), ggml_type_name(op->src[1]->type), suffix);
+    snprintf(name, 256, "%s_ssm_conv_bs=%d", base, ssm_conv_bs);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, ssm_conv_bs, FC_SSM_CONV + 0);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan(ggml_metal_library_t lib, const ggml_tensor * op)  {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
 
@@ -427,7 +459,12 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan(ggml_me
         res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
     }
 
-    res.smem = 32*sizeof(float)*nsg;
+    // Shared memory layout:
+    // - sgptg * NW floats for partial sums (nsg * 32)
+    // - sgptg floats for shared_x_dt (nsg)
+    // - sgptg floats for shared_dA (nsg)
+    // Total: nsg * (32 + 2) floats
+    res.smem = (32 + 2)*sizeof(float)*nsg;
 
     return res;
 }

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

@@ -117,6 +117,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_ad
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_tri               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv              (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext        (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1, int nsg, int nxpsg, int r1ptg);

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

@@ -77,6 +77,7 @@
 #define FC_MUL_MV                      600
 #define FC_MUL_MM                      700
 #define FC_ROPE                        800
+#define FC_SSM_CONV                    900
 
 // op-specific constants
 #define OP_FLASH_ATTN_EXT_NQPTG 8

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

@@ -221,7 +221,7 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
         }
 
         if (ctx->debug_graph > 0) {
-            GGML_LOG_DEBUG("%s: node[%5d] - %-12s %s\n", __func__, idx, ggml_op_name(node->op), is_concurrent ? "(concurrent)" : "");
+            GGML_LOG_DEBUG("%s: node[%5d] - %-12s %-12s %s\n", __func__, idx, ggml_op_name(node->op), ggml_get_name(node), is_concurrent ? "(concurrent)" : "");
         }
         if (ctx->debug_graph > 1) {
             GGML_TENSOR_LOCALS( int64_t, ne0, node->src[0], ne);
@@ -1365,15 +1365,43 @@ int ggml_metal_op_ssm_conv(ggml_metal_op_t ctx, int idx) {
         /*.nb2  =*/ nb2,
     };
 
-    auto pipeline = ggml_metal_library_get_pipeline_ssm_conv(lib, op);
+    // Use batched kernel for prefill (ne1 > 1) to reduce threadgroup dispatch overhead
+    const bool use_batched = (ne1 > 1);
 
-    ggml_metal_encoder_set_pipeline(enc, pipeline);
-    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
-    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         3);
+    if (use_batched) {
+        // Determine the smallest power of 2 that's >= ne1, but <= 256
+        int BATCH_SIZE;
+        if      (ne1 > 128) BATCH_SIZE = 256;
+        else if (ne1 > 64 ) BATCH_SIZE = 128;
+        else if (ne1 > 32 ) BATCH_SIZE = 64;
+        else if (ne1 > 16 ) BATCH_SIZE = 32;
+        else if (ne1 > 8  ) BATCH_SIZE = 16;
+        else if (ne1 > 4  ) BATCH_SIZE = 8;
+        else                BATCH_SIZE = 2;
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne1, ne02, 1, 1, 1);
+        auto pipeline = ggml_metal_library_get_pipeline_ssm_conv_batched(lib, op, BATCH_SIZE);
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         3);
+
+        // Dispatch: ne01 rows, ceil(ne1/BATCH_SIZE) token batches, ne02 sequences
+        // Each threadgroup has BATCH_SIZE threads, each handling one token
+        const int n_token_batches = (ne1 + BATCH_SIZE - 1) / BATCH_SIZE;
+        ggml_metal_encoder_dispatch_threadgroups(enc, ne01, n_token_batches, ne02, BATCH_SIZE, 1, 1);
+    } else {
+        auto pipeline = ggml_metal_library_get_pipeline_ssm_conv(lib, op);
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline);
+        ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+        ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),         3);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne1, ne02, 1, 1, 1);
+    }
 
     return 1;
 }

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

@@ -2343,7 +2343,102 @@ kernel void kernel_ssm_conv_f32_f32_4(
     x[0] = sumf;
 }
 
+constant short FC_ssm_conv_bs   [[function_constant(FC_SSM_CONV + 0)]];
+
+// Batched version: each threadgroup processes multiple tokens for better efficiency
+// Thread layout: each thread handles one token, threadgroup covers BATCH_SIZE tokens
+kernel void kernel_ssm_conv_f32_f32_batched(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_batched_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
 // ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
+// Optimized version: reduces redundant memory loads by having one thread load shared values
 kernel void kernel_ssm_scan_f32(
         constant ggml_metal_kargs_ssm_scan & args,
         device const void * src0,
@@ -2363,7 +2458,15 @@ kernel void kernel_ssm_scan_f32(
         uint3    tgpg[[threadgroups_per_grid]]) {
     constexpr short NW = N_SIMDWIDTH;
 
-    shared[tpitg.x] = 0.0f;
+    // Shared memory layout:
+    // [0..sgptg*NW-1]: partial sums for reduction (existing)
+    // [sgptg*NW..sgptg*NW+sgptg-1]: pre-computed x_dt values for each token in batch
+    // [sgptg*NW+sgptg..sgptg*NW+2*sgptg-1]: pre-computed dA values for each token in batch
+    threadgroup float * shared_sums = shared;
+    threadgroup float * shared_x_dt = shared + sgptg * NW;
+    threadgroup float * shared_dA   = shared + sgptg * NW + sgptg;
+
+    shared_sums[tpitg.x] = 0.0f;
 
     const int32_t i0 = tpitg.x;
     const int32_t i1 = tgpig.x;
@@ -2403,32 +2506,47 @@ kernel void kernel_ssm_scan_f32(
     for (int i2 = 0; i2 < n_t; i2 += sgptg) {
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        for (int t = 0; t < sgptg && i2 + t < n_t; t++) {
-            const float dt0  = dt[0];
+        // Pre-compute x_dt and dA for this batch of tokens
+        // Only first sgptg threads do the loads and expensive math
+        if (i0 < sgptg && i2 + i0 < n_t) {
+            // ns12 and ns21 are element strides (nb12/nb10, nb21/nb20)
+            device const float * x_t  = x  + i0 * args.ns12;
+            device const float * dt_t = dt + i0 * args.ns21;
+
+            const float dt0  = dt_t[0];
             const float dtsp = dt0 <= 20.0f ? log(1.0f + exp(dt0)) : dt0;
-            const float x_dt = x[0] * dtsp;
-            const float dA   = exp(dtsp * A0);
+            shared_x_dt[i0] = x_t[0] * dtsp;
+            shared_dA[i0]   = dtsp;  // Store dtsp, compute exp(dtsp * A0) per-thread since A0 varies
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (int t = 0; t < sgptg && i2 + t < n_t; t++) {
+            const float x_dt = shared_x_dt[t];
+            const float dA   = exp(shared_dA[t] * A0);
 
             s = (s0 * dA) + (B[i0] * x_dt);
 
             const float sumf = simd_sum(s * C[i0]);
 
             if (tiisg == 0) {
-                shared[t*NW + sgitg] = sumf;
+                shared_sums[t*NW + sgitg] = sumf;
             }
 
             // recurse
             s0 = s;
 
-            x  += args.ns12;
-            dt += args.ns21;
             B  += args.ns42;
             C  += args.ns52;
         }
 
+        // Advance pointers for next batch
+        x  += sgptg * args.ns12;
+        dt += sgptg * args.ns21;
+
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        const float sumf = simd_sum(shared[sgitg*NW + tiisg]);
+        const float sumf = simd_sum(shared_sums[sgitg*NW + tiisg]);
 
         if (tiisg == 0 && i2 + sgitg < n_t) {
             y[sgitg*nh*nr] = sumf;

ggml/src/ggml-sycl/softmax.cpp

@@ -277,7 +277,7 @@ static void soft_max_f32_sycl(const float *x, const T *mask,
     const int id       = get_current_device_id();
     const size_t smpbo = ggml_sycl_info().devices[id].smpbo;
 
-    if (nbytes_shared <= smpbo) {
+    if (nbytes_shared <= smpbo && ncols_x <= max_block_size) {
         launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(
             x, mask, sinks, dst, params, stream, block_dims, block_nums,
             nbytes_shared);

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

@@ -659,6 +659,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_cos_f32;
     vk_pipeline pipeline_log[2];
     vk_pipeline pipeline_tri[2];
+    vk_pipeline pipeline_diag[2];
     vk_pipeline pipeline_clamp_f32;
     vk_pipeline pipeline_pad_f32;
     vk_pipeline pipeline_roll_f32;
@@ -722,6 +723,11 @@ struct vk_device_struct {
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
     vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512;
     vk_pipeline pipeline_soft_max_back_f32;
+
+    vk_pipeline pipeline_soft_max_large1_f32, pipeline_soft_max_large1_f32_f16;
+    vk_pipeline pipeline_soft_max_large2_f32, pipeline_soft_max_large2_f32_f16;
+    vk_pipeline pipeline_soft_max_large3_f32, pipeline_soft_max_large3_f32_f16;
+
     vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16, pipeline_rope_norm_f32_f16;
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16, pipeline_rope_neox_f32_f16;
     vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16;
@@ -757,7 +763,8 @@ struct vk_device_struct {
 
     vk_pipeline pipeline_flash_attn_split_k_reduce;
 
-    vk_pipeline pipeline_topk_moe[num_topk_moe_pipelines][TOPK_MOE_COUNT];
+    // [2] is for whether to take n_experts from spec constant (0) or push constant (1)
+    vk_pipeline pipeline_topk_moe[num_topk_moe_pipelines][TOPK_MOE_COUNT][2];
 
     std::vector<vk_pipeline_ref> all_pipelines;
 
@@ -1149,6 +1156,7 @@ static_assert(sizeof(vk_op_multi_add_push_constants) <= 256);
 
 struct vk_op_topk_moe_push_constants {
     uint32_t n_rows;
+    uint32_t n_experts_push;
     uint32_t n_expert_used;
     float clamp_min;
     float clamp_max;
@@ -3730,6 +3738,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ4_XS],  "get_rows_iq4_xs",  get_rows_iq4_xs_len,  get_rows_iq4_xs_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ4_NL],  "get_rows_iq4_nl",  get_rows_iq4_nl_len,  get_rows_iq4_nl_data,  "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_MXFP4],   "get_rows_mxfp4",   get_rows_mxfp4_len,   get_rows_mxfp4_data,   "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_I32],     "get_rows_i32",     get_rows_i32_len,     get_rows_i32_data,     "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f32_f32",  get_rows_f32_f32_len,  get_rows_f32_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F16 ], "get_rows_f16_f32",  get_rows_f16_f32_len,  get_rows_f16_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
@@ -3917,6 +3926,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_tri[0], "tri_f32", tri_f32_len, tri_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_tri[1], "tri_f16", tri_f16_len, tri_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_diag[0], "diag_f32", diag_f32_len, diag_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_diag[1], "diag_f16", diag_f16_len, diag_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(device, device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);
@@ -3996,6 +4008,13 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 4, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
     ggml_vk_create_pipeline(device, device->pipeline_soft_max_back_f32, "soft_max_back_f32", soft_max_back_f32_len, soft_max_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1, true);
 
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_large1_f32,     "soft_max_large1_f32",     soft_max_large1_f32_len,     soft_max_large1_f32_data,     "main", 6, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 128, 4 }, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_large2_f32,     "soft_max_large2_f32",     soft_max_large2_f32_len,     soft_max_large2_f32_data,     "main", 6, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 128, 4 }, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_large3_f32,     "soft_max_large3_f32",     soft_max_large3_f32_len,     soft_max_large3_f32_data,     "main", 6, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 128, 4 }, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_large1_f32_f16, "soft_max_large1_f32_f16", soft_max_large1_f32_f16_len, soft_max_large1_f32_f16_data, "main", 6, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 128, 4 }, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_large2_f32_f16, "soft_max_large2_f32_f16", soft_max_large2_f32_f16_len, soft_max_large2_f32_f16_data, "main", 6, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 128, 4 }, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_large3_f32_f16, "soft_max_large3_f32_f16", soft_max_large3_f32_f16_len, soft_max_large3_f32_f16_data, "main", 6, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 128, 4 }, 1, true);
+
     ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f32, "rope_multi_f32", rope_multi_f32_len, rope_multi_f32_data, "main", 5, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
@@ -4204,10 +4223,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_whcn_f16_f32, "conv2d_dw_whcn_f16_f32", conv2d_dw_whcn_f16_f32_len, conv2d_dw_whcn_f16_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_conv2d_dw_cwhn_f16_f32, "conv2d_dw_cwhn_f16_f32", conv2d_dw_cwhn_f16_f32_len, conv2d_dw_cwhn_f16_f32_data, "main", 3, sizeof(vk_op_conv2d_dw_push_constants), {512, 1, 1}, {}, 1);
 
-    for (uint32_t i = 0; i < num_topk_moe_pipelines; ++i) {
-        ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_EARLY_SOFTMAX],      "topk_moe_f32_early_softmax_"+std::to_string(i),       topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 0, 0}, 1, true, true, device->subgroup_size);
-        ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_EARLY_SOFTMAX_NORM], "topk_moe_f32_early_softmax_norm"+std::to_string(i),   topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 1, 0}, 1, true, true, device->subgroup_size);
-        ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_LATE_SOFTMAX],       "topk_moe_f32_late_softmax"+std::to_string(i),         topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 0, 1}, 1, true, true, device->subgroup_size);
+    for (uint32_t use_push = 0; use_push < 2; ++use_push) {
+        for (uint32_t i = 0; i < num_topk_moe_pipelines; ++i) {
+            ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_EARLY_SOFTMAX][use_push],      "topk_moe_f32_early_softmax_"+std::to_string(i),       topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 0, 0, use_push}, 1, true, true, device->subgroup_size);
+            ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_EARLY_SOFTMAX_NORM][use_push], "topk_moe_f32_early_softmax_norm"+std::to_string(i),   topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 1, 0, use_push}, 1, true, true, device->subgroup_size);
+            ggml_vk_create_pipeline2(device, device->pipeline_topk_moe[i][TOPK_MOE_LATE_SOFTMAX][use_push],       "topk_moe_f32_late_softmax"+std::to_string(i),         topk_moe_f32_len, topk_moe_f32_data, "main", 3, sizeof(vk_op_topk_moe_push_constants), {1, 1, 1}, {device->subgroup_size, 1u<<i, 0, 1, use_push}, 1, true, true, device->subgroup_size);
+        }
     }
 
     for (auto &c : compiles) {
@@ -8274,6 +8295,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     switch (op) {
     case GGML_OP_GET_ROWS:
         GGML_ASSERT(src1->type == GGML_TYPE_I32);
+        if (src0->type == GGML_TYPE_I32) {
+            // i32 src only supports i32 result
+            GGML_ASSERT(dst->type == GGML_TYPE_I32);
+            return ctx->device->pipeline_get_rows[src0->type];
+        }
         if (dst->type == GGML_TYPE_F16) {
             return ctx->device->pipeline_get_rows[src0->type];
         }
@@ -8400,6 +8426,12 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_tri[dst->type == GGML_TYPE_F16];
         }
         return nullptr;
+    case GGML_OP_DIAG:
+        if (src0->type == dst->type &&
+            (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16)) {
+            return ctx->device->pipeline_diag[dst->type == GGML_TYPE_F16];
+        }
+        return nullptr;
     case GGML_OP_CLAMP:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_clamp_f32;
@@ -8554,7 +8586,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             uint32_t idx = (uint32_t)ceilf(log2f(float(dst->ne[0])));
             GGML_ASSERT(idx < num_topk_moe_pipelines);
             topk_moe_mode mode = ggml_vk_num_additional_ops_to_topk_moe_mode(ctx->num_additional_fused_ops);
-            return ctx->device->pipeline_topk_moe[idx][mode];
+            // use n_experts from push constant if it's not equal to the power of two spec constant
+            bool use_push = dst->ne[0] != (1u << idx);
+            return ctx->device->pipeline_topk_moe[idx][mode][use_push];
         }
 
         if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
@@ -9091,6 +9125,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_COS:
     case GGML_OP_LOG:
     case GGML_OP_TRI:
+    case GGML_OP_DIAG:
     case GGML_OP_CLAMP:
     case GGML_OP_PAD:
     case GGML_OP_ROLL:
@@ -9778,6 +9813,12 @@ static void ggml_vk_tri(ggml_backend_vk_context * ctx, vk_context& subctx, const
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_TRI, std::move(p));
 }
 
+static void ggml_vk_diag(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_DIAG, std::move(p));
+}
+
 static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
     p.param1 = ggml_get_op_params_f32(dst, 0);
@@ -10111,7 +10152,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_SOFT_MAX, {
+    vk_op_soft_max_push_constants pc {
         ncols,
         src1 != nullptr ? nrows_y : (uint32_t)0,
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],
@@ -10122,7 +10163,55 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
         n_head_log2,
         nrows_x,
         src2 != nullptr
-    });
+    };
+
+    if (ncols <= 16384) {
+        ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, src2, nullptr, dst, GGML_OP_SOFT_MAX, std::move(pc));
+    } else {
+
+        vk_subbuffer buf_a = ggml_vk_tensor_subbuffer(ctx, src0);
+        vk_subbuffer buf_b = src1 ? ggml_vk_tensor_subbuffer(ctx, src1) : buf_a;
+        vk_subbuffer buf_c = src2 ? ggml_vk_tensor_subbuffer(ctx, src2) : buf_a;
+        vk_subbuffer buf_d = ggml_vk_tensor_subbuffer(ctx, dst);
+
+        uint32_t elems_per_wg = 128 * 4;
+        uint32_t num_wgs = CEIL_DIV(ncols, elems_per_wg);
+        size_t tmp_size = num_wgs * nrows_x * sizeof(float);
+
+        if (ctx->prealloc_size_x < tmp_size) {
+            ctx->prealloc_size_x = tmp_size;
+            ggml_vk_preallocate_buffers(ctx, subctx);
+        }
+        if (ctx->prealloc_size_y < tmp_size) {
+            ctx->prealloc_size_y = tmp_size;
+            ggml_vk_preallocate_buffers(ctx, subctx);
+        }
+        if (ctx->prealloc_x_need_sync || ctx->prealloc_y_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+
+        vk_subbuffer buf_x = { ctx->prealloc_x, 0, tmp_size };
+        vk_subbuffer buf_y = { ctx->prealloc_y, 0, tmp_size };
+
+        std::array<uint32_t, 3> elements = { num_wgs, nrows_x, 1 };
+
+        vk_pipeline pipeline1 = src1 && src1->type == GGML_TYPE_F16 ? ctx->device->pipeline_soft_max_large1_f32_f16 : ctx->device->pipeline_soft_max_large1_f32;
+        vk_pipeline pipeline2 = src1 && src1->type == GGML_TYPE_F16 ? ctx->device->pipeline_soft_max_large2_f32_f16 : ctx->device->pipeline_soft_max_large2_f32;
+        vk_pipeline pipeline3 = src1 && src1->type == GGML_TYPE_F16 ? ctx->device->pipeline_soft_max_large3_f32_f16 : ctx->device->pipeline_soft_max_large3_f32;
+
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline1, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline2, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline3, 1);
+
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline1, { buf_a, buf_b, buf_c, buf_d, buf_x, buf_y }, pc, elements);
+        ggml_vk_sync_buffers(ctx, subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline2, { buf_a, buf_b, buf_c, buf_d, buf_x, buf_y }, pc, elements);
+        ggml_vk_sync_buffers(ctx, subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline3, { buf_a, buf_b, buf_c, buf_d, buf_x, buf_y }, pc, elements);
+
+        ctx->prealloc_x_need_sync = true;
+        ctx->prealloc_y_need_sync = true;
+    }
 }
 
 static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -10158,6 +10247,7 @@ static void ggml_vk_topk_moe(ggml_backend_vk_context * ctx, vk_context& subctx,
 
     vk_op_topk_moe_push_constants pc {};
     pc.n_rows = n_rows;
+    pc.n_experts_push = n_experts;
     pc.n_expert_used = n_expert_used;
     if (mode == TOPK_MOE_EARLY_SOFTMAX_NORM) {
         ggml_tensor * clamp = cgraph->nodes[node_idx + 7];
@@ -11857,6 +11947,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_TRI:
         ggml_vk_tri(ctx, compute_ctx, src0, node);
 
+        break;
+    case GGML_OP_DIAG:
+        ggml_vk_diag(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_CLAMP:
         ggml_vk_clamp(ctx, compute_ctx, src0, node);
@@ -12832,8 +12926,7 @@ static bool ggml_vk_can_fuse_topk_moe(ggml_backend_vk_context * ctx, const struc
     }
 
     const int n_expert = softmax->ne[0];
-    // n_expert must be a power of 2
-    if (!is_pow2(n_expert) || n_expert > (1 << (num_topk_moe_pipelines-1))) {
+    if (n_expert > (1 << (num_topk_moe_pipelines-1))) {
         return false;
     }
 
@@ -13877,6 +13970,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     case GGML_TYPE_IQ4_XS:
                     case GGML_TYPE_IQ4_NL:
                     case GGML_TYPE_MXFP4:
+                    case GGML_TYPE_I32:
                         return true;
                     default:
                         return false;
@@ -14001,6 +14095,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_LOG:
         case GGML_OP_TRI:
+        case GGML_OP_DIAG:
             return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
                    op->type == op->src[0]->type;
         case GGML_OP_ARGSORT:
@@ -14591,6 +14686,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
             tensor_clone = ggml_log(ggml_ctx, src_clone[0]);
         } else if (tensor->op == GGML_OP_TRI) {
             tensor_clone = ggml_tri(ggml_ctx, src_clone[0], ggml_get_op_params_i32(tensor, 0));
+        } else if (tensor->op == GGML_OP_DIAG) {
+            tensor_clone = ggml_diag(ggml_ctx, src_clone[0]);
         } else if (tensor->op == GGML_OP_CLAMP) {
             const float * params = (const float *)tensor->op_params;
             tensor_clone = ggml_clamp(ggml_ctx, src_clone[0], params[0], params[1]);

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

@@ -0,0 +1,29 @@
+#version 450
+
+#include "rte.glsl"
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L);
+    const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
+    const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L);
+    const uint i12_offset = i12*p.ne11*p.ne10;
+    const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L);
+    const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
+
+    if (i10 == i11) {
+        const float val = float(data_a[get_aoffset() + i13*p.nb03 + i12*p.nb02 + 0*p.nb01 + i10*p.nb00]);
+        data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val);
+    } else {
+        data_d[get_doffset() + dst_idx(idx)] = D_TYPE(0);
+    }
+}

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

@@ -26,9 +26,9 @@ void main() {
             const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
 #if defined(DATA_A_BF16)
-            FLOAT_TYPE v = FLOAT_TYPE(bf16_to_fp32(data_a[a_offset + i00]));
+            TEMP_TYPE v = TEMP_TYPE(bf16_to_fp32(data_a[a_offset + i00]));
 #else
-            FLOAT_TYPE v = FLOAT_TYPE(data_a[a_offset + i00]);
+            TEMP_TYPE v = TEMP_TYPE(data_a[a_offset + i00]);
 #endif
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
             data_d[d_offset + i00] = D_TYPE(v);

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp

@@ -7,34 +7,50 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
 
-void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
-    const uint y_idx = i * QUANT_K + 32 * ib32;
+void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i,
+                     const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint y_idx_base = i * QUANT_K + 32 * ib32;
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {    
+        const uint base_b_idx = (j * p.batch_stride_b + b_offset + y_idx_base) / 4;
+        [[unroll]] for (uint l = 0; l < 4; ++l) {            
+            const vec4 b_val_0 = vec4(data_b_v4[base_b_idx + 2 * l]);
+            const vec4 b_val_1 = vec4(data_b_v4[base_b_idx + 2 * l + 1]);
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
-    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const float d = float(data_a[ibi].d);
-        const uint qh = data_a[ibi].qh[ib32];
-        const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
-        const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+            // index for data_a
+            uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
 
-        [[unroll]] for (uint l = 0; l < 4; ++l) {
-            const uint qs = data_a[ibi].qs[4 * ib32 + l];
-            const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3);
-            const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]);
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {                
+                const float d = float(data_a[ibi].d);
+                const uint qh = data_a[ibi].qh[ib32];
 
-            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-                vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
-                vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
+                const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
+                const uint qs = data_a[ibi].qs[4 * ib32 + l];
+                const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3);              
+                const uint16_t grid = uint16_t(iq1s_grid[qs | (idxhi << 8)]);
 
-                FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-                [[unroll]] for (int k = 0; k < 4; ++k) {
-                    sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta,
-                          fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum));
-                }
-                temp[j][n] = fma(dl, sum, temp[j][n]);
+                const float delta_val = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
+                const vec4 delta_v = vec4(delta_val);				
+                const vec4 fbits0 = vec4(
+                    float(bitfieldExtract(grid, 0, 2)),
+                    float(bitfieldExtract(grid, 2, 2)),
+                    float(bitfieldExtract(grid, 4, 2)),
+                    float(bitfieldExtract(grid, 6, 2))
+                );				
+                const vec4 fbits1 = vec4(
+                    float(bitfieldExtract(grid, 8, 2)),
+                    float(bitfieldExtract(grid, 10, 2)),
+                    float(bitfieldExtract(grid, 12, 2)),
+                    float(bitfieldExtract(grid, 14, 2))
+                );
+				
+                vec4 sum_v = fma(b_val_0, fbits0 + delta_v, vec4(0.0));
+                sum_v      = fma(b_val_1, fbits1 + delta_v, sum_v);
+				FLOAT_TYPE sum = dot(sum_v, vec4(1.0));
+				
+                temp[j][n] = fma(dl, sum, temp[j][n]);				
+                ibi += num_blocks_per_row;
             }
         }
-        ibi += num_blocks_per_row;
     }
 }
 

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

@@ -244,17 +244,20 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint iqs = idx % 128;                 // 0..127
 
             const uint n = iqs / 64;                    // 0,1
-            const uint b = (iqs % 64) / 32;             // 0,1
+            const uint b = ((iqs % 64) / 32) * 4;       // 0,4
             const uint is_b = (iqs % 16) / 8;           // 0,1
             const uint qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
             const uint is = 8 * n + qhshift + is_b;     // 0..15
-            const uint qsi = n * 64 + (iqs % 32) * 2;   // 0,2,4..126
-            const uint qhi = n * 32 + (iqs % 16) * 2;   // 0,2,4..62
+            const uint qsi = n * 32 + (iqs % 32);       // 0..63
+            const uint qhi = n * 16 + (iqs % 16);       // 0..31
 
             const float dscale = float(data_a[ib].d) * float(data_a[ib].scales[is]);
 
-            buf_a[buf_idx] = FLOAT_TYPE_VEC2(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32),
-                                             dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
+            const uint ql = (uint(data_a_packed16[ib].ql[qsi]) >> b) & 0x0F0F;
+            const uint qh = (uint(data_a_packed16[ib].qh[qhi]) >> qhshift) & 0x0303;
+            const vec2 q = (vec2(unpack8(ql | (qh << 4)).xy) - 32) * dscale;
+
+            buf_a[buf_idx] = FLOAT_TYPE_VEC2(q.x, q.y);
 #elif defined(DATA_A_IQ1_S)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;

ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large1.comp

@@ -0,0 +1,62 @@
+#version 450
+
+#include "soft_max_large_common.glsl"
+
+void main() {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint rowx = gl_WorkGroupID.y;
+    const uint wg_start = gl_WorkGroupID.x * BLOCK_SIZE * num_iters;
+
+    const uint32_t i03 = rowx / (p.ne01 * p.ne02);
+    const uint32_t i02 = (rowx - i03 * p.ne01 * p.ne02) / p.ne01;
+    const uint32_t i01 = rowx % p.ne01;
+
+    uint rowy_start = 0;
+    if (p.KY > 0) {
+        rowy_start = i01 * p.nb11 + (i02 % p.ne12) * p.nb12 + (i03 % p.ne13) * p.nb13;
+    }
+
+    if (rowx >= p.nrows_x) {
+        return;
+    }
+
+    float slope = get_slope(rowx);
+
+    // Find max
+    FLOAT_TYPE max_val = p.has_sinks == 0 ? uintBitsToFloat(0xFF800000) : data_c[i02];
+
+    [[unroll]] for (uint col0 = wg_start, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
+        const uint col = col0 + tid;
+
+        FLOAT_TYPE a = FLOAT_TYPE(0);
+        if (col < p.KX) {
+            a = data_a[rowx * p.KX + col];
+        }
+
+        FLOAT_TYPE b = FLOAT_TYPE(0);
+        if (p.KY > 0 && col < p.KX) {
+            b = data_b[rowy_start + col];
+        }
+
+        FLOAT_TYPE v = a * p.scale + slope * b;
+
+        if (col < p.KX) {
+            max_val = max(max_val, v);
+        }
+    }
+
+    // reduce across the workgroup
+    vals[tid] = max_val;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            vals[tid] = max(vals[tid], vals[tid + s]);
+        }
+        barrier();
+    }
+
+    if (tid == 0) {
+        max_val = vals[0];
+        data_m[rowx * gl_NumWorkGroups.x + gl_WorkGroupID.x] = max_val;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large2.comp

@@ -0,0 +1,79 @@
+#version 450
+
+#include "soft_max_large_common.glsl"
+
+void main() {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint rowx = gl_WorkGroupID.y;
+    const uint wg_start = gl_WorkGroupID.x * BLOCK_SIZE * num_iters;
+
+    const uint32_t i03 = rowx / (p.ne01 * p.ne02);
+    const uint32_t i02 = (rowx - i03 * p.ne01 * p.ne02) / p.ne01;
+    const uint32_t i01 = rowx % p.ne01;
+
+    uint rowy_start = 0;
+    if (p.KY > 0) {
+        rowy_start = i01 * p.nb11 + (i02 % p.ne12) * p.nb12 + (i03 % p.ne13) * p.nb13;
+    }
+
+    if (rowx >= p.nrows_x) {
+        return;
+    }
+
+    float slope = get_slope(rowx);
+
+    // Find max
+    FLOAT_TYPE max_val = p.has_sinks == 0 ? uintBitsToFloat(0xFF800000) : data_c[i02];
+
+    [[unroll]] for (uint i = 0; i < gl_NumWorkGroups.x; i += BLOCK_SIZE) {
+        if (i + tid < gl_NumWorkGroups.x) {
+            max_val = max(max_val, data_m[rowx * gl_NumWorkGroups.x + i + tid]);
+        }
+    }
+
+    // reduce across the workgroup
+    vals[tid] = max_val;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            vals[tid] = max(max_val, vals[tid + s]);
+        }
+        barrier();
+    }
+
+    max_val = vals[0];
+    barrier();
+
+    FLOAT_TYPE sum = FLOAT_TYPE(0.0f);
+
+    // Compute sum{exp(x - max)}
+    [[unroll]] for (uint col0 = wg_start, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
+        const uint col = col0 + tid;
+
+        if (col >= p.KX) {
+            break;
+        }
+
+        // compute exp(a*scale+b*slope), add it to sum
+        const uint i = rowx * p.KX + col;
+        FLOAT_TYPE val;
+        val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy_start + col]) : FLOAT_TYPE(0.0f)) - max_val);
+        sum += val;
+        data_d[i] = D_TYPE(val);
+    }
+
+    // reduce across the workgroup
+    vals[tid] = sum;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            vals[tid] += vals[tid + s];
+        }
+        barrier();
+    }
+
+    if (tid == 0) {
+        sum = vals[0];
+        data_s[rowx * gl_NumWorkGroups.x + gl_WorkGroupID.x] = sum;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/soft_max_large3.comp

@@ -0,0 +1,65 @@
+#version 450
+
+#include "soft_max_large_common.glsl"
+
+shared FLOAT_TYPE sumsh[BLOCK_SIZE];
+
+void main() {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint rowx = gl_WorkGroupID.y;
+    const uint wg_start = gl_WorkGroupID.x * BLOCK_SIZE * num_iters;
+
+    const uint32_t i03 = rowx / (p.ne01 * p.ne02);
+    const uint32_t i02 = (rowx - i03 * p.ne01 * p.ne02) / p.ne01;
+    const uint32_t i01 = rowx % p.ne01;
+
+    uint rowy_start = 0;
+    if (p.KY > 0) {
+        rowy_start = i01 * p.nb11 + (i02 % p.ne12) * p.nb12 + (i03 % p.ne13) * p.nb13;
+    }
+
+    if (rowx >= p.nrows_x) {
+        return;
+    }
+
+    FLOAT_TYPE max_val = p.has_sinks == 0 ? uintBitsToFloat(0xFF800000) : data_c[i02];
+    FLOAT_TYPE sum = FLOAT_TYPE(0.0f);
+
+    [[unroll]] for (uint i = 0; i < gl_NumWorkGroups.x; i += BLOCK_SIZE) {
+        if (i + tid < gl_NumWorkGroups.x) {
+            max_val = max(max_val, data_m[rowx * gl_NumWorkGroups.x + i + tid]);
+            sum += data_s[rowx * gl_NumWorkGroups.x + i + tid];
+        }
+    }
+
+    // reduce across the workgroup
+    vals[tid] = max_val;
+    sumsh[tid] = sum;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            vals[tid] = max(max_val, vals[tid + s]);
+            sumsh[tid] += sumsh[tid + s];
+        }
+        barrier();
+    }
+
+    max_val = vals[0];
+    sum = sumsh[0];
+
+    if (p.has_sinks != 0) {
+        sum += FLOAT_TYPE(exp(FLOAT_TYPE(data_c[i02]) - max_val));
+    }
+
+    FLOAT_TYPE rcpdivisor = 1.0/sum;
+
+    [[unroll]] for (uint col0 = wg_start, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
+        const uint col = col0 + tid;
+
+        if (col >= p.KX) {
+            continue;
+        }
+
+        data_d[rowx*p.KX + col] *= D_TYPE(rcpdivisor);
+    }
+}

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

@@ -0,0 +1,53 @@
+#extension GL_EXT_control_flow_attributes : enable
+
+layout (push_constant) uniform parameter
+{
+    uint KX;
+    uint KY;
+    uint ne00;
+    uint ne01;
+    uint ne02;
+    uint ne12;
+    uint ne13;
+    uint nb11;
+    uint nb12;
+    uint nb13;
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+    uint n_head_log2;
+    uint nrows_x;
+    uint has_sinks;
+} p;
+
+#include "types.glsl"
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 128;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+layout(constant_id = 1) const uint num_iters = 4;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
+layout (binding = 2) readonly buffer Z {float data_c[];};
+layout (binding = 3) buffer D {D_TYPE data_d[];};
+layout (binding = 4) buffer M {float data_m[];};
+layout (binding = 5) buffer S {float data_s[];};
+
+shared FLOAT_TYPE vals[BLOCK_SIZE];
+
+float get_slope(uint rowx) {
+    float slope = 1.0f;
+
+    // ALiBi
+    if (p.max_bias > 0.0f) {
+        const uint h = (rowx / p.ne01) % p.ne02; // head index
+
+        const float base = h < p.n_head_log2 ? p.m0 : p.m1;
+        const uint   exp = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    return slope;
+}

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

@@ -10,6 +10,7 @@
 layout (push_constant) uniform parameter
 {
     uint n_rows;
+    uint n_experts_push;
     uint n_expert_used;
     float clamp_min;
     float clamp_max;
@@ -18,11 +19,16 @@ layout (push_constant) uniform parameter
 layout(local_size_x_id = 0, local_size_y = 4, local_size_z = 1) in;
 
 layout(constant_id = 0) const uint WARP_SIZE = 32;
-layout(constant_id = 1) const uint n_experts = 512;
+layout(constant_id = 1) const uint n_experts_spec = 512;
 layout(constant_id = 2) const bool with_norm = true;
 layout(constant_id = 3) const bool late_softmax = false;
+layout(constant_id = 4) const bool nexperts_use_push = false;
 
-const uint experts_per_thread = (n_experts > WARP_SIZE) ? n_experts / WARP_SIZE : 1;
+uint n_experts = nexperts_use_push ? n_experts_push : n_experts_spec;
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+const uint experts_per_thread = CEIL_DIV(n_experts_spec, WARP_SIZE);
 
 layout (binding = 0, std430) readonly buffer Logits {float logits[];};
 layout (binding = 1, std430) writeonly buffer Weights {float weights[];};
@@ -94,7 +100,7 @@ void main() {
     }
 
     if (!late_softmax) {
-        softmax_warp_inplace(wt, n_experts, lane, false);
+        softmax_warp_inplace(wt, n_experts, lane, nexperts_use_push);
     }
 
     // at this point, each thread holds a portion of softmax,

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

@@ -704,13 +704,15 @@ void process_shaders() {
         shader = (tname == "f32" || tname == "f16" || tname == "bf16") ? "get_rows.comp" : "get_rows_quant.comp";
 
         if (tname == "f16") {
-            string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}));
+            string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{"TEMP_TYPE", "FLOAT_TYPE"}, {data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}));
         } else {
-            string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}}));
+            string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{"TEMP_TYPE", "FLOAT_TYPE"}, {data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}}));
         }
-        string_to_spv("get_rows_" + tname + "_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}}));
+        string_to_spv("get_rows_" + tname + "_f32", shader, merge_maps(base_dict, {{"TEMP_TYPE", "FLOAT_TYPE"}, {data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}}));
     }
 
+    string_to_spv("get_rows_i32", "get_rows.comp", {{"TEMP_TYPE", "uint"}, {"A_TYPE", "uint"}, {"B_TYPE", "int"}, {"D_TYPE", "uint"}});
+
     string_to_spv("mul_mat_vec_p021_f16_f32_subgroup_add", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}});
     string_to_spv("mul_mat_vec_p021_f16_f32",              "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}});
     string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}});
@@ -854,6 +856,8 @@ void process_shaders() {
 
     string_to_spv("tri_f16",        "tri.comp",         {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("tri_f32",        "tri.comp",         {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("diag_f16",       "diag.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("diag_f32",       "diag.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
     string_to_spv("softplus_f16",   "softplus.comp",    {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("softplus_f32",   "softplus.comp",    {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
@@ -899,6 +903,13 @@ void process_shaders() {
     string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
     string_to_spv("soft_max_back_f32", "soft_max_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("soft_max_large1_f32", "soft_max_large1.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("soft_max_large2_f32", "soft_max_large2.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("soft_max_large3_f32", "soft_max_large3.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    string_to_spv("soft_max_large1_f32_f16", "soft_max_large1.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
+    string_to_spv("soft_max_large2_f32_f16", "soft_max_large2.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
+    string_to_spv("soft_max_large3_f32_f16", "soft_max_large3.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
+
     string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
     string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
     string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});

ggml/src/ggml.c

@@ -124,6 +124,13 @@ static void ggml_print_backtrace_symbols(void) {
     int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0]));
     backtrace_symbols_fd(trace, nptrs, STDERR_FILENO);
 }
+#elif defined(__APPLE__)
+#include <execinfo.h>
+static void ggml_print_backtrace_symbols(void) {
+    void * trace[100];
+    int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0]));
+    backtrace_symbols_fd(trace, nptrs, STDERR_FILENO);
+}
 #else
 static void ggml_print_backtrace_symbols(void) {
     // platform not supported
@@ -135,6 +142,20 @@ void ggml_print_backtrace(void) {
     if (GGML_NO_BACKTRACE) {
         return;
     }
+#if defined(__APPLE__)
+    // On macOS, fork+debugger attachment is problematic due to:
+    // 1. libdispatch "poisons" forked child processes
+    // 2. lldb has issues attaching to parent from forked child
+    // Use simple backtrace() instead to avoid Terminal.app crashes
+    const char * GGML_BACKTRACE_LLDB = getenv("GGML_BACKTRACE_LLDB");
+    if (!GGML_BACKTRACE_LLDB) {
+        fprintf(stderr, "WARNING: Using native backtrace. Set GGML_BACKTRACE_LLDB for more info.\n");
+        fprintf(stderr, "WARNING: GGML_BACKTRACE_LLDB may cause native MacOS Terminal.app to crash.\n");
+        fprintf(stderr, "See: https://github.com/ggml-org/llama.cpp/pull/17869\n");
+        ggml_print_backtrace_symbols();
+        return;
+    }
+#endif
 #if defined(__linux__)
     FILE * f = fopen("/proc/self/status", "r");
     size_t size = 0;
@@ -5239,8 +5260,6 @@ struct ggml_tensor * ggml_flash_attn_ext(
 
     if (mask) {
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) &&
-                "the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big");
         //GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
 
         GGML_ASSERT(q->ne[2] % mask->ne[2] == 0);

grammars/README.md

@@ -67,6 +67,30 @@ Parentheses `()` can be used to group sequences, which allows for embedding alte
 - `{m,n}` repeats the precedent symbol or sequence at between `m` and `n` times (included)
 - `{0,n}` repeats the precedent symbol or sequence at most `n` times (included)
 
+## Tokens
+
+Tokens allow grammars to match specific tokenizer tokens rather than character sequences. This is useful for constraining outputs based on special tokens (like `<think>` or `</think>`).
+
+Tokens can be specified in two ways:
+
+1. **Token ID**: Use angle brackets with the token ID in square brackets: `<[token-id]>`. For example, `<[1000]>` matches the token with ID 1000.
+
+2. **Token string**: Use angle brackets with the token text directly: `<token>`. For example, `<think>` will match the token whose text is exactly `<think>`. This only works if the string tokenizes to exactly one token in the vocabulary, otherwise the grammar will fail to parse.
+
+You can negate token matches using the `!` prefix: `!<[1000]>` or `!<think>` matches any token *except* the specified one.
+
+```
+# Match a thinking block: <think>...</think>
+# Using token strings (requires these to be single tokens in the vocab)
+root ::= <think> thinking </think> .*
+thinking ::= !</think>*
+
+# Equivalent grammar using explicit token IDs
+# Assumes token 1000 = <think>, token 1001 = </think>
+root ::= <[1000]> thinking <[1001]> .*
+thinking ::= !<[1001]>*
+```
+
 ## Comments and newlines
 
 Comments can be specified with `#`:

models/templates/mistralai-Ministral-3-14B-Reasoning-2512.jinja

@@ -0,0 +1,126 @@
+{#- Default system message if no system prompt is passed. #}
+{%- set default_system_message = '# HOW YOU SHOULD THINK AND ANSWER\n\nFirst draft your thinking process (inner monologue) until you arrive at a response. Format your response using Markdown, and use LaTeX for any mathematical equations. Write both your thoughts and the response in the same language as the input.\n\nYour thinking process must follow the template below:[THINK]Your thoughts or/and draft, like working through an exercise on scratch paper. Be as casual and as long as you want until you are confident to generate the response to the user.[/THINK]Here, provide a self-contained response.' %}
+
+{#- Begin of sequence token. #}
+{{- bos_token }}
+
+{#- Handle system prompt if it exists. #}
+{#- System prompt supports text content or text and thinking chunks. #}
+{%- if messages[0]['role'] == 'system' %}
+    {{- '[SYSTEM_PROMPT]' -}}
+    {%- if messages[0]['content'] is string %}
+        {{- messages[0]['content'] -}}
+    {%- else %}        
+        {%- for block in messages[0]['content'] %}
+            {%- if block['type'] == 'text' %}
+                {{- block['text'] }}
+            {%- elif block['type'] == 'thinking' %}
+                {{- '[THINK]' + block['thinking'] + '[/THINK]' }}
+            {%- else %}
+                {{- raise_exception('Only text and thinking chunks are supported in system message contents.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+    {{- '[/SYSTEM_PROMPT]' -}}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set loop_messages = messages %}
+    {%- if default_system_message != '' %}
+        {{- '[SYSTEM_PROMPT]' + default_system_message + '[/SYSTEM_PROMPT]' }}
+    {%- endif %}
+{%- endif %}
+
+
+{#- Tools definition #}
+{%- set tools_definition = '' %}
+{%- set has_tools = false %}
+{%- if tools is defined and tools is not none and tools|length > 0 %}
+    {%- set has_tools = true %}
+    {%- set tools_definition = '[AVAILABLE_TOOLS]' + (tools| tojson) + '[/AVAILABLE_TOOLS]' %}
+    {{- tools_definition }}
+{%- endif %}
+
+{#- Checks for alternating user/assistant messages. #}
+{%- set ns = namespace(index=0) %}
+{%- for message in loop_messages %}
+    {%- if message.role == 'user' or (message.role == 'assistant' and (message.tool_calls is not defined or message.tool_calls is none or message.tool_calls | length == 0)) %}
+        {%- if (message['role'] == 'user') != (ns.index % 2 == 0) %}
+            {{- raise_exception('After the optional system message, conversation roles must alternate user and assistant roles except for tool calls and results.') }}
+        {%- endif %}
+        {%- set ns.index = ns.index + 1 %}
+    {%- endif %}
+{%- endfor %}
+
+{#- Handle conversation messages. #}
+{%- for message in loop_messages %}
+
+    {#- User messages supports text content or text and image chunks. #}
+    {%- if message['role'] == 'user' %}
+        {%- if message['content'] is string %}
+            {{- '[INST]' + message['content'] + '[/INST]' }}
+        {%- elif message['content'] | length > 0 %}
+            {{- '[INST]' }}
+            {%- if message['content'] | length == 2 %}
+                {%- set blocks = message['content'] | sort(attribute='type') %}
+            {%- else %}
+                {%- set blocks = message['content'] %}
+            {%- endif %}
+            {%- for block in blocks %}
+                {%- if block['type'] == 'text' %}
+                    {{- block['text'] }}
+                {%- elif block['type'] in ['image', 'image_url'] %}
+                    {{- '[IMG]' }}
+                {%- else %}
+                    {{- raise_exception('Only text, image and image_url chunks are supported in user message content.') }}
+                {%- endif %}
+            {%- endfor %}
+            {{- '[/INST]' }}
+        {%- else %}
+            {{- raise_exception('User message must have a string or a list of chunks in content') }}
+        {%- endif %}
+
+    {#- Assistant messages supports text content or text, image and thinking chunks. #}
+    {%- elif message['role'] == 'assistant' %}
+        {%- if (message['content'] is none or message['content'] == '' or message['content']|length == 0) and (message['tool_calls'] is not defined or message['tool_calls'] is none or message['tool_calls']|length == 0) %}
+            {{- raise_exception('Assistant message must have a string or a list of chunks in content or a list of tool calls.') }}
+        {%- endif %}
+
+        {%- if message['content'] is string and message['content'] != '' %}
+            {{- message['content'] }}
+        {%- elif message['content'] | length > 0 %}
+            {%- for block in message['content'] %}
+                {%- if block['type'] == 'text' %}
+                    {{- block['text'] }}
+                {%- elif block['type'] == 'thinking' %}
+                    {{- '[THINK]' + block['thinking'] + '[/THINK]' }}
+                {%- else %}
+                    {{- raise_exception('Only text and thinking chunks are supported in assistant message contents.') }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        
+        {%- if message['tool_calls'] is defined and message['tool_calls'] is not none and message['tool_calls']|length > 0 %}
+            {%- for tool in message['tool_calls'] %}
+                {{- '[TOOL_CALLS]' }}
+                {%- set name = tool['function']['name'] %}
+                {%- set arguments = tool['function']['arguments'] %}
+                {%- if arguments is not string %}
+                    {%- set arguments = arguments|tojson|safe %}
+                {%- elif arguments == '' %}
+                    {%- set arguments = '{}' %}
+                {%- endif %}
+                {{- name + '[ARGS]' + arguments }}
+            {%- endfor %}
+        {%- endif %}
+
+        {{- eos_token }}
+
+    {#- Tool messages only supports text content. #}
+    {%- elif message['role'] == 'tool' %}
+        {{- '[TOOL_RESULTS]' + message['content']|string + '[/TOOL_RESULTS]' }}
+
+    {#- Raise exception for unsupported roles. #}
+    {%- else %}
+        {{- raise_exception('Only user, assistant and tool roles are supported, got ' + message['role'] + '.') }}
+    {%- endif %}
+{%- endfor %}

pyrightconfig.json

@@ -1,5 +1,5 @@
 {
-  "extraPaths": ["gguf-py"],
+  "extraPaths": ["gguf-py", "examples/model-conversion/scripts"],
   "pythonVersion": "3.9",
   "pythonPlatform": "All",
   "reportUnusedImport": "warning",

scripts/compare-logprobs.py

@@ -0,0 +1,281 @@
+import argparse
+import requests
+import json
+from pathlib import Path
+import logging
+
+logger = logging.getLogger("compare-logprobs")
+logging.basicConfig(level=logging.INFO)
+
+
+DESCRIPTION = """
+Compare logits between llama.cpp and another inference engine using OpenAI-compatible server endpoints.
+
+Unlike compare-logits.py, it allows dumping logits from a hosted API endpoint. Useful when it's not possible to run both models locally.
+
+Example usage:
+    Step 1: Dump logits from two different servers
+        python scripts/compare-logprobs.py dump logits_llama.log http://localhost:8080/v1/completions
+        python scripts/compare-logprobs.py dump logits_other.log http://other-engine:8000/v1/completions
+
+        (optionally, you can add --api-key <key> if the endpoint requires authentication)
+
+    Step 2: Compare the dumped logits
+        python scripts/compare-logprobs.py compare logits_llama.log logits_other.log report.md
+"""
+
+
+def generate_input_prompt(length: int) -> list[str]:
+    CORPUS = """
+    You are an advanced AI assistant capable of using tools to gather information, perform calculations, or execute tasks. Always think step by step before responding. If a user's query requires external data, computation, or actions beyond your internal knowledge, use the appropriate tools via function calls.
+
+    ### Tool Call Format:
+    When you need to use a tool, output the call in this exact XML format. Include the opening and closing tags. Do not escape arguments; they will be parsed as plain text.
+
+    You can make multiple calls in one go by placing them one after another.
+    """
+    words = [w.strip() for w in CORPUS.strip().split(" ")]
+    words = [w for w in words if len(w) > 0]  # filter out empty strings
+    while len(words) < length:
+        words += words
+    return words[:length]
+
+
+def dump_logits(
+    endpoint: str,
+    output_path: Path,
+    input_words: list[str],
+    pattern: list[tuple[bool, int]],
+    api_key=None,
+):
+    logger.info(f"Dumping logits to {output_path} from endpoint {endpoint}...")
+    words = input_words
+    curr_text = ""
+    n_total = sum(n for get, n in pattern if get)
+    n_done = 0
+    i_cur = 0
+    i_total = len(words)
+    with output_path.open("w") as f:
+        for get, n in pattern:
+            if not get:
+                # skip n words
+                for i in range(n):
+                    curr_text += words.pop(0) + " "
+                    i_cur += 1
+                continue
+            # get n words
+            for i in range(n):
+                curr_text += words.pop(0) + " "
+                payload = {
+                    "prompt": curr_text.strip(),
+                    "temperature": 0.0,
+                    "top_k": 1,
+                    "max_tokens": 1,
+                    "logprobs": 1,
+                    "stream": False,
+                }
+                response = requests.post(
+                    endpoint,
+                    json=payload,
+                    headers={"Authorization": f"Bearer {api_key}"} if api_key else {},
+                )
+                response.raise_for_status()
+                data = response.json()
+                data["__index"] = i_cur  # add index for easier debugging later
+                data = json.dumps(data)
+                f.write(f"{data}\n")
+                n_done += 1
+                i_cur += 1
+                logger.info(
+                    f"\n\n{data}\n\n[Step: {n_done}/{n_total} | Word: {i_cur}/{i_total}]"
+                )
+    logger.info(f"Logits dumped to {output_path}")
+
+
+def get_token_logprobs(data: dict):
+    logprobs = data["choices"][0]["logprobs"]
+    if "content" in logprobs:
+        # llama.cpp case
+        top = logprobs["content"][0]["top_logprobs"][0]
+        return top["token"], top["logprob"]
+    else:
+        # vllm case
+        tokens = logprobs["tokens"]
+        token_logprobs = logprobs["token_logprobs"]
+        return tokens[0], token_logprobs[0]
+
+
+def clean_text(text: str) -> str:
+    return (
+        "'"
+        + text.replace("\n", "\\n")
+        .replace("\t", "\\t")
+        .replace("\r", "\\r")
+        .replace("|", "\\|")
+        + "'"
+    )
+
+
+def compare_logits(input1: Path, input2: Path, output_path: Path):
+    with input1.open("r") as f1, input2.open("r") as f2, output_path.open("w") as fout:
+        lines1 = f1.readlines()
+        lines2 = f2.readlines()
+
+        tab_header = [
+            "idx",
+            input1.name,
+            "logprob_1",
+            input2.name,
+            "logprob_2",
+            "diff (abs)",
+        ]
+        tab_entries = []
+        tab_max_widths = [len(h) for h in tab_header]
+
+        assert len(lines1) == len(
+            lines2
+        ), "Input files must have the same number of lines."
+
+        fout.write("# Logits Comparison Report\n\n")
+        for i, (line1, line2) in enumerate(zip(lines1, lines2)):
+            if not line1.strip() or not line2.strip():
+                continue  # skip empty lines
+
+            data1 = json.loads(line1)
+            data2 = json.loads(line2)
+
+            idx1 = data1.get("__index", -1)
+            idx2 = data2.get("__index", -1)
+            if idx1 != idx2:
+                logger.warning(
+                    f"Warning: Mismatched indices at line {i}: {idx1} vs {idx2}"
+                )
+
+            token1, logprob1 = get_token_logprobs(data1)
+            token2, logprob2 = get_token_logprobs(data2)
+
+            token1 = clean_text(token1)
+            token2 = clean_text(token2)
+            abs_diff = abs(logprob1 - logprob2)
+
+            tab_entries.append(
+                (
+                    str(idx1 + 1),
+                    token1,
+                    f"{logprob1:.4f}",
+                    token2,
+                    f"{logprob2:.4f}",
+                    f"{(abs_diff):.4f}",
+                )
+            )
+
+        for i in range(len(tab_entries)):
+            for j in range(len(tab_header)):
+                tab_max_widths[j] = max(tab_max_widths[j], len(tab_entries[i][j]))
+
+        output = ""
+        for j in range(len(tab_header)):
+            output += f"| {tab_header[j]:<{tab_max_widths[j]}} "
+        output += "|\n"
+        for j in range(len(tab_header)):
+            output += f"|{'-' * (tab_max_widths[j] + 2)}"
+        output += "|\n"
+        for entry in tab_entries:
+            for j in range(len(tab_header)):
+                output += f"| {entry[j]:<{tab_max_widths[j]}} "
+            output += "|\n"
+
+        logger.info("\n" + output)
+        fout.write(output)
+        logger.info(f"Report written to {output_path}")
+
+
+def parse_pattern(pattern: str) -> list[tuple[bool, int]]:
+    parts = pattern.split(",")
+    result = []
+    for i, part in enumerate(parts):
+        n = int(part)
+        if i % 2 == 0:
+            result.append((True, n))  # get n words
+        else:
+            result.append((False, n))  # skip n words
+    return result
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description=DESCRIPTION, formatter_class=argparse.RawTextHelpFormatter
+    )
+    subparsers = parser.add_subparsers(
+        dest="verb", required=True, help="action to perform"
+    )
+
+    # dump subcommand
+    parser_dump = subparsers.add_parser("dump", help="dump logits from an endpoint")
+    parser_dump.add_argument(
+        "output", type=Path, help="output path for dumped logits (.log)"
+    )
+    parser_dump.add_argument(
+        "endpoint", type=str, help="OAI-compat /completions endpoint"
+    )
+    parser_dump.add_argument(
+        "--api-key",
+        type=str,
+        default=None,
+        help="API key for authentication (if required)",
+    )
+    parser_dump.add_argument(
+        "--file",
+        type=Path,
+        default=None,
+        help="File containing prompt to use instead of the default",
+    )
+    parser_dump.add_argument(
+        "--pattern",
+        type=str,
+        default="10,1000,10,4000,10",
+        help="Pattern n_get,n_skip,... where n_get is number of words to get and n_skip is number of words to skip (num of words, NOT num of tokens)",
+    )
+
+    # compare subcommand
+    parser_compare = subparsers.add_parser(
+        "compare", help="compare two dumped logits files"
+    )
+    parser_compare.add_argument("input1", type=Path, help="first input file (.log)")
+    parser_compare.add_argument("input2", type=Path, help="second input file (.log)")
+    parser_compare.add_argument(
+        "output", type=Path, help="output path for comparison report (.md)"
+    )
+
+    try:
+        return parser.parse_args()
+    except Exception as e:
+        parser.print_help()
+        raise e
+
+
+def main():
+    args = parse_args()
+
+    if args.verb == "dump":
+        pattern = parse_pattern(args.pattern)
+        input_length = sum(n for _, n in pattern)
+        input_words = generate_input_prompt(input_length)
+        if args.file is not None:
+            with args.file.open("r") as f:
+                input_words = f.read().strip().split(" ")
+                if input_length < sum(n for _, n in pattern):
+                    raise ValueError(
+                        f"Input file has only {input_length} words, but pattern requires at least {input_length} words."
+                    )
+                input_length = len(input_words)
+        logger.info(f"Using {input_length} words")
+        dump_logits(args.endpoint, args.output, input_words, pattern, args.api_key)
+    elif args.verb == "compare":
+        compare_logits(args.input1, args.input2, args.output)
+    else:
+        raise ValueError(f"Unknown verb: {args.verb}")
+
+
+if __name__ == "__main__":
+    main()

scripts/snapdragon/adb/run-cli.sh

@@ -46,7 +46,7 @@ adb $adbserial shell " \
     LD_LIBRARY_PATH=$basedir/$branch/lib   \
     ADSP_LIBRARY_PATH=$basedir/$branch/lib \
     $verbose $experimental $sched $opmask $profile $nhvx $ndev       \
-      ./$branch/bin/llama-cli --no-mmap -m $basedir/../gguf/$model   \
+      ./$branch/bin/llama-completion --no-mmap -m $basedir/../gguf/$model   \
          --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1             \
          --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on \
          -ngl 99 --device $device $cli_opts $@ \

scripts/sync-ggml.last

@@ -1 +1 @@
-55bc9320a4aae82af18e23eefd5de319a755d7b9
+130bc125a88bb57664b88932c48c38a1cb316fac

src/CMakeLists.txt

@@ -67,7 +67,7 @@ add_library(llama
             models/gemma-embedding.cpp
             models/gemma.cpp
             models/gemma2-iswa.cpp
-            models/gemma3-iswa.cpp
+            models/gemma3.cpp
             models/gemma3n-iswa.cpp
             models/glm4-moe.cpp
             models/glm4.cpp
@@ -139,6 +139,7 @@ add_library(llama
 set_target_properties(llama PROPERTIES
     VERSION ${LLAMA_INSTALL_VERSION}
     SOVERSION 0
+    MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
 )
 
 target_include_directories(llama PRIVATE .)

src/llama-batch.cpp

@@ -695,6 +695,8 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
     udata->seq_idx   .resize(LLAMA_MAX_SEQ, -1);
     udata->output    .resize(n_tokens);
 
+    udata->seq_id_data.reserve(n_tokens);
+
     seq_set_t seq_set_unq;
 
     for (size_t i = 0; i < idxs.size(); ++i) {
@@ -716,11 +718,13 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
         }
 
         udata->n_seq_id[i] = batch.n_seq_id[idxs[i]];
-        udata->seq_id[i]   = batch.seq_id[idxs[i]];
         udata->output[i]   = batch.logits[idxs[i]];
 
         for (int s = 0; s < udata->n_seq_id[i]; ++s) {
-            seq_set_unq.set(udata->seq_id[i][s]);
+            const llama_seq_id seq_id = batch.seq_id[idxs[i]][s];
+
+            udata->seq_id_data.push_back(seq_id);
+            seq_set_unq.set(seq_id);
         }
 
         if (udata->output[i]) {
@@ -728,6 +732,12 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, u
         }
     }
 
+    llama_seq_id * seq_id_ptr = udata->seq_id_data.data();
+    for (size_t i = 0; i < idxs.size(); ++i) {
+        udata->seq_id[i] = seq_id_ptr;
+        seq_id_ptr += udata->n_seq_id[i];
+    }
+
     for (uint32_t s = 0; s < n_seq_max; ++s) {
         if (seq_set_unq.test(s)) {
             udata->seq_idx[s] = udata->seq_id_unq.size();

src/llama-batch.h

@@ -56,13 +56,15 @@ struct llama_ubatch {
         std::vector<float>          embd;
         std::vector<llama_pos>      pos;
         std::vector<int32_t>        n_seq_id;
-        std::vector<llama_seq_id *> seq_id;
+        std::vector<llama_seq_id *> seq_id;      // these point into the seq_id_data below
         std::vector<llama_seq_id>   seq_id_unq;
         std::vector<int32_t>        seq_idx;
         std::vector<int8_t>         output;
+
+        std::vector<llama_seq_id> seq_id_data;
     };
 
-    // the llama_ubatch pointers above point to this data if set. otherwise - points to non-owning data
+    // the llama_ubatch pointers above point to this data if set. otherwise - point to external non-owning data
     std::shared_ptr<data_t> data;
 };
 

src/llama-context.cpp

@@ -9,6 +9,7 @@
 #include "llama-model.h"
 
 #include <cinttypes>
+#include <cmath>
 #include <cstring>
 #include <limits>
 #include <stdexcept>
@@ -72,6 +73,43 @@ llama_context::llama_context(
         cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
     }
 
+    if (cparams.yarn_ext_factor != 0) {
+        static auto get_mscale = [](float scale, float mscale) {
+            return scale <= 1.0f ? 1.0f : (0.1f * mscale * logf(scale) + 1.0f);
+        };
+
+        const float factor = 1.0f / cparams.rope_freq_scale;
+
+        // ref: https://github.com/huggingface/transformers/blob/6d00f6b0a5679c36510f203e4226e36f517c3032/src/transformers/modeling_rope_utils.py#L336-L348
+        if (hparams.rope_yarn_log_mul != 0.0f) {
+            // note: here we assume `mscale == 1.0f`
+            // TODO: start reading the actual value of mscale and handle the case where it is not 1.0f
+                  float mscale          = 1.0f;
+            const float mscale_all_dims = hparams.rope_yarn_log_mul;
+
+            // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+            // special-case DEEPSEEK v2:
+            // https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite-Chat/blob/main/config.json#L42-L43
+            if (model.arch == LLM_ARCH_DEEPSEEK2 && mscale_all_dims != 1.0f) {
+                mscale = mscale_all_dims;
+            }
+
+            cparams.yarn_attn_factor = get_mscale(factor, mscale) / get_mscale(factor, mscale_all_dims);
+
+            LLAMA_LOG_WARN("%s: setting new yarn_attn_factor = %.4f (mscale == %.1f, mscale_all_dim = %.1f)\n",
+                    __func__, cparams.yarn_attn_factor, mscale, mscale_all_dims);
+        } else {
+            cparams.yarn_attn_factor = get_mscale(factor, 1.0f);
+        }
+
+        // when YARN is applied with yarn_ext_factor != 0.0f, we need to cancel this factor:
+        // https://github.com/ggml-org/llama.cpp/blob/a81a569577cc38b32558958b048228150be63eae/ggml/src/ggml-cpu/ops.cpp#L5541-L5544
+        //
+        // ref: https://github.com/ggml-org/llama.cpp/discussions/7416
+        //      https://github.com/ggml-org/llama.cpp/pull/17945
+        cparams.yarn_attn_factor *= 1.0f / (1.0f + 0.1f * logf(factor));
+    }
+
     cparams.yarn_attn_factor *= hparams.rope_attn_factor;
 
     if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
@@ -93,14 +131,6 @@ llama_context::llama_context(
     // with causal attention, the batch size is limited by the context size
     cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
 
-    // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask
-    // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext)
-    // ref: https://github.com/ggerganov/llama.cpp/pull/5021
-    // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_memory
-    if (cparams.n_batch < GGML_KQ_MASK_PAD) {
-        LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
-        cparams.n_batch = GGML_KQ_MASK_PAD;
-    }
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
 
     cparams.op_offload = params.op_offload;
@@ -1326,6 +1356,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) {
             // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark)
             LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
+            synchronize();
             buf_output = nullptr;
             logits = nullptr;
             embd = nullptr;

src/llama-grammar.cpp

@@ -181,6 +181,52 @@ static std::pair<uint32_t, const char *> parse_char(const char * src) {
     throw std::runtime_error("unexpected end of input");
 }
 
+static std::pair<uint32_t, const char *> parse_token(const llama_vocab * vocab, const char * src) {
+    const char * pos = src;
+    if (*pos != '<') {
+        throw std::runtime_error(std::string("expecting '<' at ") + pos);
+    }
+    pos++;
+
+    // Parse <[id]>
+    if (*pos == '[') {
+        pos++;
+        const char * int_end = parse_int(pos);
+        uint32_t token_id = std::stoul(std::string(pos, int_end - pos));
+        pos = int_end;
+        if (*pos != ']') {
+            throw std::runtime_error(std::string("expecting ']' at ") + pos);
+        }
+        pos++;
+        if (*pos != '>') {
+            throw std::runtime_error(std::string("expecting '>' at ") + pos);
+        }
+        pos++;
+        return std::make_pair(token_id, pos);
+    }
+
+    if (vocab == nullptr) {
+        throw std::runtime_error(std::string("no vocab to parse token at ") + src);
+    }
+
+    // Parse <token> and tokenize to obtain the token id
+    while (*pos != 0 && *pos != '>') {
+        pos++;
+    }
+    if (*pos != '>') {
+        throw std::runtime_error(std::string("expecting '>' at ") + pos);
+    }
+    pos++;
+
+    llama_token tokens[2];
+    int32_t n_tokens = vocab->tokenize(src, static_cast<int32_t>(pos - src), tokens, 2, false, true);
+    if (n_tokens != 1) {
+        // must tokenize to exactly 1 token
+        throw std::runtime_error("invalid token '" + std::string(src, pos - src) + "'");
+    }
+    return std::make_pair(tokens[0], pos);
+}
+
 static void print_grammar_char(FILE * file, uint32_t c) {
     if (0x20 <= c && c <= 0x7f) {
         fprintf(file, "%c", static_cast<char>(c));
@@ -212,6 +258,8 @@ static void print_rule_binary(FILE * file, const llama_grammar_rule & rule) {
             case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
             case LLAMA_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
             case LLAMA_GRETYPE_CHAR_ANY:       fprintf(file, "CHAR_ANY");       break;
+            case LLAMA_GRETYPE_TOKEN:          fprintf(file, "TOKEN");          break;
+            case LLAMA_GRETYPE_TOKEN_NOT:      fprintf(file, "TOKEN_NOT");      break;
         }
         switch (elem.type) {
             case LLAMA_GRETYPE_END:
@@ -228,6 +276,17 @@ static void print_rule_binary(FILE * file, const llama_grammar_rule & rule) {
                 print_grammar_char(file, elem.value);
                 fprintf(file, "\") ");
                 break;
+            case LLAMA_GRETYPE_TOKEN:
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
+            case LLAMA_GRETYPE_TOKEN_NOT:
+                fprintf(file, "!");
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
         }
     }
     fprintf(file, "\n");
@@ -284,6 +343,17 @@ static void print_rule(
             case LLAMA_GRETYPE_CHAR_ANY:
                 fprintf(file, ".");
                 break;
+            case LLAMA_GRETYPE_TOKEN:
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
+            case LLAMA_GRETYPE_TOKEN_NOT:
+                fprintf(file, "!");
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
         }
         if (is_char_element(elem)) {
             switch (rule[i + 1].type) {
@@ -444,6 +514,17 @@ const char * llama_grammar_parser::parse_sequence(
                 }
             }
             pos = parse_space(pos + 1, is_nested);
+        } else if (*pos == '<' || *pos == '!') { // token
+            auto type = LLAMA_GRETYPE_TOKEN;
+            if (*pos == '!') { // token inverse
+                type = LLAMA_GRETYPE_TOKEN_NOT;
+                pos++;
+            }
+            auto token_pair = parse_token(vocab, pos);
+            const char * token_end  = token_pair.second;
+            last_sym_start = rule.size();
+            rule.push_back({type, token_pair.first});
+            pos = parse_space(token_end, is_nested);
         } else if (is_word_char(*pos)) { // rule reference
             const char * name_end    = parse_name(pos);
             uint32_t ref_rule_id = get_symbol_id(pos, name_end - pos);
@@ -691,6 +772,21 @@ static bool llama_grammar_match_partial_char(
     return !is_positive_char;
 }
 
+// returns true iff token matches the rule at pos (regular or inverse)
+// asserts that pos is pointing to a token element
+static bool llama_grammar_match_token(
+    const llama_grammar_element * pos,
+    const llama_token             token) {
+    GGML_ASSERT(pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT);
+    if (pos->type == LLAMA_GRETYPE_TOKEN) {
+        return pos->value == static_cast<uint32_t>(token);
+    }
+    if (pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+        return pos->value != static_cast<uint32_t>(token);
+    }
+    return false;
+}
+
 // transforms a grammar pushdown stack into N possible stacks, all ending
 // at a character range (terminal element)
 static void llama_grammar_advance_stack(
@@ -738,6 +834,8 @@ static void llama_grammar_advance_stack(
         case LLAMA_GRETYPE_CHAR:
         case LLAMA_GRETYPE_CHAR_NOT:
         case LLAMA_GRETYPE_CHAR_ANY:
+        case LLAMA_GRETYPE_TOKEN:
+        case LLAMA_GRETYPE_TOKEN_NOT:
             if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
                 // only add the stack if it's not a duplicate of one we already have
                 new_stacks.emplace_back(stack);
@@ -831,26 +929,38 @@ llama_grammar_stacks & llama_grammar_get_stacks(struct llama_grammar * grammar)
     return grammar->stacks;
 }
 
+static void llama_grammar_accept_chr(
+        struct llama_grammar       & grammar,
+        const llama_grammar_stack  & stack,
+              uint32_t               chr,
+              llama_grammar_stacks & new_stacks) {
+    if (stack.empty()) {
+        return;
+    }
+
+    const llama_grammar_element * pos = stack.back();
+
+    // ignore if this turns into a token
+    if (pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+        return;
+    }
+
+    auto match = llama_grammar_match_char(pos, chr);
+    if (match.first) {
+        llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
+        if (!llama_grammar_is_end_of_sequence(match.second)) {
+            new_stack.push_back(match.second);
+        }
+        llama_grammar_advance_stack(grammar.rules, new_stack, new_stacks);
+    }
+}
+
 void llama_grammar_accept(struct llama_grammar * grammar, uint32_t chr) {
     llama_grammar_stacks stacks_new;
     stacks_new.reserve(grammar->stacks.size());
 
     for (const auto & stack : grammar->stacks) {
-        if (stack.empty()) {
-            continue;
-        }
-
-        auto match = llama_grammar_match_char(stack.back(), chr);
-        if (match.first) {
-            const llama_grammar_element * pos = match.second;
-
-            // update top of stack to next element, if any
-            llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
-            if (!llama_grammar_is_end_of_sequence(pos)) {
-                new_stack.push_back(pos);
-            }
-            llama_grammar_advance_stack(grammar->rules, new_stack, stacks_new);
-        }
+        llama_grammar_accept_chr(*grammar, stack, chr, stacks_new);
     }
 
     grammar->stacks = std::move(stacks_new);
@@ -875,6 +985,22 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
 
     const llama_grammar_element * stack_pos = stack.back();
 
+    // if the top of the stack is a token rule, then we only need to check the token id
+    if (stack_pos->type == LLAMA_GRETYPE_TOKEN || stack_pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+        for (const auto & tok : candidates) {
+            if (*tok.code_points == 0) {
+                // reached the end of a token consumed by char rules, reject iff it ended
+                // in a partial response
+                if (tok.partial_utf8.n_remain != 0) {
+                    rejects.push_back(tok);
+                }
+            } else if (!llama_grammar_match_token(stack_pos, tok.id)) {
+                rejects.push_back(tok);
+            }
+        }
+        return rejects;
+    }
+
     llama_grammar_candidates next_candidates;
     next_candidates.reserve(candidates.size());
 
@@ -887,7 +1013,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
                 rejects.push_back(tok);
             }
         } else if (llama_grammar_match_char(stack_pos, *tok.code_points).first) {
-            next_candidates.push_back({ tok.index, tok.code_points + 1, tok.partial_utf8 });
+            next_candidates.push_back({ tok.index, tok.code_points + 1, tok.partial_utf8, tok.id });
         } else {
             rejects.push_back(tok);
         }
@@ -905,7 +1031,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
 
     auto next_rejects = llama_grammar_reject_candidates(rules, next_stacks, next_candidates);
     for (const auto & tok : next_rejects) {
-        rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8 });
+        rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8, tok.id });
     }
 
     return rejects;
@@ -972,12 +1098,13 @@ struct llama_grammar * llama_grammar_init_impl(
         vocab,
         std::move(vec_rules),
         std::move(stacks),
-        /* .partial_utf8 = */     {},
-        /* .lazy =*/              false,
-        /* .awaiting_trigger = */ false,
-        /* .trigger_buffer = */   "",
-        /* .trigger_tokens   = */ {},
-        /* .trigger_patterns    = */ {},
+        /* .partial_utf8 = */             {},
+        /* .lazy = */                     false,
+        /* .awaiting_trigger = */         false,
+        /* .trigger_buffer = */           "",
+        /* .trigger_buffer_positions = */ {},
+        /* .trigger_tokens = */           {},
+        /* .trigger_patterns = */         {},
     };
 }
 
@@ -990,7 +1117,7 @@ struct llama_grammar * llama_grammar_init_impl(
                             size_t num_trigger_patterns,
                const llama_token * trigger_tokens,
                             size_t num_trigger_tokens) {
-    llama_grammar_parser parser;
+    llama_grammar_parser parser(vocab);
 
     // if there is a grammar, parse it
     // rules will be empty (default) if there are parse errors
@@ -1077,10 +1204,11 @@ struct llama_grammar * llama_grammar_init_impl(
         vocab,
         std::move(vec_rules),
         std::move(stacks),
-        /* .partial_utf8 = */     {},
-        /* .lazy = */             lazy,
-        /* .awaiting_trigger = */ lazy,
-        /* .trigger_buffer = */   "",
+        /* .partial_utf8 = */             {},
+        /* .lazy = */                     lazy,
+        /* .awaiting_trigger = */         lazy,
+        /* .trigger_buffer = */           "",
+        /* .trigger_buffer_positions = */ {},
         std::move(vec_trigger_tokens),
         std::move(vec_trigger_patterns),
     };
@@ -1103,6 +1231,7 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
         grammar.lazy,
         grammar.awaiting_trigger,
         grammar.trigger_buffer,
+        grammar.trigger_buffer_positions,
         grammar.trigger_tokens,
         grammar.trigger_patterns,
     };
@@ -1156,7 +1285,7 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
             cur_p->data[i].logit = -INFINITY;
         } else {
             candidates_decoded.push_back(decode_utf8(piece, grammar.partial_utf8));
-            candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second });
+            candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second, id });
         }
     }
 
@@ -1175,10 +1304,12 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
         if (std::find(grammar.trigger_tokens.begin(), grammar.trigger_tokens.end(), token) != grammar.trigger_tokens.end()) {
             grammar.awaiting_trigger = false;
             grammar.trigger_buffer.clear();
-            llama_grammar_accept_str(grammar, piece);
+            llama_grammar_accept_token(grammar, token, piece);
             LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str());
             return;
         } else {
+            auto position = std::make_pair(grammar.trigger_buffer.size(), grammar.trigger_buffer.size() + piece.size());
+            grammar.trigger_buffer_positions.push_back(std::make_pair(token, position));
             grammar.trigger_buffer += piece;
 
             std::smatch match;
@@ -1196,10 +1327,23 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
                     if (start == std::string::npos) {
                         start = match.position(0);
                     }
+
+                    // replay tokens that overlap with [start, end)
+                    for (const auto & [tok, tok_pos] : grammar.trigger_buffer_positions) {
+                        auto [tok_start, tok_end] = tok_pos;
+                        if (tok_end <= start) {
+                            continue;
+                        }
+
+                        size_t piece_start = (tok_start < start) ? start : tok_start; // allow for partial token pieces
+                        size_t piece_len = tok_end - piece_start;
+                        auto tok_piece = grammar.trigger_buffer.substr(piece_start, piece_len);
+                        llama_grammar_accept_token(grammar, tok, tok_piece);
+                    }
+
                     auto constrained_str = grammar.trigger_buffer.substr(start);
-                    // std::string constrained_str(match[1].first, grammar.trigger_buffer.end());
                     grammar.trigger_buffer.clear();
-                    llama_grammar_accept_str(grammar, constrained_str);
+                    grammar.trigger_buffer_positions.clear();
                     LLAMA_LOG_DEBUG("Grammar triggered on regex: '%s'\n", constrained_str.c_str());
                     return;
                 }
@@ -1218,7 +1362,7 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
         GGML_ABORT("fatal error");
     }
 
-    llama_grammar_accept_str(grammar, piece);
+    llama_grammar_accept_token(grammar, token, piece);
 }
 
 void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string & piece) {
@@ -1235,3 +1379,59 @@ void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string
         throw std::runtime_error("Unexpected empty grammar stack after accepting piece: " + piece);
     }
 }
+
+void llama_grammar_accept_token(struct llama_grammar & grammar, llama_token token, const std::string & piece) {
+    // Note terminating 0 in decoded string
+    const auto   decoded     = decode_utf8(piece, grammar.partial_utf8);
+    const auto & code_points = decoded.first;
+
+    llama_grammar_stacks stacks_new;
+    stacks_new.reserve(grammar.stacks.size());
+
+    for (const auto & stack : grammar.stacks) {
+        if (stack.empty()) {
+            continue;
+        }
+
+        const llama_grammar_element * pos = stack.back();
+
+        if (pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+            if (llama_grammar_match_token(pos, token)) {
+                llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
+                if (!llama_grammar_is_end_of_sequence(pos + 1)) {
+                    new_stack.push_back(pos + 1);
+                }
+                llama_grammar_advance_stack(grammar.rules, new_stack, stacks_new);
+            }
+        } else {
+            llama_grammar_stacks current_stacks = {stack};
+
+            for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
+                llama_grammar_stacks next_stacks;
+
+                for (const auto & cur_stack : current_stacks) {
+                    llama_grammar_accept_chr(grammar, cur_stack, *it, next_stacks);
+                }
+
+                current_stacks = std::move(next_stacks);
+                if (current_stacks.empty()) {
+                    break;
+                }
+            }
+
+            for (auto & surviving_stack : current_stacks) {
+                if (std::find(stacks_new.begin(), stacks_new.end(), surviving_stack) == stacks_new.end()) {
+                    stacks_new.emplace_back(surviving_stack);
+                }
+            }
+        }
+    }
+
+    grammar.stacks = std::move(stacks_new);
+    grammar.partial_utf8 = decoded.second;
+
+    if (grammar.stacks.empty()) {
+        throw std::runtime_error("Unexpected empty grammar stack after accepting piece: " + piece + " (" + std::to_string(token) + ")");
+    }
+}
+

src/llama-grammar.h

@@ -36,11 +36,17 @@ enum llama_gretype {
 
     // any character (.)
     LLAMA_GRETYPE_CHAR_ANY       = 7,
+
+    // terminal element: token (<[token-id]>)
+    LLAMA_GRETYPE_TOKEN          = 8,
+
+    // inverse token (!<[token-id]>)
+    LLAMA_GRETYPE_TOKEN_NOT      = 9,
 };
 
 typedef struct llama_grammar_element {
     enum llama_gretype type;
-    uint32_t           value; // Unicode code point or rule ID
+    uint32_t           value; // Unicode code point, rule ID, or token ID
 } llama_grammar_element;
 
 struct llama_partial_utf8 {
@@ -52,6 +58,7 @@ struct llama_grammar_candidate {
     size_t               index;
     const uint32_t     * code_points;
     llama_partial_utf8   partial_utf8;
+    llama_token          id;
 };
 
 using llama_grammar_rule  = std::vector<      llama_grammar_element>;
@@ -77,10 +84,13 @@ std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_stack(
         const llama_grammar_candidates & candidates);
 
 struct llama_grammar_parser {
+    const llama_vocab * vocab;
     std::map<std::string, uint32_t> symbol_ids;
 
     llama_grammar_rules rules;
 
+    llama_grammar_parser(const struct llama_vocab * vocab = nullptr) : vocab(vocab) {}
+
     llama_grammar_stack c_rules() const;
 
     uint32_t get_symbol_id(const char * src, size_t len);
@@ -112,6 +122,9 @@ struct llama_grammar_trigger_pattern {
 };
 
 struct llama_grammar {
+    // maintain a list of llama_tokens and their positions in the trigger_buffer
+    using token_pos = std::pair<llama_token, std::pair<size_t, size_t>>;
+
     // note: allow null vocab for testing (not great)
     const llama_vocab * vocab;
 
@@ -127,6 +140,7 @@ struct llama_grammar {
     bool                     lazy             = false;
     bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
     std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
+    std::vector<token_pos>   trigger_buffer_positions; // Tokens buffered by lazy grammar. Used to replay when a trigger is found.
     std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
     std::vector<llama_grammar_trigger_pattern>
                              trigger_patterns;         // Regular expressions that trigger a lazy grammar. Must be a full match of the entire generated
@@ -171,3 +185,8 @@ void llama_grammar_accept_impl(
 void llama_grammar_accept_str(
               struct llama_grammar & grammar,
                  const std::string & piece);
+
+void llama_grammar_accept_token(
+              struct llama_grammar & grammar,
+                       llama_token   token,
+                 const std::string & piece);

src/llama-graph.cpp

@@ -385,7 +385,7 @@ bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
   //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
     res &= self_kq_mask->ne[0] == mctx->get_n_kv();
-    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+    res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
 
     return res;
 }
@@ -416,10 +416,10 @@ bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) {
   //res &= self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there
 
     res &= self_kq_mask->ne[0] == mctx->get_base()->get_n_kv();
-    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+    res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
 
     res &= self_kq_mask_swa->ne[0] == mctx->get_swa()->get_n_kv();
-    res &= self_kq_mask_swa->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);
+    res &= self_kq_mask_swa->ne[1] == params.ubatch.n_tokens;
 
     return res;
 }
@@ -452,7 +452,7 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
             }
         }
 
-        for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
+        for (int i = n_tokens; i < n_tokens; ++i) {
             for (int j = 0; j < n_enc; ++j) {
                 data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY;
             }
@@ -973,7 +973,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
         // mask out the other groups
         selection_probs = ggml_get_rows(ctx0, selection_groups, expert_groups); // [n_exp_per_group, n_group_used, n_tokens]
-        selection_probs = ggml_set_rows(ctx0, ggml_scale_bias(ctx0, selection_groups, 0.0f, -INFINITY), selection_probs, expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens]
+        selection_probs = ggml_set_rows(ctx0, ggml_fill(ctx0, selection_groups, -INFINITY), selection_probs, expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens]
         selection_probs = ggml_reshape_2d(ctx0, selection_probs, n_expert, n_tokens); // [n_expert, n_tokens]
         cb(selection_probs, "ffn_moe_probs_masked", il);
     }
@@ -1470,13 +1470,13 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
     auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);
 
     // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
-    inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+    inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens, 1, 1);
     ggml_set_input(inp->self_kq_mask);
 
     inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
 
     if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
-        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens, 1, 1);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
@@ -1558,7 +1558,7 @@ static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl(
         inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1701,7 +1701,7 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
 
     const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
-    inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
+    inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, n_tokens, 1, 1);
     ggml_set_input(inp->cross_kq_mask);
 
     inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask;
@@ -1767,7 +1767,7 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
         inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask);
 
         inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
@@ -1781,7 +1781,7 @@ llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const
         inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
         inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
 
-        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
         ggml_set_input(inp->self_kq_mask_swa);
 
         inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;

src/llama-hparams.cpp

@@ -1,6 +1,7 @@
 #include "llama-hparams.h"
 
 #include "ggml.h"
+
 #include <cassert>
 
 void llama_hparams::set_swa_pattern(uint32_t n_pattern, bool dense_first) {