Merge branch 'main' into moondream2

.buildkite/nightly-benchmarks/scripts/nightly-annotate.sh

@@ -43,7 +43,7 @@ main() {
     
 
 
-    # The figures should be genereated by a separate process outside the CI/CD pipeline
+    # The figures should be generated by a separate process outside the CI/CD pipeline
 
     # # generate figures
     # python3 -m pip install tabulate pandas matplotlib

.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh

@@ -301,6 +301,104 @@ run_serving_tests() {
   kill_gpu_processes
 }
 
+run_genai_perf_tests() {
+  # run genai-perf tests 
+
+  # $1: a json file specifying genai-perf test cases
+  local genai_perf_test_file
+  genai_perf_test_file=$1
+
+  # Iterate over genai-perf tests
+  jq -c '.[]' "$genai_perf_test_file" | while read -r params; do
+    # get the test name, and append the GPU type back to it.
+    test_name=$(echo "$params" | jq -r '.test_name')    
+    
+    # if TEST_SELECTOR is set, only run the test cases that match the selector
+    if [[ -n "$TEST_SELECTOR" ]] && [[ ! "$test_name" =~ $TEST_SELECTOR ]]; then
+      echo "Skip test case $test_name."
+      continue
+    fi
+    
+    # prepend the current serving engine to the test name
+    test_name=${CURRENT_LLM_SERVING_ENGINE}_${test_name}
+
+    # get common parameters
+    common_params=$(echo "$params" | jq -r '.common_parameters')
+    model=$(echo "$common_params" | jq -r '.model')
+    tp=$(echo "$common_params" | jq -r '.tp')
+    dataset_name=$(echo "$common_params" | jq -r '.dataset_name')
+    dataset_path=$(echo "$common_params" | jq -r '.dataset_path')
+    port=$(echo "$common_params" | jq -r '.port')
+    num_prompts=$(echo "$common_params" | jq -r '.num_prompts')
+    reuse_server=$(echo "$common_params" | jq -r '.reuse_server')
+
+    # get client and server arguments
+    server_params=$(echo "$params" | jq -r ".${CURRENT_LLM_SERVING_ENGINE}_server_parameters")
+    qps_list=$(echo "$params" | jq -r '.qps_list')
+    qps_list=$(echo "$qps_list" | jq -r '.[] | @sh')
+    echo "Running over qps list $qps_list"
+
+    # check if there is enough GPU to run the test
+    if [[ $gpu_count -lt $tp ]]; then
+      echo "Required num-shard $tp but only $gpu_count GPU found. Skip testcase $test_name."
+      continue
+    fi
+
+    if [[ $reuse_server == "true" ]]; then
+      echo "Reuse previous server for test case $test_name"
+    else
+      kill_gpu_processes
+      bash "$VLLM_SOURCE_CODE_LOC/.buildkite/nightly-benchmarks/scripts/launch-server.sh" \
+        "$server_params" "$common_params"
+    fi
+
+    if wait_for_server; then
+      echo ""
+      echo "$CURRENT_LLM_SERVING_ENGINE server is up and running."
+    else
+      echo ""
+      echo "$CURRENT_LLM_SERVING_ENGINE failed to start within the timeout period."
+      break
+    fi
+
+    # iterate over different QPS
+    for qps in $qps_list; do
+      # remove the surrounding single quote from qps
+      if [[ "$qps" == *"inf"* ]]; then
+        echo "qps was $qps"
+        qps=$num_prompts
+        echo "now qps is $qps"
+      fi
+    
+      new_test_name=$test_name"_qps_"$qps
+      backend=$CURRENT_LLM_SERVING_ENGINE
+      
+      if [[ "$backend" == *"vllm"* ]]; then
+        backend="vllm"
+      fi
+      #TODO: add output dir.
+      client_command="genai-perf profile \
+        -m $model \
+        --service-kind openai \
+        --backend vllm \
+        --endpoint-type chat \
+        --streaming \
+        --url localhost:$port \
+        --request-rate $qps \
+        --num-prompts $num_prompts \
+      "
+
+    echo "Client command: $client_command"
+
+    eval "$client_command"
+
+    #TODO: process/record outputs
+    done
+  done
+
+  kill_gpu_processes
+
+}
 
 prepare_dataset() {
 
@@ -328,12 +426,17 @@ main() {
 
   pip install -U transformers
 
+  pip install -r requirements-dev.txt
+  which genai-perf
+
   # check storage
   df -h
 
   ensure_installed wget
   ensure_installed curl
   ensure_installed jq
+  # genai-perf dependency
+  ensure_installed libb64-0d
 
   prepare_dataset
 
@@ -345,6 +448,10 @@ main() {
   # run the test
   run_serving_tests "$BENCHMARK_ROOT/tests/nightly-tests.json"
 
+  # run genai-perf tests
+  run_genai_perf_tests "$BENCHMARK_ROOT/tests/genai-perf-tests.json"
+  mv artifacts/ $RESULTS_FOLDER/
+
   # upload benchmark results to buildkite
   python3 -m pip install tabulate pandas
   python3 "$BENCHMARK_ROOT/scripts/summary-nightly-results.py"

.buildkite/nightly-benchmarks/tests/genai-perf-tests.json

@@ -0,0 +1,23 @@
+[
+    {
+        "test_name": "llama8B_tp1_genai_perf",
+        "qps_list": [4,8,16,32],
+        "common_parameters": {
+            "model": "meta-llama/Meta-Llama-3-8B-Instruct",
+            "tp": 1,
+            "port": 8000,
+            "num_prompts": 500,
+            "reuse_server": false
+        },
+        "vllm_server_parameters": {
+            "disable_log_stats": "",
+            "disable_log_requests": "",
+            "gpu_memory_utilization": 0.9,
+            "num_scheduler_steps": 10,
+            "max_num_seqs": 512,
+            "dtype": "bfloat16"
+        },
+        "genai_perf_input_parameters": {
+        }
+    }
+]

.buildkite/run-cpu-test.sh

@@ -30,7 +30,7 @@ function cpu_tests() {
   # offline inference
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
     set -e
-    python3 examples/offline_inference/offline_inference.py"
+    python3 examples/offline_inference/basic.py"
 
   # Run basic model test
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
@@ -61,7 +61,7 @@ function cpu_tests() {
     pytest -s -v -k cpu_model \
     tests/basic_correctness/test_chunked_prefill.py"  
 
-  # online inference
+  # online serving
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
     export VLLM_CPU_KVCACHE_SPACE=10 
@@ -75,8 +75,14 @@ function cpu_tests() {
       --num-prompts 20 \
       --endpoint /v1/completions \
       --tokenizer facebook/opt-125m"
+
+  # Run multi-lora tests
+  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+    set -e
+    pytest -s -v \
+    tests/lora/test_qwen2vl.py"
 }
 
-# All of CPU tests are expected to be finished less than 25 mins.
+# All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 30m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"

.buildkite/run-gh200-test.sh

@@ -24,5 +24,5 @@ remove_docker_container
 
 # Run the image and test offline inference
 docker run --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
-    python3 examples/offline_inference/offline_inference.py
+    python3 examples/offline_inference/basic.py
 '

.buildkite/run-hpu-test.sh

@@ -8,9 +8,17 @@ set -ex
 docker build -t hpu-test-env -f Dockerfile.hpu .
 
 # Setup cleanup
+# certain versions of HPU software stack have a bug that can
+# override the exit code of the script, so we need to use
+# separate remove_docker_container and remove_docker_container_and_exit
+# functions, while other platforms only need one remove_docker_container
+# function.
+EXITCODE=1
 remove_docker_container() { docker rm -f hpu-test || true; }
-trap remove_docker_container EXIT
+remove_docker_container_and_exit() { remove_docker_container; exit $EXITCODE; }
+trap remove_docker_container_and_exit EXIT
 remove_docker_container
 
 # Run the image and launch offline inference
-docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/offline_inference.py
+docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic.py
+EXITCODE=$?

.buildkite/run-neuron-test.sh

@@ -51,4 +51,4 @@ docker run --rm -it --device=/dev/neuron0 --device=/dev/neuron1 --network host \
        -e "NEURON_COMPILE_CACHE_URL=${NEURON_COMPILE_CACHE_MOUNT}" \
        --name "${container_name}" \
        ${image_name} \
-       /bin/bash -c "python3 /workspace/vllm/examples/offline_inference/offline_inference_neuron.py"
+       /bin/bash -c "python3 /workspace/vllm/examples/offline_inference/neuron.py"

.buildkite/run-openvino-test.sh

@@ -13,4 +13,4 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image and launch offline inference
-docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference/offline_inference.py
+docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference/basic.py

.buildkite/run-tpu-test.sh

@@ -23,4 +23,4 @@ docker run --privileged --net host --shm-size=16G -it \
     && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
     && python3 /workspace/vllm/tests/tpu/test_compilation.py \
     && python3 /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
-    && python3 /workspace/vllm/examples/offline_inference/offline_inference_tpu.py"
+    && python3 /workspace/vllm/examples/offline_inference/tpu.py"

.buildkite/run-xpu-test.sh

@@ -14,6 +14,6 @@ remove_docker_container
 
 # Run the image and test offline inference/tensor parallel
 docker run --name xpu-test --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --entrypoint="" xpu-test sh -c '
-    python3 examples/offline_inference/offline_inference.py
-    python3 examples/offline_inference/offline_inference_cli.py -tp 2
+    python3 examples/offline_inference/basic.py
+    python3 examples/offline_inference/cli.py -tp 2
 '

.buildkite/test-pipeline.yaml

@@ -106,7 +106,7 @@ steps:
   source_file_dependencies:
   - vllm/
   commands:
-  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
   - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
@@ -125,11 +125,15 @@ steps:
   - tests/distributed
   - tests/spec_decode/e2e/test_integration_dist_tp4
   - tests/compile
+  - examples/offline_inference/rlhf.py
   commands:
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
+  # TODO: create a dedicated test section for multi-GPU example tests
+  # when we have multiple distributed example tests
+  - python3 ../examples/offline_inference/rlhf.py
 
 - label: Metrics, Tracing Test # 10min
   num_gpus: 2 
@@ -187,19 +191,19 @@ steps:
   - examples/
   commands:
     - pip install tensorizer # for tensorizer test
-    - python3 offline_inference/offline_inference.py
+    - python3 offline_inference/basic.py
     - python3 offline_inference/cpu_offload.py
-    - python3 offline_inference/offline_inference_chat.py
-    - python3 offline_inference/offline_inference_with_prefix.py
+    - python3 offline_inference/chat.py
+    - python3 offline_inference/prefix_caching.py
     - python3 offline_inference/llm_engine_example.py
-    - python3 offline_inference/offline_inference_vision_language.py
-    - python3 offline_inference/offline_inference_vision_language_multi_image.py
+    - python3 offline_inference/vision_language.py
+    - python3 offline_inference/vision_language_multi_image.py
     - python3 other/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 other/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
-    - python3 offline_inference/offline_inference_encoder_decoder.py
-    - python3 offline_inference/offline_inference_classification.py
-    - python3 offline_inference/offline_inference_embedding.py
-    - python3 offline_inference/offline_inference_scoring.py
-    - python3 offline_inference/offline_profile.py --model facebook/opt-125m run_num_steps --num-steps 2
+    - python3 offline_inference/encoder_decoder.py
+    - python3 offline_inference/classification.py
+    - python3 offline_inference/embedding.py
+    - python3 offline_inference/scoring.py
+    - python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Prefix Caching Test # 9min
   mirror_hardwares: [amd]
@@ -214,6 +218,7 @@ steps:
   - vllm/model_executor/layers
   - vllm/sampling_metadata.py
   - tests/samplers
+  - tests/conftest.py
   commands:
     - pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
@@ -229,13 +234,15 @@ steps:
     - pytest -v -s test_logits_processor.py
     - pytest -v -s model_executor/test_guided_processors.py
 
-- label: Speculative decoding tests # 30min
+- label: Speculative decoding tests # 40min
   source_file_dependencies:
   - vllm/spec_decode
   - tests/spec_decode
+  - vllm/model_executor/models/eagle.py
   commands:
     - pytest -v -s spec_decode/e2e/test_multistep_correctness.py
     - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s spec_decode --ignore=spec_decode/e2e/test_multistep_correctness.py
+    - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
 
 - label: LoRA Test %N # 15min each
   mirror_hardwares: [amd]
@@ -367,6 +374,7 @@ steps:
   - tests/models/encoder_decoder/vision_language
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s models/multimodal
     - pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
     - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'core_model or quant_model'
     - pytest -v -s models/embedding/vision_language -m core_model
@@ -457,7 +465,10 @@ steps:
   - vllm/worker/worker_base.py
   - vllm/worker/worker.py
   - vllm/worker/model_runner.py
+  - entrypoints/llm/test_collective_rpc.py
   commands:
+  - pytest -v -s entrypoints/llm/test_collective_rpc.py
+  - torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'

.github/workflows/actionlint.yml

@@ -1,40 +0,0 @@
-name: Lint GitHub Actions workflows
-on:
-  push:
-    branches:
-      - "main"
-    paths:
-      - '.github/workflows/*.ya?ml'
-      - '.github/workflows/actionlint.*'
-      - '.github/workflows/matchers/actionlint.json'
-  pull_request:
-    branches:
-      - "main"
-    paths:
-      - '.github/workflows/*.ya?ml'
-      - '.github/workflows/actionlint.*'
-      - '.github/workflows/matchers/actionlint.json'
-
-env:
-  LC_ALL: en_US.UTF-8
-
-defaults:
-  run:
-    shell: bash
-
-permissions:
-  contents: read
-
-jobs:
-  actionlint:
-    runs-on: ubuntu-latest
-    steps:
-      - name: "Checkout"
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-
-      - name: "Run actionlint"
-        run: |
-          echo "::add-matcher::.github/workflows/matchers/actionlint.json"
-          tools/actionlint.sh -color

.github/workflows/clang-format.yml

@@ -1,53 +0,0 @@
-name: clang-format
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - '**/*.h'
-      - '**/*.cpp'
-      - '**/*.cu'
-      - '**/*.cuh'
-      - '.github/workflows/clang-format.yml'
-  pull_request:
-    branches:
-      - main
-    paths:
-      - '**/*.h'
-      - '**/*.cpp'
-      - '**/*.cu'
-      - '**/*.cuh'
-      - '.github/workflows/clang-format.yml'
-
-jobs:
-  clang-format:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.11"]
-    steps:
-    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - name: Set up Python ${{ matrix.python-version }}
-      uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-      with:
-        python-version: ${{ matrix.python-version }}
-    - name: Install dependencies
-      run: |
-        python -m pip install --upgrade pip
-        pip install clang-format==18.1.5
-    - name: Running clang-format
-      run: |
-        EXCLUDES=(
-            'csrc/moe/topk_softmax_kernels.cu'
-            'csrc/quantization/gguf/ggml-common.h'
-            'csrc/quantization/gguf/dequantize.cuh'
-            'csrc/quantization/gguf/vecdotq.cuh'
-            'csrc/quantization/gguf/mmq.cuh'
-            'csrc/quantization/gguf/mmvq.cuh'
-        )
-        find csrc/ \( -name '*.h' -o -name '*.cpp' -o -name '*.cu' -o -name '*.cuh' \) -print \
-            | grep -vFf <(printf "%s\n" "${EXCLUDES[@]}") \
-            | xargs clang-format --dry-run --Werror

.github/workflows/codespell.yml

@@ -1,45 +0,0 @@
-name: codespell
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - "**/*.md"
-      - "**/*.rst"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/codespell.yml
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - "**/*.md"
-      - "**/*.rst"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/codespell.yml
-
-jobs:
-  codespell:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - name: Set up Python ${{ matrix.python-version }}
-      uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-      with:
-        python-version: ${{ matrix.python-version }}
-    - name: Install dependencies
-      run: |
-        python -m pip install --upgrade pip
-        pip install -r requirements-lint.txt
-    - name: Spelling check with codespell
-      run: |
-        codespell --toml pyproject.toml

.github/workflows/dummy.yml

@@ -0,0 +1,20 @@
+name: dummy-checks
+
+on:
+  pull_request:
+
+jobs:
+  mypy:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.12"]
+    steps:
+      - run: echo "This is a dummy step that always passes"
+  ruff:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.12"]
+    steps:
+      - run: echo "This is a dummy step that always passes"

.github/workflows/matchers/ruff.json

@@ -1,17 +0,0 @@
-{
-    "problemMatcher": [
-      {
-        "owner": "ruff",
-        "pattern": [
-          {
-            "regexp": "^(.+?):(\\d+):(\\d+): (\\w+): (.+)$",
-            "file": 1,
-            "line": 2,
-            "column": 3,
-            "code": 4,
-            "message": 5
-          }
-        ]
-      }
-    ]
-  }

.github/workflows/mypy.yaml

@@ -1,51 +0,0 @@
-name: mypy
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - '**/*.py'
-      - '.github/workflows/mypy.yaml'
-      - 'tools/mypy.sh'
-      - 'pyproject.toml'
-  pull_request:
-    branches:
-      - main
-    # This workflow is only relevant when one of the following files changes.
-    # However, we have github configured to expect and require this workflow
-    # to run and pass before github with auto-merge a pull request. Until github
-    # allows more flexible auto-merge policy, we can just run this on every PR.
-    # It doesn't take that long to run, anyway.
-    #paths:
-    #  - '**/*.py'
-    #  - '.github/workflows/mypy.yaml'
-    #  - 'tools/mypy.sh'
-    #  - 'pyproject.toml'
-
-jobs:
-  mypy:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12"]
-    steps:
-    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-    - name: Set up Python ${{ matrix.python-version }}
-      uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-      with:
-        python-version: ${{ matrix.python-version }}
-    - name: Install dependencies
-      run: |
-        python -m pip install --upgrade pip
-        pip install mypy==1.11.1
-        pip install types-setuptools
-        pip install types-PyYAML
-        pip install types-requests
-        pip install types-setuptools
-    - name: Mypy
-      run: |
-        echo "::add-matcher::.github/workflows/matchers/mypy.json"
-        tools/mypy.sh 1 ${{ matrix.python-version }}

.github/workflows/png-lint.yml

@@ -1,37 +0,0 @@
-name: Lint PNG exports from excalidraw
-on:
-  push:
-    branches:
-      - "main"
-    paths:
-      - '*.excalidraw.png'
-      - '.github/workflows/png-lint.yml'
-  pull_request:
-    branches:
-      - "main"
-    paths:
-      - '*.excalidraw.png'
-      - '.github/workflows/png-lint.yml'
-
-env:
-  LC_ALL: en_US.UTF-8
-
-defaults:
-  run:
-    shell: bash
-
-permissions:
-  contents: read
-
-jobs:
-  actionlint:
-    runs-on: ubuntu-latest
-    steps:
-      - name: "Checkout"
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-
-      - name: "Run png-lint.sh to check excalidraw exported images"
-        run: |
-          tools/png-lint.sh

.github/workflows/pre-commit.yml

@@ -0,0 +1,17 @@
+name: pre-commit
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+jobs:
+  pre-commit:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+    - uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
+      with:
+        python-version: "3.12"
+    - run: echo "::add-matcher::.github/workflows/matchers/actionlint.json"
+    - uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd # v3.0.1

.github/workflows/ruff.yml

@@ -1,52 +0,0 @@
-name: ruff
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - pyproject.toml
-      - requirements-lint.txt
-      - .github/workflows/matchers/ruff.json
-      - .github/workflows/ruff.yml
-  pull_request:
-    branches:
-      - main
-    # This workflow is only relevant when one of the following files changes.
-    # However, we have github configured to expect and require this workflow
-    # to run and pass before github with auto-merge a pull request. Until github
-    # allows more flexible auto-merge policy, we can just run this on every PR.
-    # It doesn't take that long to run, anyway.
-    #paths:
-    #  - "**/*.py"
-    #  - pyproject.toml
-    #  - requirements-lint.txt
-    #  - .github/workflows/matchers/ruff.json
-    #  - .github/workflows/ruff.yml
-
-jobs:
-  ruff:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-lint.txt
-      - name: Analysing the code with ruff
-        run: |
-          echo "::add-matcher::.github/workflows/matchers/ruff.json"
-          ruff check --output-format github .
-      - name: Run isort
-        run: |
-          isort . --check-only

.github/workflows/shellcheck.yml

@@ -1,37 +0,0 @@
-name: Lint shell scripts
-on:
-  push:
-    branches:
-      - "main"
-    paths:
-      - '**/*.sh'
-      - '.github/workflows/shellcheck.yml'
-  pull_request:
-    branches:
-      - "main"
-    paths:
-      - '**/*.sh'
-      - '.github/workflows/shellcheck.yml'
-
-env:
-  LC_ALL: en_US.UTF-8
-
-defaults:
-  run:
-    shell: bash
-
-permissions:
-  contents: read
-
-jobs:
-  shellcheck:
-    runs-on: ubuntu-latest
-    steps:
-      - name: "Checkout"
-        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-        with:
-          fetch-depth: 0
-
-      - name: "Check shell scripts"
-        run: |
-          tools/shellcheck.sh

.github/workflows/sphinx-lint.yml

@@ -1,32 +0,0 @@
-name: Lint documentation
-
-on:
-  push:
-    branches:
-      - main
-    paths:
-      - "docs/**"
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "docs/**"
-
-jobs:
-  sphinx-lint:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install -r requirements-lint.txt
-      - name: Linting docs
-        run: tools/sphinx-lint.sh

.github/workflows/yapf.yml

@@ -1,38 +0,0 @@
-name: yapf
-
-on:
-  # Trigger the workflow on push or pull request,
-  # but only for the main branch
-  push:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - .github/workflows/yapf.yml
-  pull_request:
-    branches:
-      - main
-    paths:
-      - "**/*.py"
-      - .github/workflows/yapf.yml
-
-jobs:
-  yapf:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        python-version: ["3.12"]
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install yapf==0.32.0
-          pip install toml==0.10.2
-      - name: Running yapf
-        run: |
-          yapf --diff --recursive .

.pre-commit-config.yaml

@@ -0,0 +1,73 @@
+repos:
+- repo: https://github.com/google/yapf
+  rev: v0.32.0
+  hooks:
+  - id: yapf
+    args: [--in-place, --verbose]
+    additional_dependencies: [toml] # TODO: Remove when yapf is upgraded
+- repo: https://github.com/astral-sh/ruff-pre-commit
+  rev: v0.6.5
+  hooks:
+  - id: ruff
+    args: [--output-format, github]
+- repo: https://github.com/codespell-project/codespell
+  rev: v2.3.0
+  hooks:
+  - id: codespell
+    exclude: 'benchmarks/sonnet.txt|(build|tests/(lora/data|models/fixtures|prompts))/.*'
+- repo: https://github.com/PyCQA/isort
+  rev: 5.13.2
+  hooks:
+  - id: isort
+- repo: https://github.com/pre-commit/mirrors-clang-format
+  rev: v18.1.5
+  hooks:
+  - id: clang-format
+    exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))'
+    types_or: [c++, cuda]
+    args: [--style=file, --verbose]
+- repo: https://github.com/jackdewinter/pymarkdown
+  rev: v0.9.27
+  hooks:
+  - id: pymarkdown
+    files: docs/.*
+- repo: local
+  hooks:
+  - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.9
+    entry: tools/mypy.sh 1 "3.9"
+    language: python
+    types: [python]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, types-setuptools, types-PyYAML, types-requests]
+  - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.10
+    entry: tools/mypy.sh 1 "3.10"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+  - id: mypy-3.11 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.11
+    entry: tools/mypy.sh 1 "3.11"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+  - id: mypy-3.12 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.12
+    entry: tools/mypy.sh 1 "3.12"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
+  - id: shellcheck
+    name: Lint shell scripts
+    entry: tools/shellcheck.sh
+    language: script
+    types: [shell]
+  - id: png-lint
+    name: Lint PNG exports from excalidraw
+    entry: tools/png-lint.sh
+    language: script
+    types: [png]
+- repo: https://github.com/rhysd/actionlint
+  rev: v1.7.6
+  hooks:
+  - id: actionlint

Dockerfile.cpu

@@ -26,10 +26,10 @@ RUN pip install intel_extension_for_pytorch==2.5.0
 
 WORKDIR /workspace
 
-COPY requirements-build.txt requirements-build.txt
 ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
 ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,src=requirements-build.txt,target=requirements-build.txt \
     pip install --upgrade pip && \
     pip install -r requirements-build.txt
 
@@ -37,9 +37,9 @@ FROM cpu-test-1 AS build
 
 WORKDIR /workspace/vllm
 
-COPY requirements-common.txt requirements-common.txt
-COPY requirements-cpu.txt requirements-cpu.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
+    --mount=type=bind,src=requirements-common.txt,target=requirements-common.txt \
+    --mount=type=bind,src=requirements-cpu.txt,target=requirements-cpu.txt \
     pip install -v -r requirements-cpu.txt
 
 COPY . .

Dockerfile.hpu

@@ -1,4 +1,4 @@
-FROM vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest
+FROM vault.habana.ai/gaudi-docker/1.19.1/ubuntu22.04/habanalabs/pytorch-installer-2.5.1:latest
 
 COPY ./ /workspace/vllm
 

Dockerfile.rocm

@@ -51,10 +51,10 @@ RUN --mount=type=cache,target=/root/.cache/pip \
         *"rocm-6.2"*) \
             python3 -m pip uninstall -y torch torchvision \
             && python3 -m pip install --pre \
-                torch==2.6.0.dev20241113+rocm6.2 \
+                torch \
                 'setuptools-scm>=8' \
-                torchvision==0.20.0.dev20241113+rocm6.2 \
-                --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2;; \
+                torchvision \
+                --extra-index-url https://download.pytorch.org/whl/rocm6.2;; \
         *) ;; esac
 
 ENV LLVM_SYMBOLIZER_PATH=/opt/rocm/llvm/bin/llvm-symbolizer

README.md

@@ -38,10 +38,12 @@ The first vLLM meetup in 2025 is happening on January 22nd, Wednesday, with Goog
 ## About
 vLLM is a fast and easy-to-use library for LLM inference and serving.
 
+Originally developed in the [Sky Computing Lab](https://sky.cs.berkeley.edu) at UC Berkeley, vLLM has evloved into a community-driven project with contributions from both academia and industry.
+
 vLLM is fast with:
 
 - State-of-the-art serving throughput
-- Efficient management of attention key and value memory with **PagedAttention**
+- Efficient management of attention key and value memory with [**PagedAttention**](https://blog.vllm.ai/2023/06/20/vllm.html)
 - Continuous batching of incoming requests
 - Fast model execution with CUDA/HIP graph
 - Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), INT4, INT8, and FP8.
@@ -72,16 +74,16 @@ Find the full list of supported models [here](https://docs.vllm.ai/en/latest/mod
 
 ## Getting Started
 
-Install vLLM with `pip` or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):
+Install vLLM with `pip` or [from source](https://docs.vllm.ai/en/latest/getting_started/installation/gpu/index.html#build-wheel-from-source):
 
 ```bash
 pip install vllm
 ```
 
-Visit our [documentation](https://vllm.readthedocs.io/en/latest/) to learn more.
-- [Installation](https://vllm.readthedocs.io/en/latest/getting_started/installation.html)
-- [Quickstart](https://vllm.readthedocs.io/en/latest/getting_started/quickstart.html)
-- [List of Supported Models](https://vllm.readthedocs.io/en/latest/models/supported_models.html)
+Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
+- [Installation](https://docs.vllm.ai/en/latest/getting_started/installation/index.html)
+- [Quickstart](https://docs.vllm.ai/en/latest/getting_started/quickstart.html)
+- [List of Supported Models](https://docs.vllm.ai/en/latest/models/supported_models.html)
 
 ## Contributing
 

SECURITY.md

@@ -4,7 +4,7 @@
 
 If you believe you have found a security vulnerability in vLLM, we encourage you to let us know right away. We will investigate all legitimate reports and do our best to quickly fix the problem.
 
-Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new). Reports will then be triaged by the [vulnerability management team](https://docs.vllm.ai/contributing/vulnerability_management/).
+Please report security issues privately using [the vulnerability submission form](https://github.com/vllm-project/vllm/security/advisories/new). Reports will then be triaged by the [vulnerability management team](https://docs.vllm.ai/en/latest/contributing/vulnerability_management.html).
 
 ---
 

benchmarks/backend_request_func.py

@@ -22,6 +22,7 @@ class RequestFuncInput:
     prompt_len: int
     output_len: int
     model: str
+    model_name: Optional[str] = None
     best_of: int = 1
     logprobs: Optional[int] = None
     extra_body: Optional[dict] = None
@@ -78,7 +79,7 @@ async def async_request_tgi(
                             continue
                         chunk_bytes = chunk_bytes.decode("utf-8")
 
-                        #NOTE: Sometimes TGI returns a ping response without
+                        # NOTE: Sometimes TGI returns a ping response without
                         # any data, we should skip it.
                         if chunk_bytes.startswith(":"):
                             continue
@@ -235,7 +236,8 @@ async def async_request_openai_completions(
 
     async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
         payload = {
-            "model": request_func_input.model,
+            "model": request_func_input.model_name \
+                if request_func_input.model_name else request_func_input.model,
             "prompt": request_func_input.prompt,
             "temperature": 0.0,
             "best_of": request_func_input.best_of,
@@ -328,7 +330,8 @@ async def async_request_openai_chat_completions(
         if request_func_input.multi_modal_content:
             content.append(request_func_input.multi_modal_content)
         payload = {
-            "model": request_func_input.model,
+            "model": request_func_input.model_name \
+                if request_func_input.model_name else request_func_input.model,
             "messages": [
                 {
                     "role": "user",
@@ -417,14 +420,35 @@ def get_model(pretrained_model_name_or_path: str) -> str:
 
 
 def get_tokenizer(
-    pretrained_model_name_or_path: str, trust_remote_code: bool
+    pretrained_model_name_or_path: str,
+    tokenizer_mode: str = "auto",
+    trust_remote_code: bool = False,
+    **kwargs,
 ) -> Union[PreTrainedTokenizer, PreTrainedTokenizerFast]:
     if pretrained_model_name_or_path is not None and not os.path.exists(
             pretrained_model_name_or_path):
         pretrained_model_name_or_path = get_model(
             pretrained_model_name_or_path)
-    return AutoTokenizer.from_pretrained(pretrained_model_name_or_path,
-                                         trust_remote_code=trust_remote_code)
+    if tokenizer_mode == "slow":
+        if kwargs.get("use_fast", False):
+            raise ValueError(
+                "Cannot use the fast tokenizer in slow tokenizer mode.")
+        kwargs["use_fast"] = False
+    if tokenizer_mode == "mistral":
+        try:
+            from vllm.transformers_utils.tokenizer import MistralTokenizer
+        except ImportError as e:
+            raise ImportError("MistralTokenizer requires vllm package.\n"
+                              "Please install it with `pip install vllm` "
+                              "to use mistral tokenizer mode.") from e
+        return MistralTokenizer.from_pretrained(
+            str(pretrained_model_name_or_path))
+    else:
+        return AutoTokenizer.from_pretrained(
+            pretrained_model_name_or_path,
+            trust_remote_code=trust_remote_code,
+            **kwargs,
+        )
 
 
 ASYNC_REQUEST_FUNCS = {

benchmarks/benchmark_long_document_qa_throughput.py

@@ -2,8 +2,7 @@
 Offline benchmark to test the long document QA throughput.
 
 Example usage:
-    # This command run the vllm with 50GB CPU memory for offloading
-    # The workload samples 8 different prompts with a default input
+    # This workload samples 8 different prompts with a default input
     # length of 20000 tokens, then replicates each prompt 2 times 
     # in random order.
     python benchmark_long_document_qa_throughput.py \

benchmarks/benchmark_prefix_caching.py

@@ -10,7 +10,8 @@ Fixed example usage:
         --model meta-llama/Llama-2-7b-chat-hf \
         --enable-prefix-caching \
         --num-prompts 1 \
-        --repeat-count 100
+        --repeat-count 100 \
+        --input-length-range 128:256
 
 ShareGPT example usage:
     # This command samples 20 prompts with input lengths

benchmarks/benchmark_serving.py

@@ -525,6 +525,7 @@ async def benchmark(
     api_url: str,
     base_url: str,
     model_id: str,
+    model_name: str,
     tokenizer: PreTrainedTokenizerBase,
     input_requests: List[Tuple[str, int, int]],
     logprobs: Optional[int],
@@ -553,6 +554,7 @@ async def benchmark(
             "Multi-modal content is only supported on 'openai-chat' backend.")
     test_input = RequestFuncInput(
         model=model_id,
+        model_name=model_name,
         prompt=test_prompt,
         api_url=api_url,
         prompt_len=test_prompt_len,
@@ -573,6 +575,7 @@ async def benchmark(
     if profile:
         print("Starting profiler...")
         profile_input = RequestFuncInput(model=model_id,
+                                         model_name=model_name,
                                          prompt=test_prompt,
                                          api_url=base_url + "/start_profile",
                                          prompt_len=test_prompt_len,
@@ -616,6 +619,7 @@ async def benchmark(
     async for request in get_request(input_requests, request_rate, burstiness):
         prompt, prompt_len, output_len, mm_content = request
         request_func_input = RequestFuncInput(model=model_id,
+                                              model_name=model_name,
                                               prompt=prompt,
                                               api_url=api_url,
                                               prompt_len=prompt_len,
@@ -780,6 +784,7 @@ def main(args: argparse.Namespace):
 
     backend = args.backend
     model_id = args.model
+    model_name = args.served_model_name
     tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
     tokenizer_mode = args.tokenizer_mode
 
@@ -877,6 +882,7 @@ def main(args: argparse.Namespace):
             api_url=api_url,
             base_url=base_url,
             model_id=model_id,
+            model_name=model_name,
             tokenizer=tokenizer,
             input_requests=input_requests,
             logprobs=args.logprobs,
@@ -1222,5 +1228,12 @@ if __name__ == "__main__":
         'always use the slow tokenizer. \n* '
         '"mistral" will always use the `mistral_common` tokenizer.')
 
+    parser.add_argument("--served-model-name",
+                        type=str,
+                        default=None,
+                        help="The model name used in the API. "
+                        "If not specified, the model name will be the "
+                        "same as the ``--model`` argument. ")
+
     args = parser.parse_args()
     main(args)

benchmarks/kernels/benchmark_moe.py

@@ -1,6 +1,7 @@
 import argparse
 import time
 from datetime import datetime
+from itertools import product
 from typing import Any, Dict, List, Tuple, TypedDict
 
 import ray
@@ -11,7 +12,10 @@ from transformers import AutoConfig
 
 from vllm.model_executor.layers.fused_moe.fused_moe import *
 from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser, is_navi
+
+FP8_DTYPE = torch.float8_e4m3fnuz if current_platform.is_rocm(
+) and not is_navi() else torch.float8_e4m3fn
 
 
 class BenchmarkConfig(TypedDict):
@@ -80,8 +84,8 @@ def benchmark_config(
         a1_scale = torch.randn(1, dtype=torch.float32)
         a2_scale = torch.randn(1, dtype=torch.float32)
 
-        w1 = w1.to(torch.float8_e4m3fn)
-        w2 = w2.to(torch.float8_e4m3fn)
+        w1 = w1.to(FP8_DTYPE)
+        w2 = w2.to(FP8_DTYPE)
 
     input_gating = torch.empty(num_tokens, num_experts, dtype=torch.float32)
 
@@ -141,28 +145,172 @@ def benchmark_config(
     return avg
 
 
-def get_configs_compute_bound() -> List[Dict[str, int]]:
-    # Reduced search space for faster tuning.
-    # TODO(woosuk): Increase the search space and use a performance model to
-    # prune the search space.
+def get_rocm_tuning_space(use_fp16):
+    block_mn_range = [16, 32, 64, 128, 256]
+    block_k_range = [16, 32, 64, 128, 256]
+    if not use_fp16:
+        block_k_range.remove(16)  # BLOCK_K=16 not supported for fp8
+    num_warps_range = [1, 2, 4, 8]
+    group_m_range = [1, 4, 8, 16, 32]
+    num_stage_range = [2]
+    waves_per_eu_range = [0]
+    matrix_instr_nonkdim_range = [16, 32] if use_fp16 else []
+    kpack_range = [1, 2] if use_fp16 else []
+
+    param_ranges = {
+        "BLOCK_SIZE_M": block_mn_range,
+        "BLOCK_SIZE_N": block_mn_range,
+        "BLOCK_SIZE_K": block_k_range,
+        "GROUP_SIZE_M": group_m_range,
+        "num_warps": num_warps_range,
+        "num_stages": num_stage_range,
+        "waves_per_eu": waves_per_eu_range,
+    }
+    if use_fp16:
+        param_ranges["matrix_instr_nonkdim"] = matrix_instr_nonkdim_range
+        param_ranges["kpack"] = kpack_range
+
+    return param_ranges
+
+
+def get_configs_compute_bound(use_fp16) -> List[Dict[str, int]]:
     configs: List[BenchmarkConfig] = []
-    for num_stages in [2, 3, 4, 5]:
-        for block_m in [16, 32, 64, 128, 256]:
-            for block_k in [64, 128, 256]:
-                for block_n in [32, 64, 128, 256]:
-                    for num_warps in [4, 8]:
-                        for group_size in [1, 16, 32, 64]:
-                            configs.append({
-                                "BLOCK_SIZE_M": block_m,
-                                "BLOCK_SIZE_N": block_n,
-                                "BLOCK_SIZE_K": block_k,
-                                "GROUP_SIZE_M": group_size,
-                                "num_warps": num_warps,
-                                "num_stages": num_stages,
-                            })
+
+    if current_platform.is_rocm():
+        param_ranges = get_rocm_tuning_space(use_fp16)
+    else:
+        # Reduced search space for faster tuning.
+        # TODO(woosuk): Increase the search space and use a performance model to
+        # prune the search space.
+        block_m_range = [16, 32, 64, 128, 256]
+        block_n_range = [32, 64, 128, 256]
+        block_k_range = [64, 128, 256]
+        num_warps_range = [4, 8]
+        group_m_range = [1, 16, 32, 64]
+        num_stage_range = [2, 3, 4, 5]
+
+        param_ranges = {
+            "BLOCK_SIZE_M": block_m_range,
+            "BLOCK_SIZE_N": block_n_range,
+            "BLOCK_SIZE_K": block_k_range,
+            "GROUP_SIZE_M": group_m_range,
+            "num_warps": num_warps_range,
+            "num_stages": num_stage_range,
+        }
+
+    keys, values = zip(*param_ranges.items())
+    for config_values in product(*values):
+        config = dict(zip(keys, config_values))
+        configs.append(config)
     return configs
 
 
+def prune_rocm_search_space(num_tokens, shard_intermediate_size, hidden_size,
+                            search_space, is_fp16):
+    N1, K1 = shard_intermediate_size, hidden_size
+    N2, K2 = hidden_size, shard_intermediate_size // 2
+    pruned_space_1 = prune_rocm_configs(num_tokens * 2, N1, K1, search_space,
+                                        is_fp16)
+    pruned_space_2 = prune_rocm_configs(num_tokens * 2, N2, K2, search_space,
+                                        is_fp16)
+    search_space = merge_unique_dicts(pruned_space_1, pruned_space_2)
+    return search_space
+
+
+# The following code is inspired by ROCm/Triton GEMM tuning script:
+# https://github.com/ROCm/triton/blob/triton-mlir/scripts/amd/gemm/tune_gemm.py#L89
+def prune_rocm_configs(M, N, K, configs, is_fp16=True):
+    pruned_configs = []
+    elemBytes_a = 2 if is_fp16 else 1
+    elemBytes_b = 2 if is_fp16 else 1
+
+    mfma = 16 if M < 32 or N < 32 else 32
+
+    # TODO (zhanglx): figure out the boundary between large and small gemms
+    large_gemm = False
+    if M >= 2048 and N >= 2048:
+        large_gemm = True
+
+    for config in configs:
+        BLOCK_SIZE_M = config.get("BLOCK_SIZE_M")
+        BLOCK_SIZE_N = config.get("BLOCK_SIZE_N")
+        BLOCK_SIZE_K = config.get("BLOCK_SIZE_K")
+        num_warps = config.get("num_warps")
+
+        if is_fp16:
+            matrix_instr_nonkdim = config.get("matrix_instr_nonkdim")
+            if matrix_instr_nonkdim > mfma:
+                continue
+        if mfma == 4 and BLOCK_SIZE_K < 64:
+            continue
+        # some layouts could not work properly in case
+        # number elements per thread is less 1
+        if BLOCK_SIZE_M * BLOCK_SIZE_N < 64:
+            continue
+        SPLIT_K = config.get("SPLIT_K", 1)
+        GROUP_M = config.get("GROUP_SIZE_M")
+        if is_fp16:
+            if (matrix_instr_nonkdim > BLOCK_SIZE_M
+                    or matrix_instr_nonkdim > BLOCK_SIZE_N):
+                continue
+            if (matrix_instr_nonkdim >= M
+                    and matrix_instr_nonkdim != BLOCK_SIZE_M):
+                continue
+            if (matrix_instr_nonkdim >= N
+                    and matrix_instr_nonkdim != BLOCK_SIZE_N):
+                continue
+        # Skip BLOCK_SIZE that is too large compare to M/N
+        # unless BLOCK_SIZE is already small enough
+        if M * 2 < BLOCK_SIZE_M and BLOCK_SIZE_M != 16:
+            continue
+        if N * 2 < BLOCK_SIZE_N and BLOCK_SIZE_N != 16:
+            continue
+        # skip large split_k when not necessary
+        if SPLIT_K != 1 and not need_split_k(M, N, K):
+            continue
+        # skip split_k that leads to EVEN_K = false
+        leap = SPLIT_K * BLOCK_SIZE_K
+        modv = K % leap
+        if modv != 0:
+            continue
+        # skip large GROUP_M
+        if GROUP_M * BLOCK_SIZE_M > M and GROUP_M != 1:
+            continue
+        # out of shared memory resource
+        # TODO (zhanglx): This does not consider the LDS usage in the epilogue
+        LDS = (BLOCK_SIZE_K * BLOCK_SIZE_M * elemBytes_a +
+               BLOCK_SIZE_K * BLOCK_SIZE_N * elemBytes_b)
+        if LDS > 65536:
+            continue
+        # Skip small block sizes and num_warps for large gemm
+        # For fp16 and f8, we want to only use BLOCK_SIZE >= 64
+        if large_gemm:
+            if BLOCK_SIZE_M < 64 or BLOCK_SIZE_N < 64:
+                continue
+            if BLOCK_SIZE_K < 64:
+                continue
+            if num_warps < 4:
+                continue
+
+        pruned_configs.append(config)
+
+    return pruned_configs
+
+
+def need_split_k(SIZE_M, SIZE_N, SIZE_K):
+    return (SIZE_M < 64 or SIZE_N < 64) and SIZE_K > 1024
+
+
+def merge_unique_dicts(list1, list2):
+    result = []
+    combined_list = list1.copy()
+    combined_list.extend(list2)
+    for dictionary in combined_list:
+        if dictionary not in result:
+            result.append(dictionary)
+    return result
+
+
 @ray.remote(num_gpus=1)
 class BenchmarkWorker:
 
@@ -170,6 +318,10 @@ class BenchmarkWorker:
         torch.set_default_device("cuda")
         current_platform.seed_everything(seed)
         self.seed = seed
+        # Get the device ID to allocate tensors and kernels
+        # on the respective GPU. This is required for Ray to work
+        # correctly with multi-GPU tuning on the ROCm platform.
+        self.device_id = int(ray.get_gpu_ids()[0])
 
     def benchmark(
         self,
@@ -217,25 +369,33 @@ class BenchmarkWorker:
     ) -> Dict[str, int]:
         best_config = None
         best_time = float("inf")
-        for config in tqdm(search_space):
-            try:
-                kernel_time = benchmark_config(config,
-                                               num_tokens,
-                                               num_experts,
-                                               shard_intermediate_size,
-                                               hidden_size,
-                                               topk,
-                                               dtype,
-                                               use_fp8_w8a8,
-                                               use_int8_w8a16,
-                                               num_iters=10)
-            except triton.runtime.autotuner.OutOfResources:
-                # Some configurations may be invalid and fail to compile.
-                continue
+        if current_platform.is_rocm():
+            is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
+            search_space = prune_rocm_search_space(num_tokens,
+                                                   shard_intermediate_size,
+                                                   hidden_size, search_space,
+                                                   is_fp16)
 
-            if kernel_time < best_time:
-                best_time = kernel_time
-                best_config = config
+        with torch.cuda.device(self.device_id):
+            for config in tqdm(search_space):
+                try:
+                    kernel_time = benchmark_config(config,
+                                                   num_tokens,
+                                                   num_experts,
+                                                   shard_intermediate_size,
+                                                   hidden_size,
+                                                   topk,
+                                                   dtype,
+                                                   use_fp8_w8a8,
+                                                   use_int8_w8a16,
+                                                   num_iters=20)
+                except triton.runtime.autotuner.OutOfResources:
+                    # Some configurations may be invalid and fail to compile.
+                    continue
+
+                if kernel_time < best_time:
+                    best_time = kernel_time
+                    best_config = config
         now = datetime.now()
         print(f"{now.ctime()}] Completed tuning for batch_size={num_tokens}")
         assert best_config is not None
@@ -244,12 +404,27 @@ class BenchmarkWorker:
 
 def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
     return {
-        "BLOCK_SIZE_M": config["BLOCK_SIZE_M"],
-        "BLOCK_SIZE_N": config["BLOCK_SIZE_N"],
-        "BLOCK_SIZE_K": config["BLOCK_SIZE_K"],
-        "GROUP_SIZE_M": config["GROUP_SIZE_M"],
-        "num_warps": config["num_warps"],
-        "num_stages": config["num_stages"],
+        "BLOCK_SIZE_M":
+        config["BLOCK_SIZE_M"],
+        "BLOCK_SIZE_N":
+        config["BLOCK_SIZE_N"],
+        "BLOCK_SIZE_K":
+        config["BLOCK_SIZE_K"],
+        "GROUP_SIZE_M":
+        config["GROUP_SIZE_M"],
+        "num_warps":
+        config["num_warps"],
+        "num_stages":
+        config["num_stages"],
+        **({
+            "waves_per_eu": config["waves_per_eu"]
+        } if "waves_per_eu" in config else {}),
+        **({
+            "matrix_instr_nonkdim": config["matrix_instr_nonkdim"]
+        } if "matrix_instr_nonkdim" in config else {}),
+        **({
+            "kpack": config["kpack"]
+        } if "kpack" in config else {}),
     }
 
 
@@ -294,7 +469,7 @@ def main(args: argparse.Namespace):
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
 
     hidden_size = config.hidden_size
-    dtype = config.torch_dtype
+    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
     use_int8_w8a16 = args.dtype == "int8_w8a16"
 
@@ -322,7 +497,8 @@ def main(args: argparse.Namespace):
         return ray.get(outputs)
 
     if args.tune:
-        search_space = get_configs_compute_bound()
+        is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
+        search_space = get_configs_compute_bound(is_fp16)
         print(f"Start tuning over {len(search_space)} configurations...")
 
         start = time.time()

benchmarks/kernels/utils.py

@@ -0,0 +1,210 @@
+import dataclasses
+from typing import Any, Callable, Iterable, Optional
+
+import torch
+import torch.utils.benchmark as TBenchmark
+from torch.utils.benchmark import Measurement as TMeasurement
+
+
+@dataclasses.dataclass
+class CudaGraphBenchParams:
+    num_ops_in_cuda_graph: int
+
+
+@dataclasses.dataclass
+class ArgPool:
+    """
+    When some argument of the benchmarking function is annotated with this type,
+    the benchmarking class (BenchMM) will collapse the argument to a pick a
+    single value from the given list of values, during function invocation.
+    For every invocation during a benchmarking run, it will choose a
+    different value from the list.
+    """
+    values: Iterable[Any]
+
+    def __getitem__(self, index):
+        return self.values[index]
+
+
+class Bench:
+
+    class ArgsIterator:
+
+        def __init__(self, args_list, kwargs_list):
+            assert len(args_list) == len(kwargs_list)
+            self.args_list = args_list
+            self.kwargs_list = kwargs_list
+            self.n = len(self.args_list)
+            self.idx = 0
+
+        def __next__(self):
+            while True:
+                yield (self.args_list[self.idx], self.kwargs_list[self.idx])
+                self.idx += 1
+                self.idx = self.idx % self.n
+
+        def reset(self):
+            self.idx = 0
+
+        @property
+        def n_args(self):
+            return self.n
+
+    def __init__(self, cuda_graph_params: Optional[CudaGraphBenchParams],
+                 label: str, sub_label: str, description: str, fn: Callable,
+                 *args, **kwargs):
+
+        self.cuda_graph_params = cuda_graph_params
+        self.use_cuda_graph = self.cuda_graph_params is not None
+        self.label = label
+        self.sub_label = sub_label
+        self.description = description
+        self.fn = fn
+
+        # Process args
+        self._args = args
+        self._kwargs = kwargs
+        self.args_list, self.kwargs_list = self.collapse_argpool(
+            *args, **kwargs)
+        self.args_iterator = self.ArgsIterator(self.args_list,
+                                               self.kwargs_list)
+
+        # Cudagraph runner
+        self.g = None
+        if self.use_cuda_graph:
+            self.g = self.get_cuda_graph_runner()
+
+        # benchmark run params
+        self.min_run_time = 1
+
+    def collapse_argpool(self, *args, **kwargs):
+        argpool_args = [arg for arg in args if isinstance(arg, ArgPool)] + [
+            arg for arg in kwargs.values() if isinstance(arg, ArgPool)
+        ]
+        if len(argpool_args) == 0:
+            return [args], [kwargs]
+
+        # Make sure all argpools are of the same size
+        argpool_size = len(argpool_args[0].values)
+        assert all([argpool_size == len(arg.values) for arg in argpool_args])
+
+        # create copies of the args
+        args_list = []
+        kwargs_list = []
+        for _ in range(argpool_size):
+            args_list.append(args)
+            kwargs_list.append(kwargs.copy())
+
+        for i in range(argpool_size):
+            # collapse args; Just pick the ith value
+            args_list[i] = tuple([
+                arg[i] if isinstance(arg, ArgPool) else arg
+                for arg in args_list[i]
+            ])
+
+            # collapse kwargs
+            kwargs_i = kwargs_list[i]
+            arg_pool_keys = [
+                k for k, v in kwargs_i.items() if isinstance(v, ArgPool)
+            ]
+            for k in arg_pool_keys:
+                # again just pick the ith value
+                kwargs_i[k] = kwargs_i[k][i]
+            kwargs_list[i] = kwargs_i
+
+        return args_list, kwargs_list
+
+    def get_cuda_graph_runner(self):
+        assert self.use_cuda_graph
+        assert self.args_iterator is not None
+
+        num_graph_ops = self.cuda_graph_params.num_ops_in_cuda_graph
+
+        # warmup
+        args_it = self.args_iterator.__next__()
+        for _ in range(2):
+            args, kwargs = next(args_it)
+            self.fn(*args, **kwargs)
+
+        self.args_iterator.reset()
+        args_it = self.args_iterator.__next__()
+        stream = torch.cuda.Stream()
+        with torch.cuda.stream(stream):
+            g = torch.cuda.CUDAGraph()
+            with torch.cuda.graph(g):
+                for _ in range(num_graph_ops):
+                    args, kwargs = next(args_it)
+                    self.fn(*args, **kwargs)
+        return g
+
+    def run_cudagrah(self) -> TMeasurement:
+        assert self.use_cuda_graph
+        globals = {'g': self.g}
+
+        return TBenchmark.Timer(
+            stmt="g.replay()",
+            globals=globals,
+            label=(
+                f"{self.label}"
+                f" | cugraph {self.cuda_graph_params.num_ops_in_cuda_graph} ops"
+            ),
+            sub_label=self.sub_label,
+            description=self.description,
+        ).blocked_autorange(min_run_time=self.min_run_time)
+
+    def run_eager(self) -> TMeasurement:
+        setup = None
+        stmt = None
+        globals = None
+
+        has_arg_pool = self.args_iterator.n_args > 1
+        if has_arg_pool:
+            setup = '''
+                    args_iterator.reset()
+                    args_it = args_iterator.__next__()
+                    '''
+            stmt = '''
+                    args, kwargs = next(args_it)
+                    fn(*args, **kwargs)
+                    '''
+            globals = {'fn': self.fn, 'args_iterator': self.args_iterator}
+        else:
+            # no arg pool. Just use the args and kwargs directly
+            self.args_iterator.reset()
+            args_it = self.args_iterator.__next__()
+            args, kwargs = next(args_it)
+
+            setup = ""
+            stmt = '''
+                    fn(*args, **kwargs)
+                   '''
+            globals = {'fn': self.fn, 'args': args, 'kwargs': kwargs}
+
+        return TBenchmark.Timer(
+            stmt=stmt,
+            setup=setup,
+            globals=globals,
+            label=self.label,
+            sub_label=self.sub_label,
+            description=self.description,
+        ).blocked_autorange(min_run_time=self.min_run_time)
+
+    def run(self) -> TMeasurement:
+        timer = None
+        if self.use_cuda_graph:  # noqa SIM108
+            timer = self.run_cudagrah()
+        else:
+            timer = self.run_eager()
+        if not timer.meets_confidence() or timer.has_warnings:
+            print("Doesn't meet confidence - re-running bench ...")
+            return self.run()
+        return timer
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        if exc_type:
+            print(f"exc type {exc_type}")
+            print(f"exc value {exc_value}")
+            print(f"exc traceback {traceback}")

cmake/utils.cmake

@@ -58,8 +58,8 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
   #
   set(SRCS ${ORIG_SRCS})
   set(CXX_SRCS ${ORIG_SRCS})
-  list(FILTER SRCS EXCLUDE REGEX "\.(cc)|(cpp)$")
-  list(FILTER CXX_SRCS INCLUDE REGEX "\.(cc)|(cpp)$")
+  list(FILTER SRCS EXCLUDE REGEX "\.(cc)|(cpp)|(hip)$")
+  list(FILTER CXX_SRCS INCLUDE REGEX "\.(cc)|(cpp)|(hip)$")
 
   #
   # Generate ROCm/HIP source file names from CUDA file names.

csrc/activation_kernels.cu

@@ -9,8 +9,16 @@
 
 namespace vllm {
 
+template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
+          bool act_first>
+__device__ __forceinline__ scalar_t compute(const scalar_t& x,
+                                            const scalar_t& y) {
+  return act_first ? ACT_FN(x) * y : x * ACT_FN(y);
+}
 // Activation and gating kernel template.
-template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&)>
+
+template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
+          bool act_first>
 __global__ void act_and_mul_kernel(
     scalar_t* __restrict__ out,          // [..., d]
     const scalar_t* __restrict__ input,  // [..., 2, d]
@@ -19,7 +27,7 @@ __global__ void act_and_mul_kernel(
   for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
     const scalar_t x = VLLM_LDG(&input[token_idx * 2 * d + idx]);
     const scalar_t y = VLLM_LDG(&input[token_idx * 2 * d + d + idx]);
-    out[token_idx * d + idx] = ACT_FN(x) * y;
+    out[token_idx * d + idx] = compute<scalar_t, ACT_FN, act_first>(x, y);
   }
 }
 
@@ -55,7 +63,9 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
 }  // namespace vllm
 
 // Launch activation and gating kernel.
-#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL)                            \
+// Use ACT_FIRST (bool) indicating whether to apply the activation function
+// first.
+#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, ACT_FIRST)                 \
   int d = input.size(-1) / 2;                                            \
   int64_t num_tokens = input.numel() / input.size(-1);                   \
   dim3 grid(num_tokens);                                                 \
@@ -64,7 +74,7 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();          \
   VLLM_DISPATCH_FLOATING_TYPES(                                          \
       input.scalar_type(), "act_and_mul_kernel", [&] {                   \
-        vllm::act_and_mul_kernel<scalar_t, KERNEL<scalar_t>>             \
+        vllm::act_and_mul_kernel<scalar_t, KERNEL<scalar_t>, ACT_FIRST>  \
             <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),       \
                                          input.data_ptr<scalar_t>(), d); \
       });
@@ -72,19 +82,27 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
 void silu_and_mul(torch::Tensor& out,    // [..., d]
                   torch::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, true);
+}
+
+void mul_and_silu(torch::Tensor& out,    // [..., d]
+                  torch::Tensor& input)  // [..., 2 * d]
+{
+  // The difference between mul_and_silu and silu_and_mul is that mul_and_silu
+  // applies the silu to the latter half of the input.
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, false);
 }
 
 void gelu_and_mul(torch::Tensor& out,    // [..., d]
                   torch::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel, true);
 }
 
 void gelu_tanh_and_mul(torch::Tensor& out,    // [..., d]
                        torch::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel, true);
 }
 
 namespace vllm {

csrc/core/scalar_type.hpp

@@ -32,7 +32,7 @@ class ScalarType {
         signed_(signed_),
         bias(bias),
         finite_values_only(finite_values_only),
-        nan_repr(nan_repr){};
+        nan_repr(nan_repr) {};
 
   static constexpr ScalarType int_(uint8_t size_bits, int32_t bias = 0) {
     return ScalarType(0, size_bits - 1, true, bias);

csrc/cpu/cpu_types.hpp

@@ -2,13 +2,13 @@
 #define CPU_TYPES_HPP
 
 #if defined(__x86_64__)
-  //x86 implementation
+  // x86 implementation
   #include "cpu_types_x86.hpp"
 #elif defined(__POWER9_VECTOR__)
-  //ppc implementation
+  // ppc implementation
   #include "cpu_types_vsx.hpp"
 #elif defined(__aarch64__)
-  //arm implementation
+  // arm implementation
   #include "cpu_types_arm.hpp"
 #else
   #warning "unsupported vLLM cpu implementation"

csrc/cpu/cpu_types_arm.hpp

@@ -1,48 +1,50 @@
 #include <arm_neon.h>
-#include <torch/all.h> 
+#include <torch/all.h>
 #include <cmath>
 
 namespace vec_op {
 
 #ifdef ARM_BF16_SUPPORT
-  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)                                 \
-    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)                         \
-    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)                         \
-    AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)  
+  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
+    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
+    AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)  \
+    AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
 #else
-  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)                                 \
-    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)                         \
+  #define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
+    AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
     AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
 #endif
 
-#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...)                          \
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
   AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
 
 #ifndef CPU_OP_GUARD
-#define CPU_KERNEL_GUARD_IN(NAME)
-#define CPU_KERNEL_GUARD_OUT(NAME)
+  #define CPU_KERNEL_GUARD_IN(NAME)
+  #define CPU_KERNEL_GUARD_OUT(NAME)
 #else
-#define CPU_KERNEL_GUARD_IN(NAME)                                              \
-  std::cout << #NAME << " invoked." << std::endl;
-#define CPU_KERNEL_GUARD_OUT(NAME) std::cout << #NAME << " exit." << std::endl;
+  #define CPU_KERNEL_GUARD_IN(NAME) \
+    std::cout << #NAME << " invoked." << std::endl;
+  #define CPU_KERNEL_GUARD_OUT(NAME) \
+    std::cout << #NAME << " exit." << std::endl;
 #endif
 
 #define FORCE_INLINE __attribute__((always_inline)) inline
 
 namespace {
-  template <typename T, T... indexes, typename F>
-  constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F &&f) {
-    (f(std::integral_constant<T, indexes>{}), ...);
-  };
-}; 
+template <typename T, T... indexes, typename F>
+constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
+  (f(std::integral_constant<T, indexes>{}), ...);
+};
+};  // namespace
 
 template <typename T, T count, typename F,
           typename = std::enable_if_t<std::is_invocable_v<F, T>>>
-constexpr void unroll_loop(F &&f) {
+constexpr void unroll_loop(F&& f) {
   unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
 }
 
-template <typename T> struct Vec {
+template <typename T>
+struct Vec {
   constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; };
 };
 
@@ -54,127 +56,124 @@ struct FP16Vec8 : public Vec<FP16Vec8> {
 
   float16x8_t reg;
 
-  explicit FP16Vec8(const void *ptr)
-      : reg(vld1q_f16(static_cast<const __fp16 *>(ptr))) {};
+  explicit FP16Vec8(const void* ptr)
+      : reg(vld1q_f16(static_cast<const __fp16*>(ptr))) {};
 
-  explicit FP16Vec8(const FP32Vec8 &);
+  explicit FP16Vec8(const FP32Vec8&);
 
-  void save(void *ptr) const {
-    vst1q_f16(static_cast<__fp16 *>(ptr), reg);
-  }
+  void save(void* ptr) const { vst1q_f16(static_cast<__fp16*>(ptr), reg); }
 };
 
 struct FP16Vec16 : public Vec<FP16Vec16> {
-    constexpr static int VEC_ELEM_NUM = 16;
-    
-    float16x8x2_t reg; 
-    
-    explicit FP16Vec16(const void *ptr) {
-        reg.val[0] = vld1q_f16(reinterpret_cast<const __fp16*>(ptr));        
-        reg.val[1] = vld1q_f16(reinterpret_cast<const __fp16*>(ptr) + 8);    
-    }
-    
-    explicit FP16Vec16(const FP32Vec16& vec);
-    
-    void save(void *ptr) const {
-        vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]);       
-        vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]);   
-    }
-    
-    void save(void *ptr, const int elem_num) const {
-        int full_blocks = elem_num / 8;   
-        int remainder = elem_num % 8;     
-        
-        if (full_blocks > 0) {
-            vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]);
-            if (full_blocks > 1) {
-                vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]);
-            }
-        }
+  constexpr static int VEC_ELEM_NUM = 16;
 
-        // Note: below is the unrolled version of the following code:
-        // 
-        // for (int i = 0; i < remainder; ++i) {
-        //     reinterpret_cast<__fp16*>(ptr)[full_blocks * 8 + i] = 
-        //          vgetq_lane_f16(temp, i);
-        // }
-        // 
-        // For macOS build (Clang), the arm/neon intrinsics function 
-        // `vgetq_lane_f16` needs the parameter `i` to be constant at compile 
-        // time. 
-        
-        if (remainder > 0) {
-            float16x8_t temp = reg.val[full_blocks];
-            __fp16* fp16_ptr = reinterpret_cast<__fp16*>(ptr);
-            switch (remainder)
-            {
-            case 1:
-              fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
-              break;
-            case 2:
-              fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
-              fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
-              break;
-            case 3:
-              fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
-              fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
-              fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
-              break;
-            case 4:
-              fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
-              fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
-              fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
-              fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
-              break;
-            case 5:
-              fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
-              fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
-              fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
-              fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
-              fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4);
-              break;
-            case 6:
-              fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
-              fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
-              fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
-              fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
-              fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4);
-              fp16_ptr[full_blocks * 8 + 5] = vgetq_lane_f16(temp, 5);
-              break;
-            case 7:
-              fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
-              fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
-              fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
-              fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
-              fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4);
-              fp16_ptr[full_blocks * 8 + 5] = vgetq_lane_f16(temp, 5);
-              fp16_ptr[full_blocks * 8 + 6] = vgetq_lane_f16(temp, 6);
-              break;
-            
-            default:
-              break;
-            }
-        }
+  float16x8x2_t reg;
+
+  explicit FP16Vec16(const void* ptr) {
+    reg.val[0] = vld1q_f16(reinterpret_cast<const __fp16*>(ptr));
+    reg.val[1] = vld1q_f16(reinterpret_cast<const __fp16*>(ptr) + 8);
+  }
+
+  explicit FP16Vec16(const FP32Vec16& vec);
+
+  void save(void* ptr) const {
+    vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]);
+    vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]);
+  }
+
+  void save(void* ptr, const int elem_num) const {
+    int full_blocks = elem_num / 8;
+    int remainder = elem_num % 8;
+
+    if (full_blocks > 0) {
+      vst1q_f16(reinterpret_cast<__fp16*>(ptr), reg.val[0]);
+      if (full_blocks > 1) {
+        vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]);
+      }
     }
+
+    // Note: below is the unrolled version of the following code:
+    //
+    // for (int i = 0; i < remainder; ++i) {
+    //     reinterpret_cast<__fp16*>(ptr)[full_blocks * 8 + i] =
+    //          vgetq_lane_f16(temp, i);
+    // }
+    //
+    // For macOS build (Clang), the arm/neon intrinsics function
+    // `vgetq_lane_f16` needs the parameter `i` to be constant at compile
+    // time.
+
+    if (remainder > 0) {
+      float16x8_t temp = reg.val[full_blocks];
+      __fp16* fp16_ptr = reinterpret_cast<__fp16*>(ptr);
+      switch (remainder) {
+        case 1:
+          fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
+          break;
+        case 2:
+          fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
+          fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
+          break;
+        case 3:
+          fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
+          fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
+          fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
+          break;
+        case 4:
+          fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
+          fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
+          fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
+          fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
+          break;
+        case 5:
+          fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
+          fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
+          fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
+          fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
+          fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4);
+          break;
+        case 6:
+          fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
+          fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
+          fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
+          fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
+          fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4);
+          fp16_ptr[full_blocks * 8 + 5] = vgetq_lane_f16(temp, 5);
+          break;
+        case 7:
+          fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
+          fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
+          fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
+          fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
+          fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4);
+          fp16_ptr[full_blocks * 8 + 5] = vgetq_lane_f16(temp, 5);
+          fp16_ptr[full_blocks * 8 + 6] = vgetq_lane_f16(temp, 6);
+          break;
+
+        default:
+          break;
+      }
+    }
+  }
 };
 
-
 #ifdef ARM_BF16_SUPPORT
 struct BF16Vec8 : public Vec<BF16Vec8> {
   constexpr static int VEC_ELEM_NUM = 8;
 
   bfloat16x8_t reg;
 
-  explicit BF16Vec8(const void *ptr)
-      : reg(*reinterpret_cast<const bfloat16x8_t *>(ptr)) {};
+  explicit BF16Vec8(const void* ptr)
+      : reg(*reinterpret_cast<const bfloat16x8_t*>(ptr)) {};
 
   explicit BF16Vec8(bfloat16x8_t data) : reg(data) {};
 
-  explicit BF16Vec8(const FP32Vec8 &);
+  explicit BF16Vec8(const FP32Vec8&);
 
-  explicit BF16Vec8(float32x4x2_t v) : reg(vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[0]), v.val[1])) {};  
+  explicit BF16Vec8(float32x4x2_t v)
+      : reg(vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[0]), v.val[1])) {};
 
-  void save(void *ptr) const { *reinterpret_cast<bfloat16x8_t *>(ptr) = reg; }
+  void save(void* ptr) const { *reinterpret_cast<bfloat16x8_t*>(ptr) = reg; }
 };
 
 struct BF16Vec16 : public Vec<BF16Vec16> {
@@ -182,19 +181,18 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
 
   bfloat16x8x2_t reg;
 
-  explicit BF16Vec16(const void *ptr)
-      : reg(*reinterpret_cast<const bfloat16x8x2_t *>(ptr)) {};
+  explicit BF16Vec16(const void* ptr)
+      : reg(*reinterpret_cast<const bfloat16x8x2_t*>(ptr)) {};
 
   explicit BF16Vec16(bfloat16x8x2_t data) : reg(data) {};
 
-  explicit BF16Vec16(const FP32Vec16 &);
+  explicit BF16Vec16(const FP32Vec16&);
 
-  explicit BF16Vec16(float32x4x4_t v) : reg({
-    vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[0]), v.val[1]),
-    vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[2]), v.val[3])
-  }){};
+  explicit BF16Vec16(float32x4x4_t v)
+      : reg({vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[0]), v.val[1]),
+             vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.val[2]), v.val[3])}) {};
 
-  void save(void *ptr) const { *reinterpret_cast<bfloat16x8x2_t *>(ptr) = reg; };
+  void save(void* ptr) const { *reinterpret_cast<bfloat16x8x2_t*>(ptr) = reg; };
 };
 
 struct BF16Vec32 : public Vec<BF16Vec32> {
@@ -202,19 +200,15 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
 
   bfloat16x8x4_t reg;
 
-  explicit BF16Vec32(const void *ptr)
-      : reg(*reinterpret_cast<const bfloat16x8x4_t *>(ptr)) {};
+  explicit BF16Vec32(const void* ptr)
+      : reg(*reinterpret_cast<const bfloat16x8x4_t*>(ptr)) {};
 
   explicit BF16Vec32(bfloat16x8x4_t data) : reg(data) {};
 
-  explicit BF16Vec32(const BF16Vec8 &vec8_data) : reg({
-    vec8_data.reg,
-    vec8_data.reg,
-    vec8_data.reg,
-    vec8_data.reg
-  }) {};
+  explicit BF16Vec32(const BF16Vec8& vec8_data)
+      : reg({vec8_data.reg, vec8_data.reg, vec8_data.reg, vec8_data.reg}) {};
 
-  void save(void *ptr) const { *reinterpret_cast<bfloat16x8x4_t *>(ptr) = reg; };
+  void save(void* ptr) const { *reinterpret_cast<bfloat16x8x4_t*>(ptr) = reg; };
 };
 #endif
 
@@ -232,11 +226,11 @@ struct FP32Vec4 : public Vec<FP32Vec4> {
 
   explicit FP32Vec4() : reg(vdupq_n_f32(0.0f)) {};
 
-  explicit FP32Vec4(const float *ptr) : reg(vld1q_f32(ptr)) {};
+  explicit FP32Vec4(const float* ptr) : reg(vld1q_f32(ptr)) {};
 
   explicit FP32Vec4(float32x4_t data) : reg(data) {};
 
-  explicit FP32Vec4(const FP32Vec4 &data) : reg(data.reg) {};
+  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {};
 };
 
 struct FP32Vec8 : public Vec<FP32Vec8> {
@@ -252,32 +246,37 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
 
   explicit FP32Vec8() : reg({vmovq_n_f32(0.0), vmovq_n_f32(0.0)}) {};
 
-  explicit FP32Vec8(const float *ptr) : reg({vld1q_f32(ptr), vld1q_f32(ptr + 4)}) {};
+  explicit FP32Vec8(const float* ptr)
+      : reg({vld1q_f32(ptr), vld1q_f32(ptr + 4)}) {};
 
   explicit FP32Vec8(float32x4x2_t data) : reg(data) {};
 
-  explicit FP32Vec8(const FP32Vec8 &data) : reg(data.reg) {};
+  explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {};
 
-  explicit FP32Vec8(const FP16Vec8 &v) {
-        reg.val[0] = vcvt_f32_f16(vget_low_f16(v.reg));  
-        reg.val[1] = vcvt_f32_f16(vget_high_f16(v.reg)); 
-    };
+  explicit FP32Vec8(const FP16Vec8& v) {
+    reg.val[0] = vcvt_f32_f16(vget_low_f16(v.reg));
+    reg.val[1] = vcvt_f32_f16(vget_high_f16(v.reg));
+  };
 
-  explicit FP32Vec8(float16x8_t v) : reg({vcvt_f32_f16(vget_low_f16(v)), vcvt_f32_f16(vget_high_f16(v))}) {};
+  explicit FP32Vec8(float16x8_t v)
+      : reg({vcvt_f32_f16(vget_low_f16(v)), vcvt_f32_f16(vget_high_f16(v))}) {};
 
-  #ifdef ARM_BF16_SUPPORT
+#ifdef ARM_BF16_SUPPORT
 
-  explicit FP32Vec8(bfloat16x8_t v) : reg({vcvtq_low_f32_bf16(v), vcvtq_high_f32_bf16(v)}) {};
+  explicit FP32Vec8(bfloat16x8_t v)
+      : reg({vcvtq_low_f32_bf16(v), vcvtq_high_f32_bf16(v)}) {};
 
-  explicit FP32Vec8(const BF16Vec8 &v) : reg({vcvtq_low_f32_bf16(v.reg), vcvtq_high_f32_bf16(v.reg)}) {};
+  explicit FP32Vec8(const BF16Vec8& v)
+      : reg({vcvtq_low_f32_bf16(v.reg), vcvtq_high_f32_bf16(v.reg)}) {};
 
-  #endif
+#endif
 
   float reduce_sum() const {
     AliasReg ar;
     ar.reg = reg;
     float answer = 0;
-    unroll_loop<int, VEC_ELEM_NUM>([&answer, &ar](int i) { answer += ar.values[i]; });
+    unroll_loop<int, VEC_ELEM_NUM>(
+        [&answer, &ar](int i) { answer += ar.values[i]; });
 
     return answer;
   }
@@ -324,10 +323,14 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
     AliasReg ar;
     ar.reg = reg;
 
-    float32x2_t er_vec0 = {static_cast<float32_t>(erf(ar.values[0])), static_cast<float32_t>(erf(ar.values[1]))};
-    float32x2_t er_vec1 = {static_cast<float32_t>(erf(ar.values[2])), static_cast<float32_t>(erf(ar.values[3]))};
-    float32x2_t er_vec2 = {static_cast<float32_t>(erf(ar.values[4])), static_cast<float32_t>(erf(ar.values[5]))};
-    float32x2_t er_vec3 = {static_cast<float32_t>(erf(ar.values[6])), static_cast<float32_t>(erf(ar.values[7]))};
+    float32x2_t er_vec0 = {static_cast<float32_t>(erf(ar.values[0])),
+                           static_cast<float32_t>(erf(ar.values[1]))};
+    float32x2_t er_vec1 = {static_cast<float32_t>(erf(ar.values[2])),
+                           static_cast<float32_t>(erf(ar.values[3]))};
+    float32x2_t er_vec2 = {static_cast<float32_t>(erf(ar.values[4])),
+                           static_cast<float32_t>(erf(ar.values[5]))};
+    float32x2_t er_vec3 = {static_cast<float32_t>(erf(ar.values[6])),
+                           static_cast<float32_t>(erf(ar.values[7]))};
 
     float32x4_t result0 = vcombine_f32(er_vec0, er_vec1);
     float32x4_t result1 = vcombine_f32(er_vec2, er_vec3);
@@ -337,25 +340,29 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
     result.val[1] = result1;
 
     return FP32Vec8(result);
-  } 
-
-  FP32Vec8 operator*(const FP32Vec8 &b) const {
-    return FP32Vec8(float32x4x2_t({vmulq_f32(reg.val[0], b.reg.val[0]), vmulq_f32(reg.val[1], b.reg.val[1])}));
   }
 
-  FP32Vec8 operator+(const FP32Vec8 &b) const {
-    return FP32Vec8(float32x4x2_t({vaddq_f32(reg.val[0], b.reg.val[0]), vaddq_f32(reg.val[1], b.reg.val[1])}));
+  FP32Vec8 operator*(const FP32Vec8& b) const {
+    return FP32Vec8(float32x4x2_t({vmulq_f32(reg.val[0], b.reg.val[0]),
+                                   vmulq_f32(reg.val[1], b.reg.val[1])}));
   }
 
-  FP32Vec8 operator-(const FP32Vec8 &b) const {
-    return FP32Vec8(float32x4x2_t({vsubq_f32(reg.val[0], b.reg.val[0]), vsubq_f32(reg.val[1], b.reg.val[1])}));
+  FP32Vec8 operator+(const FP32Vec8& b) const {
+    return FP32Vec8(float32x4x2_t({vaddq_f32(reg.val[0], b.reg.val[0]),
+                                   vaddq_f32(reg.val[1], b.reg.val[1])}));
   }
 
-  FP32Vec8 operator/(const FP32Vec8 &b) const {
-    return FP32Vec8(float32x4x2_t({vdivq_f32(reg.val[0], b.reg.val[0]), vdivq_f32(reg.val[1], b.reg.val[1])}));
+  FP32Vec8 operator-(const FP32Vec8& b) const {
+    return FP32Vec8(float32x4x2_t({vsubq_f32(reg.val[0], b.reg.val[0]),
+                                   vsubq_f32(reg.val[1], b.reg.val[1])}));
   }
 
-  void save(float *ptr) const {
+  FP32Vec8 operator/(const FP32Vec8& b) const {
+    return FP32Vec8(float32x4x2_t({vdivq_f32(reg.val[0], b.reg.val[0]),
+                                   vdivq_f32(reg.val[1], b.reg.val[1])}));
+  }
+
+  void save(float* ptr) const {
     vst1q_f32(ptr, reg.val[0]);
     vst1q_f32(ptr + 4, reg.val[1]);
   }
@@ -370,103 +377,100 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   float32x4x4_t reg;
 
-  explicit FP32Vec16(float v) : reg({vmovq_n_f32(v), vmovq_n_f32(v), vmovq_n_f32(v), vmovq_n_f32(v)}) {}
+  explicit FP32Vec16(float v)
+      : reg({vmovq_n_f32(v), vmovq_n_f32(v), vmovq_n_f32(v), vmovq_n_f32(v)}) {}
 
-  explicit FP32Vec16() : reg({vmovq_n_f32(0.0), vmovq_n_f32(0.0), vmovq_n_f32(0.0), vmovq_n_f32(0.0)}) {}
+  explicit FP32Vec16()
+      : reg({vmovq_n_f32(0.0), vmovq_n_f32(0.0), vmovq_n_f32(0.0),
+             vmovq_n_f32(0.0)}) {}
 
-  explicit FP32Vec16(const float *ptr) : reg({vld1q_f32(ptr), vld1q_f32(ptr + 4), vld1q_f32(ptr + 8), vld1q_f32(ptr + 12)}) {}
+  explicit FP32Vec16(const float* ptr)
+      : reg({vld1q_f32(ptr), vld1q_f32(ptr + 4), vld1q_f32(ptr + 8),
+             vld1q_f32(ptr + 12)}) {}
 
   explicit FP32Vec16(float32x4x4_t data) : reg(data) {}
 
-  explicit FP32Vec16(const FP32Vec8 &data) {
-        reg.val[0] = data.reg.val[0]; 
-        reg.val[1] = data.reg.val[1]; 
-        reg.val[2] = data.reg.val[0]; 
-        reg.val[3] = data.reg.val[1]; 
+  explicit FP32Vec16(const FP32Vec8& data) {
+    reg.val[0] = data.reg.val[0];
+    reg.val[1] = data.reg.val[1];
+    reg.val[2] = data.reg.val[0];
+    reg.val[3] = data.reg.val[1];
   }
 
-  explicit FP32Vec16(const FP32Vec16 &data) : reg(data.reg) {}
+  explicit FP32Vec16(const FP32Vec16& data) : reg(data.reg) {}
 
-  explicit FP32Vec16(const FP16Vec8 &v) : FP32Vec16(FP32Vec8(v.reg)) {}
+  explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v.reg)) {}
 
-  #ifdef ARM_BF16_SUPPORT
-  explicit FP32Vec16(bfloat16x8x2_t v) : reg({
-    vcvtq_low_f32_bf16(v.val[0]),
-    vcvtq_high_f32_bf16(v.val[0]),
-    vcvtq_low_f32_bf16(v.val[1]),
-    vcvtq_high_f32_bf16(v.val[1])
-  }) {};
-  #endif
+#ifdef ARM_BF16_SUPPORT
+  explicit FP32Vec16(bfloat16x8x2_t v)
+      : reg({vcvtq_low_f32_bf16(v.val[0]), vcvtq_high_f32_bf16(v.val[0]),
+             vcvtq_low_f32_bf16(v.val[1]), vcvtq_high_f32_bf16(v.val[1])}) {};
+#endif
 
-  explicit FP32Vec16(const FP32Vec4 &data) {
+  explicit FP32Vec16(const FP32Vec4& data) {
     reg.val[0] = data.reg;
     reg.val[1] = data.reg;
     reg.val[2] = data.reg;
     reg.val[3] = data.reg;
   };
 
-  #ifdef ARM_BF16_SUPPORT
-  explicit FP32Vec16(const BF16Vec16 &v) : reg({
-    vcvtq_low_f32_bf16(v.reg.val[0]),
-    vcvtq_high_f32_bf16(v.reg.val[0]),
-    vcvtq_low_f32_bf16(v.reg.val[1]),
-    vcvtq_high_f32_bf16(v.reg.val[1])
-  }) {};
+#ifdef ARM_BF16_SUPPORT
+  explicit FP32Vec16(const BF16Vec16& v)
+      : reg({vcvtq_low_f32_bf16(v.reg.val[0]),
+             vcvtq_high_f32_bf16(v.reg.val[0]),
+             vcvtq_low_f32_bf16(v.reg.val[1]),
+             vcvtq_high_f32_bf16(v.reg.val[1])}) {};
 
-  explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {};
-  #endif
+  explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
+#endif
 
-  explicit FP32Vec16(const FP16Vec16 &v) {
-      reg.val[0] = vcvt_f32_f16(vget_low_f16(v.reg.val[0]));
-      reg.val[1] = vcvt_f32_f16(vget_high_f16(v.reg.val[0]));
-      reg.val[2] = vcvt_f32_f16(vget_low_f16(v.reg.val[1]));
-      reg.val[3] = vcvt_f32_f16(vget_high_f16(v.reg.val[1]));
+  explicit FP32Vec16(const FP16Vec16& v) {
+    reg.val[0] = vcvt_f32_f16(vget_low_f16(v.reg.val[0]));
+    reg.val[1] = vcvt_f32_f16(vget_high_f16(v.reg.val[0]));
+    reg.val[2] = vcvt_f32_f16(vget_low_f16(v.reg.val[1]));
+    reg.val[3] = vcvt_f32_f16(vget_high_f16(v.reg.val[1]));
   };
 
-  FP32Vec16 operator+(const FP32Vec16 &b) const {
-    return FP32Vec16(float32x4x4_t({
-        vaddq_f32(reg.val[0], b.reg.val[0]),
-        vaddq_f32(reg.val[1], b.reg.val[1]),
-        vaddq_f32(reg.val[2], b.reg.val[2]),
-        vaddq_f32(reg.val[3], b.reg.val[3])}));
+  FP32Vec16 operator+(const FP32Vec16& b) const {
+    return FP32Vec16(float32x4x4_t({vaddq_f32(reg.val[0], b.reg.val[0]),
+                                    vaddq_f32(reg.val[1], b.reg.val[1]),
+                                    vaddq_f32(reg.val[2], b.reg.val[2]),
+                                    vaddq_f32(reg.val[3], b.reg.val[3])}));
   };
 
-  FP32Vec16 operator*(const FP32Vec16 &b) const {
-    return FP32Vec16(float32x4x4_t({
-        vmulq_f32(reg.val[0], b.reg.val[0]),
-        vmulq_f32(reg.val[1], b.reg.val[1]),
-        vmulq_f32(reg.val[2], b.reg.val[2]),
-        vmulq_f32(reg.val[3], b.reg.val[3])}));
+  FP32Vec16 operator*(const FP32Vec16& b) const {
+    return FP32Vec16(float32x4x4_t({vmulq_f32(reg.val[0], b.reg.val[0]),
+                                    vmulq_f32(reg.val[1], b.reg.val[1]),
+                                    vmulq_f32(reg.val[2], b.reg.val[2]),
+                                    vmulq_f32(reg.val[3], b.reg.val[3])}));
   };
 
-  FP32Vec16 operator-(const FP32Vec16 &b) const {
-    return FP32Vec16(float32x4x4_t({
-        vsubq_f32(reg.val[0], b.reg.val[0]),
-        vsubq_f32(reg.val[1], b.reg.val[1]),
-        vsubq_f32(reg.val[2], b.reg.val[2]),
-        vsubq_f32(reg.val[3], b.reg.val[3])
-    }));
+  FP32Vec16 operator-(const FP32Vec16& b) const {
+    return FP32Vec16(float32x4x4_t({vsubq_f32(reg.val[0], b.reg.val[0]),
+                                    vsubq_f32(reg.val[1], b.reg.val[1]),
+                                    vsubq_f32(reg.val[2], b.reg.val[2]),
+                                    vsubq_f32(reg.val[3], b.reg.val[3])}));
   };
 
-  FP32Vec16 operator/(const FP32Vec16 &b) const {
-    return FP32Vec16(float32x4x4_t({
-        vdivq_f32(reg.val[0], b.reg.val[0]),
-        vdivq_f32(reg.val[1], b.reg.val[1]),
-        vdivq_f32(reg.val[2], b.reg.val[2]),
-        vdivq_f32(reg.val[3], b.reg.val[3])
-    }));
+  FP32Vec16 operator/(const FP32Vec16& b) const {
+    return FP32Vec16(float32x4x4_t({vdivq_f32(reg.val[0], b.reg.val[0]),
+                                    vdivq_f32(reg.val[1], b.reg.val[1]),
+                                    vdivq_f32(reg.val[2], b.reg.val[2]),
+                                    vdivq_f32(reg.val[3], b.reg.val[3])}));
   };
 
   float reduce_sum() const {
     AliasReg ar;
     ar.reg = reg;
     float answer = 0;
-    unroll_loop<int, VEC_ELEM_NUM>([&answer, &ar](int i) { answer += ar.values[i]; });
+    unroll_loop<int, VEC_ELEM_NUM>(
+        [&answer, &ar](int i) { answer += ar.values[i]; });
 
     return answer;
   };
 
-  template <int group_size> float reduce_sub_sum(int idx) {
+  template <int group_size>
+  float reduce_sub_sum(int idx) {
     static_assert(VEC_ELEM_NUM % group_size == 0);
 
     AliasReg ar;
@@ -479,7 +483,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return answer;
   };
 
-  void save(float *ptr) const {
+  void save(float* ptr) const {
     vst1q_f32(ptr, reg.val[0]);
     vst1q_f32(ptr + 4, reg.val[1]);
     vst1q_f32(ptr + 8, reg.val[2]);
@@ -487,43 +491,59 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   };
 };
 
-template <typename T> struct VecType { using vec_type = void; };
+template <typename T>
+struct VecType {
+  using vec_type = void;
+};
 
-template <typename T> using vec_t = typename VecType<T>::vec_type;
+template <typename T>
+using vec_t = typename VecType<T>::vec_type;
 
-template <> struct VecType<float> { using vec_type = FP32Vec8; };
+template <>
+struct VecType<float> {
+  using vec_type = FP32Vec8;
+};
 
-template <> struct VecType<c10::Half> { using vec_type = FP16Vec8; };
+template <>
+struct VecType<c10::Half> {
+  using vec_type = FP16Vec8;
+};
 
 #ifdef ARM_BF16_SUPPORT
-template <> struct VecType<c10::BFloat16> { using vec_type = BF16Vec8; };
+template <>
+struct VecType<c10::BFloat16> {
+  using vec_type = BF16Vec8;
+};
 #endif
 
-template <typename T> void storeFP32(float v, T *ptr) { *ptr = v; }
-
-template <> inline void storeFP32<c10::Half>(float v, c10::Half *ptr) {
-  *reinterpret_cast<__fp16 *>(ptr) = v;
+template <typename T>
+void storeFP32(float v, T* ptr) {
+  *ptr = v;
 }
 
-inline FP16Vec16::FP16Vec16(const FP32Vec16 &v) {
-    float16x4_t low_0 = vcvt_f16_f32(v.reg.val[0]);
-    float16x4_t high_0 = vcvt_f16_f32(v.reg.val[1]);
-    float16x4_t low_1 = vcvt_f16_f32(v.reg.val[2]);
-    float16x4_t high_1 = vcvt_f16_f32(v.reg.val[3]);
+template <>
+inline void storeFP32<c10::Half>(float v, c10::Half* ptr) {
+  *reinterpret_cast<__fp16*>(ptr) = v;
+}
 
-    reg.val[0] = vcombine_f16(low_0, high_0);
-    reg.val[1] = vcombine_f16(low_1, high_1);
+inline FP16Vec16::FP16Vec16(const FP32Vec16& v) {
+  float16x4_t low_0 = vcvt_f16_f32(v.reg.val[0]);
+  float16x4_t high_0 = vcvt_f16_f32(v.reg.val[1]);
+  float16x4_t low_1 = vcvt_f16_f32(v.reg.val[2]);
+  float16x4_t high_1 = vcvt_f16_f32(v.reg.val[3]);
+
+  reg.val[0] = vcombine_f16(low_0, high_0);
+  reg.val[1] = vcombine_f16(low_1, high_1);
 };
 
-inline FP16Vec8 :: FP16Vec8(const FP32Vec8 &v) {
-    float16x4_t lower_half = vcvt_f16_f32(v.reg.val[0]);
-    float16x4_t upper_half = vcvt_f16_f32(v.reg.val[1]);
+inline FP16Vec8 ::FP16Vec8(const FP32Vec8& v) {
+  float16x4_t lower_half = vcvt_f16_f32(v.reg.val[0]);
+  float16x4_t upper_half = vcvt_f16_f32(v.reg.val[1]);
 
-    reg = vcombine_f16(lower_half, upper_half);
+  reg = vcombine_f16(lower_half, upper_half);
 };
 
-inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) {
-
+inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) {
   acc.reg.val[0] = vfmaq_f32(acc.reg.val[0], a.reg.val[0], b.reg.val[0]);
   acc.reg.val[1] = vfmaq_f32(acc.reg.val[1], a.reg.val[1], b.reg.val[1]);
   acc.reg.val[2] = vfmaq_f32(acc.reg.val[2], a.reg.val[2], b.reg.val[2]);
@@ -531,8 +551,7 @@ inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) {
 };
 
 #ifdef ARM_BF16_SUPPORT
-inline void fma(FP32Vec16 &acc, BF16Vec32 &a, BF16Vec32 &b) {
-
+inline void fma(FP32Vec16& acc, BF16Vec32& a, BF16Vec32& b) {
   float32x4_t a0_low = vcvt_f32_bf16(vget_low_bf16(a.reg.val[0]));
   float32x4_t a0_high = vcvt_f32_bf16(vget_high_bf16(a.reg.val[0]));
   float32x4_t a1_low = vcvt_f32_bf16(vget_low_bf16(a.reg.val[1]));
@@ -551,22 +570,22 @@ inline void fma(FP32Vec16 &acc, BF16Vec32 &a, BF16Vec32 &b) {
 #endif
 
 #ifdef ARM_BF16_SUPPORT
-inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) : reg(vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[0]), v.reg.val[1])) {};
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v)
+    : reg(vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[0]), v.reg.val[1])) {
+      };
 
-inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) : reg({
-    vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[0]), v.reg.val[1]),
-    vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[2]), v.reg.val[3])
-  }){};
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v)
+    : reg({vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[0]), v.reg.val[1]),
+           vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(v.reg.val[2]),
+                               v.reg.val[3])}) {};
 #endif
 
-inline void prefetch(const void *addr) {
-    __builtin_prefetch(addr, 0, 1);
-};
+inline void prefetch(const void* addr) { __builtin_prefetch(addr, 0, 1); };
 
 #ifdef ARM_BF16_SUPPORT
 template <>
-inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) { 
-  *reinterpret_cast<__bf16 *>(ptr) = vcvth_bf16_f32(v);
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
+  *reinterpret_cast<__bf16*>(ptr) = vcvth_bf16_f32(v);
 };
 #endif
-};
+};  // namespace vec_op

csrc/cpu/cpu_types_vsx.hpp

@@ -9,38 +9,40 @@
 namespace vec_op {
 
 // FIXME: FP16 is not fully supported in Torch-CPU
-#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)                                 \
-  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)                         \
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
   AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
 
-#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...)                          \
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
   AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
 
 #ifndef CPU_OP_GUARD
-#define CPU_KERNEL_GUARD_IN(NAME)
-#define CPU_KERNEL_GUARD_OUT(NAME)
+  #define CPU_KERNEL_GUARD_IN(NAME)
+  #define CPU_KERNEL_GUARD_OUT(NAME)
 #else
-#define CPU_KERNEL_GUARD_IN(NAME)                                              \
-  std::cout << #NAME << " invoked." << std::endl;
-#define CPU_KERNEL_GUARD_OUT(NAME) std::cout << #NAME << " exit." << std::endl;
+  #define CPU_KERNEL_GUARD_IN(NAME) \
+    std::cout << #NAME << " invoked." << std::endl;
+  #define CPU_KERNEL_GUARD_OUT(NAME) \
+    std::cout << #NAME << " exit." << std::endl;
 #endif
 
 #define FORCE_INLINE __attribute__((always_inline)) inline
 
 namespace {
 template <typename T, T... indexes, typename F>
-constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F &&f) {
+constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
   (f(std::integral_constant<T, indexes>{}), ...);
 }
-}; // namespace
+};  // namespace
 
 template <typename T, T count, typename F,
           typename = std::enable_if_t<std::is_invocable_v<F, T>>>
-constexpr void unroll_loop(F &&f) {
+constexpr void unroll_loop(F&& f) {
   unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
 }
 
-template <typename T> struct Vec {
+template <typename T>
+struct Vec {
   constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; }
 };
 
@@ -68,12 +70,14 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
 
   __vector signed short reg;
 
-  explicit BF16Vec8(const void *ptr)
-      : reg((__vector signed short)vec_xl(0, (__vector signed short *)ptr)) {}
+  explicit BF16Vec8(const void* ptr)
+      : reg((__vector signed short)vec_xl(0, (__vector signed short*)ptr)) {}
 
-  explicit BF16Vec8(const FP32Vec8 &);
+  explicit BF16Vec8(const FP32Vec8&);
 
-  void save(void *ptr) const { *reinterpret_cast<__vector signed short *>(ptr) = reg; }
+  void save(void* ptr) const {
+    *reinterpret_cast<__vector signed short*>(ptr) = reg;
+  }
 };
 
 struct BF16Vec16 : public Vec<BF16Vec16> {
@@ -81,18 +85,18 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
 
   ss16x8x2_t reg;
 
-  explicit BF16Vec16(const void *ptr) {
+  explicit BF16Vec16(const void* ptr) {
     // Load 256 bits in two parts
-    reg.val[0] = (__vector signed short)vec_xl(0,  (signed short *)ptr);
-    reg.val[1] = (__vector signed short)vec_xl(16, (signed short *)ptr);
+    reg.val[0] = (__vector signed short)vec_xl(0, (signed short*)ptr);
+    reg.val[1] = (__vector signed short)vec_xl(16, (signed short*)ptr);
   }
 
-  explicit BF16Vec16(const FP32Vec16 &);
+  explicit BF16Vec16(const FP32Vec16&);
 
-  void save(void *ptr) const {
+  void save(void* ptr) const {
     // Save 256 bits in two parts
-    vec_xst(reg.val[0], 0, (signed short *)ptr);
-    vec_xst(reg.val[1], 16, (signed short *)ptr);
+    vec_xst(reg.val[0], 0, (signed short*)ptr);
+    vec_xst(reg.val[1], 16, (signed short*)ptr);
   }
 };
 
@@ -102,19 +106,15 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
   constexpr static int VEC_ELEM_NUM = 32;
 
   ss16x8x4_t reg;
-  explicit BF16Vec32(const void *ptr)
-      : reg(*reinterpret_cast<const ss16x8x4_t *>(ptr)) {}
+  explicit BF16Vec32(const void* ptr)
+      : reg(*reinterpret_cast<const ss16x8x4_t*>(ptr)) {}
 
   explicit BF16Vec32(ss16x8x4_t data) : reg(data) {}
 
-  explicit BF16Vec32(const BF16Vec8 &vec8_data) : reg({
-    vec8_data.reg,
-    vec8_data.reg,
-    vec8_data.reg,
-    vec8_data.reg
-  }) {}
+  explicit BF16Vec32(const BF16Vec8& vec8_data)
+      : reg({vec8_data.reg, vec8_data.reg, vec8_data.reg, vec8_data.reg}) {}
 
-  void save(void *ptr) const { *reinterpret_cast<ss16x8x4_t *>(ptr) = reg; }
+  void save(void* ptr) const { *reinterpret_cast<ss16x8x4_t*>(ptr) = reg; }
 };
 
 struct FP32Vec4 : public Vec<FP32Vec4> {
@@ -130,11 +130,11 @@ struct FP32Vec4 : public Vec<FP32Vec4> {
 
   explicit FP32Vec4() : reg(vec_splats(0.0f)) {}
 
-  explicit FP32Vec4(const float *ptr) : reg(vec_xl(0, ptr)) {}
+  explicit FP32Vec4(const float* ptr) : reg(vec_xl(0, ptr)) {}
 
   explicit FP32Vec4(__vector float data) : reg(data) {}
 
-  explicit FP32Vec4(const FP32Vec4 &data) : reg(data.reg) {}
+  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {}
 };
 
 struct FP32Vec8 : public Vec<FP32Vec8> {
@@ -156,19 +156,19 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
     reg.val[1] = vec_splats(0.0f);
   }
 
-  explicit FP32Vec8(const float *ptr) {
+  explicit FP32Vec8(const float* ptr) {
     reg.val[0] = vec_xl(0, ptr);
     reg.val[1] = vec_xl(16, ptr);
   }
 
   explicit FP32Vec8(f32x4x2_t data) : reg(data) {}
 
-  explicit FP32Vec8(const FP32Vec8 &data) {
+  explicit FP32Vec8(const FP32Vec8& data) {
     reg.val[0] = data.reg.val[0];
     reg.val[1] = data.reg.val[1];
   }
 
-  explicit FP32Vec8(const BF16Vec8 &v) {
+  explicit FP32Vec8(const BF16Vec8& v) {
     reg.val[0] = (__vector float)vec_mergeh(zero, v.reg);
     reg.val[1] = (__vector float)vec_mergel(zero, v.reg);
   }
@@ -177,7 +177,8 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
     AliasReg ar;
     ar.reg = reg;
     float result = 0;
-    unroll_loop<int, VEC_ELEM_NUM>([&result, &ar](int i) { result += ar.values[i]; });
+    unroll_loop<int, VEC_ELEM_NUM>(
+        [&result, &ar](int i) { result += ar.values[i]; });
 
     return result;
   }
@@ -230,23 +231,27 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
     return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
   }
 
-  FP32Vec8 operator*(const FP32Vec8 &b) const {
-    return FP32Vec8({vec_mul(reg.val[0], b.reg.val[0]), vec_mul(reg.val[1], b.reg.val[1])});
+  FP32Vec8 operator*(const FP32Vec8& b) const {
+    return FP32Vec8(
+        {vec_mul(reg.val[0], b.reg.val[0]), vec_mul(reg.val[1], b.reg.val[1])});
   }
 
-  FP32Vec8 operator+(const FP32Vec8 &b) const {
-    return FP32Vec8({vec_add(reg.val[0], b.reg.val[0]), vec_add(reg.val[1], b.reg.val[1])});
+  FP32Vec8 operator+(const FP32Vec8& b) const {
+    return FP32Vec8(
+        {vec_add(reg.val[0], b.reg.val[0]), vec_add(reg.val[1], b.reg.val[1])});
   }
 
-  FP32Vec8 operator-(const FP32Vec8 &b) const {
-    return FP32Vec8({vec_sub(reg.val[0], b.reg.val[0]), vec_sub(reg.val[1], b.reg.val[1])});
+  FP32Vec8 operator-(const FP32Vec8& b) const {
+    return FP32Vec8(
+        {vec_sub(reg.val[0], b.reg.val[0]), vec_sub(reg.val[1], b.reg.val[1])});
   }
 
-  FP32Vec8 operator/(const FP32Vec8 &b) const {
-    return FP32Vec8({vec_div(reg.val[0], b.reg.val[0]), vec_div(reg.val[1], b.reg.val[1])});
+  FP32Vec8 operator/(const FP32Vec8& b) const {
+    return FP32Vec8(
+        {vec_div(reg.val[0], b.reg.val[0]), vec_div(reg.val[1], b.reg.val[1])});
   }
 
-  void save(float *ptr) const {
+  void save(float* ptr) const {
     vec_xst(reg.val[0], 0, ptr);
     vec_xst(reg.val[1], 16, ptr);
   }
@@ -275,7 +280,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     reg.val[3] = vec_splats(0.0f);
   }
 
-  explicit FP32Vec16(const float *ptr) {
+  explicit FP32Vec16(const float* ptr) {
     reg.val[0] = vec_xl(0, ptr);
     reg.val[1] = vec_xl(16, ptr);
     reg.val[2] = vec_xl(32, ptr);
@@ -284,78 +289,76 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   explicit FP32Vec16(f32x4x4_t data) : reg(data) {}
 
-  explicit FP32Vec16(const FP32Vec16 &data) {
+  explicit FP32Vec16(const FP32Vec16& data) {
     reg.val[0] = data.reg.val[0];
     reg.val[1] = data.reg.val[1];
     reg.val[2] = data.reg.val[2];
     reg.val[3] = data.reg.val[3];
   }
 
-  explicit FP32Vec16(const FP32Vec4 &data) {
+  explicit FP32Vec16(const FP32Vec4& data) {
     reg.val[0] = data.reg;
     reg.val[1] = data.reg;
     reg.val[2] = data.reg;
     reg.val[3] = data.reg;
   }
 
-  explicit FP32Vec16(const FP32Vec8 &data) {
+  explicit FP32Vec16(const FP32Vec8& data) {
     reg.val[0] = data.reg.val[0];
     reg.val[1] = data.reg.val[1];
     reg.val[2] = data.reg.val[0];
     reg.val[3] = data.reg.val[1];
   }
 
-  explicit FP32Vec16(const BF16Vec16 &v) {
+  explicit FP32Vec16(const BF16Vec16& v) {
     reg.val[0] = (__vector float)vec_mergeh(zero, v.reg.val[0]);
     reg.val[1] = (__vector float)vec_mergel(zero, v.reg.val[0]);
     reg.val[2] = (__vector float)vec_mergeh(zero, v.reg.val[1]);
     reg.val[3] = (__vector float)vec_mergel(zero, v.reg.val[1]);
   }
 
-  explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}
+  explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
 
-  FP32Vec16 operator*(const FP32Vec16 &b) const {
-    return FP32Vec16(f32x4x4_t({
-        vec_mul(reg.val[0], b.reg.val[0]),
-        vec_mul(reg.val[1], b.reg.val[1]),
-        vec_mul(reg.val[2], b.reg.val[2]),
-        vec_mul(reg.val[3], b.reg.val[3])}));
+  FP32Vec16 operator*(const FP32Vec16& b) const {
+    return FP32Vec16(f32x4x4_t({vec_mul(reg.val[0], b.reg.val[0]),
+                                vec_mul(reg.val[1], b.reg.val[1]),
+                                vec_mul(reg.val[2], b.reg.val[2]),
+                                vec_mul(reg.val[3], b.reg.val[3])}));
   }
 
-  FP32Vec16 operator+(const FP32Vec16 &b) const {
-    return FP32Vec16(f32x4x4_t({
-        vec_add(reg.val[0], b.reg.val[0]),
-        vec_add(reg.val[1], b.reg.val[1]),
-        vec_add(reg.val[2], b.reg.val[2]),
-        vec_add(reg.val[3], b.reg.val[3])}));
+  FP32Vec16 operator+(const FP32Vec16& b) const {
+    return FP32Vec16(f32x4x4_t({vec_add(reg.val[0], b.reg.val[0]),
+                                vec_add(reg.val[1], b.reg.val[1]),
+                                vec_add(reg.val[2], b.reg.val[2]),
+                                vec_add(reg.val[3], b.reg.val[3])}));
   }
 
-  FP32Vec16 operator-(const FP32Vec16 &b) const {
-    return FP32Vec16(f32x4x4_t({
-        vec_sub(reg.val[0], b.reg.val[0]),
-        vec_sub(reg.val[1], b.reg.val[1]),
-        vec_sub(reg.val[2], b.reg.val[2]),
-        vec_sub(reg.val[3], b.reg.val[3])}));
+  FP32Vec16 operator-(const FP32Vec16& b) const {
+    return FP32Vec16(f32x4x4_t({vec_sub(reg.val[0], b.reg.val[0]),
+                                vec_sub(reg.val[1], b.reg.val[1]),
+                                vec_sub(reg.val[2], b.reg.val[2]),
+                                vec_sub(reg.val[3], b.reg.val[3])}));
   }
 
-  FP32Vec16 operator/(const FP32Vec16 &b) const {
-    return FP32Vec16(f32x4x4_t({
-        vec_div(reg.val[0], b.reg.val[0]),
-        vec_div(reg.val[1], b.reg.val[1]),
-        vec_div(reg.val[2], b.reg.val[2]),
-        vec_div(reg.val[3], b.reg.val[3])}));
+  FP32Vec16 operator/(const FP32Vec16& b) const {
+    return FP32Vec16(f32x4x4_t({vec_div(reg.val[0], b.reg.val[0]),
+                                vec_div(reg.val[1], b.reg.val[1]),
+                                vec_div(reg.val[2], b.reg.val[2]),
+                                vec_div(reg.val[3], b.reg.val[3])}));
   }
 
   float reduce_sum() const {
     AliasReg ar;
     ar.reg = reg;
     float result = 0;
-    unroll_loop<int, VEC_ELEM_NUM>([&result, &ar](int i) { result += ar.values[i]; });
+    unroll_loop<int, VEC_ELEM_NUM>(
+        [&result, &ar](int i) { result += ar.values[i]; });
 
     return result;
   }
 
-  template <int group_size> float reduce_sub_sum(int idx) {
+  template <int group_size>
+  float reduce_sub_sum(int idx) {
     static_assert(VEC_ELEM_NUM % group_size == 0);
 
     AliasReg ar;
@@ -368,7 +371,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return result;
   }
 
-  void save(float *ptr) const {
+  void save(float* ptr) const {
     vec_xst(reg.val[0], 0, ptr);
     vec_xst(reg.val[1], 16, ptr);
     vec_xst(reg.val[2], 32, ptr);
@@ -376,43 +379,62 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   }
 };
 
-template <typename T> struct VecType { using vec_type = void; };
+template <typename T>
+struct VecType {
+  using vec_type = void;
+};
 
-template <typename T> using vec_t = typename VecType<T>::vec_type;
+template <typename T>
+using vec_t = typename VecType<T>::vec_type;
 
-template <> struct VecType<float> { using vec_type = FP32Vec8; };
+template <>
+struct VecType<float> {
+  using vec_type = FP32Vec8;
+};
 
-template <> struct VecType<c10::BFloat16> { using vec_type = BF16Vec8; };
+template <>
+struct VecType<c10::BFloat16> {
+  using vec_type = BF16Vec8;
+};
 
-template <typename T> void storeFP32(float v, T *ptr) { *ptr = v; }
+template <typename T>
+void storeFP32(float v, T* ptr) {
+  *ptr = v;
+}
 
-inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) {
+inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) {
   acc = acc + a * b;
 }
 
-template <> inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) {
-  c10::BFloat16 __attribute__((__may_alias__)) *v_ptr =
-      reinterpret_cast<c10::BFloat16 *>(&v);
+template <>
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
+  c10::BFloat16 __attribute__((__may_alias__))* v_ptr =
+      reinterpret_cast<c10::BFloat16*>(&v);
   *ptr = *(v_ptr + 1);
 }
 
 #ifndef __VEC_CLASS_FP_NAN
-#define __VEC_CLASS_FP_NAN (1 << 6)
+  #define __VEC_CLASS_FP_NAN (1 << 6)
 #endif
 
-const static __vector unsigned char omask = { 0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29 };
+const static __vector unsigned char omask = {0,  1,  4,  5,  8,  9,  12, 13,
+                                             16, 17, 20, 21, 24, 25, 28, 29};
 #ifndef _ARCH_PWR10
-const static __vector unsigned int bias = { 0x00007fff, 0x00007fff, 0x00007fff, 0x00007fff };
-const static __vector unsigned int nan  = { 0x7fc00000, 0x7fc00000, 0x7fc00000, 0x7fc00000 };
-const static __vector unsigned int sh16 = { 16, 16, 16, 16 };
-const static __vector unsigned int one  = { 1, 1, 1, 1 };
+const static __vector unsigned int bias = {0x00007fff, 0x00007fff, 0x00007fff,
+                                           0x00007fff};
+const static __vector unsigned int nan = {0x7fc00000, 0x7fc00000, 0x7fc00000,
+                                          0x7fc00000};
+const static __vector unsigned int sh16 = {16, 16, 16, 16};
+const static __vector unsigned int one = {1, 1, 1, 1};
 #endif
 
-inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) {
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v) {
 #ifdef _ARCH_PWR10
   __vector signed short ret[2];
-  ret[0] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[0]);
-  ret[1] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[1]);
+  ret[0] = (__vector signed short)__builtin_vsx_xvcvspbf16(
+      (__vector unsigned char)v.reg.val[0]);
+  ret[1] = (__vector signed short)__builtin_vsx_xvcvspbf16(
+      (__vector unsigned char)v.reg.val[1]);
   reg = vec_perm(ret[0], ret[1], omask);
 #elif defined(_ARCH_PWR9)
   __vector unsigned int inp0 = (__vector unsigned int)(v.reg.val[0]);
@@ -425,8 +447,10 @@ inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) {
   __vector unsigned int rnd1 = vec_add(lsb1, bias);
   inp0 = vec_add(inp0, rnd0);
   inp1 = vec_add(inp1, rnd1);
-  __vector __bool int sel0 = vec_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN);
-  __vector __bool int sel1 = vec_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN);
+  __vector __bool int sel0 =
+      vec_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN);
+  __vector __bool int sel1 =
+      vec_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN);
   inp0 = vec_sel(inp0, nan, sel0);
   inp1 = vec_sel(inp1, nan, sel1);
   inp0 = vec_sr(inp0, sh16);
@@ -435,13 +459,17 @@ inline BF16Vec8::BF16Vec8(const FP32Vec8 &v) {
 #endif
 }
 
-inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) {
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
 #ifdef _ARCH_PWR10
   __vector signed short ret[4];
-  ret[0] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[0]);
-  ret[1] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[1]);
-  ret[2] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[2]);
-  ret[3] = (__vector signed short)__builtin_vsx_xvcvspbf16((__vector unsigned char)v.reg.val[3]);
+  ret[0] = (__vector signed short)__builtin_vsx_xvcvspbf16(
+      (__vector unsigned char)v.reg.val[0]);
+  ret[1] = (__vector signed short)__builtin_vsx_xvcvspbf16(
+      (__vector unsigned char)v.reg.val[1]);
+  ret[2] = (__vector signed short)__builtin_vsx_xvcvspbf16(
+      (__vector unsigned char)v.reg.val[2]);
+  ret[3] = (__vector signed short)__builtin_vsx_xvcvspbf16(
+      (__vector unsigned char)v.reg.val[3]);
   reg.val[0] = vec_perm(ret[0], ret[1], omask);
   reg.val[1] = vec_perm(ret[2], ret[3], omask);
 #elif defined(_ARCH_PWR9)
@@ -465,10 +493,14 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) {
   inp1 = vec_add(inp1, rnd1);
   inp2 = vec_add(inp2, rnd2);
   inp3 = vec_add(inp3, rnd3);
-  __vector __bool int sel0 = vec_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN);
-  __vector __bool int sel1 = vec_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN);
-  __vector __bool int sel2 = vec_test_data_class(v.reg.val[2], __VEC_CLASS_FP_NAN);
-  __vector __bool int sel3 = vec_test_data_class(v.reg.val[3], __VEC_CLASS_FP_NAN);
+  __vector __bool int sel0 =
+      vec_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN);
+  __vector __bool int sel1 =
+      vec_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN);
+  __vector __bool int sel2 =
+      vec_test_data_class(v.reg.val[2], __VEC_CLASS_FP_NAN);
+  __vector __bool int sel3 =
+      vec_test_data_class(v.reg.val[3], __VEC_CLASS_FP_NAN);
   inp0 = vec_sel(inp0, nan, sel0);
   inp1 = vec_sel(inp1, nan, sel1);
   inp2 = vec_sel(inp2, nan, sel2);
@@ -482,10 +514,10 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) {
 #endif
 }
 
-inline void prefetch(const void *addr) {
+inline void prefetch(const void* addr) {
   __asm__ __volatile__("dcbt 0, %0" : : "r"(addr) : "memory");
 }
 
-}; // namespace vec_op
+};  // namespace vec_op
 
 #endif

csrc/cpu/cpu_types_x86.hpp

@@ -11,39 +11,40 @@ static_assert(false, "AVX2 must be supported for the current implementation.");
 
 namespace vec_op {
 
-#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)                                 \
-  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)                         \
-  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)                      \
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)            \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)    \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
   AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
 
-#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...)                          \
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
   AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
 
 #ifndef CPU_OP_GUARD
-#define CPU_KERNEL_GUARD_IN(NAME)
-#define CPU_KERNEL_GUARD_OUT(NAME)
+  #define CPU_KERNEL_GUARD_IN(NAME)
+  #define CPU_KERNEL_GUARD_OUT(NAME)
 #else
-#define CPU_KERNEL_GUARD_IN(NAME)                                              \
-  RECORD_FUNCTION(#NAME, c10::ArrayRef<c10::IValue>({}));
-#define CPU_KERNEL_GUARD_OUT(NAME)
+  #define CPU_KERNEL_GUARD_IN(NAME) \
+    RECORD_FUNCTION(#NAME, c10::ArrayRef<c10::IValue>({}));
+  #define CPU_KERNEL_GUARD_OUT(NAME)
 #endif
 
 #define FORCE_INLINE __attribute__((always_inline)) inline
 
 namespace {
 template <typename T, T... indexes, typename F>
-constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F &&f) {
+constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
   (f(std::integral_constant<T, indexes>{}), ...);
 }
-}; // namespace
+};  // namespace
 
 template <typename T, T count, typename F,
           typename = std::enable_if_t<std::is_invocable_v<F, T>>>
-constexpr void unroll_loop(F &&f) {
+constexpr void unroll_loop(F&& f) {
   unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
 }
 
-template <typename T> struct Vec {
+template <typename T>
+struct Vec {
   constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; }
 };
 
@@ -55,12 +56,12 @@ struct FP16Vec8 : public Vec<FP16Vec8> {
 
   __m128i reg;
 
-  explicit FP16Vec8(const void *ptr)
-      : reg((__m128i)_mm_loadu_si128((__m128i *)ptr)) {}
+  explicit FP16Vec8(const void* ptr)
+      : reg((__m128i)_mm_loadu_si128((__m128i*)ptr)) {}
 
-  explicit FP16Vec8(const FP32Vec8 &);
+  explicit FP16Vec8(const FP32Vec8&);
 
-  void save(void *ptr) const { *reinterpret_cast<__m128i *>(ptr) = reg; }
+  void save(void* ptr) const { *reinterpret_cast<__m128i*>(ptr) = reg; }
 };
 
 struct FP16Vec16 : public Vec<FP16Vec16> {
@@ -68,12 +69,12 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
 
   __m256i reg;
 
-  explicit FP16Vec16(const void *ptr)
-      : reg((__m256i)_mm256_loadu_si256((__m256i *)ptr)) {}
+  explicit FP16Vec16(const void* ptr)
+      : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
-  explicit FP16Vec16(const FP32Vec16 &);
+  explicit FP16Vec16(const FP32Vec16&);
 
-  void save(void *ptr) const { *reinterpret_cast<__m256i *>(ptr) = reg; }
+  void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
 
   void save(void* ptr, const int elem_num) const {
     constexpr uint32_t M = 0xFFFFFFFF;
@@ -87,12 +88,12 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
 
   __m128i reg;
 
-  explicit BF16Vec8(const void *ptr)
-      : reg((__m128i)_mm_loadu_si128((__m128i *)ptr)) {}
+  explicit BF16Vec8(const void* ptr)
+      : reg((__m128i)_mm_loadu_si128((__m128i*)ptr)) {}
 
-  explicit BF16Vec8(const FP32Vec8 &);
+  explicit BF16Vec8(const FP32Vec8&);
 
-  void save(void *ptr) const { *reinterpret_cast<__m128i *>(ptr) = reg; }
+  void save(void* ptr) const { *reinterpret_cast<__m128i*>(ptr) = reg; }
 };
 
 struct BF16Vec16 : public Vec<BF16Vec16> {
@@ -100,12 +101,12 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
 
   __m256i reg;
 
-  explicit BF16Vec16(const void *ptr)
-      : reg((__m256i)_mm256_loadu_si256((__m256i *)ptr)) {}
+  explicit BF16Vec16(const void* ptr)
+      : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
-  explicit BF16Vec16(const FP32Vec16 &);
+  explicit BF16Vec16(const FP32Vec16&);
 
-  void save(void *ptr) const { *reinterpret_cast<__m256i *>(ptr) = reg; }
+  void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
 
   void save(void* ptr, const int elem_num) const {
     constexpr uint32_t M = 0xFFFFFFFF;
@@ -120,11 +121,11 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
 
   __m512i reg;
 
-  explicit BF16Vec32(const void *ptr) : reg((__m512i)_mm512_loadu_si512(ptr)) {}
+  explicit BF16Vec32(const void* ptr) : reg((__m512i)_mm512_loadu_si512(ptr)) {}
 
   explicit BF16Vec32(__m512i data) : reg(data) {}
 
-  explicit BF16Vec32(BF16Vec8 &vec8_data)
+  explicit BF16Vec32(BF16Vec8& vec8_data)
       : reg((__m512i)_mm512_inserti32x4(
             _mm512_inserti32x4(_mm512_inserti32x4(_mm512_castsi128_si512(
                                                       (__m128i)vec8_data.reg),
@@ -132,7 +133,7 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
                                (__m128i)vec8_data.reg, 2),
             (__m128i)vec8_data.reg, 3)) {}
 
-  void save(void *ptr) const { *reinterpret_cast<__m512i *>(ptr) = reg; }
+  void save(void* ptr) const { *reinterpret_cast<__m512i*>(ptr) = reg; }
 };
 #else
 struct BF16Vec32 : public Vec<BF16Vec32> {
@@ -141,24 +142,24 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
   __m256i reg_low;
   __m256i reg_high;
 
-  explicit BF16Vec32(const void *ptr)
-      : reg_low(_mm256_loadu_si256((__m256i const *)ptr)),
-        reg_high(_mm256_loadu_si256((__m256i const *)ptr + 1)) {}
+  explicit BF16Vec32(const void* ptr)
+      : reg_low(_mm256_loadu_si256((__m256i const*)ptr)),
+        reg_high(_mm256_loadu_si256((__m256i const*)ptr + 1)) {}
 
-  explicit BF16Vec32(__m256i low, __m256i high) : reg_low(low),
-                                                  reg_high(high) {}
+  explicit BF16Vec32(__m256i low, __m256i high)
+      : reg_low(low), reg_high(high) {}
 
-  explicit BF16Vec32(BF16Vec8 &vec8_data)
+  explicit BF16Vec32(BF16Vec8& vec8_data)
       : reg_low((__m256i)_mm256_inserti32x4(
-                _mm256_castsi128_si256((__m128i)vec8_data.reg),
-                                       (__m128i)vec8_data.reg, 1)),
+            _mm256_castsi128_si256((__m128i)vec8_data.reg),
+            (__m128i)vec8_data.reg, 1)),
         reg_high((__m256i)_mm256_inserti32x4(
-                _mm256_castsi128_si256((__m128i)vec8_data.reg),
-                                       (__m128i)vec8_data.reg, 1)) {}
+            _mm256_castsi128_si256((__m128i)vec8_data.reg),
+            (__m128i)vec8_data.reg, 1)) {}
 
-  void save(void *ptr) const {
-    *reinterpret_cast<__m256i *>(ptr) = reg_low;
-    *reinterpret_cast<__m256i *>((__m256i *)ptr + 1) = reg_high;
+  void save(void* ptr) const {
+    *reinterpret_cast<__m256i*>(ptr) = reg_low;
+    *reinterpret_cast<__m256i*>((__m256i*)ptr + 1) = reg_high;
   }
 };
 #endif
@@ -176,11 +177,11 @@ struct FP32Vec4 : public Vec<FP32Vec4> {
 
   explicit FP32Vec4() : reg(_mm_set1_ps(0.0)) {}
 
-  explicit FP32Vec4(const float *ptr) : reg(_mm_loadu_ps(ptr)) {}
+  explicit FP32Vec4(const float* ptr) : reg(_mm_loadu_ps(ptr)) {}
 
   explicit FP32Vec4(__m128 data) : reg(data) {}
 
-  explicit FP32Vec4(const FP32Vec4 &data) : reg(data.reg) {}
+  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {}
 };
 
 struct FP32Vec8 : public Vec<FP32Vec8> {
@@ -196,15 +197,15 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
 
   explicit FP32Vec8() : reg(_mm256_set1_ps(0.0)) {}
 
-  explicit FP32Vec8(const float *ptr) : reg(_mm256_loadu_ps(ptr)) {}
+  explicit FP32Vec8(const float* ptr) : reg(_mm256_loadu_ps(ptr)) {}
 
   explicit FP32Vec8(__m256 data) : reg(data) {}
 
-  explicit FP32Vec8(const FP32Vec8 &data) : reg(data.reg) {}
+  explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {}
 
-  explicit FP32Vec8(const FP16Vec8 &v) : reg(_mm256_cvtph_ps(v.reg)) {}
+  explicit FP32Vec8(const FP16Vec8& v) : reg(_mm256_cvtph_ps(v.reg)) {}
 
-  explicit FP32Vec8(const BF16Vec8 &v)
+  explicit FP32Vec8(const BF16Vec8& v)
       : reg(_mm256_castsi256_ps(
             _mm256_bslli_epi128(_mm256_cvtepu16_epi32(v.reg), 2))) {}
 
@@ -212,7 +213,8 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
     AliasReg ar;
     ar.reg = reg;
     float result = 0;
-    unroll_loop<int, VEC_ELEM_NUM>([&result, &ar](int i) { result += ar.values[i]; });
+    unroll_loop<int, VEC_ELEM_NUM>(
+        [&result, &ar](int i) { result += ar.values[i]; });
 
     return result;
   }
@@ -244,27 +246,27 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
                                   erf(ar.values[1]), erf(ar.values[0])));
   }
 
-  FP32Vec8 operator*(const FP32Vec8 &b) const {
+  FP32Vec8 operator*(const FP32Vec8& b) const {
     return FP32Vec8(_mm256_mul_ps(reg, b.reg));
   }
 
-  FP32Vec8 operator+(const FP32Vec8 &b) const {
+  FP32Vec8 operator+(const FP32Vec8& b) const {
     return FP32Vec8(_mm256_add_ps(reg, b.reg));
   }
 
-  FP32Vec8 operator-(const FP32Vec8 &b) const {
+  FP32Vec8 operator-(const FP32Vec8& b) const {
     return FP32Vec8(_mm256_sub_ps(reg, b.reg));
   }
 
-  FP32Vec8 operator/(const FP32Vec8 &b) const {
+  FP32Vec8 operator/(const FP32Vec8& b) const {
     return FP32Vec8(_mm256_div_ps(reg, b.reg));
   }
 
-  void save(float *ptr) const { _mm256_storeu_ps(ptr, reg); }
+  void save(float* ptr) const { _mm256_storeu_ps(ptr, reg); }
 };
 
 #ifdef __AVX512F__
-struct INT32Vec16: public Vec<INT32Vec16> {
+struct INT32Vec16 : public Vec<INT32Vec16> {
   constexpr static int VEC_ELEM_NUM = 16;
   union AliasReg {
     __m512i reg;
@@ -272,12 +274,11 @@ struct INT32Vec16: public Vec<INT32Vec16> {
   };
 
   __m512i reg;
-  
-  explicit INT32Vec16(const void* data_ptr) : reg(_mm512_loadu_epi32(data_ptr)) {}
 
-  void save(int32_t* ptr) const {
-    _mm512_storeu_epi32(ptr, reg);
-  }
+  explicit INT32Vec16(const void* data_ptr)
+      : reg(_mm512_loadu_epi32(data_ptr)) {}
+
+  void save(int32_t* ptr) const { _mm512_storeu_epi32(ptr, reg); }
 
   void save(int32_t* ptr, const int elem_num) const {
     constexpr uint32_t M = 0xFFFFFFFF;
@@ -301,11 +302,11 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   explicit FP32Vec16() : reg(_mm512_set1_ps(0.0)) {}
 
-  explicit FP32Vec16(const float *ptr) : reg(_mm512_loadu_ps(ptr)) {}
+  explicit FP32Vec16(const float* ptr) : reg(_mm512_loadu_ps(ptr)) {}
 
   explicit FP32Vec16(__m512 data) : reg(data) {}
 
-  explicit FP32Vec16(const FP32Vec4 &data)
+  explicit FP32Vec16(const FP32Vec4& data)
       : reg((__m512)_mm512_inserti32x4(
             _mm512_inserti32x4(
                 _mm512_inserti32x4(_mm512_castsi128_si512((__m128i)data.reg),
@@ -313,36 +314,37 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
                 (__m128i)data.reg, 2),
             (__m128i)data.reg, 3)) {}
 
-  explicit FP32Vec16(const FP32Vec8 &data)
+  explicit FP32Vec16(const FP32Vec8& data)
       : reg((__m512)_mm512_inserti32x8(
             _mm512_castsi256_si512((__m256i)data.reg), (__m256i)data.reg, 1)) {}
 
-  explicit FP32Vec16(const BF16Vec16 &v)
+  explicit FP32Vec16(const BF16Vec16& v)
       : reg(_mm512_castsi512_ps(
             _mm512_bslli_epi128(_mm512_cvtepu16_epi32(v.reg), 2))) {}
 
-  explicit FP32Vec16(const FP16Vec16 &v) : reg(_mm512_cvtph_ps(v.reg)) {}
+  explicit FP32Vec16(const FP16Vec16& v) : reg(_mm512_cvtph_ps(v.reg)) {}
 
-  explicit FP32Vec16(const FP16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}
+  explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
 
-  explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}
+  explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
 
-  explicit FP32Vec16(const INT32Vec16 &v)
-      : reg(_mm512_cvt_roundepi32_ps(v.reg, _MM_FROUND_TO_NEAREST_INT |_MM_FROUND_NO_EXC)) {}
+  explicit FP32Vec16(const INT32Vec16& v)
+      : reg(_mm512_cvt_roundepi32_ps(
+            v.reg, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)) {}
 
-  FP32Vec16 operator*(const FP32Vec16 &b) const {
+  FP32Vec16 operator*(const FP32Vec16& b) const {
     return FP32Vec16(_mm512_mul_ps(reg, b.reg));
   }
 
-  FP32Vec16 operator+(const FP32Vec16 &b) const {
+  FP32Vec16 operator+(const FP32Vec16& b) const {
     return FP32Vec16(_mm512_add_ps(reg, b.reg));
   }
 
-  FP32Vec16 operator-(const FP32Vec16 &b) const {
+  FP32Vec16 operator-(const FP32Vec16& b) const {
     return FP32Vec16(_mm512_sub_ps(reg, b.reg));
   }
 
-  FP32Vec16 operator/(const FP32Vec16 &b) const {
+  FP32Vec16 operator/(const FP32Vec16& b) const {
     return FP32Vec16(_mm512_div_ps(reg, b.reg));
   }
 
@@ -370,9 +372,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return FP32Vec16(_mm512_mask_min_ps(reg, mask, reg, b.reg));
   }
 
-  FP32Vec16 abs() const {
-    return FP32Vec16(_mm512_abs_ps(reg));
-  } 
+  FP32Vec16 abs() const { return FP32Vec16(_mm512_abs_ps(reg)); }
 
   float reduce_sum() const { return _mm512_reduce_add_ps(reg); }
 
@@ -380,14 +380,15 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   float reduce_min() const { return _mm512_reduce_min_ps(reg); }
 
-  template <int group_size> float reduce_sub_sum(int idx) {
+  template <int group_size>
+  float reduce_sub_sum(int idx) {
     static_assert(VEC_ELEM_NUM % group_size == 0);
     constexpr uint32_t base_mask = (0xFFFF >> (16 - group_size));
     __mmask16 mask = _cvtu32_mask16(base_mask << (idx * group_size));
     return _mm512_mask_reduce_add_ps(mask, reg);
   }
 
-  void save(float *ptr) const { _mm512_storeu_ps(ptr, reg); }
+  void save(float* ptr) const { _mm512_storeu_ps(ptr, reg); }
 
   void save(float* ptr, const int elem_num) const {
     constexpr uint32_t M = 0xFFFFFFFF;
@@ -407,32 +408,30 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   __m256 reg_low;
   __m256 reg_high;
 
-  explicit FP32Vec16(float v) : reg_low(_mm256_set1_ps(v)),
-                                reg_high(_mm256_set1_ps(v)) {}
+  explicit FP32Vec16(float v)
+      : reg_low(_mm256_set1_ps(v)), reg_high(_mm256_set1_ps(v)) {}
 
-  explicit FP32Vec16() : reg_low(_mm256_set1_ps(0.0)),
-                         reg_high(_mm256_set1_ps(0.0)) {}
+  explicit FP32Vec16()
+      : reg_low(_mm256_set1_ps(0.0)), reg_high(_mm256_set1_ps(0.0)) {}
 
-  explicit FP32Vec16(const float *ptr) : reg_low(_mm256_loadu_ps(ptr)),
-                                         reg_high(_mm256_loadu_ps(ptr + 8)) {}
+  explicit FP32Vec16(const float* ptr)
+      : reg_low(_mm256_loadu_ps(ptr)), reg_high(_mm256_loadu_ps(ptr + 8)) {}
 
   explicit FP32Vec16(__m256 low, __m256 high) : reg_low(low), reg_high(high) {}
 
-  explicit FP32Vec16(const FP32Vec16 &data) : reg_low(data.reg_low),
-                                              reg_high(data.reg_high) {}
+  explicit FP32Vec16(const FP32Vec16& data)
+      : reg_low(data.reg_low), reg_high(data.reg_high) {}
 
-  explicit FP32Vec16(const FP32Vec4 &data)
+  explicit FP32Vec16(const FP32Vec4& data)
       : reg_low((__m256)_mm256_inserti128_si256(
-                _mm256_castsi128_si256((__m128i)data.reg),
-                                       (__m128i)data.reg, 1)),
+            _mm256_castsi128_si256((__m128i)data.reg), (__m128i)data.reg, 1)),
         reg_high((__m256)_mm256_inserti128_si256(
-                 _mm256_castsi128_si256((__m128i)data.reg),
-                                       (__m128i)data.reg, 1)) {}
+            _mm256_castsi128_si256((__m128i)data.reg), (__m128i)data.reg, 1)) {}
 
-  explicit FP32Vec16(const FP32Vec8 &data)
+  explicit FP32Vec16(const FP32Vec8& data)
       : reg_low(data.reg), reg_high(data.reg) {}
 
-  explicit FP32Vec16(const FP16Vec16 &v) {
+  explicit FP32Vec16(const FP16Vec16& v) {
     __m128i low = _mm256_extractf128_si256(v.reg, 0);
     __m128i high = _mm256_extractf128_si256(v.reg, 1);
 
@@ -440,9 +439,9 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     reg_high = _mm256_cvtph_ps(high);
   }
 
-  explicit FP32Vec16(const FP16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}
+  explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
 
-  explicit FP32Vec16(const BF16Vec16 &v) {
+  explicit FP32Vec16(const BF16Vec16& v) {
     __m128i low = _mm256_extractf128_si256(v.reg, 0);
     __m128i high = _mm256_extractf128_si256(v.reg, 1);
 
@@ -456,24 +455,24 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     reg_high = _mm256_castsi256_ps(v_high_shifted);
   }
 
-  explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}
+  explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
 
-  FP32Vec16 operator*(const FP32Vec16 &b) const {
+  FP32Vec16 operator*(const FP32Vec16& b) const {
     return FP32Vec16(_mm256_mul_ps(reg_low, b.reg_low),
                      _mm256_mul_ps(reg_high, b.reg_high));
   }
 
-  FP32Vec16 operator+(const FP32Vec16 &b) const {
+  FP32Vec16 operator+(const FP32Vec16& b) const {
     return FP32Vec16(_mm256_add_ps(reg_low, b.reg_low),
                      _mm256_add_ps(reg_high, b.reg_high));
   }
 
-  FP32Vec16 operator-(const FP32Vec16 &b) const {
+  FP32Vec16 operator-(const FP32Vec16& b) const {
     return FP32Vec16(_mm256_sub_ps(reg_low, b.reg_low),
                      _mm256_sub_ps(reg_high, b.reg_high));
   }
 
-  FP32Vec16 operator/(const FP32Vec16 &b) const {
+  FP32Vec16 operator/(const FP32Vec16& b) const {
     return FP32Vec16(_mm256_div_ps(reg_low, b.reg_low),
                      _mm256_div_ps(reg_high, b.reg_high));
   }
@@ -484,7 +483,8 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return low.reduce_sum() + high.reduce_sum();
   }
 
-  template <int group_size> float reduce_sub_sum(int idx) {
+  template <int group_size>
+  float reduce_sub_sum(int idx) {
     float sum = 0.0;
     static_assert(VEC_ELEM_NUM % group_size == 0);
     constexpr uint32_t base_mask = (0xFFFF >> (16 - group_size));
@@ -507,7 +507,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return sum;
   }
 
-  void save(float *ptr) const {
+  void save(float* ptr) const {
     _mm256_storeu_ps(ptr, reg_low);
     _mm256_storeu_ps(ptr + 8, reg_high);
   }
@@ -515,7 +515,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 #endif
 
 #ifdef __AVX512F__
-struct INT8Vec16: public Vec<INT8Vec16> {
+struct INT8Vec16 : public Vec<INT8Vec16> {
   constexpr static int VEC_ELEM_NUM = 16;
   union AliasReg {
     __m128i reg;
@@ -523,14 +523,12 @@ struct INT8Vec16: public Vec<INT8Vec16> {
   };
 
   __m128i reg;
-  
-  explicit INT8Vec16(const FP32Vec16& vec) : reg(
-    _mm512_cvtepi32_epi8(_mm512_cvt_roundps_epi32(vec.reg, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))
-  ) {}
 
-  void save(int8_t* ptr) const {
-    _mm_storeu_epi8(ptr, reg);
-  }
+  explicit INT8Vec16(const FP32Vec16& vec)
+      : reg(_mm512_cvtepi32_epi8(_mm512_cvt_roundps_epi32(
+            vec.reg, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))) {}
+
+  void save(int8_t* ptr) const { _mm_storeu_epi8(ptr, reg); }
 
   void save(int8_t* ptr, const int elem_num) const {
     constexpr uint32_t M = 0xFFFFFFFF;
@@ -540,71 +538,92 @@ struct INT8Vec16: public Vec<INT8Vec16> {
 };
 #endif
 
-template <typename T> struct VecType { using vec_type = void; };
+template <typename T>
+struct VecType {
+  using vec_type = void;
+};
 
-template <typename T> using vec_t = typename VecType<T>::vec_type;
+template <typename T>
+using vec_t = typename VecType<T>::vec_type;
 
-template <> struct VecType<float> { using vec_type = FP32Vec8; };
+template <>
+struct VecType<float> {
+  using vec_type = FP32Vec8;
+};
 
-template <> struct VecType<c10::Half> { using vec_type = FP16Vec8; };
+template <>
+struct VecType<c10::Half> {
+  using vec_type = FP16Vec8;
+};
 
-template <> struct VecType<c10::BFloat16> { using vec_type = BF16Vec8; };
+template <>
+struct VecType<c10::BFloat16> {
+  using vec_type = BF16Vec8;
+};
 
-template <typename T> void storeFP32(float v, T *ptr) { *ptr = v; }
+template <typename T>
+void storeFP32(float v, T* ptr) {
+  *ptr = v;
+}
 
-inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) {
+inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) {
   acc = acc + a * b;
 }
 
-template <> inline void storeFP32<c10::Half>(float v, c10::Half *ptr) {
-  *reinterpret_cast<unsigned short *>(ptr) =
+template <>
+inline void storeFP32<c10::Half>(float v, c10::Half* ptr) {
+  *reinterpret_cast<unsigned short*>(ptr) =
       _cvtss_sh(v, _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC);
 }
 
-inline FP16Vec8::FP16Vec8(const FP32Vec8 &v)
+inline FP16Vec8::FP16Vec8(const FP32Vec8& v)
     : reg(_mm256_cvtps_ph(v.reg,
                           _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)) {}
 
 #ifdef __AVX512F__
-inline FP16Vec16::FP16Vec16(const FP32Vec16 &v)
+inline FP16Vec16::FP16Vec16(const FP32Vec16& v)
     : reg(_mm512_cvtps_ph(v.reg,
                           _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)) {}
 #else
-inline FP16Vec16::FP16Vec16(const FP32Vec16 &v)
-    : reg(_mm256_insertf128_si256(_mm256_castsi128_si256(FP16Vec8(FP32Vec8(v.reg_low)).reg), FP16Vec8(FP32Vec8(v.reg_low)).reg, 1)) {}
+inline FP16Vec16::FP16Vec16(const FP32Vec16& v)
+    : reg(_mm256_insertf128_si256(
+          _mm256_castsi128_si256(FP16Vec8(FP32Vec8(v.reg_low)).reg),
+          FP16Vec8(FP32Vec8(v.reg_low)).reg, 1)) {}
 #endif
 
 #ifdef __AVX512BF16__
-template <> inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) {
-  *reinterpret_cast<__bfloat16 *>(ptr) = _mm_cvtness_sbh(v);
+template <>
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
+  *reinterpret_cast<__bfloat16*>(ptr) = _mm_cvtness_sbh(v);
 }
 
-inline BF16Vec8::BF16Vec8(const FP32Vec8 &v)
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v)
     : reg((__m128i)_mm256_cvtneps_pbh(v.reg)) {}
 
-inline BF16Vec16::BF16Vec16(const FP32Vec16 &v)
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v)
     : reg((__m256i)_mm512_cvtneps_pbh(v.reg)) {}
 
-inline void fma(FP32Vec16 &acc, BF16Vec32 &a, BF16Vec32 &b) {
+inline void fma(FP32Vec16& acc, BF16Vec32& a, BF16Vec32& b) {
   acc.reg = _mm512_dpbf16_ps(acc.reg, (__m512bh)a.reg, (__m512bh)b.reg);
 }
 #else
-template <> inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) {
-  c10::BFloat16 __attribute__((__may_alias__)) *v_ptr =
-      reinterpret_cast<c10::BFloat16 *>(&v);
+template <>
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
+  c10::BFloat16 __attribute__((__may_alias__))* v_ptr =
+      reinterpret_cast<c10::BFloat16*>(&v);
   *ptr = *(v_ptr + 1);
 }
 
-#ifdef __AVX512F__
-inline BF16Vec8::BF16Vec8(const FP32Vec8 &v)
+  #ifdef __AVX512F__
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v)
     : reg(_mm256_cvtepi32_epi16(
           _mm256_bsrli_epi128(_mm256_castps_si256(v.reg), 2))) {}
 
-inline BF16Vec16::BF16Vec16(const FP32Vec16 &v)
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v)
     : reg(_mm512_cvtepi32_epi16(
           _mm512_bsrli_epi128(_mm512_castps_si512(v.reg), 2))) {}
-#else
-namespace{
+  #else
+namespace {
 __m128i FP32Vec8_to_BF16Vec8_avx2(__m256 a) {
   __m256i ai = _mm256_castps_si256(a);
   ai = _mm256_srli_epi32(ai, 16);
@@ -612,21 +631,21 @@ __m128i FP32Vec8_to_BF16Vec8_avx2(__m256 a) {
   ai = _mm256_permute4x64_epi64(ai, 0b00111001);
   return _mm256_extracti128_si256(ai, 0);
 }
-}
+}  // namespace
 
-inline BF16Vec8::BF16Vec8(const FP32Vec8 &v)
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v)
     : reg(FP32Vec8_to_BF16Vec8_avx2(v.reg)) {}
 
-inline BF16Vec16::BF16Vec16(const FP32Vec16 &v) {
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
   BF16Vec8 low = BF16Vec8(FP32Vec8(v.reg_low));
   BF16Vec8 high = BF16Vec8(FP32Vec8(v.reg_high));
   reg = _mm256_insertf128_si256(_mm256_castsi128_si256(low.reg), high.reg, 1);
 }
-#endif // __AVX512F__
-#endif // __AVX512BF16__
+  #endif  // __AVX512F__
+#endif    // __AVX512BF16__
 
-inline void prefetch(const void *addr) { _mm_prefetch(addr, _MM_HINT_T1); }
+inline void prefetch(const void* addr) { _mm_prefetch(addr, _MM_HINT_T1); }
 
-}; // namespace vec_op
+};  // namespace vec_op
 
 #endif

csrc/cutlass_extensions/common.hpp

@@ -27,8 +27,7 @@
 inline int get_cuda_max_shared_memory_per_block_opt_in(int const device) {
   int max_shared_mem_per_block_opt_in = 0;
   cudaDeviceGetAttribute(&max_shared_mem_per_block_opt_in,
-                        cudaDevAttrMaxSharedMemoryPerBlockOptin,
-                        device);
+                         cudaDevAttrMaxSharedMemoryPerBlockOptin, device);
   return max_shared_mem_per_block_opt_in;
 }
 

csrc/ops.h

@@ -86,6 +86,8 @@ void batched_rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
 
 void silu_and_mul(torch::Tensor& out, torch::Tensor& input);
 
+void mul_and_silu(torch::Tensor& out, torch::Tensor& input);
+
 void gelu_and_mul(torch::Tensor& out, torch::Tensor& input);
 
 void gelu_tanh_and_mul(torch::Tensor& out, torch::Tensor& input);

csrc/prepare_inputs/advance_step.cu

@@ -95,6 +95,16 @@ __global__ void advance_step_flashinfer_kernel(
     long* input_positions_ptr, int* seq_lens_ptr, long* slot_mapping_ptr,
     int const* block_tables_ptr, int64_t const block_tables_stride,
     int* paged_kv_last_page_len_ptr, int* block_table_bound_ptr) {
+  int const n_pad = num_seqs - num_queries;
+  if (n_pad && blockIdx.x == 0) {
+    // Handle cuda graph padding
+    int const offset = num_queries;
+    for (int i = threadIdx.x; i < n_pad; i += blockDim.x) {
+      input_tokens_ptr[offset + i] = 0;
+      input_positions_ptr[offset + i] = 0;
+      slot_mapping_ptr[offset + i] = -1;
+    }
+  }
   int num_query_blocks = div_ceil(num_queries, num_threads);
 
   if (blockIdx.x < num_query_blocks) {

csrc/torch_bindings.cpp

@@ -55,6 +55,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("silu_and_mul(Tensor! out, Tensor input) -> ()");
   ops.impl("silu_and_mul", torch::kCUDA, &silu_and_mul);
 
+  ops.def("mul_and_silu(Tensor! out, Tensor input) -> ()");
+  ops.impl("mul_and_silu", torch::kCUDA, &mul_and_silu);
+
   // Activation function used in GeGLU with `none` approximation.
   ops.def("gelu_and_mul(Tensor! out, Tensor input) -> ()");
   ops.impl("gelu_and_mul", torch::kCUDA, &gelu_and_mul);

docs/README.md

@@ -16,4 +16,5 @@ make html
 ```bash
 python -m http.server -d build/html/
 ```
+
 Launch your browser and open localhost:8000.

docs/requirements-docs.txt

@@ -3,6 +3,7 @@ sphinx-book-theme==1.0.1
 sphinx-copybutton==0.5.2
 myst-parser==3.0.1
 sphinx-argparse==0.4.0
+sphinx-design==0.6.1
 sphinx-togglebutton==0.3.2
 msgspec
 cloudpickle

docs/source/api/model/index.md

@@ -9,4 +9,3 @@ interfaces_base
 interfaces
 adapters
 ```
-

docs/source/api/multimodal/index.md

@@ -7,7 +7,7 @@ vLLM provides experimental support for multi-modal models through the {mod}`vllm
 Multi-modal inputs can be passed alongside text and token prompts to [supported models](#supported-mm-models)
 via the `multi_modal_data` field in {class}`vllm.inputs.PromptType`.
 
-Looking to add your own multi-modal model? Please follow the instructions listed [here](#enabling-multimodal-inputs).
+Looking to add your own multi-modal model? Please follow the instructions listed [here](#supports-multimodal).
 
 ## Module Contents
 

docs/source/api/multimodal/inputs.md

@@ -3,7 +3,7 @@
 ## User-facing inputs
 
 ```{eval-rst}
-.. autodata:: vllm.multimodal.MultiModalDataDict
+.. autodata:: vllm.multimodal.inputs.MultiModalDataDict
 ```
 
 ## Internal data structures

docs/source/community/sponsors.md

@@ -6,6 +6,7 @@ vLLM is a community project. Our compute resources for development and testing a
 <!-- Note: Please keep these consistent with README.md. -->
 
 Cash Donations:
+
 - a16z
 - Dropbox
 - Sequoia Capital
@@ -13,6 +14,7 @@ Cash Donations:
 - ZhenFund
 
 Compute Resources:
+
 - AMD
 - Anyscale
 - AWS

docs/source/conf.py

@@ -43,6 +43,7 @@ extensions = [
     "sphinx.ext.autosummary",
     "myst_parser",
     "sphinxarg.ext",
+    "sphinx_design",
     "sphinx_togglebutton",
 ]
 myst_enable_extensions = [
@@ -55,7 +56,7 @@ templates_path = ['_templates']
 # List of patterns, relative to source directory, that match files and
 # directories to ignore when looking for source files.
 # This pattern also affects html_static_path and html_extra_path.
-exclude_patterns: List[str] = ["**/*.template.md"]
+exclude_patterns: List[str] = ["**/*.template.md", "**/*.inc.md"]
 
 # Exclude the prompt "$" when copying code
 copybutton_prompt_text = r"\$ "

docs/source/contributing/model/basic.md

@@ -1,6 +1,6 @@
 (new-model-basic)=
 
-# Basic Implementation
+# Implementing a Basic Model
 
 This guide walks you through the steps to implement a basic vLLM model.
 
@@ -57,7 +57,17 @@ class MyModelForCausalLM(nn.Module):
 
 ### Computation Code
 
-Rewrite the {meth}`~torch.nn.Module.forward` method of your model to remove any unnecessary code, such as training-specific code. Modify the input parameters to treat `input_ids` and `positions` as flattened tensors with a single batch size dimension, without a max-sequence length dimension.
+- Add a `get_input_embeddings` method inside `MyModel` module that returns the text embeddings given `input_ids`. This is equivalent to directly calling the text embedding layer, but provides a unified interface in case `MyModel` is used within a composite multimodal model.
+
+```python
+class MyModel(nn.Module):
+        ...
+
+    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+        ... 
+```
+
+- Rewrite the {meth}`~torch.nn.Module.forward` method of your model to remove any unnecessary code, such as training-specific code. Modify the input parameters to treat `input_ids` and `positions` as flattened tensors with a single batch size dimension, without a max-sequence length dimension.
 
 ```python
 def forward(

docs/source/contributing/model/index.md

@@ -2,7 +2,7 @@
 
 # Adding a New Model
 
-This section provides more information on how to integrate a [HuggingFace Transformers](https://github.com/huggingface/transformers) model into vLLM.
+This section provides more information on how to integrate a [PyTorch](https://pytorch.org/) model into vLLM.
 
 ```{toctree}
 :caption: Contents
@@ -10,6 +10,7 @@ This section provides more information on how to integrate a [HuggingFace Transf
 
 basic
 registration
+tests
 multimodal
 ```
 

docs/source/contributing/model/multimodal.md

@@ -1,6 +1,6 @@
-(enabling-multimodal-inputs)=
+(supports-multimodal)=
 
-# Enabling Multimodal Inputs
+# Multi-Modal Support
 
 This document walks you through the steps to extend a basic model so that it accepts [multi-modal inputs](#multimodal-inputs).
 
@@ -9,7 +9,78 @@ This document walks you through the steps to extend a basic model so that it acc
 It is assumed that you have already implemented the model in vLLM according to [these steps](#new-model-basic).
 Further update the model as follows:
 
-- Implement the {class}`~vllm.model_executor.models.interfaces.SupportsMultiModal` interface.
+- Reserve a keyword parameter in {meth}`~torch.nn.Module.forward` for each input tensor that corresponds to a multi-modal input, as shown in the following example:
+
+  ```diff
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        kv_caches: List[torch.Tensor],
+        attn_metadata: AttentionMetadata,
+  +     pixel_values: torch.Tensor,
+    ) -> SamplerOutput:
+  ```
+  
+  More conveniently, you can simply pass `**kwargs` to the {meth}`~torch.nn.Module.forward` method and retrieve the keyword parameters for multimodal inputs from it.
+
+- Implement {meth}`~vllm.model_executor.models.interfaces.SupportsMultiModal.get_multimodal_embeddings` that returns the embeddings from running the multimodal inputs through the multimodal tokenizer of the model. Below we provide a boilerplate of a typical implementation pattern, but feel free to adjust it to your own needs.
+
+    ```python
+    class YourModelForImage2Seq(nn.Module):
+        ...
+
+        def _process_image_input(self, image_input: YourModelImageInputs) -> torch.Tensor:
+
+            assert self.vision_encoder is not None
+            image_features = self.vision_encoder(image_input)
+            return self.multi_modal_projector(image_features)
+
+        def get_multimodal_embeddings(self, **kwargs: object) -> Optional[NestedTensors]:
+
+            # Validate the multimodal input keyword arguments
+            image_input = self._parse_and_validate_image_input(**kwargs)
+            if image_input is None:
+                return None
+
+            # Run multimodal inputs through encoder and projector
+            vision_embeddings = self._process_image_input(image_input)
+            return vision_embeddings
+    ```
+
+    ```{important}
+    The returned `multimodal_embeddings` must be either a **3D {class}`torch.Tensor`** of shape `(num_items, feature_size, hidden_size)`, or a **list / tuple of 2D {class}`torch.Tensor`'s** of shape `(feature_size, hidden_size)`, so that `multimodal_embeddings[i]` retrieves the embeddings generated from the `i`-th multimodal data item (e.g, image) of the request.
+    ```
+
+- Implement {meth}`~vllm.model_executor.models.interfaces.SupportsMultiModal.get_input_embeddings` to merge `multimodal_embeddings` with text embeddings from the `input_ids`. If input processing for the model is implemented correctly (see sections below), then you can leverage the utility function we provide to easily merge the embeddings.
+
+    ```python
+    from .utils import merge_multimodal_embeddings
+
+    class YourModelForImage2Seq(nn.Module):
+        ...
+
+        def get_input_embeddings(
+            self,
+            input_ids: torch.Tensor,
+            multimodal_embeddings: Optional[NestedTensors] = None,
+        ) -> torch.Tensor:
+
+            # `get_input_embeddings` should already be implemented for the language 
+            # model as one of the requirements of basic vLLM model implementation.
+            inputs_embeds = self.language_model.get_input_embeddings(input_ids)
+
+            if multimodal_embeddings is not None:
+                inputs_embeds = merge_multimodal_embeddings(
+                    input_ids=input_ids, 
+                    inputs_embeds=inputs_embeds, 
+                    multimodal_embeddings=multimodal_embeddings,
+                    placeholder_token_id=self.config.image_token_index)
+
+            return inputs_embeds
+    ```
+
+- Once the above steps are done, update the model class with the {class}`~vllm.model_executor.models.interfaces.SupportsMultiModal` interface.
 
   ```diff
   + from vllm.model_executor.models.interfaces import SupportsMultiModal
@@ -23,117 +94,359 @@ Further update the model as follows:
   Check out [the HuggingFace Transformers documentation](https://huggingface.co/docs/transformers/model_doc/auto#multimodal) for some examples.
   ```
 
-- If you haven't already done so, reserve a keyword parameter in {meth}`~torch.nn.Module.forward`
-  for each input tensor that corresponds to a multi-modal input, as shown in the following example:
+## 2. Specify processing information
 
-  ```diff
-    def forward(
-        self,
-        input_ids: torch.Tensor,
-        positions: torch.Tensor,
-        kv_caches: List[torch.Tensor],
-        attn_metadata: AttentionMetadata,
-  +     pixel_values: torch.Tensor,
-    ) -> SamplerOutput:
-  ```
+Next, create a subclass of {class}`~vllm.multimodal.processing.BaseProcessingInfo`
+to provide basic information related to HF processing.
 
-## 2. Register input mappers
+### Maximum number of input items
 
-For each modality type that the model accepts as input, decorate the model class with {meth}`MULTIMODAL_REGISTRY.register_input_mapper <vllm.multimodal.MultiModalRegistry.register_input_mapper>`.
-This decorator accepts a function that maps multi-modal inputs to the keyword arguments you have previously defined in {meth}`~torch.nn.Module.forward`.
+You need to override the abstract method {meth}`~vllm.multimodal.processing.BaseProcessingInfo.get_supported_mm_limits`
+to return the maximum number of input items for each modality supported by the model.
+
+For example, if the model supports any number of images but only one video per prompt:
+
+```python
+def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
+    return {"image": None, "video": 1}
+```
+
+### Maximum number of placeholder feature tokens
+
+Also, override the abstract method {meth}`~vllm.multimodal.processing.BaseProcessingInfo.get_mm_max_tokens_per_item`
+to return the maximum number of placeholder feature tokens per input item for each modality.
+
+When calling the model, the output embeddings from the visual encoder are assigned to the input positions
+containing placeholder feature tokens. Therefore, the number of placeholder feature tokens should be equal
+to the size of the output embeddings.
+
+::::{tab-set}
+:::{tab-item} Basic example: LLaVA
+:sync: llava
+
+Looking at the code of HF's `LlavaForConditionalGeneration`:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/modeling_llava.py#L530-L544
+n_image_tokens = (input_ids == self.config.image_token_index).sum().item()
+n_image_features = image_features.shape[0] * image_features.shape[1]
+
+if n_image_tokens != n_image_features:
+    raise ValueError(
+        f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+    )
+special_image_mask = (
+    (input_ids == self.config.image_token_index)
+    .unsqueeze(-1)
+    .expand_as(inputs_embeds)
+    .to(inputs_embeds.device)
+)
+image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
+```
+
+The number of placeholder feature tokens per image is `image_features.shape[1]`.
+`image_features` is calculated inside the `get_image_features` method:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/modeling_llava.py#L290-L300
+image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
+
+selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
+if vision_feature_select_strategy == "default":
+    selected_image_feature = selected_image_feature[:, 1:]
+elif vision_feature_select_strategy == "full":
+    selected_image_feature = selected_image_feature
+else:
+    raise ValueError(f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}")
+image_features = self.multi_modal_projector(selected_image_feature)
+return image_features
+```
+
+We can infer that `image_features.shape[1]` is based on `image_outputs.hidden_states.shape[1]` from the vision tower
+(`CLIPVisionModel` for the [`llava-hf/llava-1.5-7b-hf`](https://huggingface.co/llava-hf/llava-1.5-7b-hf) model).
+Moreover, we only need the sequence length (the second dimension of the tensor) to get `image_features.shape[1]`.
+The sequence length is determined by the initial hidden states in `CLIPVisionTransformer` since the attention
+mechanism doesn't change the sequence length of the output hidden states.
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/clip/modeling_clip.py#L1094-L1102
+hidden_states = self.embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)
+hidden_states = self.pre_layrnorm(hidden_states)
+
+encoder_outputs = self.encoder(
+    inputs_embeds=hidden_states,
+    output_attentions=output_attentions,
+    output_hidden_states=output_hidden_states,
+    return_dict=return_dict,
+)
+```
+
+To find the sequence length, we turn to the code of `CLIPVisionEmbeddings`:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/clip/modeling_clip.py#L247-L257
+target_dtype = self.patch_embedding.weight.dtype
+patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype))  # shape = [*, width, grid, grid]
+patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+class_embeds = self.class_embedding.expand(batch_size, 1, -1)
+embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+if interpolate_pos_encoding:
+    embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
+else:
+    embeddings = embeddings + self.position_embedding(self.position_ids)
+return embeddings
+```
+
+We can infer that `embeddings.shape[1] == self.num_positions`, where
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/clip/modeling_clip.py#L195-L196
+self.num_patches = (self.image_size // self.patch_size) ** 2
+self.num_positions = self.num_patches + 1
+```
+
+Overall, the number of placeholder feature tokens for an image can be calculated as:
+
+```python
+def get_num_image_tokens(
+    self,
+    *,
+    image_width: int,
+    image_height: int,
+) -> int:
+    hf_config = self.get_hf_config()
+    hf_processor = self.get_hf_processor()
+
+    image_size = hf_config.vision_config.image_size
+    patch_size = hf_config.vision_config.patch_size
+
+    num_image_tokens = (image_size // patch_size) ** 2 + 1
+    if hf_processor.vision_feature_select_strategy == "default":
+        num_image_tokens -= 1
+
+    return num_image_tokens
+```
+
+Notice that the number of image tokens doesn't depend on the image width and height.
+So, we can calculate the maximum number of image tokens using any image size:
+
+```python
+def get_image_size_with_most_features(self) -> ImageSize:
+    hf_config = self.get_hf_config()
+    width = height = hf_config.image_size
+    return ImageSize(width=width, height=height)
+
+def get_max_image_tokens(self) -> int:
+    target_width, target_height = self.get_image_size_with_most_features()
+
+    return self.get_num_image_tokens(
+        image_width=target_width,
+        image_height=target_height,
+    )
+```
+
+And thus, we can override the method as:
+
+```python
+def get_mm_max_tokens_per_item(self, seq_len: int) -> Mapping[str, int]:
+    return {"image": self.get_max_image_tokens()}
+```
+
+```{note}
+Our [actual code](gh-file:vllm/model_executor/models/llava.py) is more abstracted to support vision encoders other than CLIP.
+```
+
+:::
+::::
+
+## 3. Specify dummy inputs
+
+Then, inherit {class}`~vllm.multimodal.profiling.BaseDummyInputsBuilder` to construct dummy inputs for
+HF processing as well as memory profiling.
+
+### For memory profiling
+
+Override the abstract method {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_processor_inputs`
+to construct dummy inputs for memory profiling. This dummy input should result in the worst-case memory usage of
+the model so that vLLM can reserve the correct amount of memory for it.
+
+Assuming that the memory usage increases with the number of tokens, the dummy input can be constructed based
+on the code for {meth}`~vllm.multimodal.processing.BaseProcessingInfo.get_mm_max_tokens_per_item`.
+
+::::{tab-set}
+:::{tab-item} Basic example: LLaVA
+:sync: llava
+Making use of the `get_image_size_with_most_features` method implemented in the previous section:
+
+```python
+def get_dummy_processor_inputs(
+    self,
+    seq_len: int,
+    mm_counts: Mapping[str, int],
+) -> ProcessorInputs:
+    num_images = mm_counts.get("image", 0)
+
+    processor = self.info.get_hf_processor()
+    image_token = processor.image_token
+  
+    hf_config = self.get_hf_config()
+    target_width, target_height = self.info.get_image_size_with_most_features()
+
+    mm_data = {
+        "image":
+        self._get_dummy_images(width=target_width,
+                               height=target_height,
+                               num_images=num_images)
+    }
+
+    return ProcessorInputs(
+        prompt_text=image_token * num_images,
+        mm_data=mm_data,
+    )
+```
+
+:::
+::::
+
+## 4. Specify processing details
+
+Afterwards, create a subclass of {class}`~vllm.multimodal.processing.BaseMultiModalProcessor`
+to fill in the missing details about HF processing.
+
+```{seealso}
+[Multi-Modal Data Processing](#mm-processing)
+```
+
+### Multi-modal fields
+
+Override {class}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_mm_fields_config` to
+return a schema of the tensors outputted by the HF processor that are related to the input multi-modal items.
+
+::::{tab-set}
+:::{tab-item} Basic example: LLaVA
+:sync: llava
+
+Looking at the model's `forward` method:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/modeling_llava.py#L387-L404
+def forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    pixel_values: torch.FloatTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    vision_feature_layer: Optional[int] = None,
+    vision_feature_select_strategy: Optional[str] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    num_logits_to_keep: int = 0,
+) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
+```
+
+The only related keyword argument is `pixel_values` which directly corresponds to input images.
+The shape of `pixel_values` is `(N, C, H, W)` where `N` is the number of images.
+So, we override the method as follows:
+
+```python
+def _get_mm_fields_config(
+    self,
+    hf_inputs: BatchFeature,
+    hf_processor_mm_kwargs: Mapping[str, object],
+) -> Mapping[str, MultiModalFieldConfig]:
+    return dict(
+        pixel_values=MultiModalFieldConfig.batched("image"),
+    )
+```
+
+```{note}
+Our [actual code](gh-file:vllm/model_executor/models/llava.py) additionally supports
+pre-computed image embeddings, which can be passed to be model via the `image_embeds` argument.
+```
+
+:::
+::::
+
+### Prompt replacements
+
+Override {class}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements` to
+return a list of {class}`~vllm.multimodal.processing.PromptReplacement` instances.
+
+Each {class}`~vllm.multimodal.processing.PromptReplacement` instance specifies a find-and-replace
+operation performed by the HF processor.
+
+::::{tab-set}
+:::{tab-item} Basic example: LLaVA
+:sync: llava
+
+Looking at HF's `LlavaProcessor`:
+
+```python
+# https://github.com/huggingface/transformers/blob/v4.47.1/src/transformers/models/llava/processing_llava.py#L167-L170
+prompt_strings = []
+for sample in text:
+    sample = sample.replace(self.image_token, self.image_token * num_image_tokens)
+    prompt_strings.append(sample)
+```
+
+It simply repeats each input `image_token` a number of times equal to the number of placeholder feature tokens (`num_image_tokens`).
+Based on this, we override the method as follows:
+
+```python
+def _get_prompt_replacements(
+    self,
+    mm_items: MultiModalDataItems,
+    hf_processor_mm_kwargs: Mapping[str, object],
+    out_mm_kwargs: MultiModalKwargs,
+) -> list[PromptReplacement]:
+    hf_config = self.info.get_hf_config()
+    image_token_id = hf_config.image_token_index
+
+    def get_replacement(item_idx: int):
+        images = mm_items.get_items("image", ImageProcessorItems)
+
+        image_size = images.get_image_size(item_idx)
+        num_image_tokens = self.info.get_num_image_tokens(
+            image_width=image_size.width,
+            image_height=image_size.height,
+        )
+
+        return [image_token_id] * num_image_tokens
+
+    return [
+        PromptReplacement(
+            modality="image",
+            target=[image_token_id],
+            replacement=get_replacement,
+        ),
+    ]
+```
+
+:::
+::::
+
+## 5. Register processor-related classes
+
+After you have defined {class}`~vllm.multimodal.processing.BaseProcessingInfo` (Step 2),
+{class}`~vllm.multimodal.profiling.BaseDummyInputsBuilder` (Step 3),
+and {class}`~vllm.multimodal.processing.BaseMultiModalProcessor` (Step 4),
+decorate the model class with {meth}`MULTIMODAL_REGISTRY.register_processor <vllm.multimodal.registry.MultiModalRegistry.register_processor>`
+to register them to the multi-modal registry:
 
 ```diff
   from vllm.model_executor.models.interfaces import SupportsMultiModal
 + from vllm.multimodal import MULTIMODAL_REGISTRY
 
-+ @MULTIMODAL_REGISTRY.register_image_input_mapper()
++ @MULTIMODAL_REGISTRY.register_processor(YourMultiModalProcessor,
++                                         info=YourProcessingInfo,
++                                         dummy_inputs=YourDummyInputsBuilder)
   class YourModelForImage2Seq(nn.Module, SupportsMultiModal):
 ```
-
-A default mapper is available for each modality in the core vLLM library. This input mapper will be used if you do not provide your own function.
-
-```{seealso}
-[Input Processing Pipeline](#input-processing-pipeline)
-```
-
-## 3. Register maximum number of multi-modal tokens
-
-For each modality type that the model accepts as input, calculate the maximum possible number of tokens per data item
-and register it via {meth}`INPUT_REGISTRY.register_dummy_data <vllm.inputs.registry.InputRegistry.register_max_multimodal_tokens>`.
-
-```diff
-  from vllm.inputs import INPUT_REGISTRY
-  from vllm.model_executor.models.interfaces import SupportsMultiModal
-  from vllm.multimodal import MULTIMODAL_REGISTRY
-
-  @MULTIMODAL_REGISTRY.register_image_input_mapper()
-+ @MULTIMODAL_REGISTRY.register_max_image_tokens(<your_calculation>)
-  @INPUT_REGISTRY.register_dummy_data(<your_dummy_data_factory>)
-  class YourModelForImage2Seq(nn.Module, SupportsMultiModal):
-```
-
-Here are some examples:
-
-- Image inputs (static feature size): [LLaVA-1.5 Model](gh-file:vllm/model_executor/models/llava.py)
-- Image inputs (dynamic feature size): [LLaVA-NeXT Model](gh-file:vllm/model_executor/models/llava_next.py)
-
-```{seealso}
-[Input Processing Pipeline](#input-processing-pipeline)
-```
-
-## 4. (Optional) Register dummy data
-
-During startup, dummy data is passed to the vLLM model to allocate memory. This only consists of text input by default, which may not be applicable to multi-modal models.
-In such cases, you can define your own dummy data by registering a factory method via {meth}`INPUT_REGISTRY.register_dummy_data <vllm.inputs.registry.InputRegistry.register_dummy_data>`.
-
-```diff
-  from vllm.inputs import INPUT_REGISTRY
-  from vllm.model_executor.models.interfaces import SupportsMultiModal
-  from vllm.multimodal import MULTIMODAL_REGISTRY
-
-  @MULTIMODAL_REGISTRY.register_image_input_mapper()
-  @MULTIMODAL_REGISTRY.register_max_image_tokens(<your_calculation>)
-+ @INPUT_REGISTRY.register_dummy_data(<your_dummy_data_factory>)
-  class YourModelForImage2Seq(nn.Module, SupportsMultiModal):
-```
-
-```{note}
-The dummy data should have the maximum possible number of multi-modal tokens, as described in the previous step.
-```
-
-Here are some examples:
-
-- Image inputs (static feature size): [LLaVA-1.5 Model](gh-file:vllm/model_executor/models/llava.py)
-- Image inputs (dynamic feature size): [LLaVA-NeXT Model](gh-file:vllm/model_executor/models/llava_next.py)
-
-```{seealso}
-[Input Processing Pipeline](#input-processing-pipeline)
-```
-
-## 5. (Optional) Register input processor
-
-Sometimes, there is a need to process inputs at the {class}`~vllm.LLMEngine` level before they are passed to the model executor.
-This is often due to the fact that unlike implementations in HuggingFace Transformers, the reshaping and/or expansion of multi-modal embeddings needs to take place outside model's {meth}`~torch.nn.Module.forward` call.
-You can register input processors via {meth}`INPUT_REGISTRY.register_input_processor <vllm.inputs.registry.InputRegistry.register_input_processor>`.
-
-```diff
-  from vllm.inputs import INPUT_REGISTRY
-  from vllm.model_executor.models.interfaces import SupportsMultiModal
-  from vllm.multimodal import MULTIMODAL_REGISTRY
-
-  @MULTIMODAL_REGISTRY.register_image_input_mapper()
-  @MULTIMODAL_REGISTRY.register_max_image_tokens(<your_calculation>)
-  @INPUT_REGISTRY.register_dummy_data(<your_dummy_data_factory>)
-+ @INPUT_REGISTRY.register_input_processor(<your_input_processor>)
-  class YourModelForImage2Seq(nn.Module, SupportsMultiModal):
-```
-
-A common use case of input processors is inserting placeholder tokens to leverage the vLLM framework for attention mask generation.
-Here are some examples:
-
-- Insert static number of image tokens: [LLaVA-1.5 Model](gh-file:vllm/model_executor/models/llava.py)
-- Insert dynamic number of image tokens: [LLaVA-NeXT Model](gh-file:vllm/model_executor/models/llava_next.py)
-
-```{seealso}
-[Input Processing Pipeline](#input-processing-pipeline)
-```

docs/source/contributing/model/registration.md

@@ -1,6 +1,6 @@
 (new-model-registration)=
 
-# Model Registration
+# Registering a Model to vLLM
 
 vLLM relies on a model registry to determine how to run each model.
 A list of pre-registered architectures can be found [here](#supported-models).
@@ -15,7 +15,6 @@ This gives you the ability to modify the codebase and test your model.
 
 After you have implemented your model (see [tutorial](#new-model-basic)), put it into the <gh-dir:vllm/model_executor/models> directory.
 Then, add your model class to `_VLLM_MODELS` in <gh-file:vllm/model_executor/models/registry.py> so that it is automatically registered upon importing vLLM.
-You should also include an example HuggingFace repository for this model in <gh-file:tests/models/registry.py> to run the unit tests.
 Finally, update our [list of supported models](#supported-models) to promote your model!
 
 ```{important}
@@ -48,7 +47,7 @@ ModelRegistry.register_model("YourModelForCausalLM", "your_code:YourModelForCaus
 
 ```{important}
 If your model is a multimodal model, ensure the model class implements the {class}`~vllm.model_executor.models.interfaces.SupportsMultiModal` interface.
-Read more about that [here](#enabling-multimodal-inputs).
+Read more about that [here](#supports-multimodal).
 ```
 
 ```{note}

docs/source/contributing/model/tests.md

@@ -0,0 +1,63 @@
+(new-model-tests)=
+
+# Writing Unit Tests
+
+This page explains how to write unit tests to verify the implementation of your model.
+
+## Required Tests
+
+These tests are necessary to get your PR merged into vLLM library.
+Without them, the CI for your PR will fail.
+
+### Model loading
+
+Include an example HuggingFace repository for your model in <gh-file:tests/models/registry.py>.
+This enables a unit test that loads dummy weights to ensure that the model can be initialized in vLLM.
+
+```{important}
+The list of models in each section should be maintained in alphabetical order.
+```
+
+```{tip}
+If your model requires a development version of HF Transformers, you can set
+`min_transformers_version` to skip the test in CI until the model is released.
+```
+
+## Optional Tests
+
+These tests are optional to get your PR merged into vLLM library.
+Passing these tests provides more confidence that your implementation is correct, and helps avoid future regressions.
+
+### Model correctness
+
+These tests compare the model outputs of vLLM against [HF Transformers](https://github.com/huggingface/transformers). You can add new tests under the subdirectories of <gh-dir:tests/models>.
+
+#### Generative models
+
+For [generative models](#generative-models), there are two levels of correctness tests, as defined in <gh-file:tests/models/utils.py>:
+
+- Exact correctness (`check_outputs_equal`): The text outputted by vLLM should exactly match the text outputted by HF.
+- Logprobs similarity (`check_logprobs_close`): The logprobs outputted by vLLM should be in the top-k logprobs outputted by HF, and vice versa.
+
+#### Pooling models
+
+For [pooling models](#pooling-models), we simply check the cosine similarity, as defined in <gh-file:tests/models/embedding/utils.py>.
+
+(mm-processing-tests)=
+
+### Multi-modal processing
+
+#### Common tests
+
+Adding your model to <gh-file:tests/models/multimodal/processing/test_common.py> verifies that the following input combinations result in the same outputs:
+
+- Text + multi-modal data
+- Tokens + multi-modal data
+- Text + cached multi-modal data
+- Tokens + cached multi-modal data
+
+#### Model-specific tests
+
+You can add a new file under <gh-dir:tests/models/multimodal/processing> to run tests that only apply to your model.
+
+For example, if the HF processor for your model accepts user-specified keyword arguments, you can verify that the keyword arguments are being applied correctly, such as in <gh-file:tests/models/multimodal/processing/test_phi3v.py>.

docs/source/contributing/overview.md

@@ -25,10 +25,12 @@ Check out the [building from source](#build-from-source) documentation for detai
 ```bash
 pip install -r requirements-dev.txt
 
-# linting and formatting
-bash format.sh
-# Static type checking
-mypy
+# Linting, formatting and static type checking
+pre-commit install
+
+# You can manually run pre-commit with
+pre-commit run --all-files
+
 # Unit tests
 pytest tests/
 ```
@@ -37,8 +39,6 @@ pytest tests/
 Currently, the repository is not fully checked by `mypy`.
 ```
 
-# Contribution Guidelines
-
 ## Issues
 
 If you encounter a bug or have a feature request, please [search existing issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue) first to see if it has already been reported. If not, please [file a new issue](https://github.com/vllm-project/vllm/issues/new/choose), providing as much relevant information as possible.
@@ -90,7 +90,8 @@ If the PR spans more than one category, please include all relevant prefixes.
 The PR needs to meet the following code quality standards:
 
 - We adhere to [Google Python style guide](https://google.github.io/styleguide/pyguide.html) and [Google C++ style guide](https://google.github.io/styleguide/cppguide.html).
-- Pass all linter checks. Please use <gh-file:format.sh> to format your code.
+- Pass all linter checks. Please use `pre-commit` to format your code. See
+  <https://pre-commit.com/#usage> if `pre-commit` is new to you.
 - The code needs to be well-documented to ensure future contributors can easily
   understand the code.
 - Include sufficient tests to ensure the project stays correct and robust. This

docs/source/contributing/profiling/profiling_index.md

@@ -26,7 +26,7 @@ Set the env variable VLLM_RPC_TIMEOUT to a big number before you start the serve
 
 ### Offline Inference
 
-Refer to <gh-file:examples/offline_inference/offline_inference_with_profiler.py> for an example.
+Refer to <gh-file:examples/offline_inference/simple_profiling.py> for an example.
 
 ### OpenAI Server
 

docs/source/deployment/docker.md

@@ -2,6 +2,8 @@
 
 # Using Docker
 
+(deployment-docker-pre-built-image)=
+
 ## Use vLLM's Official Docker Image
 
 vLLM offers an official Docker image for deployment.
@@ -17,25 +19,32 @@ $ docker run --runtime nvidia --gpus all \
     --model mistralai/Mistral-7B-v0.1
 ```
 
+You can add any other <project:#engine-args> you need after the image tag (`vllm/vllm-openai:latest`).
+
 ```{note}
 You can either use the `ipc=host` flag or `--shm-size` flag to allow the
 container to access the host's shared memory. vLLM uses PyTorch, which uses shared
 memory to share data between processes under the hood, particularly for tensor parallel inference.
 ```
 
+(deployment-docker-build-image-from-source)=
+
 ## Building vLLM's Docker Image from Source
 
 You can build and run vLLM from source via the provided <gh-file:Dockerfile>. To build vLLM:
 
 ```console
-$ # optionally specifies: --build-arg max_jobs=8 --build-arg nvcc_threads=2
-$ DOCKER_BUILDKIT=1 docker build . --target vllm-openai --tag vllm/vllm-openai
+# optionally specifies: --build-arg max_jobs=8 --build-arg nvcc_threads=2
+DOCKER_BUILDKIT=1 docker build . --target vllm-openai --tag vllm/vllm-openai
 ```
 
 ```{note}
 By default vLLM will build for all GPU types for widest distribution. If you are just building for the
 current GPU type the machine is running on, you can add the argument `--build-arg torch_cuda_arch_list=""`
 for vLLM to find the current GPU type and build for that.
+
+If you are using Podman instead of Docker, you might need to disable SELinux labeling by
+adding `--security-opt label=disable` when running `podman build` command to avoid certain [existing issues](https://github.com/containers/buildah/discussions/4184).
 ```
 
 ## Building for Arm64/aarch64

docs/source/deployment/frameworks/cerebrium.md

@@ -13,14 +13,14 @@ vLLM can be run on a cloud based GPU machine with [Cerebrium](https://www.cerebr
 To install the Cerebrium client, run:
 
 ```console
-$ pip install cerebrium
-$ cerebrium login
+pip install cerebrium
+cerebrium login
 ```
 
 Next, create your Cerebrium project, run:
 
 ```console
-$ cerebrium init vllm-project
+cerebrium init vllm-project
 ```
 
 Next, to install the required packages, add the following to your cerebrium.toml:
@@ -58,10 +58,10 @@ def run(prompts: list[str], temperature: float = 0.8, top_p: float = 0.95):
 Then, run the following code to deploy it to the cloud:
 
 ```console
-$ cerebrium deploy
+cerebrium deploy
 ```
 
-If successful, you should be returned a CURL command that you can call inference against. Just remember to end the url with the function name you are calling (in our case` /run`)
+If successful, you should be returned a CURL command that you can call inference against. Just remember to end the url with the function name you are calling (in our case`/run`)
 
 ```python
 curl -X POST https://api.cortex.cerebrium.ai/v4/p-xxxxxx/vllm/run \

docs/source/deployment/frameworks/dstack.md

@@ -13,16 +13,16 @@ vLLM can be run on a cloud based GPU machine with [dstack](https://dstack.ai/),
 To install dstack client, run:
 
 ```console
-$ pip install "dstack[all]
-$ dstack server
+pip install "dstack[all]
+dstack server
 ```
 
 Next, to configure your dstack project, run:
 
 ```console
-$ mkdir -p vllm-dstack
-$ cd vllm-dstack
-$ dstack init
+mkdir -p vllm-dstack
+cd vllm-dstack
+dstack init
 ```
 
 Next, to provision a VM instance with LLM of your choice (`NousResearch/Llama-2-7b-chat-hf` for this example), create the following `serve.dstack.yml` file for the dstack `Service`:

docs/source/deployment/frameworks/skypilot.md

@@ -334,12 +334,12 @@ run: |
 
 1. Start the chat web UI:
 
-```console
-sky launch -c gui ./gui.yaml --env ENDPOINT=$(sky serve status --endpoint vllm)
-```
+    ```console
+    sky launch -c gui ./gui.yaml --env ENDPOINT=$(sky serve status --endpoint vllm)
+    ```
 
 2. Then, we can access the GUI at the returned gradio link:
 
-```console
-| INFO | stdout | Running on public URL: https://6141e84201ce0bb4ed.gradio.live
-```
+    ```console
+    | INFO | stdout | Running on public URL: https://6141e84201ce0bb4ed.gradio.live
+    ```

docs/source/deployment/integrations/llamastack.md

@@ -7,7 +7,7 @@ vLLM is also available via [Llama Stack](https://github.com/meta-llama/llama-sta
 To install Llama Stack, run
 
 ```console
-$ pip install llama-stack -q
+pip install llama-stack -q
 ```
 
 ## Inference using OpenAI Compatible API

docs/source/deployment/k8s.md

@@ -14,234 +14,235 @@ Before you begin, ensure that you have the following:
 
 ## Deployment Steps
 
-1. **Create a PVC , Secret and Deployment for vLLM**
+1. Create a PVC, Secret and Deployment for vLLM
 
-PVC is used to store the model cache and it is optional, you can use hostPath or other storage options
+      PVC is used to store the model cache and it is optional, you can use hostPath or other storage options
 
-```yaml
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: mistral-7b
-  namespace: default
-spec:
-  accessModes:
-  - ReadWriteOnce
-  resources:
-    requests:
-      storage: 50Gi
-  storageClassName: default
-  volumeMode: Filesystem
-```
-
-Secret is optional and only required for accessing gated models, you can skip this step if you are not using gated models
-
-```yaml
-apiVersion: v1
-kind: Secret
-metadata:
-  name: hf-token-secret
-  namespace: default
-type: Opaque
-stringData:
-  token: "REPLACE_WITH_TOKEN"
-```
-
-Next to create the deployment file for vLLM to run the model server. The following example deploys the `Mistral-7B-Instruct-v0.3` model.
-
-Here are two examples for using NVIDIA GPU and AMD GPU. 
-
-- NVIDIA GPU
-
-```yaml
-apiVersion: apps/v1
-kind: Deployment
-metadata:
-  name: mistral-7b
-  namespace: default
-  labels:
-    app: mistral-7b
-spec:
-  replicas: 1
-  selector:
-    matchLabels:
-      app: mistral-7b
-  template:
-    metadata:
-      labels:
-        app: mistral-7b
-    spec:
-      volumes:
-      - name: cache-volume
-        persistentVolumeClaim:
-          claimName: mistral-7b
-      # vLLM needs to access the host's shared memory for tensor parallel inference.
-      - name: shm
-        emptyDir:
-          medium: Memory
-          sizeLimit: "2Gi"
-      containers:
-      - name: mistral-7b
-        image: vllm/vllm-openai:latest
-        command: ["/bin/sh", "-c"]
-        args: [
-          "vllm serve mistralai/Mistral-7B-Instruct-v0.3 --trust-remote-code --enable-chunked-prefill --max_num_batched_tokens 1024"
-        ]
-        env:
-        - name: HUGGING_FACE_HUB_TOKEN
-          valueFrom:
-            secretKeyRef:
-              name: hf-token-secret
-              key: token
-        ports:
-        - containerPort: 8000
+      ```yaml
+      apiVersion: v1
+      kind: PersistentVolumeClaim
+      metadata:
+        name: mistral-7b
+        namespace: default
+      spec:
+        accessModes:
+        - ReadWriteOnce
         resources:
-          limits:
-            cpu: "10"
-            memory: 20G
-            nvidia.com/gpu: "1"
           requests:
-            cpu: "2"
-            memory: 6G
-            nvidia.com/gpu: "1"
-        volumeMounts:
-        - mountPath: /root/.cache/huggingface
-          name: cache-volume
-        - name: shm
-          mountPath: /dev/shm
-        livenessProbe:
-          httpGet:
-            path: /health
-            port: 8000
-          initialDelaySeconds: 60
-          periodSeconds: 10
-        readinessProbe:
-          httpGet:
-            path: /health
-            port: 8000
-          initialDelaySeconds: 60
-          periodSeconds: 5
-```
+            storage: 50Gi
+        storageClassName: default
+        volumeMode: Filesystem
+      ```
 
-- AMD GPU
+      Secret is optional and only required for accessing gated models, you can skip this step if you are not using gated models
 
-You can refer to the `deployment.yaml` below if using AMD ROCm GPU like MI300X.
+      ```yaml
+      apiVersion: v1
+      kind: Secret
+      metadata:
+        name: hf-token-secret
+        namespace: default
+      type: Opaque
+      stringData:
+        token: "REPLACE_WITH_TOKEN"
+      ```
 
-```yaml
-apiVersion: apps/v1
-kind: Deployment
-metadata:
-  name: mistral-7b
-  namespace: default
-  labels:
-    app: mistral-7b
-spec:
-  replicas: 1
-  selector:
-    matchLabels:
-      app: mistral-7b
-  template:
-    metadata:
-      labels:
-        app: mistral-7b
-    spec:
-      volumes:
-      # PVC
-      - name: cache-volume
-        persistentVolumeClaim:
-          claimName: mistral-7b
-      # vLLM needs to access the host's shared memory for tensor parallel inference.
-      - name: shm
-        emptyDir:
-          medium: Memory
-          sizeLimit: "8Gi"
-      hostNetwork: true
-      hostIPC: true
-      containers:
-      - name: mistral-7b
-        image: rocm/vllm:rocm6.2_mi300_ubuntu20.04_py3.9_vllm_0.6.4
-        securityContext:
-          seccompProfile:
-            type: Unconfined
-          runAsGroup: 44
-          capabilities:
-            add:
-            - SYS_PTRACE
-        command: ["/bin/sh", "-c"]
-        args: [
-          "vllm serve mistralai/Mistral-7B-v0.3 --port 8000 --trust-remote-code --enable-chunked-prefill --max_num_batched_tokens 1024"
-        ]
-        env:
-        - name: HUGGING_FACE_HUB_TOKEN
-          valueFrom:
-            secretKeyRef:
-              name: hf-token-secret
-              key: token
+      Next to create the deployment file for vLLM to run the model server. The following example deploys the `Mistral-7B-Instruct-v0.3` model.
+
+      Here are two examples for using NVIDIA GPU and AMD GPU.
+
+      NVIDIA GPU:
+
+      ```yaml
+      apiVersion: apps/v1
+      kind: Deployment
+      metadata:
+        name: mistral-7b
+        namespace: default
+        labels:
+          app: mistral-7b
+      spec:
+        replicas: 1
+        selector:
+          matchLabels:
+            app: mistral-7b
+        template:
+          metadata:
+            labels:
+              app: mistral-7b
+          spec:
+            volumes:
+            - name: cache-volume
+              persistentVolumeClaim:
+                claimName: mistral-7b
+            # vLLM needs to access the host's shared memory for tensor parallel inference.
+            - name: shm
+              emptyDir:
+                medium: Memory
+                sizeLimit: "2Gi"
+            containers:
+            - name: mistral-7b
+              image: vllm/vllm-openai:latest
+              command: ["/bin/sh", "-c"]
+              args: [
+                "vllm serve mistralai/Mistral-7B-Instruct-v0.3 --trust-remote-code --enable-chunked-prefill --max_num_batched_tokens 1024"
+              ]
+              env:
+              - name: HUGGING_FACE_HUB_TOKEN
+                valueFrom:
+                  secretKeyRef:
+                    name: hf-token-secret
+                    key: token
+              ports:
+              - containerPort: 8000
+              resources:
+                limits:
+                  cpu: "10"
+                  memory: 20G
+                  nvidia.com/gpu: "1"
+                requests:
+                  cpu: "2"
+                  memory: 6G
+                  nvidia.com/gpu: "1"
+              volumeMounts:
+              - mountPath: /root/.cache/huggingface
+                name: cache-volume
+              - name: shm
+                mountPath: /dev/shm
+              livenessProbe:
+                httpGet:
+                  path: /health
+                  port: 8000
+                initialDelaySeconds: 60
+                periodSeconds: 10
+              readinessProbe:
+                httpGet:
+                  path: /health
+                  port: 8000
+                initialDelaySeconds: 60
+                periodSeconds: 5
+      ```
+
+      AMD GPU:
+
+      You can refer to the `deployment.yaml` below if using AMD ROCm GPU like MI300X.
+
+      ```yaml
+      apiVersion: apps/v1
+      kind: Deployment
+      metadata:
+        name: mistral-7b
+        namespace: default
+        labels:
+          app: mistral-7b
+      spec:
+        replicas: 1
+        selector:
+          matchLabels:
+            app: mistral-7b
+        template:
+          metadata:
+            labels:
+              app: mistral-7b
+          spec:
+            volumes:
+            # PVC
+            - name: cache-volume
+              persistentVolumeClaim:
+                claimName: mistral-7b
+            # vLLM needs to access the host's shared memory for tensor parallel inference.
+            - name: shm
+              emptyDir:
+                medium: Memory
+                sizeLimit: "8Gi"
+            hostNetwork: true
+            hostIPC: true
+            containers:
+            - name: mistral-7b
+              image: rocm/vllm:rocm6.2_mi300_ubuntu20.04_py3.9_vllm_0.6.4
+              securityContext:
+                seccompProfile:
+                  type: Unconfined
+                runAsGroup: 44
+                capabilities:
+                  add:
+                  - SYS_PTRACE
+              command: ["/bin/sh", "-c"]
+              args: [
+                "vllm serve mistralai/Mistral-7B-v0.3 --port 8000 --trust-remote-code --enable-chunked-prefill --max_num_batched_tokens 1024"
+              ]
+              env:
+              - name: HUGGING_FACE_HUB_TOKEN
+                valueFrom:
+                  secretKeyRef:
+                    name: hf-token-secret
+                    key: token
+              ports:
+              - containerPort: 8000
+              resources:
+                limits:
+                  cpu: "10"
+                  memory: 20G
+                  amd.com/gpu: "1"
+                requests:
+                  cpu: "6"
+                  memory: 6G
+                  amd.com/gpu: "1"
+              volumeMounts:
+              - name: cache-volume
+                mountPath: /root/.cache/huggingface
+              - name: shm
+                mountPath: /dev/shm
+      ```
+
+      You can get the full example with steps and sample yaml files from <https://github.com/ROCm/k8s-device-plugin/tree/master/example/vllm-serve>.
+
+2. Create a Kubernetes Service for vLLM
+
+      Next, create a Kubernetes Service file to expose the `mistral-7b` deployment:
+
+      ```yaml
+      apiVersion: v1
+      kind: Service
+      metadata:
+        name: mistral-7b
+        namespace: default
+      spec:
         ports:
-        - containerPort: 8000
-        resources:
-          limits:
-            cpu: "10"
-            memory: 20G
-            amd.com/gpu: "1"
-          requests:
-            cpu: "6"
-            memory: 6G
-            amd.com/gpu: "1"
-        volumeMounts:
-        - name: cache-volume
-          mountPath: /root/.cache/huggingface
-        - name: shm
-          mountPath: /dev/shm
-```
-You can get the full example with steps and sample yaml files from <https://github.com/ROCm/k8s-device-plugin/tree/master/example/vllm-serve>.
+        - name: http-mistral-7b
+          port: 80
+          protocol: TCP
+          targetPort: 8000
+        # The label selector should match the deployment labels & it is useful for prefix caching feature
+        selector:
+          app: mistral-7b
+        sessionAffinity: None
+        type: ClusterIP
+      ```
 
-2. **Create a Kubernetes Service for vLLM**
+3. Deploy and Test
 
-Next, create a Kubernetes Service file to expose the `mistral-7b` deployment:
+      Apply the deployment and service configurations using `kubectl apply -f <filename>`:
 
-```yaml
-apiVersion: v1
-kind: Service
-metadata:
-  name: mistral-7b
-  namespace: default
-spec:
-  ports:
-  - name: http-mistral-7b
-    port: 80
-    protocol: TCP
-    targetPort: 8000
-  # The label selector should match the deployment labels & it is useful for prefix caching feature
-  selector:
-    app: mistral-7b
-  sessionAffinity: None
-  type: ClusterIP
-```
+      ```console
+      kubectl apply -f deployment.yaml
+      kubectl apply -f service.yaml
+      ```
 
-3. **Deploy and Test**
+      To test the deployment, run the following `curl` command:
 
-Apply the deployment and service configurations using `kubectl apply -f <filename>`:
+      ```console
+      curl http://mistral-7b.default.svc.cluster.local/v1/completions \
+        -H "Content-Type: application/json" \
+        -d '{
+              "model": "mistralai/Mistral-7B-Instruct-v0.3",
+              "prompt": "San Francisco is a",
+              "max_tokens": 7,
+              "temperature": 0
+            }'
+      ```
 
-```console
-kubectl apply -f deployment.yaml
-kubectl apply -f service.yaml
-```
-
-To test the deployment, run the following `curl` command:
-
-```console
-curl http://mistral-7b.default.svc.cluster.local/v1/completions \
-  -H "Content-Type: application/json" \
-  -d '{
-        "model": "mistralai/Mistral-7B-Instruct-v0.3",
-        "prompt": "San Francisco is a",
-        "max_tokens": 7,
-        "temperature": 0
-      }'
-```
-
-If the service is correctly deployed, you should receive a response from the vLLM model.
+      If the service is correctly deployed, you should receive a response from the vLLM model.
 
 ## Conclusion
 

docs/source/design/automatic_prefix_caching.md

@@ -2,11 +2,11 @@
 
 # Automatic Prefix Caching
 
-The core idea of [PagedAttention](#design-paged-attention) is to partition the KV cache of each request into KV Blocks. Each block contains the attention keys and values for a fixed number of tokens. The PagedAttention algorithm allows these blocks to be stored in non-contiguous physical memory so that we can eliminate memory fragmentation by allocating the memory on demand.
+The core idea of [PagedAttention](https://blog.vllm.ai/2023/06/20/vllm.html) is to partition the KV cache of each request into KV Blocks. Each block contains the attention keys and values for a fixed number of tokens. The PagedAttention algorithm allows these blocks to be stored in non-contiguous physical memory so that we can eliminate memory fragmentation by allocating the memory on demand.
 
 To automatically cache the KV cache, we utilize the following key observation: Each KV block can be uniquely identified by the tokens within the block and the tokens in the prefix before the block.
 
-```
+```text
                     Block 1                  Block 2                  Block 3
          [A gentle breeze stirred] [the leaves as children] [laughed in the distance]
 Block 1: |<--- block tokens ---->|
@@ -14,19 +14,16 @@ Block 2: |<------- prefix ------>| |<--- block tokens --->|
 Block 3: |<------------------ prefix -------------------->| |<--- block tokens ---->|
 ```
 
-
 In the example above, the KV cache in the first block can be uniquely identified with the tokens “A gentle breeze stirred”. The third block can be uniquely identified with the tokens in the block “laughed in the distance”, along with the prefix tokens “A gentle breeze stirred the leaves as children”. Therefore, we can build the following one-to-one mapping:
 
-```
+```text
 hash(prefix tokens + block tokens) <--> KV Block
 ```
 
 With this mapping, we can add another indirection in vLLM’s KV cache management. Previously, each sequence in vLLM maintained a mapping from their logical KV blocks to physical blocks. To achieve automatic caching of KV blocks, we map the logical KV blocks to their hash value and maintain a global hash table of all the physical blocks. In this way, all the KV blocks sharing the same hash value (e.g., shared prefix blocks across two requests) can be mapped to the same physical block and share the memory space.
 
-
 This design achieves automatic prefix caching without the need of maintaining a tree structure among the KV blocks. More specifically, all of the blocks are independent of each other and can be allocated and freed by itself, which enables us to manages the KV cache as ordinary caches in operating system.
 
-
 ## Generalized Caching Policy
 
 Keeping all the KV blocks in a hash table enables vLLM to cache KV blocks from earlier requests to save memory and accelerate the computation of future requests. For example, if a new request shares the system prompt with the previous request, the KV cache of the shared prompt can directly be used for the new request without recomputation. However, the total KV cache space is limited and we have to decide which KV blocks to keep or evict when the cache is full.
@@ -41,5 +38,5 @@ Note that this eviction policy effectively implements the exact policy as in [Ra
 
 However, the hash-based KV cache management gives us the flexibility to handle more complicated serving scenarios and implement more complicated eviction policies beyond the policy above:
 
-- Multi-LoRA serving. When serving requests for multiple LoRA adapters, we can simply let the hash of each KV block to also include the LoRA ID the request is querying for to enable caching for all adapters. In this way, we can jointly manage the KV blocks for different adapters, which simplifies the system implementation and improves the global cache hit rate and efficiency.
-- Multi-modal models. When the user input includes more than just discrete tokens, we can use different hashing methods to handle the caching of inputs of different modalities. For example, perceptual hashing for images to cache similar input images.
+* Multi-LoRA serving. When serving requests for multiple LoRA adapters, we can simply let the hash of each KV block to also include the LoRA ID the request is querying for to enable caching for all adapters. In this way, we can jointly manage the KV blocks for different adapters, which simplifies the system implementation and improves the global cache hit rate and efficiency.
+* Multi-modal models. When the user input includes more than just discrete tokens, we can use different hashing methods to handle the caching of inputs of different modalities. For example, perceptual hashing for images to cache similar input images.

docs/source/design/input_processing/input_processing_pipeline.md

@@ -1,19 +0,0 @@
-(input-processing-pipeline)=
-
-# Input Processing Pipeline
-
-1. Input data is passed to {class}`~vllm.LLMEngine` (or {class}`~vllm.AsyncLLMEngine`).
-
-2. Tokenize the data if necessary.
-
-3. Process the inputs using {meth}`INPUT_REGISTRY.process_input <vllm.inputs.registry.InputRegistry.process_input>`.
-
-   - For example, add placeholder tokens to reserve KV cache for multi-modal embeddings.
-
-4. Send the processed inputs to {class}`~vllm.executor.executor_base.ExecutorBase`.
-
-5. Distribute the inputs via {class}`~vllm.worker.worker_base.WorkerBase` to {class}`~vllm.worker.model_runner_base.ModelRunnerBase`.
-
-6. If the data contains multi-modal data, convert it into keyword arguments using {meth}`MULTIMODAL_REGISTRY.map_input <vllm.multimodal.MultiModalRegistry.map_input>`.
-
-   - For example, convert a {class}`PIL.Image.Image` input to its pixel values for a vision model.

docs/source/design/input_processing/model_inputs_index.md

@@ -1,43 +0,0 @@
-(input-processing)=
-
-# Input Processing
-
-```{eval-rst}
-.. currentmodule:: vllm.inputs
-```
-
-Each model can override parts of vLLM's [input processing pipeline](#input-processing-pipeline) via
-{data}`~vllm.inputs.INPUT_REGISTRY` and {data}`~vllm.multimodal.MULTIMODAL_REGISTRY`.
-
-Currently, this mechanism is only utilized in [multi-modal](#multi-modality) models for preprocessing multi-modal input
-data in addition to input prompt, but it can be extended to text-only language models when needed.
-
-## Guides
-
-```{toctree}
-:maxdepth: 1
-
-input_processing_pipeline
-```
-
-## Module Contents
-
-### LLM Engine Inputs
-
-```{eval-rst}
-.. autoclass:: vllm.inputs.DecoderOnlyInputs
-    :members:
-    :show-inheritance:
-```
-
-### Registry
-
-```{eval-rst}
-.. autodata:: vllm.inputs.INPUT_REGISTRY
-```
-
-```{eval-rst}
-.. automodule:: vllm.inputs.registry
-    :members:
-    :show-inheritance:
-```

docs/source/design/mm_processing.md

@@ -0,0 +1,64 @@
+(mm-processing)=
+
+# Multi-Modal Data Processing
+
+To enable various optimizations in vLLM such as [chunked prefill](#chunked-prefill) and [prefix caching](#automatic-prefix-caching), we use {class}`~vllm.multimodal.processing.BaseMultiModalProcessor` to provide the correspondence between placeholder feature tokens (e.g. `<image>`) and multi-modal inputs (e.g. the raw input image) based on the outputs of HF processor.
+
+Here are the main features of {class}`~vllm.multimodal.processing.BaseMultiModalProcessor`:
+
+## Prompt Replacement Detection
+
+One of the main responsibilies of HF processor is to replace input placeholder tokens (e.g. `<image>` for a single image) with feature placeholder tokens (e.g. `<image><image>...<image>`, the number of which equals to the feature size). The information about which tokens have been replaced is key to finding the correspondence between placeholder feature tokens and multi-modal inputs.
+
+In vLLM, this information is specified using {class}`~vllm.multimodal.processing.PromptReplacement` in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements`. Given this specification, we can automatically detect whether HF has replaced the input placeholder tokens by checking whether the feature placeholder tokens exist in the prompt.
+
+## Tokenized Prompt Inputs
+
+To enable tokenization in a separate process, we support passing input token IDs alongside multi-modal data.
+
+### The problem
+
+Consider that HF processors follow these main steps:
+
+1. Tokenize the text
+2. Process multi-modal inputs
+3. Perform prompt replacement
+
+And we require that:
+
+- For text + multi-modal inputs, apply all steps 1--3.
+- For tokenized + multi-modal inputs, apply only steps 2--3.
+
+How can we achieve this without rewriting HF processors? We can try to call the HF processor several times on different inputs:
+
+- For text + multi-modal inputs, simply call the HF processor directly.
+- For tokenized + multi-modal inputs, call the processor only on the multi-modal inputs.
+
+While HF processors support text + multi-modal inputs natively, this is not so for tokenized + multi-modal inputs: an error is thrown if the number of input placeholder tokens do not correspond to the number of multi-modal inputs.
+
+Moreover, since the tokenized text has not passed through the HF processor, we have to apply Step 3 by ourselves to keep the output tokens and multi-modal data consistent with each other.
+
+(mm-dummy-text)=
+
+### Dummy text
+
+We work around the first issue by requiring each model to define how to generate dummy text based on the number of multi-modal inputs, via {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_processor_inputs`. This lets us generate dummy text corresponding to the multi-modal inputs and input them together to obtain the processed multi-modal data.
+
+(mm-automatic-prompt-replacement)=
+
+### Automatic prompt replacement
+
+We address the second issue by implementing model-agnostic code in
+{meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_prompt_replacements` to automatically replace input placeholder tokens with feature placeholder tokens based on the specification outputted by {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._get_prompt_replacements`.
+
+### Summary
+
+With the help of dummy text and automatic prompt replacement, our multi-modal processor can finally accept both text and token prompts with multi-modal data. The detailed logic is shown in {meth}`~vllm.multimodal.processing.BaseMultiModalProcessor._apply_hf_processor_main`.
+
+## Processor Output Caching
+
+Some HF processors, such as the one for Qwen2-VL, are [very slow](gh-issue:9238). To alleviate this problem, we cache the multi-modal outputs of HF processor to avoid processing the same multi-modal input (e.g. image) again.
+
+When new data is passed in, we first check which items are in the cache, and which ones are missing. The missing items are passed into the HF processor in a single batch and cached, before being merged with the existing items in the cache.
+
+Since we only process the missing multi-modal data items, the number of input placeholder tokens no longer corresponds to the number of the multi-modal inputs, so they can't be passed alongside the text prompt to HF processor. Therefore, we process the text and multi-modal inputs separately, using [dummy text](#mm-dummy-text) to avoid HF errors. Since this skips HF's prompt replacement code, we apply [automatic prompt replacement](#mm-automatic-prompt-replacement) afterwards to keep the output tokens and multi-modal data consistent with each other.

docs/source/features/compatibility_matrix.md

@@ -322,7 +322,9 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
 
 ```
 
-### Feature x Hardware
+(feature-x-hardware)=
+
+## Feature x Hardware
 
 ```{list-table}
    :header-rows: 1
@@ -359,7 +361,7 @@ Check the '✗' with links to see tracking issue for unsupported feature/hardwar
      - ✅
      - ✅
      - ✅
-     - [✗](gh-pr:4830)
+     - ✅
      - ✅
    * - <abbr title="Prompt Adapter">prmpt adptr</abbr>
      - ✅

docs/source/features/quantization/auto_awq.md

@@ -15,7 +15,7 @@ The main benefits are lower latency and memory usage.
 You can quantize your own models by installing AutoAWQ or picking one of the [400+ models on Huggingface](https://huggingface.co/models?sort=trending&search=awq).
 
 ```console
-$ pip install autoawq
+pip install autoawq
 ```
 
 After installing AutoAWQ, you are ready to quantize a model. Here is an example of how to quantize `mistralai/Mistral-7B-Instruct-v0.2`:
@@ -47,7 +47,7 @@ print(f'Model is quantized and saved at "{quant_path}"')
 To run an AWQ model with vLLM, you can use [TheBloke/Llama-2-7b-Chat-AWQ](https://huggingface.co/TheBloke/Llama-2-7b-Chat-AWQ) with the following command:
 
 ```console
-$ python examples/offline_inference/llm_engine_example.py --model TheBloke/Llama-2-7b-Chat-AWQ --quantization awq
+python examples/offline_inference/llm_engine_example.py --model TheBloke/Llama-2-7b-Chat-AWQ --quantization awq
 ```
 
 AWQ models are also supported directly through the LLM entrypoint:

docs/source/features/quantization/bnb.md

@@ -9,7 +9,7 @@ Compared to other quantization methods, BitsAndBytes eliminates the need for cal
 Below are the steps to utilize BitsAndBytes with vLLM.
 
 ```console
-$ pip install bitsandbytes>=0.45.0
+pip install bitsandbytes>=0.45.0
 ```
 
 vLLM reads the model's config file and supports both in-flight quantization and pre-quantized checkpoint.
@@ -17,7 +17,7 @@ vLLM reads the model's config file and supports both in-flight quantization and
 You can find bitsandbytes quantized models on <https://huggingface.co/models?other=bitsandbytes>.
 And usually, these repositories have a config.json file that includes a quantization_config section.
 
-## Read quantized checkpoint.
+## Read quantized checkpoint
 
 ```python
 from vllm import LLM
@@ -37,10 +37,11 @@ model_id = "huggyllama/llama-7b"
 llm = LLM(model=model_id, dtype=torch.bfloat16, trust_remote_code=True, \
 quantization="bitsandbytes", load_format="bitsandbytes")
 ```
+
 ## OpenAI Compatible Server
 
 Append the following to your 4bit model arguments:
 
-```
+```console
 --quantization bitsandbytes --load-format bitsandbytes
 ```

docs/source/features/quantization/fp8.md

@@ -41,7 +41,7 @@ Currently, we load the model at original precision before quantizing down to 8-b
 To produce performant FP8 quantized models with vLLM, you'll need to install the [llm-compressor](https://github.com/vllm-project/llm-compressor/) library:
 
 ```console
-$ pip install llmcompressor
+pip install llmcompressor
 ```
 
 ## Quantization Process
@@ -54,16 +54,15 @@ The quantization process involves three main steps:
 
 ### 1. Loading the Model
 
-Use `SparseAutoModelForCausalLM`, which wraps `AutoModelForCausalLM`, for saving and loading quantized models:
+Load your model and tokenizer using the standard `transformers` AutoModel classes:
 
 ```python
-from llmcompressor.transformers import SparseAutoModelForCausalLM
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
 MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
-
-model = SparseAutoModelForCausalLM.from_pretrained(
-  MODEL_ID, device_map="auto", torch_dtype="auto")
+model = AutoModelForCausalLM.from_pretrained(
+    MODEL_ID, device_map="auto", torch_dtype="auto",
+)
 tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
 ```
 
@@ -98,7 +97,7 @@ tokenizer.save_pretrained(SAVE_DIR)
 Install `vllm` and `lm-evaluation-harness`:
 
 ```console
-$ pip install vllm lm-eval==0.4.4
+pip install vllm lm-eval==0.4.4
 ```
 
 Load and run the model in `vllm`:

docs/source/features/quantization/fp8_e4m3_kvcache.md

@@ -17,7 +17,7 @@ unquantized model through a quantizer tool (e.g. AMD quantizer or NVIDIA AMMO).
 To install AMMO (AlgorithMic Model Optimization):
 
 ```console
-$ pip install --no-cache-dir --extra-index-url https://pypi.nvidia.com nvidia-ammo
+pip install --no-cache-dir --extra-index-url https://pypi.nvidia.com nvidia-ammo
 ```
 
 Studies have shown that FP8 E4M3 quantization typically only minimally degrades inference accuracy. The most recent silicon

docs/source/features/quantization/gguf.md

@@ -13,16 +13,16 @@ Currently, vllm only supports loading single-file GGUF models. If you have a mul
 To run a GGUF model with vLLM, you can download and use the local GGUF model from [TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF](https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF) with the following command:
 
 ```console
-$ wget https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/main/tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf
-$ # We recommend using the tokenizer from base model to avoid long-time and buggy tokenizer conversion.
-$ vllm serve ./tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf --tokenizer TinyLlama/TinyLlama-1.1B-Chat-v1.0
+wget https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/main/tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf
+# We recommend using the tokenizer from base model to avoid long-time and buggy tokenizer conversion.
+vllm serve ./tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf --tokenizer TinyLlama/TinyLlama-1.1B-Chat-v1.0
 ```
 
 You can also add `--tensor-parallel-size 2` to enable tensor parallelism inference with 2 GPUs:
 
 ```console
-$ # We recommend using the tokenizer from base model to avoid long-time and buggy tokenizer conversion.
-$ vllm serve ./tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf --tokenizer TinyLlama/TinyLlama-1.1B-Chat-v1.0 --tensor-parallel-size 2
+# We recommend using the tokenizer from base model to avoid long-time and buggy tokenizer conversion.
+vllm serve ./tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf --tokenizer TinyLlama/TinyLlama-1.1B-Chat-v1.0 --tensor-parallel-size 2
 ```
 
 ```{warning}

docs/source/features/quantization/int8.md

@@ -16,7 +16,7 @@ INT8 computation is supported on NVIDIA GPUs with compute capability > 7.5 (Turi
 To use INT8 quantization with vLLM, you'll need to install the [llm-compressor](https://github.com/vllm-project/llm-compressor/) library:
 
 ```console
-$ pip install llmcompressor
+pip install llmcompressor
 ```
 
 ## Quantization Process
@@ -30,14 +30,13 @@ The quantization process involves four main steps:
 
 ### 1. Loading the Model
 
-Use `SparseAutoModelForCausalLM`, which wraps `AutoModelForCausalLM`, for saving and loading quantized models:
+Load your model and tokenizer using the standard `transformers` AutoModel classes:
 
 ```python
-from llmcompressor.transformers import SparseAutoModelForCausalLM
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
 MODEL_ID = "meta-llama/Meta-Llama-3-8B-Instruct"
-model = SparseAutoModelForCausalLM.from_pretrained(
+model = AutoModelForCausalLM.from_pretrained(
     MODEL_ID, device_map="auto", torch_dtype="auto",
 )
 tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)

docs/source/features/quantization/supported_hardware.md

@@ -113,7 +113,7 @@ The table below shows the compatibility of various quantization implementations
   - ✅︎
   - ✅︎
   - ✅︎
-  - ✗
+  - ✅︎
   - ✗
   - ✗
   - ✗

docs/source/features/spec_decode.md

@@ -192,11 +192,11 @@ A few important things to consider when using the EAGLE based draft models:
 
 1. The EAGLE draft models available in the [HF repository for EAGLE models](https://huggingface.co/yuhuili) cannot be
    used directly with vLLM due to differences in the expected layer names and model definition.
-   To use these models with vLLM, use the [following script](https://gist.github.com/abhigoyal1997/1e7a4109ccb7704fbc67f625e86b2d6d) 
+   To use these models with vLLM, use the [following script](https://gist.github.com/abhigoyal1997/1e7a4109ccb7704fbc67f625e86b2d6d)
    to convert them. Note that this script does not modify the model's weights.
 
    In the above example, use the script to first convert
-   the [yuhuili/EAGLE-LLaMA3-Instruct-8B](https://huggingface.co/yuhuili/EAGLE-LLaMA3-Instruct-8B) model 
+   the [yuhuili/EAGLE-LLaMA3-Instruct-8B](https://huggingface.co/yuhuili/EAGLE-LLaMA3-Instruct-8B) model
    and then use the converted checkpoint as the draft model in vLLM.
 
 2. The EAGLE based draft models need to be run without tensor parallelism
@@ -207,7 +207,6 @@ A few important things to consider when using the EAGLE based draft models:
    reported in the reference implementation [here](https://github.com/SafeAILab/EAGLE). This issue is under
    investigation and tracked here: [https://github.com/vllm-project/vllm/issues/9565](https://github.com/vllm-project/vllm/issues/9565).
 
-
 A variety of EAGLE draft models are available on the Hugging Face hub:
 
 | Base Model                                                           | EAGLE on Hugging Face                     | # EAGLE Parameters |
@@ -224,7 +223,6 @@ A variety of EAGLE draft models are available on the Hugging Face hub:
 | Qwen2-7B-Instruct                                                    | yuhuili/EAGLE-Qwen2-7B-Instruct          | 0.26B              |
 | Qwen2-72B-Instruct                                                   | yuhuili/EAGLE-Qwen2-72B-Instruct         | 1.05B              |
 
-
 ## Lossless guarantees of Speculative Decoding
 
 In vLLM, speculative decoding aims to enhance inference efficiency while maintaining accuracy. This section addresses the lossless guarantees of
@@ -250,8 +248,6 @@ speculative decoding, breaking down the guarantees into three key areas:
    same request across runs. For more details, see the FAQ section
    titled *Can the output of a prompt vary across runs in vLLM?* in the [FAQs](#faq).
 
-**Conclusion**
-
 While vLLM strives to ensure losslessness in speculative decoding, variations in generated outputs with and without speculative decoding
 can occur due to following factors:
 
@@ -259,8 +255,6 @@ can occur due to following factors:
 - **Batch Size and Numerical Stability**: Changes in batch size may cause variations in logprobs and output probabilities, potentially
   due to non-deterministic behavior in batched operations or numerical instability.
 
-**Mitigation Strategies**
-
 For mitigation strategies, please refer to the FAQ entry *Can the output of a prompt vary across runs in vLLM?* in the [FAQs](#faq).
 
 ## Resources for vLLM contributors

docs/source/features/structured_outputs.md

@@ -5,7 +5,7 @@
 vLLM supports the generation of structured outputs using [outlines](https://github.com/dottxt-ai/outlines), [lm-format-enforcer](https://github.com/noamgat/lm-format-enforcer), or [xgrammar](https://github.com/mlc-ai/xgrammar) as backends for the guided decoding.
 This document shows you some examples of the different options that are available to generate structured outputs.
 
-## Online Inference (OpenAI API)
+## Online Serving (OpenAI API)
 
 You can generate structured outputs using the OpenAI's [Completions](https://platform.openai.com/docs/api-reference/completions) and [Chat](https://platform.openai.com/docs/api-reference/chat) API.
 
@@ -239,7 +239,7 @@ The main available options inside `GuidedDecodingParams` are:
 - `backend`
 - `whitespace_pattern`
 
-These parameters can be used in the same way as the parameters from the Online Inference examples above.
+These parameters can be used in the same way as the parameters from the Online Serving examples above.
 One example for the usage of the `choices` parameter is shown below:
 
 ```python
@@ -257,4 +257,4 @@ outputs = llm.generate(
 print(outputs[0].outputs[0].text)
 ```
 
-Full example: <gh-file:examples/offline_inference/offline_inference_structured_outputs.py>
+Full example: <gh-file:examples/offline_inference/structured_outputs.py>

docs/source/features/tool_calling.md

@@ -55,21 +55,24 @@ print(f"Result: {get_weather(**json.loads(tool_call.arguments))}")
 ```
 
 Example output:
-```
+
+```text
 Function called: get_weather
 Arguments: {"location": "San Francisco, CA", "unit": "fahrenheit"}
 Result: Getting the weather for San Francisco, CA in fahrenheit...
 ```
 
 This example demonstrates:
-- Setting up the server with tool calling enabled
-- Defining an actual function to handle tool calls
-- Making a request with `tool_choice="auto"`
-- Handling the structured response and executing the corresponding function
+
+* Setting up the server with tool calling enabled
+* Defining an actual function to handle tool calls
+* Making a request with `tool_choice="auto"`
+* Handling the structured response and executing the corresponding function
 
 You can also specify a particular function using named function calling by setting `tool_choice={"type": "function", "function": {"name": "get_weather"}}`. Note that this will use the guided decoding backend - so the first time this is used, there will be several seconds of latency (or more) as the FSM is compiled for the first time before it is cached for subsequent requests.
 
 Remember that it's the callers responsibility to:
+
 1. Define appropriate tools in the request
 2. Include relevant context in the chat messages
 3. Handle the tool calls in your application logic
@@ -77,20 +80,21 @@ Remember that it's the callers responsibility to:
 For more advanced usage, including parallel tool calls and different model-specific parsers, see the sections below.
 
 ## Named Function Calling
+
 vLLM supports named function calling in the chat completion API by default. It does so using Outlines through guided decoding, so this is
 enabled by default, and will work with any supported model. You are guaranteed a validly-parsable function call - not a
 high-quality one.
 
-vLLM will use guided decoding to ensure the response matches the tool parameter object defined by the JSON schema in the `tools` parameter. 
+vLLM will use guided decoding to ensure the response matches the tool parameter object defined by the JSON schema in the `tools` parameter.
 For best results, we recommend ensuring that the expected output format / schema is specified in the prompt to ensure that the model's intended generation is aligned with the schema that it's being forced to generate by the guided decoding backend.
 
 To use a named function, you need to define the functions in the `tools` parameter of the chat completion request, and
 specify the `name` of one of the tools in the `tool_choice` parameter of the chat completion request.
 
-
 ## Automatic Function Calling
 
 To enable this feature, you should set the following flags:
+
 * `--enable-auto-tool-choice` -- **mandatory** Auto tool choice. tells vLLM that you want to enable the model to generate its own tool calls when it
 deems appropriate.
 * `--tool-call-parser` -- select the tool parser to use (listed below). Additional tool parsers
@@ -104,28 +108,28 @@ from HuggingFace; and you can find an example of this in a `tokenizer_config.jso
 
 If your favorite tool-calling model is not supported, please feel free to contribute a parser & tool use chat template!
 
-
 ### Hermes Models (`hermes`)
 
 All Nous Research Hermes-series models newer than Hermes 2 Pro should be supported.
+
 * `NousResearch/Hermes-2-Pro-*`
 * `NousResearch/Hermes-2-Theta-*`
 * `NousResearch/Hermes-3-*`
 
-
 _Note that the Hermes 2 **Theta** models are known to have degraded tool call quality & capabilities due to the merge
 step in their creation_.
 
 Flags: `--tool-call-parser hermes`
 
-
 ### Mistral Models (`mistral`)
 
 Supported models:
+
 * `mistralai/Mistral-7B-Instruct-v0.3` (confirmed)
 * Additional mistral function-calling models are compatible as well.
 
 Known issues:
+
 1. Mistral 7B struggles to generate parallel tool calls correctly.
 2. Mistral's `tokenizer_config.json` chat template requires tool call IDs that are exactly 9 digits, which is
 much shorter than what vLLM generates. Since an exception is thrown when this condition
@@ -136,13 +140,12 @@ it works with vLLM's tool call IDs (provided `tool_call_id` fields are truncated
 * `examples/tool_chat_template_mistral_parallel.jinja` - this is a "better" version that adds a tool-use system prompt
 when tools are provided, that results in much better reliability when working with parallel tool calling.
 
-
 Recommended flags: `--tool-call-parser mistral --chat-template examples/tool_chat_template_mistral_parallel.jinja`
 
-
 ### Llama Models (`llama3_json`)
 
 Supported models:
+
 * `meta-llama/Meta-Llama-3.1-8B-Instruct`
 * `meta-llama/Meta-Llama-3.1-70B-Instruct`
 * `meta-llama/Meta-Llama-3.1-405B-Instruct`
@@ -152,6 +155,7 @@ The tool calling that is supported is the [JSON based tool calling](https://llam
 Other tool calling formats like the built in python tool calling or custom tool calling are not supported.
 
 Known issues:
+
 1. Parallel tool calls are not supported.
 2. The model can generate parameters with a wrong format, such as generating
    an array serialized as string instead of an array.
@@ -164,6 +168,7 @@ Recommended flags: `--tool-call-parser llama3_json --chat-template examples/tool
 #### IBM Granite
 
 Supported models:
+
 * `ibm-granite/granite-3.0-8b-instruct`
 
 Recommended flags: `--tool-call-parser granite --chat-template examples/tool_chat_template_granite.jinja`
@@ -182,42 +187,45 @@ Recommended flags: `--tool-call-parser granite-20b-fc --chat-template examples/t
 
 `examples/tool_chat_template_granite_20b_fc.jinja`: this is a modified chat template from the original on Huggingface, which is not vLLM compatible. It blends function description elements from the Hermes template and follows the same system prompt as "Response Generation" mode from [the paper](https://arxiv.org/abs/2407.00121). Parallel function calls are supported.
 
-
 ### InternLM Models (`internlm`)
 
 Supported models:
+
 * `internlm/internlm2_5-7b-chat` (confirmed)
 * Additional internlm2.5 function-calling models are compatible as well
 
 Known issues:
+
 * Although this implementation also supports InternLM2, the tool call results are not stable when testing with the `internlm/internlm2-chat-7b` model.
 
 Recommended flags: `--tool-call-parser internlm --chat-template examples/tool_chat_template_internlm2_tool.jinja`
 
-
 ### Jamba Models (`jamba`)
+
 AI21's Jamba-1.5 models are supported.
+
 * `ai21labs/AI21-Jamba-1.5-Mini`
 * `ai21labs/AI21-Jamba-1.5-Large`
 
-
 Flags: `--tool-call-parser jamba`
 
-
 ### Models with Pythonic Tool Calls (`pythonic`)
 
 A growing number of models output a python list to represent tool calls instead of using JSON. This has the advantage of inherently supporting parallel tool calls and removing ambiguity around the JSON schema required for tool calls. The `pythonic` tool parser can support such models.
 
 As a concrete example, these models may look up the weather in San Francisco and Seattle by generating:
+
 ```python
 [get_weather(city='San Francisco', metric='celsius'), get_weather(city='Seattle', metric='celsius')]
 ```
 
 Limitations:
+
 * The model must not generate both text and tool calls in the same generation. This may not be hard to change for a specific model, but the community currently lacks consensus on which tokens to emit when starting and ending tool calls.  (In particular, the Llama 3.2 models emit no such tokens.)
 * Llama's smaller models struggle to use tools effectively.
 
 Example supported models:
+
 * `meta-llama/Llama-3.2-1B-Instruct`\* (use with `examples/tool_chat_template_llama3.2_pythonic.jinja`)
 * `meta-llama/Llama-3.2-3B-Instruct`\* (use with `examples/tool_chat_template_llama3.2_pythonic.jinja`)
 * `Team-ACE/ToolACE-8B` (use with `examples/tool_chat_template_toolace.jinja`)
@@ -231,7 +239,6 @@ Llama's smaller models frequently fail to emit tool calls in the correct format.
 
 ---
 
-
 ## How to write a tool parser plugin
 
 A tool parser plugin is a Python file containing one or more ToolParser implementations. You can write a ToolParser similar to the `Hermes2ProToolParser` in vllm/entrypoints/openai/tool_parsers/hermes_tool_parser.py.
@@ -284,7 +291,8 @@ class ExampleToolParser(ToolParser):
 ```
 
 Then you can use this plugin in the command line like this.
-```
+
+```console
     --enable-auto-tool-choice \
     --tool-parser-plugin <absolute path of the plugin file>
     --tool-call-parser example \

docs/source/getting_started/faq.md

@@ -30,7 +30,7 @@ changes in batch size, or batch expansion in speculative decoding. These batchin
 can lead to slightly different logit/logprob values at each step. Such differences can accumulate, potentially resulting in
 different tokens being sampled. Once a different token is sampled, further divergence is likely.
 
-**Mitigation Strategies**
+## Mitigation Strategies
 
 - For improved stability and reduced variance, use `float32`. Note that this will require more memory.
 - If using `bfloat16`, switching to `float16` can also help.

docs/source/getting_started/installation/ai_accelerator/hpu-gaudi.inc.md

@@ -1,10 +1,13 @@
-(installation-gaudi)=
+# Installation
 
-# Installation for Intel® Gaudi®
+This tab provides instructions on running vLLM with Intel Gaudi devices.
 
-This README provides instructions on running vLLM with Intel Gaudi devices.
+## Requirements
 
-## Requirements and Installation
+- OS: Ubuntu 22.04 LTS
+- Python: 3.10
+- Intel Gaudi accelerator
+- Intel Gaudi software version 1.18.0
 
 Please follow the instructions provided in the [Gaudi Installation
 Guide](https://docs.habana.ai/en/latest/Installation_Guide/index.html)
@@ -12,42 +15,24 @@ to set up the execution environment. To achieve the best performance,
 please follow the methods outlined in the [Optimizing Training Platform
 Guide](https://docs.habana.ai/en/latest/PyTorch/Model_Optimization_PyTorch/Optimization_in_Training_Platform.html).
 
-### Requirements
+## Configure a new environment
 
-- OS: Ubuntu 22.04 LTS
-- Python: 3.10
-- Intel Gaudi accelerator
-- Intel Gaudi software version 1.18.0
-
-### Quick start using Dockerfile
-
-```console
-$ docker build -f Dockerfile.hpu -t vllm-hpu-env  .
-$ docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --rm vllm-hpu-env
-```
-
-```{tip}
-If you're observing the following error: `docker: Error response from daemon: Unknown runtime specified habana.`, please refer to "Install Using Containers" section of [Intel Gaudi Software Stack and Driver Installation](https://docs.habana.ai/en/v1.18.0/Installation_Guide/Bare_Metal_Fresh_OS.html). Make sure you have `habana-container-runtime` package installed and that `habana` container runtime is registered.
-```
-
-### Build from source
-
-#### Environment verification
+### Environment verification
 
 To verify that the Intel Gaudi software was correctly installed, run:
 
 ```console
-$ hl-smi # verify that hl-smi is in your PATH and each Gaudi accelerator is visible
-$ apt list --installed | grep habana # verify that habanalabs-firmware-tools, habanalabs-graph, habanalabs-rdma-core, habanalabs-thunk and habanalabs-container-runtime are installed
-$ pip list | grep habana # verify that habana-torch-plugin, habana-torch-dataloader, habana-pyhlml and habana-media-loader are installed
-$ pip list | grep neural # verify that neural_compressor is installed
+hl-smi # verify that hl-smi is in your PATH and each Gaudi accelerator is visible
+apt list --installed | grep habana # verify that habanalabs-firmware-tools, habanalabs-graph, habanalabs-rdma-core, habanalabs-thunk and habanalabs-container-runtime are installed
+pip list | grep habana # verify that habana-torch-plugin, habana-torch-dataloader, habana-pyhlml and habana-media-loader are installed
+pip list | grep neural # verify that neural_compressor is installed
 ```
 
 Refer to [Intel Gaudi Software Stack
 Verification](https://docs.habana.ai/en/latest/Installation_Guide/SW_Verification.html#platform-upgrade)
 for more details.
 
-#### Run Docker Image
+### Run Docker Image
 
 It is highly recommended to use the latest Docker image from Intel Gaudi
 vault. Refer to the [Intel Gaudi
@@ -57,33 +42,58 @@ for more details.
 Use the following commands to run a Docker image:
 
 ```console
-$ docker pull vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest
-$ docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest
+docker pull vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest
+docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --ipc=host vault.habana.ai/gaudi-docker/1.18.0/ubuntu22.04/habanalabs/pytorch-installer-2.4.0:latest
 ```
 
-#### Build and Install vLLM
+## Set up using Python
+
+### Pre-built wheels
+
+Currently, there are no pre-built Intel Gaudi wheels.
+
+### Build wheel from source
 
 To build and install vLLM from source, run:
 
 ```console
-$ git clone https://github.com/vllm-project/vllm.git
-$ cd vllm
-$ python setup.py develop
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+python setup.py develop
 ```
 
 Currently, the latest features and performance optimizations are developed in Gaudi's [vLLM-fork](https://github.com/HabanaAI/vllm-fork) and we periodically upstream them to vLLM main repo. To install latest [HabanaAI/vLLM-fork](https://github.com/HabanaAI/vllm-fork), run the following:
 
 ```console
-$ git clone https://github.com/HabanaAI/vllm-fork.git
-$ cd vllm-fork
-$ git checkout habana_main
-$ python setup.py develop
+git clone https://github.com/HabanaAI/vllm-fork.git
+cd vllm-fork
+git checkout habana_main
+python setup.py develop
 ```
 
-## Supported Features
+## Set up using Docker
+
+### Pre-built images
+
+Currently, there are no pre-built Intel Gaudi images.
+
+### Build image from source
+
+```console
+docker build -f Dockerfile.hpu -t vllm-hpu-env  .
+docker run -it --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_MCA_btl_vader_single_copy_mechanism=none --cap-add=sys_nice --net=host --rm vllm-hpu-env
+```
+
+```{tip}
+If you're observing the following error: `docker: Error response from daemon: Unknown runtime specified habana.`, please refer to "Install Using Containers" section of [Intel Gaudi Software Stack and Driver Installation](https://docs.habana.ai/en/v1.18.0/Installation_Guide/Bare_Metal_Fresh_OS.html). Make sure you have `habana-container-runtime` package installed and that `habana` container runtime is registered.
+```
+
+## Extra information
+
+## Supported features
 
 - [Offline inference](#offline-inference)
-- Online inference via [OpenAI-Compatible Server](#openai-compatible-server)
+- Online serving via [OpenAI-Compatible Server](#openai-compatible-server)
 - HPU autodetection - no need to manually select device within vLLM
 - Paged KV cache with algorithms enabled for Intel Gaudi accelerators
 - Custom Intel Gaudi implementations of Paged Attention, KV cache ops,
@@ -94,14 +104,14 @@ $ python setup.py develop
   for accelerating low-batch latency and throughput
 - Attention with Linear Biases (ALiBi)
 
-## Unsupported Features
+## Unsupported features
 
 - Beam search
 - LoRA adapters
 - Quantization
 - Prefill chunking (mixed-batch inferencing)
 
-## Supported Configurations
+## Supported configurations
 
 The following configurations have been validated to be function with
 Gaudi2 devices. Configurations that are not listed may or may not work.
@@ -137,7 +147,7 @@ Gaudi2 devices. Configurations that are not listed may or may not work.
 - [meta-llama/Meta-Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3.1-70B-Instruct)
   with tensor parallelism on 8x HPU, BF16 datatype with random or greedy sampling
 
-## Performance Tuning
+## Performance tuning
 
 ### Execution modes
 
@@ -181,7 +191,7 @@ Bucketing allows us to reduce the number of required graphs significantly, but i
 
 Bucketing ranges are determined with 3 parameters - `min`, `step` and `max`. They can be set separately for prompt and decode phase, and for batch size and sequence length dimension. These parameters can be observed in logs during vLLM startup:
 
-```
+```text
 INFO 08-01 21:37:59 hpu_model_runner.py:493] Prompt bucket config (min, step, max_warmup) bs:[1, 32, 4], seq:[128, 128, 1024]
 INFO 08-01 21:37:59 hpu_model_runner.py:499] Generated 24 prompt buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024)]
 INFO 08-01 21:37:59 hpu_model_runner.py:504] Decode bucket config (min, step, max_warmup) bs:[1, 128, 4], seq:[128, 128, 2048]
@@ -192,7 +202,7 @@ INFO 08-01 21:37:59 hpu_model_runner.py:509] Generated 48 decode buckets: [(1, 1
 
 Example (with ramp-up)
 
-```
+```text
 min = 2, step = 32, max = 64
 => ramp_up = (2, 4, 8, 16)
 => stable = (32, 64)
@@ -201,7 +211,7 @@ min = 2, step = 32, max = 64
 
 Example (without ramp-up)
 
-```
+```text
 min = 128, step = 128, max = 512
 => ramp_up = ()
 => stable = (128, 256, 384, 512)
@@ -224,7 +234,7 @@ Bucketing is transparent to a client -- padding in sequence length dimension is
 
 Warmup is an optional, but highly recommended step occurring before vLLM server starts listening. It executes a forward pass for each bucket with dummy data. The goal is to pre-compile all graphs and not incur any graph compilation overheads within bucket boundaries during server runtime. Each warmup step is logged during vLLM startup:
 
-```
+```text
 INFO 08-01 22:26:47 hpu_model_runner.py:1066] [Warmup][Prompt][1/24] batch_size:4 seq_len:1024 free_mem:79.16 GiB
 INFO 08-01 22:26:47 hpu_model_runner.py:1066] [Warmup][Prompt][2/24] batch_size:4 seq_len:896 free_mem:55.43 GiB
 INFO 08-01 22:26:48 hpu_model_runner.py:1066] [Warmup][Prompt][3/24] batch_size:4 seq_len:768 free_mem:55.43 GiB
@@ -273,7 +283,7 @@ When there's large amount of requests pending, vLLM scheduler will attempt to fi
 
 Each described step is logged by vLLM server, as follows (negative values correspond to memory being released):
 
-```
+```text
 INFO 08-02 17:37:44 hpu_model_runner.py:493] Prompt bucket config (min, step, max_warmup) bs:[1, 32, 4], seq:[128, 128, 1024]
 INFO 08-02 17:37:44 hpu_model_runner.py:499] Generated 24 prompt buckets: [(1, 128), (1, 256), (1, 384), (1, 512), (1, 640), (1, 768), (1, 896), (1, 1024), (2, 128), (2, 256), (2, 384), (2, 512), (2, 640), (2, 768), (2, 896), (2, 1024), (4, 128), (4, 256), (4, 384), (4, 512), (4, 640), (4, 768), (4, 896), (4, 1024)]
 INFO 08-02 17:37:44 hpu_model_runner.py:504] Decode bucket config (min, step, max_warmup) bs:[1, 128, 4], seq:[128, 128, 2048]
@@ -349,26 +359,26 @@ INFO 08-02 17:38:43 hpu_executor.py:91] init_cache_engine took 37.92 GiB of devi
   - Default values:
 
     - Prompt:
-      : - batch size min (`VLLM_PROMPT_BS_BUCKET_MIN`): `1`
-        - batch size step (`VLLM_PROMPT_BS_BUCKET_STEP`): `min(max_num_seqs, 32)`
-        - batch size max (`VLLM_PROMPT_BS_BUCKET_MAX`): `min(max_num_seqs, 64)`
-        - sequence length min (`VLLM_PROMPT_SEQ_BUCKET_MIN`): `block_size`
-        - sequence length step (`VLLM_PROMPT_SEQ_BUCKET_STEP`): `block_size`
-        - sequence length max (`VLLM_PROMPT_SEQ_BUCKET_MAX`): `max_model_len`
+      - batch size min (`VLLM_PROMPT_BS_BUCKET_MIN`): `1`
+      - batch size step (`VLLM_PROMPT_BS_BUCKET_STEP`): `min(max_num_seqs, 32)`
+      - batch size max (`VLLM_PROMPT_BS_BUCKET_MAX`): `min(max_num_seqs, 64)`
+      - sequence length min (`VLLM_PROMPT_SEQ_BUCKET_MIN`): `block_size`
+      - sequence length step (`VLLM_PROMPT_SEQ_BUCKET_STEP`): `block_size`
+      - sequence length max (`VLLM_PROMPT_SEQ_BUCKET_MAX`): `max_model_len`
     - Decode:
-      : - batch size min (`VLLM_DECODE_BS_BUCKET_MIN`): `1`
-        - batch size step (`VLLM_DECODE_BS_BUCKET_STEP`): `min(max_num_seqs, 32)`
-        - batch size max (`VLLM_DECODE_BS_BUCKET_MAX`): `max_num_seqs`
-        - sequence length min (`VLLM_DECODE_BLOCK_BUCKET_MIN`): `block_size`
-        - sequence length step (`VLLM_DECODE_BLOCK_BUCKET_STEP`): `block_size`
-        - sequence length max (`VLLM_DECODE_BLOCK_BUCKET_MAX`): `max(128, (max_num_seqs*max_model_len)/block_size)`
+      - batch size min (`VLLM_DECODE_BS_BUCKET_MIN`): `1`
+      - batch size step (`VLLM_DECODE_BS_BUCKET_STEP`): `min(max_num_seqs, 32)`
+      - batch size max (`VLLM_DECODE_BS_BUCKET_MAX`): `max_num_seqs`
+      - sequence length min (`VLLM_DECODE_BLOCK_BUCKET_MIN`): `block_size`
+      - sequence length step (`VLLM_DECODE_BLOCK_BUCKET_STEP`): `block_size`
+      - sequence length max (`VLLM_DECODE_BLOCK_BUCKET_MAX`): `max(128, (max_num_seqs*max_model_len)/block_size)`
 
 Additionally, there are HPU PyTorch Bridge environment variables impacting vLLM execution:
 
 - `PT_HPU_LAZY_MODE`: if `0`, PyTorch Eager backend for Gaudi will be used, if `1` PyTorch Lazy backend for Gaudi will be used, `1` is default
 - `PT_HPU_ENABLE_LAZY_COLLECTIVES`: required to be `true` for tensor parallel inference with HPU Graphs
 
-## Troubleshooting: Tweaking HPU Graphs
+## Troubleshooting: tweaking HPU graphs
 
 If you experience device out-of-memory issues or want to attempt
 inference at higher batch sizes, try tweaking HPU Graphs by following
@@ -385,5 +395,5 @@ the below:
   completely. With HPU Graphs disabled, you are trading latency and
   throughput at lower batches for potentially higher throughput on
   higher batches. You can do that by adding `--enforce-eager` flag to
-  server (for online inference), or by passing `enforce_eager=True`
+  server (for online serving), or by passing `enforce_eager=True`
   argument to LLM constructor (for offline inference).

docs/source/getting_started/installation/ai_accelerator/index.md

@@ -0,0 +1,375 @@
+# Other AI accelerators
+
+vLLM is a Python library that supports the following AI accelerators. Select your AI accelerator type to see vendor specific instructions:
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} TPU
+:sync: tpu
+
+```{include} tpu.inc.md
+:start-after: "# Installation"
+:end-before: "## Requirements"
+```
+
+:::
+
+:::{tab-item} Intel Gaudi
+:sync: hpu-gaudi
+
+```{include} hpu-gaudi.inc.md
+:start-after: "# Installation"
+:end-before: "## Requirements"
+```
+
+:::
+
+:::{tab-item} Neuron
+:sync: neuron
+
+```{include} neuron.inc.md
+:start-after: "# Installation"
+:end-before: "## Requirements"
+```
+
+:::
+
+:::{tab-item} OpenVINO
+:sync: openvino
+
+```{include} openvino.inc.md
+:start-after: "# Installation"
+:end-before: "## Requirements"
+```
+
+:::
+
+::::
+
+## Requirements
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} TPU
+:sync: tpu
+
+```{include} tpu.inc.md
+:start-after: "## Requirements"
+:end-before: "## Configure a new environment"
+```
+
+:::
+
+:::{tab-item} Intel Gaudi
+:sync: hpu-gaudi
+
+```{include} hpu-gaudi.inc.md
+:start-after: "## Requirements"
+:end-before: "## Configure a new environment"
+```
+
+:::
+
+:::{tab-item} Neuron
+:sync: neuron
+
+```{include} neuron.inc.md
+:start-after: "## Requirements"
+:end-before: "## Configure a new environment"
+```
+
+:::
+
+:::{tab-item} OpenVINO
+:sync: openvino
+
+```{include} openvino.inc.md
+:start-after: "## Requirements"
+:end-before: "## Set up using Python"
+```
+
+:::
+
+::::
+
+## Configure a new environment
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} TPU
+:sync: tpu
+
+```{include} tpu.inc.md
+:start-after: "## Configure a new environment"
+:end-before: "## Set up using Python"
+```
+
+:::
+
+:::{tab-item} Intel Gaudi
+:sync: hpu-gaudi
+
+```{include} hpu-gaudi.inc.md
+:start-after: "## Configure a new environment"
+:end-before: "## Set up using Python"
+```
+
+:::
+
+:::{tab-item} Neuron
+:sync: neuron
+
+```{include} neuron.inc.md
+:start-after: "## Configure a new environment"
+:end-before: "## Set up using Python"
+```
+
+:::
+
+:::{tab-item} OpenVINO
+:sync: openvino
+
+```{include} ../python_env_setup.inc.md
+```
+
+:::
+
+::::
+
+## Set up using Python
+
+### Pre-built wheels
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} TPU
+:sync: tpu
+
+```{include} tpu.inc.md
+:start-after: "### Pre-built wheels"
+:end-before: "### Build wheel from source"
+```
+
+:::
+
+:::{tab-item} Intel Gaudi
+:sync: hpu-gaudi
+
+```{include} hpu-gaudi.inc.md
+:start-after: "### Pre-built wheels"
+:end-before: "### Build wheel from source"
+```
+
+:::
+
+:::{tab-item} Neuron
+:sync: neuron
+
+```{include} neuron.inc.md
+:start-after: "### Pre-built wheels"
+:end-before: "### Build wheel from source"
+```
+
+:::
+
+:::{tab-item} OpenVINO
+:sync: openvino
+
+```{include} openvino.inc.md
+:start-after: "### Pre-built wheels"
+:end-before: "### Build wheel from source"
+```
+
+:::
+
+::::
+
+### Build wheel from source
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} TPU
+:sync: tpu
+
+```{include} tpu.inc.md
+:start-after: "### Build wheel from source"
+:end-before: "## Set up using Docker"
+```
+
+:::
+
+:::{tab-item} Intel Gaudi
+:sync: hpu-gaudi
+
+```{include} hpu-gaudi.inc.md
+:start-after: "### Build wheel from source"
+:end-before: "## Set up using Docker"
+```
+
+:::
+
+:::{tab-item} Neuron
+:sync: neuron
+
+```{include} neuron.inc.md
+:start-after: "### Build wheel from source"
+:end-before: "## Set up using Docker"
+```
+
+:::
+
+:::{tab-item} OpenVINO
+:sync: openvino
+
+```{include} openvino.inc.md
+:start-after: "### Build wheel from source"
+:end-before: "## Set up using Docker"
+```
+
+:::
+
+::::
+
+## Set up using Docker
+
+### Pre-built images
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} TPU
+:sync: tpu
+
+```{include} tpu.inc.md
+:start-after: "### Pre-built images"
+:end-before: "### Build image from source"
+```
+
+:::
+
+:::{tab-item} Intel Gaudi
+:sync: hpu-gaudi
+
+```{include} hpu-gaudi.inc.md
+:start-after: "### Pre-built images"
+:end-before: "### Build image from source"
+```
+
+:::
+
+:::{tab-item} Neuron
+:sync: neuron
+
+```{include} neuron.inc.md
+:start-after: "### Pre-built images"
+:end-before: "### Build image from source"
+```
+
+:::
+
+:::{tab-item} OpenVINO
+:sync: openvino
+
+```{include} openvino.inc.md
+:start-after: "### Pre-built images"
+:end-before: "### Build image from source"
+```
+
+:::
+
+::::
+
+### Build image from source
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} TPU
+:sync: tpu
+
+```{include} tpu.inc.md
+:start-after: "### Build image from source"
+:end-before: "## Extra information"
+```
+
+:::
+
+:::{tab-item} Intel Gaudi
+:sync: hpu-gaudi
+
+```{include} hpu-gaudi.inc.md
+:start-after: "### Build image from source"
+:end-before: "## Extra information"
+```
+
+:::
+
+:::{tab-item} Neuron
+:sync: neuron
+
+```{include} neuron.inc.md
+:start-after: "### Build image from source"
+:end-before: "## Extra information"
+```
+
+:::
+
+:::{tab-item} OpenVINO
+:sync: openvino
+
+```{include} openvino.inc.md
+:start-after: "### Build image from source"
+:end-before: "## Extra information"
+```
+
+:::
+
+::::
+
+## Extra information
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} TPU
+:sync: tpu
+
+```{include} tpu.inc.md
+:start-after: "## Extra information"
+```
+
+:::
+
+:::{tab-item} Intel Gaudi
+:sync: hpu-gaudi
+
+```{include} hpu-gaudi.inc.md
+:start-after: "## Extra information"
+```
+
+:::
+
+:::{tab-item} Neuron
+:sync: neuron
+
+```{include} neuron.inc.md
+:start-after: "## Extra information"
+```
+
+:::
+
+:::{tab-item} OpenVINO
+:sync: openvino
+
+```{include} openvino.inc.md
+:start-after: "## Extra information"
+```
+
+:::
+
+::::

docs/source/getting_started/installation/ai_accelerator/neuron.inc.md

@@ -1,6 +1,4 @@
-(installation-neuron)=
-
-# Installation for Neuron
+# Installation
 
 vLLM 0.3.3 onwards supports model inferencing and serving on AWS Trainium/Inferentia with Neuron SDK with continuous batching.
 Paged Attention and Chunked Prefill are currently in development and will be available soon.
@@ -14,28 +12,9 @@ Data types currently supported in Neuron SDK are FP16 and BF16.
 - Pytorch 2.0.1/2.1.1
 - AWS Neuron SDK 2.16/2.17 (Verified on python 3.8)
 
-Installation steps:
+## Configure a new environment
 
-- [Build from source](#build-from-source-neuron)
-
-  - [Step 0. Launch Trn1/Inf2 instances](#launch-instances)
-  - [Step 1. Install drivers and tools](#install-drivers)
-  - [Step 2. Install transformers-neuronx and its dependencies](#install-tnx)
-  - [Step 3. Install vLLM from source](#install-vllm)
-
-(build-from-source-neuron)=
-
-```{note}
-The currently supported version of Pytorch for Neuron installs `triton` version `2.1.0`. This is incompatible with `vllm >= 0.5.3`. You may see an error `cannot import name 'default_dump_dir...`. To work around this, run a `pip install --upgrade triton==3.0.0` after installing the vLLM wheel.
-```
-
-## Build from source
-
-Following instructions are applicable to Neuron SDK 2.16 and beyond.
-
-(launch-instances)=
-
-### Step 0. Launch Trn1/Inf2 instances
+### Launch Trn1/Inf2 instances
 
 Here are the steps to launch trn1/inf2 instances, in order to install [PyTorch Neuron ("torch-neuronx") Setup on Ubuntu 22.04 LTS](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/general/setup/neuron-setup/pytorch/neuronx/ubuntu/torch-neuronx-ubuntu22.html).
 
@@ -45,9 +24,7 @@ Here are the steps to launch trn1/inf2 instances, in order to install [PyTorch N
 - When launching a Trn1/Inf2, please adjust your primary EBS volume size to a minimum of 512GB.
 - After launching the instance, follow the instructions in [Connect to your instance](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/AccessingInstancesLinux.html) to connect to the instance
 
-(install-drivers)=
-
-### Step 1. Install drivers and tools
+### Install drivers and tools
 
 The installation of drivers and tools wouldn't be necessary, if [Deep Learning AMI Neuron](https://docs.aws.amazon.com/dlami/latest/devguide/appendix-ami-release-notes.html) is installed. In case the drivers and tools are not installed on the operating system, follow the steps below:
 
@@ -82,9 +59,21 @@ sudo apt-get install aws-neuronx-tools=2.* -y
 export PATH=/opt/aws/neuron/bin:$PATH
 ```
 
-(install-tnx)=
+## Set up using Python
 
-### Step 2. Install transformers-neuronx and its dependencies
+### Pre-built wheels
+
+Currently, there are no pre-built Neuron wheels.
+
+### Build wheel from source
+
+```{note}
+The currently supported version of Pytorch for Neuron installs `triton` version `2.1.0`. This is incompatible with `vllm >= 0.5.3`. You may see an error `cannot import name 'default_dump_dir...`. To work around this, run a `pip install --upgrade triton==3.0.0` after installing the vLLM wheel.
+```
+
+Following instructions are applicable to Neuron SDK 2.16 and beyond.
+
+#### Install transformers-neuronx and its dependencies
 
 [transformers-neuronx](https://github.com/aws-neuron/transformers-neuronx) will be the backend to support inference on trn1/inf2 instances.
 Follow the steps below to install transformer-neuronx package and its dependencies.
@@ -116,17 +105,31 @@ python -m pip install awscli
 python -m pip install --upgrade neuronx-cc==2.* --pre torch-neuronx==2.1.* torchvision transformers-neuronx
 ```
 
-(install-vllm)=
-
-### Step 3. Install vLLM from source
+#### Install vLLM from source
 
 Once neuronx-cc and transformers-neuronx packages are installed, we will be able to install vllm as follows:
 
 ```console
-$ git clone https://github.com/vllm-project/vllm.git
-$ cd vllm
-$ pip install -U -r requirements-neuron.txt
-$ VLLM_TARGET_DEVICE="neuron" pip install .
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+pip install -U -r requirements-neuron.txt
+VLLM_TARGET_DEVICE="neuron" pip install .
 ```
 
 If neuron packages are detected correctly in the installation process, `vllm-0.3.0+neuron212` will be installed.
+
+## Set up using Docker
+
+### Pre-built images
+
+Currently, there are no pre-built Neuron images.
+
+### Build image from source
+
+See <project:#deployment-docker-build-image-from-source> for instructions on building the Docker image.
+
+Make sure to use <gh-file:Dockerfile.neuron> in place of the default Dockerfile.
+
+## Extra information
+
+There is no extra information for this device.

docs/source/getting_started/installation/ai_accelerator/openvino.inc.md

@@ -1,63 +1,65 @@
-(installation-openvino)=
+# Installation
 
-# Installation for OpenVINO
-
-vLLM powered by OpenVINO supports all LLM models from [vLLM supported models list](#supported-models) and can perform optimal model serving on all x86-64 CPUs with, at least, AVX2 support, as well as on both integrated and discrete Intel® GPUs ([the list of supported GPUs](https://docs.openvino.ai/2024/about-openvino/release-notes-openvino/system-requirements.html#gpu)). OpenVINO vLLM backend supports the following advanced vLLM features:
-
-- Prefix caching (`--enable-prefix-caching`)
-- Chunked prefill (`--enable-chunked-prefill`)
-
-**Table of contents**:
-
-- [Requirements](#openvino-backend-requirements)
-- [Quick start using Dockerfile](#openvino-backend-quick-start-dockerfile)
-- [Build from source](#install-openvino-backend-from-source)
-- [Performance tips](#openvino-backend-performance-tips)
-- [Limitations](#openvino-backend-limitations)
-
-(openvino-backend-requirements)=
+vLLM powered by OpenVINO supports all LLM models from [vLLM supported models list](#supported-models) and can perform optimal model serving on all x86-64 CPUs with, at least, AVX2 support, as well as on both integrated and discrete Intel® GPUs ([the list of supported GPUs](https://docs.openvino.ai/2024/about-openvino/release-notes-openvino/system-requirements.html#gpu)).
 
 ## Requirements
 
 - OS: Linux
 - Instruction set architecture (ISA) requirement: at least AVX2.
 
-(openvino-backend-quick-start-dockerfile)=
+## Set up using Python
 
-## Quick start using Dockerfile
+### Pre-built wheels
+
+Currently, there are no pre-built OpenVINO wheels.
+
+### Build wheel from source
+
+First, install Python. For example, on Ubuntu 22.04, you can run:
 
 ```console
-$ docker build -f Dockerfile.openvino -t vllm-openvino-env .
-$ docker run -it --rm vllm-openvino-env
+sudo apt-get update  -y
+sudo apt-get install python3
 ```
 
-(install-openvino-backend-from-source)=
+Second, install prerequisites vLLM OpenVINO backend installation:
 
-## Install from source
+```console
+pip install --upgrade pip
+pip install -r requirements-build.txt --extra-index-url https://download.pytorch.org/whl/cpu
+```
 
-- First, install Python. For example, on Ubuntu 22.04, you can run:
+Finally, install vLLM with OpenVINO backend:
 
-  ```console
-  $ sudo apt-get update  -y
-  $ sudo apt-get install python3
-  ```
+```console
+PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" VLLM_TARGET_DEVICE=openvino python -m pip install -v .
+```
 
-- Second, install prerequisites vLLM OpenVINO backend installation:
+:::{tip}
+To use vLLM OpenVINO backend with a GPU device, ensure your system is properly set up. Follow the instructions provided here: [https://docs.openvino.ai/2024/get-started/configurations/configurations-intel-gpu.html](https://docs.openvino.ai/2024/get-started/configurations/configurations-intel-gpu.html).
+:::
 
-  ```console
-  $ pip install --upgrade pip
-  $ pip install -r requirements-build.txt --extra-index-url https://download.pytorch.org/whl/cpu
-  ```
+## Set up using Docker
 
-- Finally, install vLLM with OpenVINO backend:
+### Pre-built images
 
-  ```console
-  $ PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" VLLM_TARGET_DEVICE=openvino python -m pip install -v .
-  ```
+Currently, there are no pre-built OpenVINO images.
 
-- [Optional] To use vLLM OpenVINO backend with a GPU device, ensure your system is properly set up. Follow the instructions provided here: [https://docs.openvino.ai/2024/get-started/configurations/configurations-intel-gpu.html](https://docs.openvino.ai/2024/get-started/configurations/configurations-intel-gpu.html).
+### Build image from source
 
-(openvino-backend-performance-tips)=
+```console
+docker build -f Dockerfile.openvino -t vllm-openvino-env .
+docker run -it --rm vllm-openvino-env
+```
+
+## Extra information
+
+## Supported features
+
+OpenVINO vLLM backend supports the following advanced vLLM features:
+
+- Prefix caching (`--enable-prefix-caching`)
+- Chunked prefill (`--enable-chunked-prefill`)
 
 ## Performance tips
 
@@ -95,8 +97,6 @@ $ VLLM_OPENVINO_DEVICE=GPU VLLM_OPENVINO_ENABLE_QUANTIZED_WEIGHTS=ON \
     python3 vllm/benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-chat-hf --dataset vllm/benchmarks/ShareGPT_V3_unfiltered_cleaned_split.json
 ```
 
-(openvino-backend-limitations)=
-
 ## Limitations
 
 - LoRA serving is not supported.

docs/source/getting_started/installation/ai_accelerator/tpu.inc.md

@@ -1,6 +1,4 @@
-(installation-tpu)=
-
-# Installation for TPUs
+# Installation
 
 Tensor Processing Units (TPUs) are Google's custom-developed application-specific
 integrated circuits (ASICs) used to accelerate machine learning workloads. TPUs
@@ -54,7 +52,16 @@ In all of the following commands, replace the ALL CAPS parameter names with
 appropriate values. See the parameter descriptions table for more information.
 ```
 
-## Provision a Cloud TPU with the queued resource API
+### Provision Cloud TPUs with GKE
+
+For more information about using TPUs with GKE, see:
+- <https://cloud.google.com/kubernetes-engine/docs/how-to/tpus>
+- <https://cloud.google.com/kubernetes-engine/docs/concepts/tpus>
+- <https://cloud.google.com/kubernetes-engine/docs/concepts/plan-tpus>
+
+## Configure a new environment
+
+### Provision a Cloud TPU with the queued resource API
 
 Create a TPU v5e with 4 TPU chips:
 
@@ -102,6 +109,14 @@ Connect to your TPU using SSH:
 gcloud compute tpus tpu-vm ssh TPU_NAME --zone ZONE
 ```
 
+## Set up using Python
+
+### Pre-built wheels
+
+Currently, there are no pre-built TPU wheels.
+
+### Build wheel from source
+
 Install Miniconda:
 
 ```bash
@@ -142,28 +157,25 @@ Run the setup script:
 VLLM_TARGET_DEVICE="tpu" python setup.py develop
 ```
 
-## Provision Cloud TPUs with GKE
+## Set up using Docker
 
-For more information about using TPUs with GKE, see
-<https://cloud.google.com/kubernetes-engine/docs/how-to/tpus>
-<https://cloud.google.com/kubernetes-engine/docs/concepts/tpus>
-<https://cloud.google.com/kubernetes-engine/docs/concepts/plan-tpus>
+### Pre-built images
 
-(build-docker-tpu)=
+See <project:#deployment-docker-pre-built-image> for instructions on using the official Docker image, making sure to substitute the image name `vllm/vllm-openai` with `vllm/vllm-tpu`.
 
-## Build a docker image with {code}`Dockerfile.tpu`
+### Build image from source
 
 You can use <gh-file:Dockerfile.tpu> to build a Docker image with TPU support.
 
 ```console
-$ docker build -f Dockerfile.tpu -t vllm-tpu .
+docker build -f Dockerfile.tpu -t vllm-tpu .
 ```
 
 Run the Docker image with the following command:
 
 ```console
-$ # Make sure to add `--privileged --net host --shm-size=16G`.
-$ docker run --privileged --net host --shm-size=16G -it vllm-tpu
+# Make sure to add `--privileged --net host --shm-size=16G`.
+docker run --privileged --net host --shm-size=16G -it vllm-tpu
 ```
 
 ```{note}
@@ -189,3 +201,7 @@ Install OpenBLAS with the following command:
 $ sudo apt-get install libopenblas-base libopenmpi-dev libomp-dev
 ```
 ````
+
+## Extra information
+
+There is no extra information for this device.

docs/source/getting_started/installation/cpu-arm.md

@@ -1,46 +0,0 @@
-(installation-arm)=
-
-# Installation for ARM CPUs
-
-vLLM has been adapted to work on ARM64 CPUs with NEON support, leveraging the CPU backend initially developed for the x86 platform. This guide provides installation instructions specific to ARM (which also apply to Apple Silicon, see [Installation for macOS](#installation-apple) for more). For additional details on supported features, refer to the [x86 CPU documentation](#installation-x86) covering:
-
-- CPU backend inference capabilities
-- Relevant runtime environment variables
-- Performance optimization tips
-
-ARM CPU backend currently supports Float32, FP16 and BFloat16 datatypes.
-Contents:
-
-1. [Requirements](#arm-backend-requirements)
-2. [Quick Start with Dockerfile](#arm-backend-quick-start-dockerfile)
-3. [Building from Source](#build-arm-backend-from-source)
-
-(arm-backend-requirements)=
-
-## Requirements
-
-- **Operating System**: Linux or macOS
-- **Compilers**: `gcc/g++ >= 12.3.0` (optional, but recommended) or `Apple Clang >= 15.0.0` for macOS
-- **Instruction Set Architecture (ISA)**: NEON support is required
-
-(arm-backend-quick-start-dockerfile)=
-
-## Quick Start with Dockerfile
-
-You can quickly set up vLLM on ARM using Docker:
-
-```console
-$ docker build -f Dockerfile.arm -t vllm-cpu-env --shm-size=4g .
-$ docker run -it \
-             --rm \
-             --network=host \
-             --cpuset-cpus=<cpu-id-list, optional> \
-             --cpuset-mems=<memory-node, optional> \
-             vllm-cpu-env
-```
-
-(build-arm-backend-from-source)=
-
-## Building from Source
-
-To build vLLM from source on Ubuntu 22.04 or other Linux distributions, follow a similar process as with x86. Testing has been conducted on AWS Graviton3 instances for compatibility.

docs/source/getting_started/installation/cpu/apple.inc.md

@@ -1,42 +1,40 @@
-(installation-apple)=
+# Installation
 
-# Installation for macOS
-
-vLLM has experimental support for macOS with Apple Silicon. For now, users shall build from the source vLLM to natively run on macOS. For more details, like running on vLLM in a docker container, see [ARM CPU Documentation](installation-arm)
+vLLM has experimental support for macOS with Apple silicon. For now, users shall build from the source vLLM to natively run on macOS.
 
 Currently the CPU implementation for macOS supports FP32 and FP16 datatypes.
 
 ## Requirements
 
-- **Operating System**: `macOS Sonoma` or later
-- **SDK** `XCode 15.4` or later with Command Line Tools
-- **Compilers**: `Apple Clang >= 15.0.0`
+- OS: `macOS Sonoma` or later
+- SDK: `XCode 15.4` or later with Command Line Tools
+- Compiler: `Apple Clang >= 15.0.0`
 
-<!-- (arm-backend-quick-start-dockerfile)= -->
+## Set up using Python
 
-## Build and installation
+### Pre-built wheels
+
+### Build wheel from source
 
 After installation of XCode and the Command Line Tools, which include Apple Clang, execute the following commands to build and install vLLM from the source.
 
-```
-$ git clone https://github.com/vllm-project/vllm.git
-$ cd vllm
-$ pip install -r requirements-cpu.txt
-$ pip install -e . 
+```console
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+pip install -r requirements-cpu.txt
+pip install -e . 
 ```
 
 ```{note}
 On macOS the `VLLM_TARGET_DEVICE` is automatically set to `cpu`, which currently is the only supported device.
 ```
 
+#### Troubleshooting
 
-
-## Troubleshooting
-
-If the build has error like the following snippet where standard C++ headers cannot be found, try to remove and reinstall your 
+If the build has error like the following snippet where standard C++ headers cannot be found, try to remove and reinstall your
 [Command Line Tools for Xcode](https://developer.apple.com/download/all/).
 
-```
+```text
 [...] fatal error: 'map' file not found
           1 | #include <map>
             |          ^~~~~
@@ -49,3 +47,10 @@ If the build has error like the following snippet where standard C++ headers can
       1 error generated.
 ```
 
+## Set up using Docker
+
+### Pre-built images
+
+### Build image from source
+
+## Extra information

docs/source/getting_started/installation/cpu/arm.inc.md

@@ -0,0 +1,30 @@
+# Installation
+
+vLLM has been adapted to work on ARM64 CPUs with NEON support, leveraging the CPU backend initially developed for the x86 platform.
+
+ARM CPU backend currently supports Float32, FP16 and BFloat16 datatypes.
+
+## Requirements
+
+- OS: Linux
+- Compiler: `gcc/g++ >= 12.3.0` (optional, recommended)
+- Instruction Set Architecture (ISA): NEON support is required
+
+## Set up using Python
+
+### Pre-built wheels
+
+### Build wheel from source
+
+:::{include} build.inc.md
+:::
+
+Testing has been conducted on AWS Graviton3 instances for compatibility.
+
+## Set up using Docker
+
+### Pre-built images
+
+### Build image from source
+
+## Extra information

docs/source/getting_started/installation/cpu/build.inc.md

@@ -0,0 +1,21 @@
+First, install recommended compiler. We recommend to use `gcc/g++ >= 12.3.0` as the default compiler to avoid potential problems. For example, on Ubuntu 22.4, you can run:
+
+```console
+sudo apt-get update  -y
+sudo apt-get install -y gcc-12 g++-12 libnuma-dev
+sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
+```
+
+Second, install Python packages for vLLM CPU backend building:
+
+```console
+pip install --upgrade pip
+pip install cmake>=3.26 wheel packaging ninja "setuptools-scm>=8" numpy
+pip install -v -r requirements-cpu.txt --extra-index-url https://download.pytorch.org/whl/cpu
+```
+
+Finally, build and install vLLM CPU backend:
+
+```console
+VLLM_TARGET_DEVICE=cpu python setup.py install
+```

docs/source/getting_started/installation/cpu/index.md

@@ -1,35 +1,136 @@
-(installation-x86)=
+# CPU
 
-# Installation for x86 CPUs
+vLLM is a Python library that supports the following CPU variants. Select your CPU type to see vendor specific instructions:
 
-vLLM initially supports basic model inferencing and serving on x86 CPU platform, with data types FP32, FP16 and BF16. vLLM CPU backend supports the following vLLM features:
+::::{tab-set}
+:sync-group: device
 
-- Tensor Parallel
-- Model Quantization (`INT8 W8A8, AWQ`)
-- Chunked-prefill
-- Prefix-caching
-- FP8-E5M2 KV-Caching (TODO)
+:::{tab-item} x86
+:sync: x86
 
-Table of contents:
+```{include} x86.inc.md
+:start-after: "# Installation"
+:end-before: "## Requirements"
+```
 
-1. [Requirements](#cpu-backend-requirements)
-2. [Quick start using Dockerfile](#cpu-backend-quick-start-dockerfile)
-3. [Build from source](#build-cpu-backend-from-source)
-4. [Related runtime environment variables](#env-intro)
-5. [Intel Extension for PyTorch](#ipex-guidance)
-6. [Performance tips](#cpu-backend-performance-tips)
+:::
 
-(cpu-backend-requirements)=
+:::{tab-item} ARM
+:sync: arm
+
+```{include} arm.inc.md
+:start-after: "# Installation"
+:end-before: "## Requirements"
+```
+
+:::
+
+:::{tab-item} Apple silicon
+:sync: apple
+
+```{include} apple.inc.md
+:start-after: "# Installation"
+:end-before: "## Requirements"
+```
+
+:::
+
+::::
 
 ## Requirements
 
-- OS: Linux
-- Compiler: `gcc/g++>=12.3.0` (optional, recommended)
-- Instruction set architecture (ISA) requirement: AVX512 (optional, recommended)
+- Python: 3.9 -- 3.12
 
-(cpu-backend-quick-start-dockerfile)=
+::::{tab-set}
+:sync-group: device
 
-## Quick start using Dockerfile
+:::{tab-item} x86
+:sync: x86
+
+```{include} x86.inc.md
+:start-after: "## Requirements"
+:end-before: "## Set up using Python"
+```
+
+:::
+
+:::{tab-item} ARM
+:sync: arm
+
+```{include} arm.inc.md
+:start-after: "## Requirements"
+:end-before: "## Set up using Python"
+```
+
+:::
+
+:::{tab-item} Apple silicon
+:sync: apple
+
+```{include} apple.inc.md
+:start-after: "## Requirements"
+:end-before: "## Set up using Python"
+```
+
+:::
+
+::::
+
+## Set up using Python
+
+### Create a new Python environment
+
+```{include} ../python_env_setup.inc.md
+```
+
+### Pre-built wheels
+
+Currently, there are no pre-built CPU wheels.
+
+### Build wheel from source
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} x86
+:sync: x86
+
+```{include} x86.inc.md
+:start-after: "### Build wheel from source"
+:end-before: "## Set up using Docker"
+```
+
+:::
+
+:::{tab-item} ARM
+:sync: arm
+
+```{include} arm.inc.md
+:start-after: "### Build wheel from source"
+:end-before: "## Set up using Docker"
+```
+
+:::
+
+:::{tab-item} Apple silicon
+:sync: apple
+
+```{include} apple.inc.md
+:start-after: "### Build wheel from source"
+:end-before: "## Set up using Docker"
+```
+
+:::
+
+::::
+
+## Set up using Docker
+
+### Pre-built images
+
+Currently, there are no pre-build CPU images.
+
+### Build image from source
 
 ```console
 $ docker build -f Dockerfile.cpu -t vllm-cpu-env --shm-size=4g .
@@ -41,69 +142,42 @@ $ docker run -it \
              vllm-cpu-env
 ```
 
-(build-cpu-backend-from-source)=
+:::{tip}
+For ARM or Apple silicon, use `Dockerfile.arm`
+:::
 
-## Build from source
+## Supported features
 
-- First, install recommended compiler. We recommend to use `gcc/g++ >= 12.3.0` as the default compiler to avoid potential problems. For example, on Ubuntu 22.4, you can run:
+vLLM CPU backend supports the following vLLM features:
 
-```console
-$ sudo apt-get update  -y
-$ sudo apt-get install -y gcc-12 g++-12 libnuma-dev
-$ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
-```
-
-- Second, install Python packages for vLLM CPU backend building:
-
-```console
-$ pip install --upgrade pip
-$ pip install cmake>=3.26 wheel packaging ninja "setuptools-scm>=8" numpy
-$ pip install -v -r requirements-cpu.txt --extra-index-url https://download.pytorch.org/whl/cpu
-```
-
-- Finally, build and install vLLM CPU backend:
-
-```console
-$ VLLM_TARGET_DEVICE=cpu python setup.py install
-```
-
-```{note}
-- AVX512_BF16 is an extension ISA provides native BF16 data type conversion and vector product instructions, will brings some performance improvement compared with pure AVX512. The CPU backend build script will check the host CPU flags to determine whether to enable AVX512_BF16.
-- If you want to force enable AVX512_BF16 for the cross-compilation, please set environment variable `VLLM_CPU_AVX512BF16=1` before the building.
-```
-
-(env-intro)=
+- Tensor Parallel
+- Model Quantization (`INT8 W8A8, AWQ, GPTQ`)
+- Chunked-prefill
+- Prefix-caching
+- FP8-E5M2 KV-Caching (TODO)
 
 ## Related runtime environment variables
 
 - `VLLM_CPU_KVCACHE_SPACE`: specify the KV Cache size (e.g, `VLLM_CPU_KVCACHE_SPACE=40` means 40 GB space for KV cache), larger setting will allow vLLM running more requests in parallel. This parameter should be set based on the hardware configuration and memory management pattern of users.
 - `VLLM_CPU_OMP_THREADS_BIND`: specify the CPU cores dedicated to the OpenMP threads. For example, `VLLM_CPU_OMP_THREADS_BIND=0-31` means there will be 32 OpenMP threads bound on 0-31 CPU cores. `VLLM_CPU_OMP_THREADS_BIND=0-31|32-63` means there will be 2 tensor parallel processes, 32 OpenMP threads of rank0 are bound on 0-31 CPU cores, and the OpenMP threads of rank1 are bound on 32-63 CPU cores.
 
-(ipex-guidance)=
-
-## Intel Extension for PyTorch
-
-- [Intel Extension for PyTorch (IPEX)](https://github.com/intel/intel-extension-for-pytorch) extends PyTorch with up-to-date features optimizations for an extra performance boost on Intel hardware.
-
-(cpu-backend-performance-tips)=
-
 ## Performance tips
 
 - We highly recommend to use TCMalloc for high performance memory allocation and better cache locality. For example, on Ubuntu 22.4, you can run:
 
 ```console
-$ sudo apt-get install libtcmalloc-minimal4 # install TCMalloc library
-$ find / -name *libtcmalloc* # find the dynamic link library path
-$ export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$LD_PRELOAD # prepend the library to LD_PRELOAD
-$ python examples/offline_inference/offline_inference.py # run vLLM
+sudo apt-get install libtcmalloc-minimal4 # install TCMalloc library
+find / -name *libtcmalloc* # find the dynamic link library path
+export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$LD_PRELOAD # prepend the library to LD_PRELOAD
+python examples/offline_inference/basic.py # run vLLM
 ```
 
 - When using the online serving, it is recommended to reserve 1-2 CPU cores for the serving framework to avoid CPU oversubscription. For example, on a platform with 32 physical CPU cores, reserving CPU 30 and 31 for the framework and using CPU 0-29 for OpenMP:
 
 ```console
-$ export VLLM_CPU_KVCACHE_SPACE=40
-$ export VLLM_CPU_OMP_THREADS_BIND=0-29
-$ vllm serve facebook/opt-125m
+export VLLM_CPU_KVCACHE_SPACE=40
+export VLLM_CPU_OMP_THREADS_BIND=0-29
+vllm serve facebook/opt-125m
 ```
 
 - If using vLLM CPU backend on a machine with hyper-threading, it is recommended to bind only one OpenMP thread on each physical CPU core using `VLLM_CPU_OMP_THREADS_BIND`. On a hyper-threading enabled platform with 16 logical CPU cores / 8 physical CPU cores:
@@ -132,23 +206,23 @@ CPU NODE SOCKET CORE L1d:L1i:L2:L3 ONLINE    MAXMHZ   MINMHZ      MHZ
 
 # On this platform, it is recommend to only bind openMP threads on logical CPU cores 0-7 or 8-15
 $ export VLLM_CPU_OMP_THREADS_BIND=0-7
-$ python examples/offline_inference/offline_inference.py
+$ python examples/offline_inference/basic.py
 ```
 
 - If using vLLM CPU backend on a multi-socket machine with NUMA, be aware to set CPU cores using `VLLM_CPU_OMP_THREADS_BIND` to avoid cross NUMA node memory access.
 
-## CPU Backend Considerations
+## Other considerations
 
 - The CPU backend significantly differs from the GPU backend since the vLLM architecture was originally optimized for GPU use. A number of optimizations are needed to enhance its performance.
 
 - Decouple the HTTP serving components from the inference components. In a GPU backend configuration, the HTTP serving and tokenization tasks operate on the CPU, while inference runs on the GPU, which typically does not pose a problem. However, in a CPU-based setup, the HTTP serving and tokenization can cause significant context switching and reduced cache efficiency. Therefore, it is strongly recommended to segregate these two components for improved performance.
 
-- On CPU based setup with NUMA enabled, the memory access performance may be largely impacted by the [topology](https://github.com/intel/intel-extension-for-pytorch/blob/main/docs/tutorials/performance_tuning/tuning_guide.md#non-uniform-memory-access-numa). For NUMA architecture, two optimizations are to recommended: Tensor Parallel or Data Parallel.
+- On CPU based setup with NUMA enabled, the memory access performance may be largely impacted by the [topology](https://github.com/intel/intel-extension-for-pytorch/blob/main/docs/tutorials/performance_tuning/tuning_guide.inc.md#non-uniform-memory-access-numa). For NUMA architecture, two optimizations are to recommended: Tensor Parallel or Data Parallel.
 
   - Using Tensor Parallel for a latency constraints deployment: following GPU backend design, a Megatron-LM's parallel algorithm will be used to shard the model, based on the number of NUMA nodes (e.g. TP = 2 for a two NUMA node system). With [TP feature on CPU](gh-pr:6125) merged, Tensor Parallel is supported for serving and offline inferencing. In general each NUMA node is treated as one GPU card. Below is the example script to enable Tensor Parallel = 2 for serving:
 
     ```console
-    $ VLLM_CPU_KVCACHE_SPACE=40 VLLM_CPU_OMP_THREADS_BIND="0-31|32-63" vllm serve meta-llama/Llama-2-7b-chat-hf -tp=2 --distributed-executor-backend mp
+    VLLM_CPU_KVCACHE_SPACE=40 VLLM_CPU_OMP_THREADS_BIND="0-31|32-63" vllm serve meta-llama/Llama-2-7b-chat-hf -tp=2 --distributed-executor-backend mp
     ```
 
-  - Using Data Parallel for maximum throughput: to launch an LLM serving endpoint on each NUMA node along with one additional load balancer to dispatch the requests to those endpoints. Common solutions like [Nginx](#nginxloadbalancer) or HAProxy are recommended. Anyscale Ray project provides the feature on LLM [serving](https://docs.ray.io/en/latest/serve/index.html). Here is the example to setup a scalable LLM serving with [Ray Serve](https://github.com/intel/llm-on-ray/blob/main/docs/setup.md).
+  - Using Data Parallel for maximum throughput: to launch an LLM serving endpoint on each NUMA node along with one additional load balancer to dispatch the requests to those endpoints. Common solutions like [Nginx](#nginxloadbalancer) or HAProxy are recommended. Anyscale Ray project provides the feature on LLM [serving](https://docs.ray.io/en/latest/serve/index.html). Here is the example to setup a scalable LLM serving with [Ray Serve](https://github.com/intel/llm-on-ray/blob/main/docs/setup.inc.md).

docs/source/getting_started/installation/cpu/x86.inc.md

@@ -0,0 +1,35 @@
+# Installation
+
+vLLM initially supports basic model inferencing and serving on x86 CPU platform, with data types FP32, FP16 and BF16.
+
+## Requirements
+
+- OS: Linux
+- Compiler: `gcc/g++ >= 12.3.0` (optional, recommended)
+- Instruction Set Architecture (ISA): AVX512 (optional, recommended)
+
+## Set up using Python
+
+### Pre-built wheels
+
+### Build wheel from source
+
+:::{include} build.inc.md
+:::
+
+```{note}
+- AVX512_BF16 is an extension ISA provides native BF16 data type conversion and vector product instructions, will brings some performance improvement compared with pure AVX512. The CPU backend build script will check the host CPU flags to determine whether to enable AVX512_BF16.
+- If you want to force enable AVX512_BF16 for the cross-compilation, please set environment variable `VLLM_CPU_AVX512BF16=1` before the building.
+```
+
+## Set up using Docker
+
+### Pre-built images
+
+### Build image from source
+
+## Extra information
+
+## Intel Extension for PyTorch
+
+- [Intel Extension for PyTorch (IPEX)](https://github.com/intel/intel-extension-for-pytorch) extends PyTorch with up-to-date features optimizations for an extra performance boost on Intel hardware.

docs/source/getting_started/installation/device.template.md

@@ -0,0 +1,17 @@
+# Installation
+
+## Requirements
+
+## Set up using Python
+
+### Pre-built wheels
+
+### Build wheel from source
+
+## Set up using Docker
+
+### Pre-built images
+
+### Build image from source
+
+## Extra information

docs/source/getting_started/installation/gpu-rocm.md

@@ -1,163 +0,0 @@
-(installation-rocm)=
-
-# Installation for ROCm
-
-vLLM supports AMD GPUs with ROCm 6.2.
-
-## Requirements
-
-- OS: Linux
-- Python: 3.9 -- 3.12
-- GPU: MI200s (gfx90a), MI300 (gfx942), Radeon RX 7900 series (gfx1100)
-- ROCm 6.2
-
-Installation options:
-
-1. [Build from source with docker](#build-from-source-docker-rocm)
-2. [Build from source](#build-from-source-rocm)
-
-(build-from-source-docker-rocm)=
-
-## Option 1: Build from source with docker (recommended)
-
-You can build and install vLLM from source.
-
-First, build a docker image from <gh-file:Dockerfile.rocm> and launch a docker container from the image.
-It is important that the user kicks off the docker build using buildkit. Either the user put DOCKER_BUILDKIT=1 as environment variable when calling docker build command, or the user needs to setup buildkit in the docker daemon configuration /etc/docker/daemon.json as follows and restart the daemon:
-
-```console
-{
-    "features": {
-        "buildkit": true
-    }
-}
-```
-
-<gh-file:Dockerfile.rocm> uses ROCm 6.2 by default, but also supports ROCm 5.7, 6.0 and 6.1 in older vLLM branches.
-It provides flexibility to customize the build of docker image using the following arguments:
-
-- `BASE_IMAGE`: specifies the base image used when running `docker build`, specifically the PyTorch on ROCm base image.
-- `BUILD_FA`: specifies whether to build CK flash-attention. The default is 1. For [Radeon RX 7900 series (gfx1100)](https://rocm.docs.amd.com/projects/radeon/en/latest/index.html), this should be set to 0 before flash-attention supports this target.
-- `FX_GFX_ARCHS`: specifies the GFX architecture that is used to build CK flash-attention, for example, `gfx90a;gfx942` for MI200 and MI300. The default is `gfx90a;gfx942`
-- `FA_BRANCH`: specifies the branch used to build the CK flash-attention in [ROCm's flash-attention repo](https://github.com/ROCmSoftwarePlatform/flash-attention). The default is `ae7928c`
-- `BUILD_TRITON`: specifies whether to build triton flash-attention. The default value is 1.
-
-Their values can be passed in when running `docker build` with `--build-arg` options.
-
-To build vllm on ROCm 6.2 for MI200 and MI300 series, you can use the default:
-
-```console
-$ DOCKER_BUILDKIT=1 docker build -f Dockerfile.rocm -t vllm-rocm .
-```
-
-To build vllm on ROCm 6.2 for Radeon RX7900 series (gfx1100), you should specify `BUILD_FA` as below:
-
-```console
-$ DOCKER_BUILDKIT=1 docker build --build-arg BUILD_FA="0" -f Dockerfile.rocm -t vllm-rocm .
-```
-
-To run the above docker image `vllm-rocm`, use the below command:
-
-```console
-$ docker run -it \
-   --network=host \
-   --group-add=video \
-   --ipc=host \
-   --cap-add=SYS_PTRACE \
-   --security-opt seccomp=unconfined \
-   --device /dev/kfd \
-   --device /dev/dri \
-   -v <path/to/model>:/app/model \
-   vllm-rocm \
-   bash
-```
-
-Where the `<path/to/model>` is the location where the model is stored, for example, the weights for llama2 or llama3 models.
-
-(build-from-source-rocm)=
-
-## Option 2: Build from source
-
-0. Install prerequisites (skip if you are already in an environment/docker with the following installed):
-
-- [ROCm](https://rocm.docs.amd.com/en/latest/deploy/linux/index.html)
-- [PyTorch](https://pytorch.org/)
-
-For installing PyTorch, you can start from a fresh docker image, e.g, `rocm/pytorch:rocm6.2_ubuntu20.04_py3.9_pytorch_release_2.3.0`, `rocm/pytorch-nightly`.
-
-Alternatively, you can install PyTorch using PyTorch wheels. You can check PyTorch installation guide in PyTorch [Getting Started](https://pytorch.org/get-started/locally/)
-
-1. Install [Triton flash attention for ROCm](https://github.com/ROCm/triton)
-
-Install ROCm's Triton flash attention (the default triton-mlir branch) following the instructions from [ROCm/triton](https://github.com/ROCm/triton/blob/triton-mlir/README.md)
-
-```console
-$ python3 -m pip install ninja cmake wheel pybind11
-$ pip uninstall -y triton
-$ git clone https://github.com/OpenAI/triton.git
-$ cd triton
-$ git checkout e192dba
-$ cd python
-$ pip3 install .
-$ cd ../..
-```
-
-```{note}
-- If you see HTTP issue related to downloading packages during building triton, please try again as the HTTP error is intermittent.
-```
-
-2. Optionally, if you choose to use CK flash attention, you can install [flash attention for ROCm](https://github.com/ROCm/flash-attention/tree/ck_tile)
-
-Install ROCm's flash attention (v2.5.9.post1) following the instructions from [ROCm/flash-attention](https://github.com/ROCm/flash-attention/tree/ck_tile#amd-gpurocm-support)
-Alternatively, wheels intended for vLLM use can be accessed under the releases.
-
-For example, for ROCm 6.2, suppose your gfx arch is `gfx90a`. To get your gfx architecture, run `rocminfo |grep gfx`.
-
-```console
-$ git clone https://github.com/ROCm/flash-attention.git
-$ cd flash-attention
-$ git checkout 3cea2fb
-$ git submodule update --init
-$ GPU_ARCHS="gfx90a" python3 setup.py install
-$ cd ..
-```
-
-```{note}
-- You might need to downgrade the "ninja" version to 1.10 it is not used when compiling flash-attention-2 (e.g. `pip install ninja==1.10.2.4`)
-```
-
-3. Build vLLM. For example, vLLM on ROCM 6.2 can be built with the following steps:
-
-```bash
-$ pip install --upgrade pip
-
-# Install PyTorch
-$ pip uninstall torch -y
-$ pip install --no-cache-dir --pre torch==2.6.0.dev20241024 --index-url https://download.pytorch.org/whl/nightly/rocm6.2
-
-# Build & install AMD SMI
-$ pip install /opt/rocm/share/amd_smi
-
-# Install dependencies
-$ pip install --upgrade numba scipy huggingface-hub[cli]
-$ pip install "numpy<2"
-$ pip install -r requirements-rocm.txt
-
-# Build vLLM for MI210/MI250/MI300.
-$ export PYTORCH_ROCM_ARCH="gfx90a;gfx942"
-$ python3 setup.py develop
-```
-
-This may take 5-10 minutes. Currently, `pip install .` does not work for ROCm installation.
-
-```{tip}
-- Triton flash attention is used by default. For benchmarking purposes, it is recommended to run a warm up step before collecting perf numbers.
-- Triton flash attention does not currently support sliding window attention. If using half precision, please use CK flash-attention for sliding window support.
-- To use CK flash-attention or PyTorch naive attention, please use this flag `export VLLM_USE_TRITON_FLASH_ATTN=0` to turn off triton flash attention.
-- The ROCm version of PyTorch, ideally, should match the ROCm driver version.
-```
-
-```{tip}
-- For MI300x (gfx942) users, to achieve optimal performance, please refer to [MI300x tuning guide](https://rocm.docs.amd.com/en/latest/how-to/tuning-guides/mi300x/index.html) for performance optimization and tuning tips on system and workflow level.
-  For vLLM, please refer to [vLLM performance optimization](https://rocm.docs.amd.com/en/latest/how-to/tuning-guides/mi300x/workload.html#vllm-performance-optimization).
-```

docs/source/getting_started/installation/gpu/cuda.inc.md

@@ -1,72 +1,52 @@
-(installation-cuda)=
+# Installation
 
-# Installation for CUDA
-
-vLLM is a Python library that also contains pre-compiled C++ and CUDA (12.1) binaries.
+vLLM contains pre-compiled C++ and CUDA (12.1) binaries.
 
 ## Requirements
 
-- OS: Linux
-- Python: 3.9 -- 3.12
 - GPU: compute capability 7.0 or higher (e.g., V100, T4, RTX20xx, A100, L4, H100, etc.)
 
-## Install released versions
+## Set up using Python
 
 ### Create a new Python environment
 
-You can create a new Python environment using `conda`:
-
-```console
-$ # (Recommended) Create a new conda environment.
-$ conda create -n myenv python=3.12 -y
-$ conda activate myenv
-```
-
 ```{note}
-[PyTorch has deprecated the conda release channel](https://github.com/pytorch/pytorch/issues/138506). If you use `conda`, please only use it to create Python environment rather than installing packages. In particular, the PyTorch installed via `conda` will statically link `NCCL` library, which can cause issues when vLLM tries to use `NCCL`. See <gh-issue:8420> for more details.
-```
-
-Or you can create a new Python environment using [uv](https://docs.astral.sh/uv/), a very fast Python environment manager. Please follow the [documentation](https://docs.astral.sh/uv/#getting-started) to install `uv`. After installing `uv`, you can create a new Python environment using the following command:
-
-```console
-$ # (Recommended) Create a new uv environment. Use `--seed` to install `pip` and `setuptools` in the environment.
-$ uv venv myenv --python 3.12 --seed
-$ source myenv/bin/activate
+PyTorch installed via `conda` will statically link `NCCL` library, which can cause issues when vLLM tries to use `NCCL`. See <gh-issue:8420> for more details.
 ```
 
 In order to be performant, vLLM has to compile many cuda kernels. The compilation unfortunately introduces binary incompatibility with other CUDA versions and PyTorch versions, even for the same PyTorch version with different building configurations.
 
 Therefore, it is recommended to install vLLM with a **fresh new** environment. If either you have a different CUDA version or you want to use an existing PyTorch installation, you need to build vLLM from source. See [below](#build-from-source) for more details.
 
-### Install vLLM
+### Pre-built wheels
 
 You can install vLLM using either `pip` or `uv pip`:
 
 ```console
-$ # Install vLLM with CUDA 12.1.
-$ pip install vllm # If you are using pip.
-$ uv pip install vllm # If you are using uv.
+# Install vLLM with CUDA 12.1.
+pip install vllm # If you are using pip.
+uv pip install vllm # If you are using uv.
 ```
 
 As of now, vLLM's binaries are compiled with CUDA 12.1 and public PyTorch release versions by default. We also provide vLLM binaries compiled with CUDA 11.8 and public PyTorch release versions:
 
 ```console
-$ # Install vLLM with CUDA 11.8.
-$ export VLLM_VERSION=0.6.1.post1
-$ export PYTHON_VERSION=310
-$ pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cu118-cp${PYTHON_VERSION}-cp${PYTHON_VERSION}-manylinux1_x86_64.whl --extra-index-url https://download.pytorch.org/whl/cu118
+# Install vLLM with CUDA 11.8.
+export VLLM_VERSION=0.6.1.post1
+export PYTHON_VERSION=310
+pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cu118-cp${PYTHON_VERSION}-cp${PYTHON_VERSION}-manylinux1_x86_64.whl --extra-index-url https://download.pytorch.org/whl/cu118
 ```
 
 (install-the-latest-code)=
 
-## Install the latest code
+#### Install the latest code
 
 LLM inference is a fast-evolving field, and the latest code may contain bug fixes, performance improvements, and new features that are not released yet. To allow users to try the latest code without waiting for the next release, vLLM provides wheels for Linux running on a x86 platform with CUDA 12 for every commit since `v0.5.3`.
 
-### Install the latest code using `pip`
+##### Install the latest code using `pip`
 
 ```console
-$ pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
+pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
 ```
 
 `--pre` is required for `pip` to consider pre-released versions.
@@ -74,82 +54,65 @@ $ pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
 If you want to access the wheels for previous commits (e.g. to bisect the behavior change, performance regression), due to the limitation of `pip`, you have to specify the full URL of the wheel file by embedding the commit hash in the URL:
 
 ```console
-$ export VLLM_COMMIT=33f460b17a54acb3b6cc0b03f4a17876cff5eafd # use full commit hash from the main branch
-$ pip install https://wheels.vllm.ai/${VLLM_COMMIT}/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl
+export VLLM_COMMIT=33f460b17a54acb3b6cc0b03f4a17876cff5eafd # use full commit hash from the main branch
+pip install https://wheels.vllm.ai/${VLLM_COMMIT}/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl
 ```
 
 Note that the wheels are built with Python 3.8 ABI (see [PEP 425](https://peps.python.org/pep-0425/) for more details about ABI), so **they are compatible with Python 3.8 and later**. The version string in the wheel file name (`1.0.0.dev`) is just a placeholder to have a unified URL for the wheels, the actual versions of wheels are contained in the wheel metadata (the wheels listed in the extra index url have correct versions). Although we don't support Python 3.8 any more (because PyTorch 2.5 dropped support for Python 3.8), the wheels are still built with Python 3.8 ABI to keep the same wheel name as before.
 
-### Install the latest code using `uv`
+##### Install the latest code using `uv`
 
 Another way to install the latest code is to use `uv`:
 
 ```console
-$ uv pip install vllm --extra-index-url https://wheels.vllm.ai/nightly
+uv pip install vllm --extra-index-url https://wheels.vllm.ai/nightly
 ```
 
 If you want to access the wheels for previous commits (e.g. to bisect the behavior change, performance regression), you can specify the commit hash in the URL:
 
 ```console
-$ export VLLM_COMMIT=72d9c316d3f6ede485146fe5aabd4e61dbc59069 # use full commit hash from the main branch
-$ uv pip install vllm --extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT}
+export VLLM_COMMIT=72d9c316d3f6ede485146fe5aabd4e61dbc59069 # use full commit hash from the main branch
+uv pip install vllm --extra-index-url https://wheels.vllm.ai/${VLLM_COMMIT}
 ```
 
 The `uv` approach works for vLLM `v0.6.6` and later and offers an easy-to-remember command. A unique feature of `uv` is that packages in `--extra-index-url` have [higher priority than the default index](https://docs.astral.sh/uv/pip/compatibility/#packages-that-exist-on-multiple-indexes). If the latest public release is `v0.6.6.post1`, `uv`'s behavior allows installing a commit before `v0.6.6.post1` by specifying the `--extra-index-url`. In contrast, `pip` combines packages from `--extra-index-url` and the default index, choosing only the latest version, which makes it difficult to install a development version prior to the released version.
 
-### Install the latest code using `docker`
+### Build wheel from source
 
-Another way to access the latest code is to use the docker images:
-
-```console
-$ export VLLM_COMMIT=33f460b17a54acb3b6cc0b03f4a17876cff5eafd # use full commit hash from the main branch
-$ docker pull public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:${VLLM_COMMIT}
-```
-
-These docker images are used for CI and testing only, and they are not intended for production use. They will be expired after several days.
-
-The latest code can contain bugs and may not be stable. Please use it with caution.
-
-(build-from-source)=
-
-## Build from source
-
-(python-only-build)=
-
-### Python-only build (without compilation)
+#### Set up using Python-only build (without compilation)
 
 If you only need to change Python code, you can build and install vLLM without compilation. Using `pip`'s [`--editable` flag](https://pip.pypa.io/en/stable/topics/local-project-installs/#editable-installs), changes you make to the code will be reflected when you run vLLM:
 
 ```console
-$ git clone https://github.com/vllm-project/vllm.git
-$ cd vllm
-$ VLLM_USE_PRECOMPILED=1 pip install --editable .
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+VLLM_USE_PRECOMPILED=1 pip install --editable .
 ```
 
-This will download the latest nightly wheel from https://wheels.vllm.ai/nightly/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl and use the compiled libraries from there in the installation.
+This will download the [latest nightly wheel](https://wheels.vllm.ai/nightly/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl) and use the compiled libraries from there in the installation.
 
 The `VLLM_PRECOMPILED_WHEEL_LOCATION` environment variable can be used instead of `VLLM_USE_PRECOMPILED` to specify a custom path or URL to the wheel file. For example, to use the [0.6.1.post1 PyPi wheel](https://pypi.org/project/vllm/#files):
 
 ```console
-$ export VLLM_PRECOMPILED_WHEEL_LOCATION=https://files.pythonhosted.org/packages/4a/4c/ee65ba33467a4c0de350ce29fbae39b9d0e7fcd887cc756fa993654d1228/vllm-0.6.3.post1-cp38-abi3-manylinux1_x86_64.whl
-$ pip install --editable .
+export VLLM_PRECOMPILED_WHEEL_LOCATION=https://files.pythonhosted.org/packages/4a/4c/ee65ba33467a4c0de350ce29fbae39b9d0e7fcd887cc756fa993654d1228/vllm-0.6.3.post1-cp38-abi3-manylinux1_x86_64.whl
+pip install --editable .
 ```
 
-You can find more information about vLLM's wheels [above](#install-the-latest-code).
+You can find more information about vLLM's wheels in <project:#install-the-latest-code>.
 
 ```{note}
 There is a possibility that your source code may have a different commit ID compared to the latest vLLM wheel, which could potentially lead to unknown errors.
-It is recommended to use the same commit ID for the source code as the vLLM wheel you have installed. Please refer to [the section above](#install-the-latest-code) for instructions on how to install a specified wheel.
+It is recommended to use the same commit ID for the source code as the vLLM wheel you have installed. Please refer to <project:#install-the-latest-code> for instructions on how to install a specified wheel.
 ```
 
-### Full build (with compilation)
+#### Full build (with compilation)
 
 If you want to modify C++ or CUDA code, you'll need to build vLLM from source. This can take several minutes:
 
 ```console
-$ git clone https://github.com/vllm-project/vllm.git
-$ cd vllm
-$ pip install -e .
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+pip install -e .
 ```
 
 ```{tip}
@@ -162,7 +125,7 @@ As long as `which ccache` command can find the `ccache` binary, it will be used
 The following environment variables can be set to configure the vLLM `sccache` remote: `SCCACHE_BUCKET=vllm-build-sccache SCCACHE_REGION=us-west-2 SCCACHE_S3_NO_CREDENTIALS=1`. We also recommend setting `SCCACHE_IDLE_TIMEOUT=0`.
 ```
 
-#### Use an existing PyTorch installation
+##### Use an existing PyTorch installation
 
 There are scenarios where the PyTorch dependency cannot be easily installed via pip, e.g.:
 
@@ -172,32 +135,32 @@ There are scenarios where the PyTorch dependency cannot be easily installed via
 To build vLLM using an existing PyTorch installation:
 
 ```console
-$ git clone https://github.com/vllm-project/vllm.git
-$ cd vllm
-$ python use_existing_torch.py
-$ pip install -r requirements-build.txt
-$ pip install -e . --no-build-isolation
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+python use_existing_torch.py
+pip install -r requirements-build.txt
+pip install -e . --no-build-isolation
 ```
 
-#### Use the local cutlass for compilation
+##### Use the local cutlass for compilation
 
 Currently, before starting the build process, vLLM fetches cutlass code from GitHub. However, there may be scenarios where you want to use a local version of cutlass instead.
 To achieve this, you can set the environment variable VLLM_CUTLASS_SRC_DIR to point to your local cutlass directory.
 
 ```console
-$ git clone https://github.com/vllm-project/vllm.git
-$ cd vllm
-$ VLLM_CUTLASS_SRC_DIR=/path/to/cutlass pip install -e .
+git clone https://github.com/vllm-project/vllm.git
+cd vllm
+VLLM_CUTLASS_SRC_DIR=/path/to/cutlass pip install -e .
 ```
 
-#### Troubleshooting
+##### Troubleshooting
 
 To avoid your system being overloaded, you can limit the number of compilation jobs
 to be run simultaneously, via the environment variable `MAX_JOBS`. For example:
 
 ```console
-$ export MAX_JOBS=6
-$ pip install -e .
+export MAX_JOBS=6
+pip install -e .
 ```
 
 This is especially useful when you are building on less powerful machines. For example, when you use WSL it only [assigns 50% of the total memory by default](https://learn.microsoft.com/en-us/windows/wsl/wsl-config#main-wsl-settings), so using `export MAX_JOBS=1` can avoid compiling multiple files simultaneously and running out of memory.
@@ -206,31 +169,56 @@ A side effect is a much slower build process.
 Additionally, if you have trouble building vLLM, we recommend using the NVIDIA PyTorch Docker image.
 
 ```console
-$ # Use `--ipc=host` to make sure the shared memory is large enough.
-$ docker run --gpus all -it --rm --ipc=host nvcr.io/nvidia/pytorch:23.10-py3
+# Use `--ipc=host` to make sure the shared memory is large enough.
+docker run --gpus all -it --rm --ipc=host nvcr.io/nvidia/pytorch:23.10-py3
 ```
 
 If you don't want to use docker, it is recommended to have a full installation of CUDA Toolkit. You can download and install it from [the official website](https://developer.nvidia.com/cuda-toolkit-archive). After installation, set the environment variable `CUDA_HOME` to the installation path of CUDA Toolkit, and make sure that the `nvcc` compiler is in your `PATH`, e.g.:
 
 ```console
-$ export CUDA_HOME=/usr/local/cuda
-$ export PATH="${CUDA_HOME}/bin:$PATH"
+export CUDA_HOME=/usr/local/cuda
+export PATH="${CUDA_HOME}/bin:$PATH"
 ```
 
 Here is a sanity check to verify that the CUDA Toolkit is correctly installed:
 
 ```console
-$ nvcc --version # verify that nvcc is in your PATH
-$ ${CUDA_HOME}/bin/nvcc --version # verify that nvcc is in your CUDA_HOME
+nvcc --version # verify that nvcc is in your PATH
+${CUDA_HOME}/bin/nvcc --version # verify that nvcc is in your CUDA_HOME
 ```
 
-### Unsupported OS build
+#### Unsupported OS build
 
 vLLM can fully run only on Linux but for development purposes, you can still build it on other systems (for example, macOS), allowing for imports and a more convenient development environment. The binaries will not be compiled and won't work on non-Linux systems.
 
 Simply disable the `VLLM_TARGET_DEVICE` environment variable before installing:
 
 ```console
-$ export VLLM_TARGET_DEVICE=empty
-$ pip install -e .
+export VLLM_TARGET_DEVICE=empty
+pip install -e .
 ```
+
+## Set up using Docker
+
+### Pre-built images
+
+See <project:#deployment-docker-pre-built-image> for instructions on using the official Docker image.
+
+Another way to access the latest code is to use the docker images:
+
+```console
+export VLLM_COMMIT=33f460b17a54acb3b6cc0b03f4a17876cff5eafd # use full commit hash from the main branch
+docker pull public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:${VLLM_COMMIT}
+```
+
+These docker images are used for CI and testing only, and they are not intended for production use. They will be expired after several days.
+
+The latest code can contain bugs and may not be stable. Please use it with caution.
+
+### Build image from source
+
+See <project:#deployment-docker-build-image-from-source> for instructions on building the Docker image.
+
+## Supported features
+
+See <project:#feature-x-hardware> compatibility matrix for feature support information.

docs/source/getting_started/installation/gpu/index.md

@@ -0,0 +1,300 @@
+# GPU
+
+vLLM is a Python library that supports the following GPU variants. Select your GPU type to see vendor specific instructions:
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} CUDA
+:sync: cuda
+
+```{include} cuda.inc.md
+:start-after: "# Installation"
+:end-before: "## Requirements"
+```
+
+:::
+
+:::{tab-item} ROCm
+:sync: rocm
+
+```{include} rocm.inc.md
+:start-after: "# Installation"
+:end-before: "## Requirements"
+```
+
+:::
+
+:::{tab-item} XPU
+:sync: xpu
+
+```{include} xpu.inc.md
+:start-after: "# Installation"
+:end-before: "## Requirements"
+```
+
+:::
+
+::::
+
+## Requirements
+
+- OS: Linux
+- Python: 3.9 -- 3.12
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} CUDA
+:sync: cuda
+
+```{include} cuda.inc.md
+:start-after: "## Requirements"
+:end-before: "## Set up using Python"
+```
+
+:::
+
+:::{tab-item} ROCm
+:sync: rocm
+
+```{include} rocm.inc.md
+:start-after: "## Requirements"
+:end-before: "## Set up using Python"
+```
+
+:::
+
+:::{tab-item} XPU
+:sync: xpu
+
+```{include} xpu.inc.md
+:start-after: "## Requirements"
+:end-before: "## Set up using Python"
+```
+
+:::
+
+::::
+
+## Set up using Python
+
+### Create a new Python environment
+
+```{include} ../python_env_setup.inc.md
+```
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} CUDA
+:sync: cuda
+
+```{include} cuda.inc.md
+:start-after: "## Create a new Python environment"
+:end-before: "### Pre-built wheels"
+```
+
+:::
+
+:::{tab-item} ROCm
+:sync: rocm
+
+There is no extra information on creating a new Python environment for this device.
+
+:::
+
+:::{tab-item} XPU
+:sync: xpu
+
+There is no extra information on creating a new Python environment for this device.
+
+:::
+
+::::
+
+### Pre-built wheels
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} CUDA
+:sync: cuda
+
+```{include} cuda.inc.md
+:start-after: "### Pre-built wheels"
+:end-before: "### Build wheel from source"
+```
+
+:::
+
+:::{tab-item} ROCm
+:sync: rocm
+
+```{include} rocm.inc.md
+:start-after: "### Pre-built wheels"
+:end-before: "### Build wheel from source"
+```
+
+:::
+
+:::{tab-item} XPU
+:sync: xpu
+
+```{include} xpu.inc.md
+:start-after: "### Pre-built wheels"
+:end-before: "### Build wheel from source"
+```
+
+:::
+
+::::
+
+(build-from-source)=
+
+### Build wheel from source
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} CUDA
+:sync: cuda
+
+```{include} cuda.inc.md
+:start-after: "### Build wheel from source"
+:end-before: "## Set up using Docker"
+```
+
+:::
+
+:::{tab-item} ROCm
+:sync: rocm
+
+```{include} rocm.inc.md
+:start-after: "### Build wheel from source"
+:end-before: "## Set up using Docker"
+```
+
+:::
+
+:::{tab-item} XPU
+:sync: xpu
+
+```{include} xpu.inc.md
+:start-after: "### Build wheel from source"
+:end-before: "## Set up using Docker"
+```
+
+:::
+
+::::
+
+## Set up using Docker
+
+### Pre-built images
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} CUDA
+:sync: cuda
+
+```{include} cuda.inc.md
+:start-after: "### Pre-built images"
+:end-before: "### Build image from source"
+```
+
+:::
+
+:::{tab-item} ROCm
+:sync: rocm
+
+```{include} rocm.inc.md
+:start-after: "### Pre-built images"
+:end-before: "### Build image from source"
+```
+
+:::
+
+:::{tab-item} XPU
+:sync: xpu
+
+```{include} xpu.inc.md
+:start-after: "### Pre-built images"
+:end-before: "### Build image from source"
+```
+
+:::
+
+::::
+
+### Build image from source
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} CUDA
+:sync: cuda
+
+```{include} cuda.inc.md
+:start-after: "### Build image from source"
+:end-before: "## Supported features"
+```
+
+:::
+
+:::{tab-item} ROCm
+:sync: rocm
+
+```{include} rocm.inc.md
+:start-after: "### Build image from source"
+:end-before: "## Supported features"
+```
+
+:::
+
+:::{tab-item} XPU
+:sync: xpu
+
+```{include} xpu.inc.md
+:start-after: "### Build image from source"
+:end-before: "## Supported features"
+```
+
+:::
+
+::::
+
+## Supported features
+
+::::{tab-set}
+:sync-group: device
+
+:::{tab-item} CUDA
+:sync: cuda
+
+```{include} cuda.inc.md
+:start-after: "## Supported features"
+```
+
+:::
+
+:::{tab-item} ROCm
+:sync: rocm
+
+```{include} rocm.inc.md
+:start-after: "## Supported features"
+```
+
+:::
+
+:::{tab-item} XPU
+:sync: xpu
+
+```{include} xpu.inc.md
+:start-after: "## Supported features"
+```
+
+:::
+
+::::

docs/source/getting_started/installation/gpu/rocm.inc.md

@@ -0,0 +1,166 @@
+# Installation
+
+vLLM supports AMD GPUs with ROCm 6.2.
+
+## Requirements
+
+- GPU: MI200s (gfx90a), MI300 (gfx942), Radeon RX 7900 series (gfx1100)
+- ROCm 6.2
+
+## Set up using Python
+
+### Pre-built wheels
+
+Currently, there are no pre-built ROCm wheels.
+
+### Build wheel from source
+
+0. Install prerequisites (skip if you are already in an environment/docker with the following installed):
+
+- [ROCm](https://rocm.docs.amd.com/en/latest/deploy/linux/index.html)
+- [PyTorch](https://pytorch.org/)
+
+    For installing PyTorch, you can start from a fresh docker image, e.g, `rocm/pytorch:rocm6.2_ubuntu20.04_py3.9_pytorch_release_2.3.0`, `rocm/pytorch-nightly`.
+
+    Alternatively, you can install PyTorch using PyTorch wheels. You can check PyTorch installation guide in PyTorch [Getting Started](https://pytorch.org/get-started/locally/)
+
+1. Install [Triton flash attention for ROCm](https://github.com/ROCm/triton)
+
+    Install ROCm's Triton flash attention (the default triton-mlir branch) following the instructions from [ROCm/triton](https://github.com/ROCm/triton/blob/triton-mlir/README.md)
+
+    ```console
+    python3 -m pip install ninja cmake wheel pybind11
+    pip uninstall -y triton
+    git clone https://github.com/OpenAI/triton.git
+    cd triton
+    git checkout e192dba
+    cd python
+    pip3 install .
+    cd ../..
+    ```
+
+    ```{note}
+    - If you see HTTP issue related to downloading packages during building triton, please try again as the HTTP error is intermittent.
+    ```
+
+2. Optionally, if you choose to use CK flash attention, you can install [flash attention for ROCm](https://github.com/ROCm/flash-attention/tree/ck_tile)
+
+    Install ROCm's flash attention (v2.5.9.post1) following the instructions from [ROCm/flash-attention](https://github.com/ROCm/flash-attention/tree/ck_tile#amd-gpurocm-support)
+    Alternatively, wheels intended for vLLM use can be accessed under the releases.
+
+    For example, for ROCm 6.2, suppose your gfx arch is `gfx90a`. To get your gfx architecture, run `rocminfo |grep gfx`.
+
+    ```console
+    git clone https://github.com/ROCm/flash-attention.git
+    cd flash-attention
+    git checkout 3cea2fb
+    git submodule update --init
+    GPU_ARCHS="gfx90a" python3 setup.py install
+    cd ..
+    ```
+
+    ```{note}
+    - You might need to downgrade the "ninja" version to 1.10 it is not used when compiling flash-attention-2 (e.g. `pip install ninja==1.10.2.4`)
+    ```
+
+3. Build vLLM. For example, vLLM on ROCM 6.2 can be built with the following steps:
+
+    ```bash
+    $ pip install --upgrade pip
+
+    # Install PyTorch
+    $ pip uninstall torch -y
+    $ pip install --no-cache-dir --pre torch --index-url https://download.pytorch.org/whl/rocm6.2
+
+    # Build & install AMD SMI
+    $ pip install /opt/rocm/share/amd_smi
+
+    # Install dependencies
+    $ pip install --upgrade numba scipy huggingface-hub[cli]
+    $ pip install "numpy<2"
+    $ pip install -r requirements-rocm.txt
+
+    # Build vLLM for MI210/MI250/MI300.
+    $ export PYTORCH_ROCM_ARCH="gfx90a;gfx942"
+    $ python3 setup.py develop
+    ```
+
+    This may take 5-10 minutes. Currently, `pip install .` does not work for ROCm installation.
+
+    ```{tip}
+    - Triton flash attention is used by default. For benchmarking purposes, it is recommended to run a warm up step before collecting perf numbers.
+    - Triton flash attention does not currently support sliding window attention. If using half precision, please use CK flash-attention for sliding window support.
+    - To use CK flash-attention or PyTorch naive attention, please use this flag `export VLLM_USE_TRITON_FLASH_ATTN=0` to turn off triton flash attention.
+    - The ROCm version of PyTorch, ideally, should match the ROCm driver version.
+    ```
+
+```{tip}
+- For MI300x (gfx942) users, to achieve optimal performance, please refer to [MI300x tuning guide](https://rocm.docs.amd.com/en/latest/how-to/tuning-guides/mi300x/index.html) for performance optimization and tuning tips on system and workflow level.
+  For vLLM, please refer to [vLLM performance optimization](https://rocm.docs.amd.com/en/latest/how-to/tuning-guides/mi300x/workload.html#vllm-performance-optimization).
+```
+
+## Set up using Docker
+
+### Pre-built images
+
+Currently, there are no pre-built ROCm images.
+
+### Build image from source
+
+Building the Docker image from source is the recommended way to use vLLM with ROCm.
+
+First, build a docker image from <gh-file:Dockerfile.rocm> and launch a docker container from the image.
+It is important that the user kicks off the docker build using buildkit. Either the user put DOCKER_BUILDKIT=1 as environment variable when calling docker build command, or the user needs to setup buildkit in the docker daemon configuration /etc/docker/daemon.json as follows and restart the daemon:
+
+```console
+{
+    "features": {
+        "buildkit": true
+    }
+}
+```
+
+<gh-file:Dockerfile.rocm> uses ROCm 6.2 by default, but also supports ROCm 5.7, 6.0 and 6.1 in older vLLM branches.
+It provides flexibility to customize the build of docker image using the following arguments:
+
+- `BASE_IMAGE`: specifies the base image used when running `docker build`, specifically the PyTorch on ROCm base image.
+- `BUILD_FA`: specifies whether to build CK flash-attention. The default is 1. For [Radeon RX 7900 series (gfx1100)](https://rocm.docs.amd.com/projects/radeon/en/latest/index.html), this should be set to 0 before flash-attention supports this target.
+- `FX_GFX_ARCHS`: specifies the GFX architecture that is used to build CK flash-attention, for example, `gfx90a;gfx942` for MI200 and MI300. The default is `gfx90a;gfx942`
+- `FA_BRANCH`: specifies the branch used to build the CK flash-attention in [ROCm's flash-attention repo](https://github.com/ROCmSoftwarePlatform/flash-attention). The default is `ae7928c`
+- `BUILD_TRITON`: specifies whether to build triton flash-attention. The default value is 1.
+
+Their values can be passed in when running `docker build` with `--build-arg` options.
+
+To build vllm on ROCm 6.2 for MI200 and MI300 series, you can use the default:
+
+```console
+DOCKER_BUILDKIT=1 docker build -f Dockerfile.rocm -t vllm-rocm .
+```
+
+To build vllm on ROCm 6.2 for Radeon RX7900 series (gfx1100), you should specify `BUILD_FA` as below:
+
+```console
+DOCKER_BUILDKIT=1 docker build --build-arg BUILD_FA="0" -f Dockerfile.rocm -t vllm-rocm .
+```
+
+To run the above docker image `vllm-rocm`, use the below command:
+
+```console
+docker run -it \
+   --network=host \
+   --group-add=video \
+   --ipc=host \
+   --cap-add=SYS_PTRACE \
+   --security-opt seccomp=unconfined \
+   --device /dev/kfd \
+   --device /dev/dri \
+   -v <path/to/model>:/app/model \
+   vllm-rocm \
+   bash
+```
+
+Where the `<path/to/model>` is the location where the model is stored, for example, the weights for llama2 or llama3 models.
+
+## Supported features
+
+See <project:#feature-x-hardware> compatibility matrix for feature support information.

docs/source/getting_started/installation/gpu/xpu.inc.md

@@ -1,26 +1,47 @@
-(installation-xpu)=
-
-# Installation for XPUs
+# Installation
 
 vLLM initially supports basic model inferencing and serving on Intel GPU platform.
 
-Table of contents:
-
-1. [Requirements](#xpu-backend-requirements)
-2. [Quick start using Dockerfile](#xpu-backend-quick-start-dockerfile)
-3. [Build from source](#build-xpu-backend-from-source)
-
-(xpu-backend-requirements)=
-
 ## Requirements
 
-- OS: Linux
 - Supported Hardware: Intel Data Center GPU, Intel ARC GPU
 - OneAPI requirements: oneAPI 2024.2
 
-(xpu-backend-quick-start-dockerfile)=
+## Set up using Python
 
-## Quick start using Dockerfile
+### Pre-built wheels
+
+Currently, there are no pre-built XPU wheels.
+
+### Build wheel from source
+
+- First, install required driver and intel OneAPI 2024.2 or later.
+- Second, install Python packages for vLLM XPU backend building:
+
+```console
+source /opt/intel/oneapi/setvars.sh
+pip install --upgrade pip
+pip install -v -r requirements-xpu.txt
+```
+
+- Finally, build and install vLLM XPU backend:
+
+```console
+VLLM_TARGET_DEVICE=xpu python setup.py install
+```
+
+```{note}
+- FP16 is the default data type in the current XPU backend. The BF16 data
+  type will be supported in the future.
+```
+
+## Set up using Docker
+
+### Pre-built images
+
+Currently, there are no pre-built XPU images.
+
+### Build image from source
 
 ```console
 $ docker build -f Dockerfile.xpu -t vllm-xpu-env --shm-size=4g .
@@ -32,43 +53,19 @@ $ docker run -it \
              vllm-xpu-env
 ```
 
-(build-xpu-backend-from-source)=
-
-## Build from source
-
-- First, install required driver and intel OneAPI 2024.2 or later.
-- Second, install Python packages for vLLM XPU backend building:
-
-```console
-$ source /opt/intel/oneapi/setvars.sh
-$ pip install --upgrade pip
-$ pip install -v -r requirements-xpu.txt
-```
-
-- Finally, build and install vLLM XPU backend:
-
-```console
-$ VLLM_TARGET_DEVICE=xpu python setup.py install
-```
-
-```{note}
-- FP16 is the default data type in the current XPU backend. The BF16 data
-  type will be supported in the future.
-```
-
-## Distributed inference and serving
+## Supported features
 
 XPU platform supports tensor-parallel inference/serving and also supports pipeline parallel as a beta feature for online serving. We requires Ray as the distributed runtime backend. For example, a reference execution likes following:
 
 ```console
-$ python -m vllm.entrypoints.openai.api_server \
-$      --model=facebook/opt-13b \
-$      --dtype=bfloat16 \
-$      --device=xpu \
-$      --max_model_len=1024 \
-$      --distributed-executor-backend=ray \
-$      --pipeline-parallel-size=2 \
-$      -tp=8
+python -m vllm.entrypoints.openai.api_server \
+     --model=facebook/opt-13b \
+     --dtype=bfloat16 \
+     --device=xpu \
+     --max_model_len=1024 \
+     --distributed-executor-backend=ray \
+     --pipeline-parallel-size=2 \
+     -tp=8
 ```
 
 By default, a ray instance will be launched automatically if no existing one is detected in system, with `num-gpus` equals to `parallel_config.world_size`. We recommend properly starting a ray cluster before execution, referring to the <gh-file:examples/online_serving/run_cluster.sh> helper script.