Prune Ray v1 non-SPMD code paths

.buildkite/nightly-benchmarks/scripts/convert-results-json-to-markdown.py

@@ -368,7 +368,7 @@ if __name__ == "__main__":
         # The GPUs sometimes come in format of "GPUTYPE\nGPUTYPE\n...",
         # we want to turn it into "8xGPUTYPE"
         df["GPU"] = df["GPU"].apply(
-            lambda x: f"{len(x.splitlines())}x{x.splitlines()[0]}"
+            lambda x: f"{len(x.split('\n'))}x{x.split('\n')[0]}"
         )
 
     # get markdown tables

.buildkite/nightly-benchmarks/scripts/launch-server.sh

@@ -181,14 +181,18 @@ launch_vllm_server() {
   if echo "$common_params" | jq -e 'has("fp8")' >/dev/null; then
     echo "Key 'fp8' exists in common params. Use neuralmagic fp8 model for convenience."
     model=$(echo "$common_params" | jq -r '.neuralmagic_quantized_model')
-    server_command="vllm serve $model \
+    server_command="python3 \
+        -m vllm.entrypoints.openai.api_server \
         -tp $tp \
+        --model $model \
         --port $port \
         $server_args"
   else
     echo "Key 'fp8' does not exist in common params."
-    server_command="vllm serve $model \
+    server_command="python3 \
+        -m vllm.entrypoints.openai.api_server \
         -tp $tp \
+        --model $model \
         --port $port \
         $server_args"
   fi

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

@@ -365,7 +365,8 @@ run_serving_tests() {
       continue
     fi
 
-    server_command="$server_envs vllm serve \
+    server_command="$server_envs python3 \
+      -m vllm.entrypoints.openai.api_server \
       $server_args"
 
     # run the server
@@ -454,6 +455,11 @@ main() {
   fi
   check_hf_token
 
+  # Set to v1 to run v1 benchmark
+  if [[ "${ENGINE_VERSION:-v0}" == "v1" ]]; then
+    export VLLM_USE_V1=1
+  fi
+
   # dependencies
   (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
   (which jq) || (apt-get update && apt-get -y install jq)

.buildkite/pyproject.toml

@@ -0,0 +1,46 @@
+# This local pyproject file is part of the migration from yapf to ruff format.
+# It uses the same core rules as the main pyproject.toml file, but with the
+# following differences:
+# - ruff line length is overridden to 88
+# - deprecated typing ignores (UP006, UP035) have been removed
+
+[tool.ruff]
+line-length = 88
+
+[tool.ruff.lint.per-file-ignores]
+"vllm/third_party/**" = ["ALL"]
+"vllm/version.py" = ["F401"]
+"vllm/_version.py" = ["ALL"]
+
+[tool.ruff.lint]
+select = [
+    # pycodestyle
+    "E",
+    # Pyflakes
+    "F",
+    # pyupgrade
+    "UP",
+    # flake8-bugbear
+    "B",
+    # flake8-simplify
+    "SIM",
+    # isort
+    "I",
+    # flake8-logging-format
+    "G",
+]
+ignore = [
+    # star imports
+    "F405", "F403",
+    # lambda expression assignment
+    "E731",
+    # Loop control variable not used within loop body
+    "B007",
+    # f-string format
+    "UP032",
+    # Can remove once 3.10+ is the minimum Python version
+    "UP007",
+]
+
+[tool.ruff.format]
+docstring-code-format = true

.buildkite/release-pipeline.yaml

@@ -48,7 +48,7 @@ steps:
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='7.0 7.5 8.0 8.9 9.0+PTX' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
       - "bash .buildkite/scripts/upload-wheels.sh"
@@ -76,7 +76,7 @@ steps:
       queue: arm64_cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
 
   # Add job to create multi-arch manifest
@@ -150,16 +150,11 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64"
-      - "docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64"
-      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-x86_64 vllm/vllm-openai:nightly-x86_64"
-      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-aarch64 vllm/vllm-openai:nightly-aarch64"
-      - "docker push vllm/vllm-openai:nightly-x86_64"
-      - "docker push vllm/vllm-openai:nightly-aarch64"
-      - "docker manifest create vllm/vllm-openai:nightly vllm/vllm-openai:nightly-x86_64 vllm/vllm-openai:nightly-aarch64 --amend"
-      - "docker manifest create vllm/vllm-openai:nightly-$BUILDKITE_COMMIT vllm/vllm-openai:nightly-x86_64 vllm/vllm-openai:nightly-aarch64 --amend"
-      - "docker manifest push vllm/vllm-openai:nightly"
-      - "docker manifest push vllm/vllm-openai:nightly-$BUILDKITE_COMMIT"
+      - "docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
+      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT vllm/vllm-openai:nightly"
+      - "docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT vllm/vllm-openai:nightly-$BUILDKITE_COMMIT"
+      - "docker push vllm/vllm-openai:nightly"
+      - "docker push vllm/vllm-openai:nightly-$BUILDKITE_COMMIT"
       # Clean up old nightly builds (keep only last 14)
       - "bash .buildkite/scripts/cleanup-nightly-builds.sh"
     plugins:
@@ -168,4 +163,3 @@ steps:
           password-env: DOCKERHUB_TOKEN
     env:
       DOCKER_BUILDKIT: "1"
-      DOCKERHUB_USERNAME: "vllmbot"

.buildkite/scripts/cleanup-nightly-builds.sh

@@ -8,41 +8,20 @@ set -ex
 # DockerHub API endpoint for vllm/vllm-openai repository
 REPO_API_URL="https://hub.docker.com/v2/repositories/vllm/vllm-openai/tags"
 
-# Get DockerHub credentials from environment
+# Get DockerHub token from environment
 if [ -z "$DOCKERHUB_TOKEN" ]; then
     echo "Error: DOCKERHUB_TOKEN environment variable is not set"
     exit 1
 fi
 
-if [ -z "$DOCKERHUB_USERNAME" ]; then
-    echo "Error: DOCKERHUB_USERNAME environment variable is not set"
-    exit 1
-fi
-
-# Get DockerHub bearer token
-echo "Getting DockerHub bearer token..."
-set +x
-BEARER_TOKEN=$(curl -s -X POST \
-    -H "Content-Type: application/json" \
-    -d "{\"username\": \"$DOCKERHUB_USERNAME\", \"password\": \"$DOCKERHUB_TOKEN\"}" \
-    "https://hub.docker.com/v2/users/login" | jq -r '.token')
-set -x
-
-if [ -z "$BEARER_TOKEN" ] || [ "$BEARER_TOKEN" = "null" ]; then
-    echo "Error: Failed to get DockerHub bearer token"
-    exit 1
-fi
-
 # Function to get all tags from DockerHub
 get_all_tags() {
     local page=1
     local all_tags=""
     
     while true; do
-        set +x
-        local response=$(curl -s -H "Authorization: Bearer $BEARER_TOKEN" \
+        local response=$(curl -s -H "Authorization: Bearer $DOCKERHUB_TOKEN" \
             "$REPO_API_URL?page=$page&page_size=100")
-        set -x
         
         # Get both last_updated timestamp and tag name, separated by |
         local tags=$(echo "$response" | jq -r '.results[] | select(.name | startswith("nightly-")) | "\(.last_updated)|\(.name)"')
@@ -64,9 +43,7 @@ delete_tag() {
     echo "Deleting tag: $tag_name"
     
     local delete_url="https://hub.docker.com/v2/repositories/vllm/vllm-openai/tags/$tag_name"
-    set +x
-    local response=$(curl -s -X DELETE -H "Authorization: Bearer $BEARER_TOKEN" "$delete_url")
-    set -x
+    local response=$(curl -s -X DELETE -H "Authorization: Bearer $DOCKERHUB_TOKEN" "$delete_url")
     
     if echo "$response" | jq -e '.detail' > /dev/null 2>&1; then
         echo "Warning: Failed to delete tag $tag_name: $(echo "$response" | jq -r '.detail')"

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -86,6 +86,10 @@ if [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
   commands=${commands//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
 fi
 
+if [[ $commands == *"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'"* ]]; then
+  commands=${commands//"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'"/"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2 and not BambaForCausalLM and not Gemma2ForCausalLM and not Grok1ModelForCausalLM and not Zamba2ForCausalLM and not Gemma2Model and not GritLM'"}
+fi
+
 if [[ $commands == *"pytest -v -s compile/test_basic_correctness.py"* ]]; then
   commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s compile/test_basic_correctness.py"}
 fi

.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh

@@ -25,28 +25,25 @@ function cpu_tests() {
 
   # offline inference
   podman exec -it "$container_id" bash -c "
-    set -xve
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m" >> $HOME/test_basic.log
+    set -e
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
 
   # Run basic model test
   podman exec -it "$container_id" bash -c "
-    set -evx
+    set -e
     pip install pytest pytest-asyncio einops peft Pillow soundfile transformers_stream_generator matplotlib
     pip install sentence-transformers datamodel_code_generator
-
-    # Note: disable Bart until supports V1
-    # pytest -v -s tests/models/language/generation/test_bart.py -m cpu_model
+    pytest -v -s tests/models/language/generation/test_bart.py -m cpu_model
     pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-openai-community/gpt2]
     pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-facebook/opt-125m]
     pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-google/gemma-1.1-2b-it]
     pytest -v -s tests/models/language/pooling/test_classification.py::test_models[float-jason9693/Qwen2.5-1.5B-apeach]
-    # TODO: Below test case tests/models/language/pooling/test_embedding.py::test_models[True-ssmits/Qwen2-7B-Instruct-embed-base] fails on ppc64le. Disabling it for time being.
-    # pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model" >> $HOME/test_rest.log
+    pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model"
 }
 
 # All of CPU tests are expected to be finished less than 40 mins.
 
 export container_id
 export -f cpu_tests
-timeout 120m bash -c cpu_tests
+timeout 40m bash -c cpu_tests
 

.buildkite/scripts/hardware_ci/run-cpu-test.sh

@@ -58,8 +58,11 @@ function cpu_tests() {
     # pytest -x -v -s tests/kernels/attention/test_cache.py -m cpu_model
     # pytest -x -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
 
-    pytest -x -v -s tests/models/language/generation -m cpu_model
-    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model
+    # Note: disable Bart until supports V1
+    pytest -x -v -s tests/models/language/generation -m cpu_model \
+                --ignore=tests/models/language/generation/test_bart.py
+    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model \
+                --ignore=tests/models/language/generation/test_bart.py
 
     pytest -x -v -s tests/models/language/pooling -m cpu_model
     pytest -x -v -s tests/models/multimodal/generation \

.buildkite/scripts/hardware_ci/run-npu-test.sh

@@ -1,191 +0,0 @@
-#!/bin/bash
-
-# This script build the Ascend NPU docker image and run the offline inference inside the container.
-# It serves a sanity check for compilation and basic model usage.
-set -ex
-
-# Base ubuntu image with basic ascend development libraries and python installed
-VLLM_ASCEND_REPO="https://github.com/vllm-project/vllm-ascend.git"
-CONFIG_FILE_REMOTE_PATH="tests/e2e/vllm_interface/vllm_test.cfg"
-TEST_RUN_CONFIG_FILE="vllm_test.cfg"
-VLLM_ASCEND_TMP_DIR=
-# Get the test run configuration file from the vllm-ascend repository
-fetch_vllm_test_cfg() {
-    VLLM_ASCEND_TMP_DIR=$(mktemp -d)
-    # Ensure that the temporary directory is cleaned up when an exception occurs during configuration file retrieval
-    cleanup() {
-        rm -rf "${VLLM_ASCEND_TMP_DIR}"
-    }
-    trap cleanup EXIT
-
-    GIT_TRACE=1 git clone -v --depth 1 "${VLLM_ASCEND_REPO}" "${VLLM_ASCEND_TMP_DIR}"
-    if [ ! -f "${VLLM_ASCEND_TMP_DIR}/${CONFIG_FILE_REMOTE_PATH}" ]; then
-        echo "Error: file '${CONFIG_FILE_REMOTE_PATH}' does not exist in the warehouse" >&2
-        exit 1
-    fi
-
-    # If the file already exists locally, just overwrite it
-    cp "${VLLM_ASCEND_TMP_DIR}/${CONFIG_FILE_REMOTE_PATH}" "${TEST_RUN_CONFIG_FILE}"
-    echo "Copied ${CONFIG_FILE_REMOTE_PATH} to ${TEST_RUN_CONFIG_FILE}"
-
-    # Since the trap will be overwritten later, and when it is executed here, the task of cleaning up resources
-    # when the trap is abnormal has been completed, so the temporary resources are manually deleted here.
-    rm -rf "${VLLM_ASCEND_TMP_DIR}"
-    trap - EXIT
-}
-
-# Downloads test run configuration file from a remote URL.
-# Loads the configuration into the current script environment.
-get_config() {
-    if [ ! -f "${TEST_RUN_CONFIG_FILE}" ]; then
-        echo "Error: file '${TEST_RUN_CONFIG_FILE}' does not exist in the warehouse" >&2
-        exit 1
-    fi
-    source "${TEST_RUN_CONFIG_FILE}"
-    echo "Base docker image name that get from configuration: ${BASE_IMAGE_NAME}"
-    return 0
-}
-
-# get test running configuration.
-fetch_vllm_test_cfg
-get_config
-# Check if the function call was successful. If not, exit the script.
-if [ $? -ne 0 ]; then
-  exit 1
-fi
-
-image_name="npu/vllm-ci:${BUILDKITE_COMMIT}_${EPOCHSECONDS}"
-container_name="npu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
-
-# BUILDKITE_AGENT_NAME format is {hostname}-{agent_idx}-{npu_card_num}cards
-agent_idx=$(echo "${BUILDKITE_AGENT_NAME}" | awk -F'-' '{print $(NF-1)}')
-echo "agent_idx: ${agent_idx}"
-builder_name="cachebuilder${agent_idx}"
-builder_cache_dir="/mnt/docker-cache${agent_idx}"
-mkdir -p ${builder_cache_dir}
-
-# Try building the docker image
-cat <<EOF | DOCKER_BUILDKIT=1 docker build \
-    --add-host cache-service-vllm.nginx-pypi-cache.svc.cluster.local:${PYPI_CACHE_HOST} \
-    --builder ${builder_name} --cache-from type=local,src=${builder_cache_dir} \
-                           --cache-to type=local,dest=${builder_cache_dir},mode=max \
-    --progress=plain --load -t ${image_name} -f - .
-FROM ${BASE_IMAGE_NAME}
-
-# Define environments
-ENV DEBIAN_FRONTEND=noninteractive
-
-RUN pip config set global.index-url http://cache-service-vllm.nginx-pypi-cache.svc.cluster.local:${PYPI_CACHE_PORT}/pypi/simple && \
-    pip config set global.trusted-host cache-service-vllm.nginx-pypi-cache.svc.cluster.local && \
-    apt-get update -y && \
-    apt-get install -y python3-pip git vim wget net-tools gcc g++ cmake libnuma-dev && \
-    rm -rf /var/cache/apt/* && \
-    rm -rf /var/lib/apt/lists/*
-
-# Install for pytest to make the docker build cache layer always valid
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install pytest>=6.0  modelscope
-
-WORKDIR /workspace/vllm
-
-# Install vLLM dependencies in advance. Effect: As long as common.txt remains unchanged, the docker cache layer will be valid.
-COPY requirements/common.txt /workspace/vllm/requirements/common.txt
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements/common.txt
-
-COPY . .
-
-# Install vLLM
-RUN --mount=type=cache,target=/root/.cache/pip \
-    VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /workspace/vllm/ --extra-index https://download.pytorch.org/whl/cpu/ && \
-    python3 -m pip uninstall -y triton
-
-# Install vllm-ascend
-WORKDIR /workspace
-ARG VLLM_ASCEND_REPO=https://github.com/vllm-project/vllm-ascend.git
-ARG VLLM_ASCEND_TAG=main
-RUN git config --global url."https://gh-proxy.test.osinfra.cn/https://github.com/".insteadOf "https://github.com/" && \
-    git clone --depth 1 \$VLLM_ASCEND_REPO --branch \$VLLM_ASCEND_TAG /workspace/vllm-ascend
-
-# Install vllm dependencies in advance. Effect: As long as common.txt remains unchanged, the docker cache layer will be valid.
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r /workspace/vllm-ascend/requirements.txt
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    export PIP_EXTRA_INDEX_URL=https://mirrors.huaweicloud.com/ascend/repos/pypi && \
-    source /usr/local/Ascend/ascend-toolkit/set_env.sh && \
-    source /usr/local/Ascend/nnal/atb/set_env.sh && \
-    export LD_LIBRARY_PATH=\$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/`uname -i`-linux/devlib && \
-    python3 -m pip install -v -e /workspace/vllm-ascend/ --extra-index https://download.pytorch.org/whl/cpu/
-
-ENV VLLM_WORKER_MULTIPROC_METHOD=spawn
-ENV VLLM_USE_MODELSCOPE=True
-
-WORKDIR /workspace/vllm-ascend
-
-CMD ["/bin/bash"]
-
-EOF
-
-# Setup cleanup
-remove_docker_container() {
-  docker rm -f "${container_name}" || true;
-  docker image rm -f "${image_name}" || true;
-  docker system prune -f || true;
-}
-trap remove_docker_container EXIT
-
-# Generate corresponding --device args based on BUILDKITE_AGENT_NAME
-# Ascend NPU BUILDKITE_AGENT_NAME format is {hostname}-{agent_idx}-{npu_card_num}cards, and agent_idx starts from 1.
-#   e.g. atlas-a2-001-1-2cards means this is the 1-th agent on atlas-a2-001 host, and it has 2 NPU cards.
-#   returns --device /dev/davinci0 --device /dev/davinci1
-parse_and_gen_devices() {
-    local input="$1"
-    local index cards_num
-    if [[ "$input" =~ ([0-9]+)-([0-9]+)cards$ ]]; then
-        index="${BASH_REMATCH[1]}"
-        cards_num="${BASH_REMATCH[2]}"
-    else
-        echo "parse error" >&2
-        return 1
-    fi
-
-    local devices=""
-    local i=0
-    while (( i < cards_num )); do
-        local dev_idx=$(((index - 1)*cards_num + i ))
-        devices="$devices --device /dev/davinci${dev_idx}"
-        ((i++))
-    done
-
-    # trim leading space
-    devices="${devices#"${devices%%[![:space:]]*}"}"
-    # Output devices: assigned to the caller variable
-    printf '%s' "$devices"
-}
-
-devices=$(parse_and_gen_devices "${BUILDKITE_AGENT_NAME}") || exit 1
-
-# Run the image and execute the Out-Of-Tree (OOT) platform interface test case on Ascend NPU hardware.
-# This test checks whether the OOT platform interface is functioning properly in conjunction with
-# the hardware plugin vllm-ascend.
-model_cache_dir=/mnt/modelscope${agent_idx}
-mkdir -p ${model_cache_dir}
-docker run \
-    ${devices} \
-    --device /dev/davinci_manager \
-    --device /dev/devmm_svm \
-    --device /dev/hisi_hdc \
-    -v /usr/local/dcmi:/usr/local/dcmi \
-    -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-    -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
-    -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-    -v /etc/ascend_install.info:/etc/ascend_install.info \
-    -v ${model_cache_dir}:/root/.cache/modelscope \
-    --entrypoint="" \
-    --name "${container_name}" \
-    "${image_name}" \
-    bash -c '
-    set -e
-    pytest -v -s tests/e2e/vllm_interface/
-'

.buildkite/scripts/hardware_ci/run-tpu-v1-test-part2.sh

@@ -62,11 +62,12 @@ echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
     && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
-    && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
+    && python3 -m pip install --progress-bar off hf-transfer
 echo "--- Python dependencies installed ---"
-
+export VLLM_USE_V1=1
 export VLLM_XLA_CHECK_RECOMPILATION=1
 export VLLM_XLA_CACHE_PATH=
+echo "Using VLLM V1"
 
 echo "--- Hardware Information ---"
 # tpu-info

.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh

@@ -62,11 +62,12 @@ echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
     && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
-    && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
+    && python3 -m pip install --progress-bar off hf-transfer
 echo "--- Python dependencies installed ---"
-
+export VLLM_USE_V1=1
 export VLLM_XLA_CHECK_RECOMPILATION=1
 export VLLM_XLA_CACHE_PATH=
+echo "Using VLLM V1"
 
 echo "--- Hardware Information ---"
 # tpu-info

.buildkite/scripts/hardware_ci/run-xpu-test.sh

@@ -35,14 +35,16 @@ docker run \
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -O.cudagraph_mode=NONE
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
-    VLLM_ATTENTION_BACKEND=TRITON_ATTN python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    VLLM_ATTENTION_BACKEND=TRITON_ATTN_VLLM_V1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
     cd tests
     pytest -v -s v1/core
     pytest -v -s v1/engine
     pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
     pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
     pytest -v -s v1/structured_output
-    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py
+    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_eagle.py --ignore=v1/spec_decode/test_tree_attention.py
     pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_shared_storage_connector.py
     pytest -v -s v1/test_serial_utils.py
+    pytest -v -s v1/test_utils.py
+    pytest -v -s v1/test_metrics_reader.py
 '

.buildkite/scripts/run-benchmarks.sh

@@ -18,7 +18,7 @@ vllm bench throughput --input-len 256 --output-len 256 --output-json throughput_
 bench_throughput_exit_code=$?
 
 # run server-based benchmarks and upload the result to buildkite
-vllm serve meta-llama/Llama-2-7b-chat-hf &
+python3 -m vllm.entrypoints.openai.api_server --model meta-llama/Llama-2-7b-chat-hf &
 server_pid=$!
 wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
 

.buildkite/scripts/run-prime-rl-test.sh

@@ -1,59 +0,0 @@
-#!/bin/bash
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-# Setup script for Prime-RL integration tests
-# This script prepares the environment for running Prime-RL tests with nightly vLLM
-
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd "${SCRIPT_DIR}/../.." && pwd)"
-PRIME_RL_REPO="https://github.com/PrimeIntellect-ai/prime-rl.git"
-PRIME_RL_DIR="${REPO_ROOT}/prime-rl"
-
-echo "Setting up Prime-RL integration test environment..."
-
-# Clean up any existing Prime-RL directory
-if [ -d "${PRIME_RL_DIR}" ]; then
-    echo "Removing existing Prime-RL directory..."
-    rm -rf "${PRIME_RL_DIR}"
-fi
-
-# Install UV if not available
-if ! command -v uv &> /dev/null; then
-    echo "Installing UV package manager..."
-    curl -LsSf https://astral.sh/uv/install.sh | sh
-    source $HOME/.local/bin/env
-fi
-
-# Clone Prime-RL repository at specific branch for reproducible tests
-PRIME_RL_BRANCH="integ-vllm-main"
-echo "Cloning Prime-RL repository at branch: ${PRIME_RL_BRANCH}..."
-git clone --branch "${PRIME_RL_BRANCH}" --single-branch "${PRIME_RL_REPO}" "${PRIME_RL_DIR}"
-cd "${PRIME_RL_DIR}"
-
-echo "Setting up UV project environment..."
-export UV_PROJECT_ENVIRONMENT=/usr/local
-ln -s /usr/bin/python3 /usr/local/bin/python
-
-# Remove vllm pin from pyproject.toml
-echo "Removing vllm pin from pyproject.toml..."
-sed -i '/vllm==/d' pyproject.toml
-
-# Sync Prime-RL dependencies
-echo "Installing Prime-RL dependencies..."
-uv sync --inexact && uv sync --inexact --all-extras
-
-# Verify installation
-echo "Verifying installations..."
-uv run python -c "import vllm; print(f'vLLM version: {vllm.__version__}')"
-uv run python -c "import prime_rl; print('Prime-RL imported successfully')"
-
-echo "Prime-RL integration test environment setup complete!"
-
-echo "Running Prime-RL integration tests..."
-export WANDB_MODE=offline # this makes this test not require a WANDB_API_KEY
-uv run pytest -vs tests/integration/test_rl.py -m gpu
-
-echo "Prime-RL integration tests completed!"

.buildkite/scripts/tpu/quantized_v6e_1.env

@@ -9,6 +9,6 @@ MAX_NUM_BATCHED_TOKENS=1024
 TENSOR_PARALLEL_SIZE=1
 MAX_MODEL_LEN=2048
 DOWNLOAD_DIR=/mnt/disks/persist
-EXPECTED_THROUGHPUT=8.7
+EXPECTED_THROUGHPUT=10.0
 INPUT_LEN=1800
 OUTPUT_LEN=128

.buildkite/scripts/tpu/run_bm.sh

@@ -42,7 +42,7 @@ echo "lanching vllm..."
 echo "logging to $VLLM_LOG"
 echo
 
-vllm serve $MODEL \
+VLLM_USE_V1=1 vllm serve $MODEL \
  --seed 42 \
  --max-num-seqs $MAX_NUM_SEQS \
  --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \

.buildkite/test-pipeline.yaml

@@ -6,28 +6,24 @@
 # to generate the final pipeline yaml file.
 
 # Documentation
-# label(str): the name of the test. emojis allowed.
-# fast_check(bool): whether to run this on each commit on the fastcheck pipeline.
-# torch_nightly(bool): whether to run this on vllm against the torch nightly pipeline.
-# fast_check_only(bool): run this test on the fastcheck pipeline only
-# optional(bool): never run this test by default (i.e. need to unblock manually) unless it's a scheduled nightly run.
-# soft_fail(bool): allow this step to fail without failing the entire pipeline (useful for flaky or experimental tests).
+# label(str): the name of the test. emoji allowed.
+# fast_check(bool): whether to run this on each commit on fastcheck pipeline.
+# torch_nightly(bool): whether to run this on vllm against torch nightly pipeline.
+# fast_check_only(bool): run this test on fastcheck pipeline only
+# optional(bool): never run this test by default (i.e. need to unblock manually) unless it's scheduled nightly run.
 # command(str): the single command to run for tests. incompatible with commands.
-# commands(list): the list of commands to run for the test. incompatible with command.
-# mirror_hardwares(list): the list of hardware to run the test on as well. currently only supports [amdexperimental]
-# gpu(str): override the GPU selection for the test. default is L4 GPUs. supports a100, b200, h200
-# num_gpus(int): override the number of GPUs for the test. defaults to 1 GPU. currently supports 2,4.
-# num_nodes(int): whether to simulate multi-node setup by launching multiple containers on one host,
-#     in this case, commands must be specified. the first command runs on the first host, the second
+# commands(list): the list of commands to run for test. incompatbile with command.
+# mirror_hardwares(list): the list of hardwares to run the test on as well. currently only supports [amd]
+# gpu(str): override the GPU selection for the test. default is on L4 GPUs. currently only supports a100
+# num_gpus(int): override the number of GPUs for the test. default to 1 GPU. currently support 2,4.
+# num_nodes(int): whether to simulate multi-node setup by launch multiple containers on one host,
+#     in this case, commands must be specified. the first command runs on first host, the second
 #     command runs on the second host.
-# timeout_in_minutes(int): sets a timeout for the step in minutes. if not specified, uses the default timeout.
-# parallelism(int): number of parallel jobs to run for this step. enables test sharding using $$BUILDKITE_PARALLEL_JOB
-#     and $$BUILDKITE_PARALLEL_JOB_COUNT environment variables.
-# working_dir(str): specify the place where the command should execute, default to /vllm-workspace/tests
-# source_file_dependencies(list): the list of prefixes to opt-in the test for, if empty, the test will always run.
+# working_dir(str): specify the place where command should execute, default to /vllm-workspace/tests
+# source_file_dependencies(list): the list of prefix to opt-in the test for, if empty, the test will always run.
 
 # When adding a test
-# - If the test belongs to an existing group, add it there
+# - If the test belong to an existing group, add it there
 # - If the test is short, add to any existing step
 # - If the test takes more than 10min, then it is okay to create a new step.
 #   Note that all steps execute in parallel.
@@ -50,28 +46,19 @@ steps:
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
-  - tests/multimodal
-  - tests/utils_
-  commands:
-  - pytest -v -s -m 'not cpu_test' multimodal
-  - pytest -v -s utils_
-
-- label: Async Engine, Inputs, Utils, Worker Test (CPU) # 4 mins
-  timeout_in_minutes: 10
-  source_file_dependencies:
-  - vllm/
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/multimodal
+  - tests/utils_
   - tests/standalone_tests/lazy_imports.py
   - tests/transformers_utils
-  no_gpu: true
   commands:
   - python3 standalone_tests/lazy_imports.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
-  - pytest -v -s -m 'cpu_test' multimodal
-  - pytest -v -s transformers_utils
+  - pytest -v -s multimodal
+  - pytest -v -s utils_ # Utils
+  - pytest -v -s transformers_utils # transformers_utils
 
 - label: Python-only Installation Test # 10min
   timeout_in_minutes: 20
@@ -123,7 +110,7 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Entrypoints Integration Test (API Server) # 100min
   timeout_in_minutes: 130
@@ -161,6 +148,7 @@ steps:
   num_gpus: 4
   source_file_dependencies:
   - vllm/distributed/
+  - vllm/core/
   - tests/distributed/test_utils
   - tests/distributed/test_pynccl
   - tests/distributed/test_events
@@ -168,34 +156,28 @@ steps:
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
-  - tests/v1/distributed
+  - tests/v1/test_async_llm_dp.py
+  - tests/v1/test_external_lb_dp.py
+  - tests/v1/test_internal_lb_dp.py
+  - tests/v1/test_hybrid_lb_dp.py
   - tests/v1/engine/test_engine_core_client.py
-  - tests/distributed/test_symm_mem_allreduce.py
   commands:
-  # test with torchrun tp=2 and external_dp=2
+  # test with tp=2 and external_dp=2
+  - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
   - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
-  # test with torchrun tp=2 and pp=2
+  # test with tp=2 and pp=2
   - PP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
-  # test with torchrun tp=4 and dp=1
-  - TP_SIZE=4 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
-  # test with torchrun tp=2, pp=2 and dp=1
-  - PP_SIZE=2 TP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
-  # test with torchrun tp=1 and dp=4 with ep
-  - DP_SIZE=4 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
-  # test with torchrun tp=2 and dp=2 with ep
-  - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
   # test with internal dp
   - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
-  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
-  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
-  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_internal_lb_dp.py
-  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py
+  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
+  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_external_lb_dp.py
+  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/test_internal_lb_dp.py
+  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/test_hybrid_lb_dp.py
   - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
   - 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 distributed/test_events.py
-  - pytest -v -s distributed/test_symm_mem_allreduce.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
   - pushd ../examples/offline_inference
@@ -296,35 +278,24 @@ steps:
     - tests/v1
   commands:
     # split the test to avoid interference
-    - pytest -v -s -m 'not cpu_test' v1/core
+    - pytest -v -s v1/core
     - pytest -v -s v1/executor
     - pytest -v -s v1/kv_offload
     - pytest -v -s v1/sample
     - pytest -v -s v1/logits_processors
     - pytest -v -s v1/worker
+    - pytest -v -s v1/structured_output
     - pytest -v -s v1/spec_decode
-    - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
-    - pytest -v -s -m 'not cpu_test' v1/metrics
+    - pytest -v -s v1/kv_connector/unit
+    - pytest -v -s v1/metrics
+    - pytest -v -s v1/test_serial_utils.py
+    - pytest -v -s v1/test_utils.py
     - pytest -v -s v1/test_oracle.py
-    - pytest -v -s v1/test_request.py
+    - pytest -v -s v1/test_metrics_reader.py
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
-- label: V1 Test others (CPU) # 5 mins
-  source_file_dependencies:
-    - vllm/
-    - tests/v1
-  no_gpu: true
-  commands:
-    # split the test to avoid interference
-    - pytest -v -s -m 'cpu_test' v1/core
-    - pytest -v -s v1/structured_output
-    - pytest -v -s v1/test_serial_utils.py
-    - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
-    - pytest -v -s -m 'cpu_test' v1/metrics
-
-
 - label: Examples Test # 30min
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental]
@@ -343,13 +314,12 @@ steps:
     - python3 offline_inference/vision_language.py --seed 0
     - python3 offline_inference/vision_language_pooling.py --seed 0
     - python3 offline_inference/vision_language_multi_image.py --seed 0
-    - python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
+    - VLLM_USE_V1=0 python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
     - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
     - python3 offline_inference/basic/classify.py
     - python3 offline_inference/basic/embed.py
     - python3 offline_inference/basic/score.py
-    - python3 offline_inference/spec_decode.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
-    - python3 offline_inference/spec_decode.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
+    - VLLM_USE_V1=0 python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Platform Tests (CUDA) # 4min
   timeout_in_minutes: 15
@@ -398,12 +368,12 @@ steps:
     - pytest -v -s compile/test_pass_manager.py
     - pytest -v -s compile/test_fusion.py
     - pytest -v -s compile/test_fusion_attn.py
-    - pytest -v -s compile/test_functionalization.py
     - pytest -v -s compile/test_silu_mul_quant_fusion.py
+    - pytest -v -s compile/test_sequence_parallelism.py
+    - pytest -v -s compile/test_async_tp.py
     - pytest -v -s compile/test_fusion_all_reduce.py
     - pytest -v -s compile/test_decorator.py
     - pytest -v -s compile/test_noop_elimination.py
-    - pytest -v -s compile/test_aot_compile.py
 
 - label: PyTorch Fullgraph Smoke Test # 15min
   timeout_in_minutes: 30
@@ -432,9 +402,8 @@ steps:
   source_file_dependencies:
   - csrc/
   - tests/kernels/core
-  - tests/kernels/test_top_k_per_row.py
   commands:
-    - pytest -v -s kernels/core kernels/test_top_k_per_row.py
+    - pytest -v -s kernels/core
 
 - label: Kernels Attention Test %N # 23min
   timeout_in_minutes: 35
@@ -478,23 +447,33 @@ steps:
   source_file_dependencies:
   - csrc/mamba/
   - tests/kernels/mamba
-  - vllm/model_executor/layers/mamba/ops
   commands:
     - pytest -v -s kernels/mamba
 
-- label: Model Executor Test # 23min
-  timeout_in_minutes: 35
+- label: Tensorizer Test # 14min
+  timeout_in_minutes: 25
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
-  - vllm/model_executor
-  - tests/model_executor
+  - vllm/model_executor/model_loader
+  - tests/tensorizer_loader
   - tests/entrypoints/openai/test_tensorizer_entrypoint.py
   commands:
     - apt-get update && apt-get install -y curl libsodium23
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pytest -v -s model_executor
+    - pytest -v -s tensorizer_loader
     - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py
 
+- label: Model Executor Test # 7min
+  timeout_in_minutes: 20
+  mirror_hardwares: [amdexperimental]
+  source_file_dependencies:
+  - vllm/model_executor
+  - tests/model_executor
+  commands:
+    - apt-get update && apt-get install -y curl libsodium23
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -v -s model_executor
+
 - label: Benchmarks # 11min
   timeout_in_minutes: 20
   mirror_hardwares: [amdexperimental]
@@ -528,7 +507,7 @@ steps:
   # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now
   # we can only upgrade after this is resolved
   - pip install --pre torchao==0.13.0.dev20250814 --index-url https://download.pytorch.org/whl/nightly/cu128
-  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/
+  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization
 
 - label: LM Eval Small Models # 53min
   timeout_in_minutes: 75
@@ -556,17 +535,10 @@ steps:
   source_file_dependencies:
     - vllm/
     - tests/tool_use
+    - tests/mistral_tool_use
   commands:
-    - pytest -v -s -m 'not cpu_test' tool_use
-
-- label: OpenAI-Compatible Tool Use (CPU) # 5 mins
-  timeout_in_minutes: 10
-  source_file_dependencies:
-    - vllm/
-    - tests/tool_use
-  no_gpu: true
-  commands:
-    - pytest -v -s -m 'cpu_test' tool_use
+    - pytest -v -s tool_use
+    - pytest -v -s mistral_tool_use
 
 #####  models test  #####
 
@@ -606,19 +578,13 @@ steps:
   - vllm/
   - tests/models/test_transformers.py
   - tests/models/test_registry.py
-  commands:
-    - pytest -v -s models/test_transformers.py models/test_registry.py
-
-- label: Basic Models Test (Other CPU) # 5min
-  timeout_in_minutes: 10
-  torch_nightly: true
-  source_file_dependencies:
-  - vllm/
   - tests/models/test_utils.py
   - tests/models/test_vision.py
-  no_gpu: true
   commands:
-    - pytest -v -s models/test_utils.py models/test_vision.py
+    - pytest -v -s models/test_transformers.py \
+                   models/test_registry.py \
+                   models/test_utils.py \
+                   models/test_vision.py
 
 - label: Language Models Tests (Standard)
   timeout_in_minutes: 25
@@ -788,13 +754,11 @@ steps:
   commands:
     - pip install --upgrade git+https://github.com/huggingface/transformers
     - pytest -v -s tests/models/test_initialization.py
-    - pytest -v -s tests/models/test_transformers.py
     - pytest -v -s tests/models/multimodal/processing/
     - pytest -v -s tests/models/multimodal/test_mapping.py
     - python3 examples/offline_inference/basic/chat.py
+    - python3 examples/offline_inference/audio_language.py --model-type whisper
     - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
-    # Whisper needs spawn method to avoid deadlock
-    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
 
 - label: Blackwell Test # 38 min
   timeout_in_minutes: 60
@@ -829,20 +793,18 @@ steps:
     - pytest -v -s tests/kernels/quantization/test_flashinfer_scaled_mm.py
     - pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py
     - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
-    - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
+    - pytest -v -s tests/kernels/moe/test_mxfp4_moe.py
     # Fusion
     - pytest -v -s tests/compile/test_fusion_all_reduce.py
     - pytest -v -s tests/compile/test_fusion_attn.py::test_attention_quant_pattern
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
-    - pytest -v -s tests/kernels/quantization/test_nvfp4_qutlass.py
-    - pytest -v -s tests/kernels/quantization/test_mxfp4_qutlass.py
 
-- label: Blackwell GPT-OSS Eval
+- label: GPT-OSS Eval (Blackwell)
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
   gpu: b200
-  optional: true # run on nightlies
+  optional: true # disable while debugging
   source_file_dependencies:
   - tests/evals/gpt_oss
   - vllm/model_executor/models/gpt_oss.py
@@ -850,34 +812,7 @@ steps:
   - vllm/v1/attention/backends/flashinfer.py
   commands:
     - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
-
-- label: Blackwell Quantized MoE Test
-  timeout_in_minutes: 60
-  working_dir: "/vllm-workspace/"
-  gpu: b200
-  source_file_dependencies:
-  - tests/quantization/test_blackwell_moe.py
-  - vllm/model_executor/models/deepseek_v2.py
-  - vllm/model_executor/models/gpt_oss.py
-  - vllm/model_executor/models/llama4.py
-  - vllm/model_executor/layers/fused_moe
-  - vllm/model_executor/layers/quantization/compressed_tensors
-  - vllm/model_executor/layers/quantization/modelopt.py
-  - vllm/model_executor/layers/quantization/mxfp4.py
-  - vllm/v1/attention/backends/flashinfer.py
-  commands:
-    - pytest -s -v tests/quantization/test_blackwell_moe.py
-
-- label: Blackwell LM Eval Small Models
-  timeout_in_minutes: 120
-  gpu: b200
-  optional: true # run on nightlies
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt --tp-size=1
+    - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58 --server-args '--tensor-parallel-size 2'
 
 #####  1 GPU test  #####
 #####  multi gpus test  #####
@@ -921,58 +856,47 @@ steps:
     - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
     - python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=1 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
 
-- label: Distributed Tests (2 GPUs) # 68min
-  timeout_in_minutes: 90
+- label: Distributed Tests (2 GPUs) # 110min
+  timeout_in_minutes: 150
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
-  - vllm/compilation/
   - vllm/distributed/
   - vllm/engine/
   - vllm/executor/
-  - vllm/worker/worker_base.py
-  - vllm/v1/engine/
-  - vllm/v1/worker/
-  - tests/compile/test_basic_correctness.py
-  - tests/compile/test_wrapper.py
+  - vllm/model_executor/models/
   - tests/distributed/
-  - tests/entrypoints/llm/test_collective_rpc.py
-  - tests/v1/distributed
+  - vllm/compilation
+  - vllm/worker/worker_base.py
+  - vllm/worker/worker.py
+  - vllm/worker/model_runner.py
+  - entrypoints/llm/test_collective_rpc.py
+  - tests/v1/test_async_llm_dp.py
+  - tests/v1/test_external_lb_dp.py
   - tests/v1/entrypoints/openai/test_multi_api_servers.py
-  - tests/v1/shutdown
-  - tests/v1/worker/test_worker_memory_snapshot.py
+  - vllm/v1/engine/
   commands:
-  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
-  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_external_lb_dp.py
   - DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py
   - pytest -v -s entrypoints/llm/test_collective_rpc.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'
-  - pytest -v -s distributed/test_sequence_parallel.py
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
-  - pytest -v -s v1/worker/test_worker_memory_snapshot.py
-
-- label: Distributed Model Tests (2 GPUs) # 37min
-  timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
-  working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
-  source_file_dependencies:
-  - vllm/model_executor/model_loader/sharded_state_loader.py
-  - vllm/model_executor/models/
-  - tests/basic_correctness/
-  - tests/model_executor/model_loader/test_sharded_state_loader.py
-  - tests/models/
-  commands:
   - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py
   # Avoid importing model tests that cause CUDA reinitialization error
   - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
   - pytest models/language -v -s -m 'distributed(num_gpus=2)'
   - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py
   - VLLM_WORKER_MULTIPROC_METHOD=spawn pytest models/multimodal/generation/test_whisper.py -v -s -m 'distributed(num_gpus=2)'
+  # test sequence parallel
+  - pytest -v -s distributed/test_sequence_parallel.py
+  # this test fails consistently.
+  # TODO: investigate and fix
+  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
+  - pytest -v -s models/multimodal/generation/test_maverick.py
 
 - label: Plugin Tests (2 GPUs) # 40min
   timeout_in_minutes: 60
@@ -1095,8 +1019,6 @@ steps:
   working_dir: "/vllm-workspace/"
   num_gpus: 2
   commands:
-    - pytest -v -s tests/compile/test_async_tp.py
-    - pytest -v -s tests/compile/test_sequence_parallelism.py
     - pytest -v -s tests/distributed/test_context_parallel.py
     - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
 
@@ -1108,16 +1030,3 @@ steps:
   num_gpus: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
-    - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
-
-##### RL Integration Tests #####
-- label: Prime-RL Integration Test # 15min
-  timeout_in_minutes: 30
-  optional: true
-  num_gpus: 2
-  working_dir: "/vllm-workspace"
-  source_file_dependencies:
-  - vllm/
-  - .buildkite/scripts/run-prime-rl-test.sh
-  commands:
-    - bash .buildkite/scripts/run-prime-rl-test.sh

.github/CODEOWNERS

@@ -4,14 +4,19 @@
 # This lists cover the "core" components of vLLM that require careful review
 /vllm/attention @LucasWilkinson
 /vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/core @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @22quinn
 /vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @22quinn
+/vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/model_executor/layers/fused_moe @mgoin
 /vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @NickLucche
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256
 /vllm/model_executor/layers/mamba @tdoublep
 /vllm/model_executor/model_loader @22quinn
 /vllm/multimodal @DarkLight1337 @ywang96 @NickLucche
+/vllm/v1/attention @LucasWilkinson
+/vllm/v1/sample @22quinn @houseroad
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
 /vllm/reasoning @aarnphm @chaunceyjiang
@@ -23,17 +28,14 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
 /vllm/config @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
-/vllm/config/cache.py @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg @heheda12345
 
 # vLLM V1
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
-/vllm/v1/attention @LucasWilkinson
+/vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
+/vllm/v1/spec_decode @benchislett @luccafong
 /vllm/v1/attention/backends/flashinfer.py @mgoin
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
 /vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat @heheda12345 @ApostaC
-/vllm/v1/sample @22quinn @houseroad @njhill
-/vllm/v1/spec_decode @benchislett @luccafong
-/vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
 /vllm/v1/kv_cache_interface.py @heheda12345
 /vllm/v1/offloading @ApostaC
 
@@ -55,7 +57,7 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/weight_loading @mgoin @youkaichao @yewentao256
 /tests/lora @jeejeelee
 /tests/models/language/generation/test_hybrid.py @tdoublep
-/tests/v1/kv_connector/nixl_integration @NickLucche
+/tests/v1/kv_connector/nixl_integration @NickLucche 
 /tests/v1/kv_connector @ApostaC
 /tests/v1/offloading @ApostaC
 
@@ -64,26 +66,18 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/models/test_transformers.py @hmellor
 
 # Docs
-/docs/mkdocs @hmellor
-/docs/**/*.yml @hmellor
-/requirements/docs.txt @hmellor
-.readthedocs.yaml @hmellor
+/docs @hmellor
 mkdocs.yaml @hmellor
 
-# Linting
-.markdownlint.yaml @hmellor
-.pre-commit-config.yaml @hmellor
-/tools/pre_commit @hmellor
-
 # CPU
-/vllm/v1/worker/cpu* @bigPYJ1151
+/vllm/v1/worker/^cpu @bigPYJ1151
 /csrc/cpu @bigPYJ1151
 /vllm/platforms/cpu.py @bigPYJ1151
 /cmake/cpu_extension.cmake @bigPYJ1151
 /docker/Dockerfile.cpu @bigPYJ1151
 
 # Intel GPU
-/vllm/v1/worker/xpu* @jikunshang
+/vllm/v1/worker/^xpu @jikunshang
 /vllm/platforms/xpu.py @jikunshang
 /docker/Dockerfile.xpu @jikunshang
 

.github/ISSUE_TEMPLATE/750-RFC.yml

@@ -43,6 +43,10 @@ body:
       Any other things you would like to mention.
   validations:
     required: false
+- type: markdown
+  attributes:
+    value: >
+      Thanks for contributing 🎉! The vLLM core team hosts a biweekly RFC review session at 9:30AM Pacific Time, while most RFCs can be discussed online, you can optionally sign up for a slot to discuss your RFC online [here](https://docs.google.com/document/d/1CiLVBZeIVfR7_PNAKVSusxpceywkoOOB78qoWqHvSZc/edit).
 - type: checkboxes
   id: askllm
   attributes:

.github/mergify.yml

@@ -2,7 +2,6 @@ pull_request_rules:
 - name: label-documentation
   description: Automatically apply documentation label
   conditions:
-    - label != stale
     - or:
       - files~=^[^/]+\.md$
       - files~=^docs/
@@ -11,13 +10,10 @@ pull_request_rules:
     label:
       add:
         - documentation
-    comment:
-      message: "Documentation preview: https://vllm--{{number}}.org.readthedocs.build/en/{{number}}/"
 
 - name: label-ci-build
   description: Automatically apply ci/build label
   conditions:
-    - label != stale
     - or:
       - files~=^\.github/
       - files~=\.buildkite/
@@ -34,7 +30,6 @@ pull_request_rules:
 - name: label-deepseek
   description: Automatically apply deepseek label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*deepseek.*\.py
       - files~=^tests/.*deepseek.*\.py
@@ -51,7 +46,6 @@ pull_request_rules:
 - name: label-frontend
   description: Automatically apply frontend label
   conditions:
-    - label != stale
     - files~=^vllm/entrypoints/
   actions:
     label:
@@ -61,7 +55,6 @@ pull_request_rules:
 - name: label-llama
   description: Automatically apply llama label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*llama.*\.py
       - files~=^tests/.*llama.*\.py
@@ -77,7 +70,6 @@ pull_request_rules:
 - name: label-multi-modality
   description: Automatically apply multi-modality label
   conditions:
-    - label != stale
     - or:
       - files~=^vllm/multimodal/
       - files~=^tests/multimodal/
@@ -91,7 +83,6 @@ pull_request_rules:
 - name: label-new-model
   description: Automatically apply new-model label
   conditions:
-    - label != stale
     - and:
       - files~=^vllm/model_executor/models/
       - files=vllm/model_executor/models/registry.py
@@ -103,7 +94,6 @@ pull_request_rules:
 - name: label-performance
   description: Automatically apply performance label
   conditions:
-    - label != stale
     - or:
       - files~=^benchmarks/
       - files~=^vllm/benchmarks/
@@ -117,7 +107,6 @@ pull_request_rules:
 - name: label-qwen
   description: Automatically apply qwen label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*qwen.*\.py
       - files~=^tests/.*qwen.*\.py
@@ -132,7 +121,6 @@ pull_request_rules:
 - name: label-gpt-oss
   description: Automatically apply gpt-oss label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*gpt[-_]?oss.*\.py
       - files~=^tests/.*gpt[-_]?oss.*\.py
@@ -154,7 +142,6 @@ pull_request_rules:
 - name: label-rocm
   description: Automatically apply rocm label
   conditions:
-    - label != stale
     - or:
       - files~=^csrc/rocm/
       - files~=^docker/Dockerfile.rocm
@@ -175,7 +162,6 @@ pull_request_rules:
 - name: label-structured-output
   description: Automatically apply structured-output label
   conditions:
-    - label != stale
     - or:
       - files~=^benchmarks/structured_schemas/
       - files=benchmarks/benchmark_serving_structured_output.py
@@ -195,7 +181,6 @@ pull_request_rules:
 - name: label-speculative-decoding
   description: Automatically apply speculative-decoding label
   conditions:
-    - label != stale
     - or:
       - files~=^vllm/v1/spec_decode/
       - files~=^tests/v1/spec_decode/
@@ -211,7 +196,6 @@ pull_request_rules:
 - name: label-v1
   description: Automatically apply v1 label
   conditions:
-    - label != stale
     - or:
       - files~=^vllm/v1/
       - files~=^tests/v1/
@@ -224,7 +208,6 @@ pull_request_rules:
   description: Automatically apply tpu label
   # Keep this list in sync with `label-tpu-remove` conditions
   conditions:
-    - label != stale
     - or:
       - files~=tpu.py
       - files~=_tpu
@@ -240,7 +223,6 @@ pull_request_rules:
   description: Automatically remove tpu label
   # Keep this list in sync with `label-tpu` conditions
   conditions:
-    - label != stale
     - and:
       - -files~=tpu.py
       - -files~=_tpu
@@ -255,9 +237,9 @@ pull_request_rules:
 - name: label-tool-calling
   description: Automatically add tool-calling label
   conditions:
-    - label != stale
     - or:
       - files~=^tests/tool_use/
+      - files~=^tests/mistral_tool_use/
       - files~=^tests/entrypoints/openai/tool_parsers/
       - files=tests/entrypoints/openai/test_chat_with_tool_reasoning.py
       - files~=^vllm/entrypoints/openai/tool_parsers/
@@ -274,9 +256,8 @@ pull_request_rules:
 
 - name: ping author on conflicts and add 'needs-rebase' label
   conditions:
-    - label != stale
-    - conflict
-    - -closed
+      - conflict
+      - -closed
   actions:
     label:
       add:
@@ -290,12 +271,10 @@ pull_request_rules:
 
 - name: assign reviewer for tensorizer changes
   conditions:
-    - label != stale
-    - or:
       - files~=^vllm/model_executor/model_loader/tensorizer.py
       - files~=^vllm/model_executor/model_loader/tensorizer_loader.py
       - files~=^tests/entrypoints/openai/test_tensorizer_entrypoint.py
-      - files~=^tests/model_executor/model_loader/tensorizer_loader/
+      - files~=^tests/tensorizer_loader/
   actions:
     assign:
       users:
@@ -303,7 +282,6 @@ pull_request_rules:
 
 - name: assign reviewer for modelopt changes
   conditions:
-    - label != stale
     - or:
         - files~=^vllm/model_executor/layers/quantization/modelopt\.py$
         - files~=^vllm/model_executor/layers/quantization/__init__\.py$
@@ -318,8 +296,8 @@ pull_request_rules:
 
 - name: remove 'needs-rebase' label when conflict is resolved
   conditions:
-    - -conflict
-    - -closed
+      - -conflict
+      - -closed
   actions:
     label:
       remove:
@@ -328,7 +306,6 @@ pull_request_rules:
 - name: label-kv-connector
   description: Automatically apply kv-connector label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/online_serving/disaggregated[^/]*/.*
       - files~=^examples/offline_inference/disaggregated[^/]*/.*

.github/workflows/stale.yml

@@ -13,7 +13,7 @@ jobs:
       actions: write
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/stale@5f858e3efba33a5ca4407a664cc011ad407f2008 # v10.1.0
+      - uses: actions/stale@3a9db7e6a41a89f618792c92c0e97cc736e1b13f # v10.0.0
         with:
           # Increasing this value ensures that changes to this workflow
           # propagate to all issues and PRs in days rather than months

.pre-commit-config.yaml

@@ -6,18 +6,30 @@ default_stages:
   - manual # Run in CI
 exclude: 'vllm/third_party/.*'
 repos:
-- repo: https://github.com/astral-sh/ruff-pre-commit
-  rev: v0.14.0
+- repo: https://github.com/google/yapf
+  rev: v0.43.0
   hooks:
-  - id: ruff-check
+  - id: yapf
+    args: [--in-place, --verbose]
+    # Keep the same list from yapfignore here to avoid yapf failing without any inputs
+    exclude: '(.buildkite|benchmarks|build|examples)/.*'
+- repo: https://github.com/astral-sh/ruff-pre-commit
+  rev: v0.11.7
+  hooks:
+  - id: ruff
     args: [--output-format, github, --fix]
   - id: ruff-format
+    files: ^(.buildkite|benchmarks|examples)/.*
 - repo: https://github.com/crate-ci/typos
-  rev: v1.38.1
+  rev: v1.35.5
   hooks:
   - id: typos
+- repo: https://github.com/PyCQA/isort
+  rev: 6.0.1
+  hooks:
+  - id: isort
 - repo: https://github.com/pre-commit/mirrors-clang-format
-  rev: v21.1.2
+  rev: v20.1.3
   hooks:
   - id: clang-format
     exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
@@ -34,10 +46,10 @@ repos:
   hooks:
   - id: actionlint
 - repo: https://github.com/astral-sh/uv-pre-commit
-  rev: 0.9.1
+  rev: 0.6.17
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128, --python-platform, x86_64-manylinux_2_28]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
@@ -48,32 +60,38 @@ repos:
     files: ^requirements/test\.(in|txt)$
   - id: mypy-local
     name: Run mypy for local Python installation
-    entry: python tools/pre_commit/mypy.py 0 "local"
+    entry: tools/mypy.sh 0 "local"
+    language: python
+    types: [python]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, types-setuptools, types-PyYAML, types-requests, pydantic]
     stages: [pre-commit] # Don't run in CI
-    <<: &mypy_common
-      language: python
-      types_or: [python, pyi]
-      require_serial: true
-      additional_dependencies: [mypy==1.11.1, regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
+  - 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
+    stages: [manual] # Only run in CI
   - 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: python tools/pre_commit/mypy.py 1 "3.10"
-    <<: *mypy_common
+    entry: tools/mypy.sh 1 "3.10"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
   - 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: python tools/pre_commit/mypy.py 1 "3.11"
-    <<: *mypy_common
+    entry: tools/mypy.sh 1 "3.11"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
   - 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: python tools/pre_commit/mypy.py 1 "3.12"
-    <<: *mypy_common
-    stages: [manual] # Only run in CI
-  - id: mypy-3.13 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
-    name: Run mypy for Python 3.13
-    entry: python tools/pre_commit/mypy.py 1 "3.13"
-    <<: *mypy_common
+    entry: tools/mypy.sh 1 "3.12"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
   - id: shellcheck
     name: Lint shell scripts
@@ -137,10 +155,11 @@ repos:
     additional_dependencies: [regex]
   - id: check-pickle-imports
     name: Prevent new pickle/cloudpickle imports
-    entry: python tools/pre_commit/check_pickle_imports.py
+    entry: python tools/check_pickle_imports.py
     language: python
     types: [python]
-    additional_dependencies: [regex]
+    pass_filenames: false
+    additional_dependencies: [pathspec, regex]
   - id: validate-config
     name: Validate configuration has default values and that each field has a docstring
     entry: python tools/validate_config.py

.readthedocs.yaml

@@ -13,7 +13,6 @@ build:
 
 mkdocs:
   configuration: mkdocs.yaml
-  fail_on_warning: true
 
 # Optionally declare the Python requirements required to build your docs
 python:

CMakeLists.txt

@@ -34,10 +34,10 @@ install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
 #
-set(PYTHON_SUPPORTED_VERSIONS "3.10" "3.11" "3.12" "3.13")
+set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12" "3.13")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -86,9 +86,6 @@ find_package(Torch REQUIRED)
 # Supported NVIDIA architectures.
 # This check must happen after find_package(Torch) because that's when CMAKE_CUDA_COMPILER_VERSION gets defined
 if(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
-   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
-  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;11.0;12.0")
-elseif(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
    CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8)
   set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0")
 else()
@@ -178,15 +175,6 @@ if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
 endif()
 
-#
-# Set compression mode for CUDA >=13.x.
-#
-if(VLLM_GPU_LANG STREQUAL "CUDA" AND
-   DEFINED CMAKE_CUDA_COMPILER_VERSION AND
-   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
-  list(APPEND VLLM_GPU_FLAGS "--compress-mode=size")
-endif()
-
 #
 # Set CUDA include flags for CXX compiler.
 #
@@ -269,8 +257,8 @@ set(VLLM_EXT_SRC
   "csrc/sampler.cu"
   "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
-  "csrc/quantization/w8a8/int8/scaled_quant.cu"
-  "csrc/quantization/w8a8/fp8/common.cu"
+  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
+  "csrc/quantization/fp8/common.cu"
   "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
   "csrc/quantization/gguf/gguf_kernel.cu"
   "csrc/quantization/activation_kernels.cu"
@@ -282,7 +270,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
   # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
-  set(CUTLASS_REVISION "v4.2.1" CACHE STRING "CUTLASS revision to use")
+  set(CUTLASS_REVISION "v4.0.0" CACHE STRING "CUTLASS revision to use")
 
   # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
   if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
@@ -314,13 +302,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_EXT_SRC
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/permute_cols.cu"
-    "csrc/quantization/w8a8/cutlass/scaled_mm_entry.cu"
+    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
     "csrc/quantization/fp4/nvfp4_quant_entry.cu"
     "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
+    "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu"
     "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
     "csrc/cutlass_extensions/common.cpp"
-    "csrc/quantization/w8a8/fp8/per_token_group_quant.cu"
-    "csrc/quantization/w8a8/int8/per_token_group_quant.cu")
+    "csrc/quantization/fp8/per_token_group_quant.cu")
 
   set_gencode_flags_for_srcs(
     SRCS "${VLLM_EXT_SRC}"
@@ -424,11 +412,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
     set(SRCS
-       "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm90.cu"
-       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_fp8.cu"
-       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_int8.cu"
-       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_azp_sm90_int8.cu"
-       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm90_fp8.cu")
+       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
+       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -452,16 +440,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # The cutlass_scaled_mm kernels for Geforce Blackwell SM120 (c3x, i.e. CUTLASS 3.x) require
   # CUDA 12.8 or later
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
-      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm120.cu"
-      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm120_fp8.cu"
-      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm120_fp8.cu"
     )
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -486,16 +470,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
   # require CUDA 12.8 or later
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
-      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm100.cu"
-      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm100_fp8.cu"
-      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm100_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
     )
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -526,7 +506,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # subtract out the archs that are already built for 3x
   list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
   if (SCALED_MM_2X_ARCHS)
-    set(SRCS "csrc/quantization/w8a8/cutlass/scaled_mm_c2x.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_2X_ARCHS}")
@@ -570,11 +550,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # The nvfp4_scaled_mm_sm120 kernels for Geforce Blackwell SM120 require
   # CUDA 12.8 or later
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(FP4_ARCHS "12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(FP4_ARCHS "12.0a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(FP4_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
@@ -593,11 +569,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # FP4 Archs and flags
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(FP4_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(FP4_ARCHS "10.0a;10.1a;12.0a;12.1a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
@@ -619,11 +591,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # CUTLASS MLA Archs and flags
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(MLA_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(MLA_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
     set(SRCS
       "csrc/attention/mla/sm100_cutlass_mla_kernel.cu")
@@ -649,7 +617,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # if it's possible to compile MoE kernels that use its output.
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm90.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm90.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -667,13 +635,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm100.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -692,13 +656,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # moe_data.cu is used by all CUTLASS MoE kernels.
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
-    set(SRCS "csrc/quantization/w8a8/cutlass/moe/moe_data.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
@@ -715,13 +675,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
-  else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
-  endif()
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/w8a8/cutlass/moe/blockwise_scaled_group_mm_sm100.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -1007,7 +963,6 @@ endif()
 # For CUDA we also build and ship some external projects.
 if (VLLM_GPU_LANG STREQUAL "CUDA")
     include(cmake/external_projects/flashmla.cmake)
-    include(cmake/external_projects/qutlass.cmake)
 
     # vllm-flash-attn should be last as it overwrites some CMake functions
     include(cmake/external_projects/vllm_flash_attn.cmake)

README.md

@@ -21,7 +21,6 @@ Join us at the [PyTorch Conference, October 22-23](https://events.linuxfoundatio
 
 *Latest News* 🔥
 
-- [2025/09] We hosted [vLLM Toronto Meetup](https://luma.com/e80e0ymm) focused on tackling inference at scale and speculative decoding with speakers from NVIDIA and Red Hat! Please find the meetup slides [here](https://docs.google.com/presentation/d/1IYJYmJcu9fLpID5N5RbW_vO0XLo0CGOR14IXOjB61V8/edit?usp=sharing).
 - [2025/08] We hosted [vLLM Shenzhen Meetup](https://mp.weixin.qq.com/s/k8ZBO1u2_2odgiKWH_GVTQ) focusing on the ecosystem around vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Ua2SVKVSu-wp5vou_6ElraDt2bnKhiEA).
 - [2025/08] We hosted [vLLM Singapore Meetup](https://www.sginnovate.com/event/vllm-sg-meet). We shared V1 updates, disaggregated serving and MLLM speedups with speakers from Embedded LLM, AMD, WekaIO, and A*STAR. Please find the meetup slides [here](https://drive.google.com/drive/folders/1ncf3GyqLdqFaB6IeB834E5TZJPLAOiXZ?usp=sharing).
 - [2025/08] We hosted [vLLM Shanghai Meetup](https://mp.weixin.qq.com/s/pDmAXHcN7Iqc8sUKgJgGtg) focusing on building, developing, and integrating with vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1OvLx39wnCGy_WKq8SiVKf7YcxxYI3WCH).
@@ -149,7 +148,6 @@ Compute Resources:
 - Trainy
 - UC Berkeley
 - UC San Diego
-- Volcengine
 
 Slack Sponsor: Anyscale
 

benchmarks/auto_tune/auto_tune.sh

@@ -74,7 +74,7 @@ start_server() {
     local vllm_log=$4
     local profile_dir=$5
 
-    pkill -if "vllm serve" || true
+    pkill -if vllm
 
     # Define the common arguments as a bash array.
     # Each argument and its value are separate elements.
@@ -96,22 +96,17 @@ start_server() {
     # This correctly passes each element as a separate argument.
     if [[ -n "$profile_dir" ]]; then
         # Start server with profiling enabled
-        VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir \
+        VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir \
             vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
     else
         # Start server without profiling
-        VLLM_SERVER_DEV_MODE=1 \
+        VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 \
             vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
     fi
-    local server_pid=$!
 
     # wait for 10 minutes...
     server_started=0
     for i in {1..60}; do
-        # This line checks whether the server is still alive or not,
-        # since that we should always have permission to send signal to the server process.
-        kill -0 $server_pid 2> /dev/null || break
-
         RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout)
         STATUS_CODE=$(echo "$RESPONSE" | tail -n 1)
         if [[ "$STATUS_CODE" -eq 200 ]]; then
@@ -123,7 +118,7 @@ start_server() {
     done
 
     if (( ! server_started )); then
-        echo "server did not start within 10 minutes or crashed. Please check server log at $vllm_log".
+        echo "server did not start within 10 minutes. Please check server log at $vllm_log".
         return 1
     else
         return 0
@@ -139,7 +134,7 @@ run_benchmark() {
     echo "vllm_log: $vllm_log"
     echo
     rm -f $vllm_log
-    pkill -if "vllm serve" || true
+    pkill -if vllm
 
     echo "starting server..."
     # Call start_server without a profile_dir to avoid profiling overhead
@@ -232,7 +227,7 @@ run_benchmark() {
 
     echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
 
-    pkill -if "vllm serve" || true
+    pkill -if vllm
     sleep 10
     echo "===================="
     return 0
@@ -308,6 +303,6 @@ if (( $(echo "$best_throughput > 0" | bc -l) )); then
 else
     echo "No configuration met the latency requirements. Skipping final profiling run."
 fi
-pkill -if "vllm serve" || true
+pkill -if vllm
 echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH"
 echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT"

benchmarks/benchmark_block_pool.py

@@ -2,9 +2,9 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import gc
 
-from benchmark_utils import TimeCollector
 from tabulate import tabulate
 
+from benchmark_utils import TimeCollector
 from vllm.utils import FlexibleArgumentParser
 from vllm.v1.core.block_pool import BlockPool
 

benchmarks/benchmark_ngram_proposer.py

@@ -1,31 +1,17 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import gc
-import time
-from unittest import mock
 
 import numpy as np
-from benchmark_utils import TimeCollector
 from tabulate import tabulate
 
-from vllm.config import (
-    CacheConfig,
-    DeviceConfig,
-    LoadConfig,
-    ModelConfig,
-    ParallelConfig,
-    SchedulerConfig,
-    SpeculativeConfig,
-    VllmConfig,
-)
-from vllm.platforms import current_platform
+from benchmark_utils import TimeCollector
+from vllm.config import ModelConfig, SpeculativeConfig, VllmConfig
 from vllm.utils import FlexibleArgumentParser
 from vllm.v1.spec_decode.ngram_proposer import NgramProposer
-from vllm.v1.worker.gpu_input_batch import InputBatch
-from vllm.v1.worker.gpu_model_runner import GPUModelRunner
 
 
-def benchmark_propose(args):
+def main(args):
     rows = []
     for max_ngram in args.max_ngram:
         collector = TimeCollector(TimeCollector.US)
@@ -83,88 +69,10 @@ def benchmark_propose(args):
     )
 
 
-def benchmark_batched_propose(args):
-    NUM_SPECULATIVE_TOKENS_NGRAM = 10
-    PROMPT_LOOKUP_MIN = 5
-    PROMPT_LOOKUP_MAX = 15
-    MAX_MODEL_LEN = int(1e7)
-    DEVICE = current_platform.device_type
-
-    model_config = ModelConfig(model="facebook/opt-125m", runner="generate")
-
-    speculative_config = SpeculativeConfig(
-        target_model_config=model_config,
-        target_parallel_config=ParallelConfig(),
-        method="ngram",
-        num_speculative_tokens=NUM_SPECULATIVE_TOKENS_NGRAM,
-        prompt_lookup_max=PROMPT_LOOKUP_MAX,
-        prompt_lookup_min=PROMPT_LOOKUP_MIN,
-    )
-
-    vllm_config = VllmConfig(
-        model_config=model_config,
-        cache_config=CacheConfig(),
-        speculative_config=speculative_config,
-        device_config=DeviceConfig(device=current_platform.device_type),
-        parallel_config=ParallelConfig(),
-        load_config=LoadConfig(),
-        scheduler_config=SchedulerConfig(),
-    )
-
-    # monkey patch vllm.v1.worker.gpu_model_runner.get_pp_group
-    mock_pp_group = mock.MagicMock()
-    mock_pp_group.world_size = 1
-    with mock.patch(
-        "vllm.v1.worker.gpu_model_runner.get_pp_group", return_value=mock_pp_group
-    ):
-        runner = GPUModelRunner(vllm_config, DEVICE)
-
-        # hack max model len
-        runner.max_model_len = MAX_MODEL_LEN
-        runner.drafter.max_model_len = MAX_MODEL_LEN
-
-        dummy_input_batch = InputBatch(
-            max_num_reqs=args.num_req,
-            max_model_len=MAX_MODEL_LEN,
-            max_num_batched_tokens=args.num_req * args.num_token,
-            device=DEVICE,
-            pin_memory=False,
-            vocab_size=256000,
-            block_sizes=[16],
-        )
-        dummy_input_batch._req_ids = list(str(id) for id in range(args.num_req))
-        dummy_input_batch.spec_decode_unsupported_reqs = ()
-        dummy_input_batch.num_tokens_no_spec = [args.num_token] * args.num_req
-        dummy_input_batch.token_ids_cpu = np.random.randint(
-            0, 20, (args.num_req, args.num_token)
-        )
-
-        runner.input_batch = dummy_input_batch
-
-        sampled_token_ids = [[0]] * args.num_req
-
-        print("Starting benchmark")
-        # first run is warmup so ignore it
-        for _ in range(args.num_iteration):
-            start = time.time()
-            runner.drafter.propose(
-                sampled_token_ids,
-                dummy_input_batch.req_ids,
-                dummy_input_batch.num_tokens_no_spec,
-                dummy_input_batch.token_ids_cpu,
-                dummy_input_batch.spec_decode_unsupported_reqs,
-            )
-            end = time.time()
-            print(f"Iteration time (s): {end - start}")
-
-
 def invoke_main() -> None:
     parser = FlexibleArgumentParser(
         description="Benchmark the performance of N-gram speculative decode drafting"
     )
-    parser.add_argument(
-        "--batched", action="store_true", help="consider time to prepare batch"
-    )
     parser.add_argument(
         "--num-iteration",
         type=int,
@@ -197,17 +105,8 @@ def invoke_main() -> None:
         help="Number of speculative tokens to generate",
     )
     args = parser.parse_args()
-
-    if not args.batched:
-        benchmark_propose(args)
-    else:
-        benchmark_batched_propose(args)
+    main(args)
 
 
-"""
-# Example command lines:
-# time python3 benchmarks/benchmark_ngram_proposer.py
-# time python3 benchmarks/benchmark_ngram_proposer.py --batched --num-iteration 4 --num-token 1000000 --num-req 128
-"""  # noqa: E501
 if __name__ == "__main__":
     invoke_main()  # pragma: no cover

benchmarks/benchmark_serving_structured_output.py

@@ -37,13 +37,14 @@ from typing import Optional
 import datasets
 import numpy as np
 import pandas as pd
+from tqdm.asyncio import tqdm
+from transformers import PreTrainedTokenizerBase
+
 from backend_request_func import (
     ASYNC_REQUEST_FUNCS,
     RequestFuncInput,
     RequestFuncOutput,
 )
-from tqdm.asyncio import tqdm
-from transformers import PreTrainedTokenizerBase
 
 try:
     from vllm.transformers_utils.tokenizer import get_tokenizer
@@ -448,8 +449,7 @@ async def benchmark(
     def prepare_extra_body(request) -> dict:
         extra_body = {}
         # Add the schema to the extra_body
-        extra_body["structured_outputs"] = {}
-        extra_body["structured_outputs"][request.structure_type] = request.schema
+        extra_body[request.structure_type] = request.schema
         return extra_body
 
     print("Starting initial single prompt test run...")
@@ -909,13 +909,13 @@ def create_argument_parser():
     parser.add_argument(
         "--tokenizer",
         type=str,
-        help="Name or path of the tokenizer, if not using the default tokenizer.",
+        help="Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
     parser.add_argument(
         "--tokenizer-mode",
         type=str,
         default="auto",
-        help="Name or path of the tokenizer, if not using the default tokenizer.",
+        help="Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
     parser.add_argument(
         "--num-prompts",

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -17,7 +17,7 @@ from weight_shapes import WEIGHT_SHAPES
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    w8a8_triton_block_scaled_mm,
+    w8a8_block_fp8_matmul,
 )
 from vllm.utils import FlexibleArgumentParser, cdiv
 
@@ -158,7 +158,7 @@ def bench_fp8(
         "cutlass_fp8_fp8_fp16_scaled_mm_bias": lambda: ops.cutlass_scaled_mm(
             a, b, scale_a, scale_b, torch.float16, bias.to(dtype=torch.float16)
         ),
-        "triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_triton_block_scaled_mm(
+        "triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_block_fp8_matmul(
             a_cont, b.t(), block_scale_a, block_scale_b.t(), (128, 128)
         ),
         "cutlass_fp8_fp8_fp16_scaled_mm_blockwise": lambda: ops.cutlass_scaled_mm(

benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh

@@ -55,7 +55,9 @@ benchmark() {
   output_len=$2
 
 
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=0 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8100 \
     --max-model-len 10000 \
     --gpu-memory-utilization 0.6 \
@@ -63,7 +65,9 @@ benchmark() {
     '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
 
 
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=1 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8200 \
     --max-model-len 10000 \
     --gpu-memory-utilization 0.6 \

benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh

@@ -38,12 +38,16 @@ wait_for_server() {
 launch_chunked_prefill() {
   model="meta-llama/Meta-Llama-3.1-8B-Instruct"
   # disagg prefill
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=0 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8100 \
     --max-model-len 10000 \
     --enable-chunked-prefill \
     --gpu-memory-utilization 0.6 &
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=1 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8200 \
     --max-model-len 10000 \
     --enable-chunked-prefill \
@@ -58,14 +62,18 @@ launch_chunked_prefill() {
 launch_disagg_prefill() {
   model="meta-llama/Meta-Llama-3.1-8B-Instruct"
   # disagg prefill
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=0 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8100 \
     --max-model-len 10000 \
     --gpu-memory-utilization 0.6 \
     --kv-transfer-config \
     '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
 
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
+  CUDA_VISIBLE_DEVICES=1 python3 \
+    -m vllm.entrypoints.openai.api_server \
+    --model $model \
     --port 8200 \
     --max-model-len 10000 \
     --gpu-memory-utilization 0.6 \

benchmarks/kernels/bench_mxfp4_qutlass.py

@@ -1,191 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-#
-# Copyright (C) 2025 Roberto L. Castro (Roberto.LopezCastro@ist.ac.at).
-# All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#       http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-
-import argparse
-import copy
-import itertools
-
-import torch
-from compressed_tensors.transform.utils.hadamard import deterministic_hadamard_matrix
-from weight_shapes import WEIGHT_SHAPES
-
-from vllm._custom_ops import fusedQuantizeMx, matmul_mxf4_bf16_tn
-from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
-from vllm.triton_utils import triton
-
-PROVIDER_CFGS = {
-    "torch-bf16": dict(enabled=True),
-    "mxfp4": dict(no_a_quant=False, enabled=True),
-    "mxfp4-noquant": dict(no_a_quant=True, enabled=True),
-}
-
-_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
-
-
-def get_hadamard_matrix(group_size: int, dtype: torch.dtype, device: torch.device):
-    return (
-        deterministic_hadamard_matrix(group_size, dtype=dtype, device=device)
-        * group_size**-0.5
-    )
-
-
-def _quant_weight_mxfp4(
-    b: torch.Tensor, forward_hadamard_matrix: torch.Tensor, device: str
-):
-    weight_hf_e2m1, weight_hf_e8m0 = fusedQuantizeMx(
-        b, forward_hadamard_matrix, method="abs_max"
-    )
-    weight_hf_scale_block = to_blocked(weight_hf_e8m0, backend="triton")
-    return weight_hf_e2m1, weight_hf_scale_block
-
-
-def build_mxfp4_runner(cfg, a, b, forward_hadamard_matrix, dtype, device):
-    weight_hf_e2m1, weight_hf_scale_block = _quant_weight_mxfp4(
-        b, forward_hadamard_matrix, device
-    )
-    alpha = torch.tensor([1.0], device="cuda")
-
-    if cfg["no_a_quant"]:
-        # Pre-quantize activation
-        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeMx(
-            a, forward_hadamard_matrix, method="abs_max"
-        )
-        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton")
-
-        def run():
-            return matmul_mxf4_bf16_tn(
-                input_hf_e2m1,
-                weight_hf_e2m1,
-                input_hf_scale_block,
-                weight_hf_scale_block,
-                alpha,
-            )
-
-        return run
-
-    # Quantize activation on-the-fly
-    def run():
-        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeMx(
-            a, forward_hadamard_matrix, method="abs_max"
-        )
-        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton")
-        return matmul_mxf4_bf16_tn(
-            input_hf_e2m1,
-            weight_hf_e2m1,
-            input_hf_scale_block,
-            weight_hf_scale_block,
-            alpha,
-        )
-
-    return run
-
-
-@triton.testing.perf_report(
-    triton.testing.Benchmark(
-        x_names=["batch_size"],
-        x_vals=[
-            1,
-            4,
-            8,
-            16,
-            32,
-            64,
-            128,
-            256,
-            512,
-            1024,
-            2048,
-            4096,
-            8192,
-            16384,
-            24576,
-            32768,
-        ],
-        x_log=False,
-        line_arg="provider",
-        line_vals=_enabled,
-        line_names=_enabled,
-        ylabel="TFLOP/s (larger is better)",
-        plot_name="BF16 vs MXFP4 GEMMs",
-        args={},
-    )
-)
-def benchmark(batch_size, provider, N, K, had_size):
-    M = batch_size
-    device = "cuda"
-    dtype = torch.bfloat16
-
-    a = torch.randn((M, K), device=device, dtype=dtype)
-    b = torch.randn((N, K), device=device, dtype=dtype)
-    forward_hadamard_matrix = get_hadamard_matrix(had_size, dtype, device)
-
-    quantiles = [0.5, 0.2, 0.8]
-
-    if provider == "torch-bf16":
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: torch.nn.functional.linear(a, b), rep=200, quantiles=quantiles
-        )
-    else:
-        cfg = PROVIDER_CFGS[provider]
-        run_quant = build_mxfp4_runner(
-            cfg, a, b, forward_hadamard_matrix, dtype, device
-        )
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: run_quant(), rep=200, quantiles=quantiles
-        )
-
-    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
-    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
-
-
-def prepare_shapes(args):
-    out = []
-    for model, tp_size in itertools.product(args.models, args.tp_sizes):
-        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
-            KN[tp_dim] //= tp_size
-            KN.append(model)
-            out.append(KN)
-    return out
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--models",
-        nargs="+",
-        type=str,
-        default=["meta-llama/Llama-3.3-70B-Instruct"],
-        choices=list(WEIGHT_SHAPES.keys()),
-    )
-    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
-    args = parser.parse_args()
-
-    for K, N, model in prepare_shapes(args):
-        for had_size in [32, 64, 128]:
-            print(f"{model}, N={N} K={K}, HAD={had_size}, BF16 vs MXFP4 GEMMs TFLOP/s:")
-            benchmark.run(
-                print_data=True,
-                show_plots=True,
-                save_path=f"bench_mxfp4_res_n{N}_k{K}",
-                N=N,
-                K=K,
-                had_size=had_size,
-            )
-
-    print("Benchmark finished!")

benchmarks/kernels/bench_nvfp4_gemm.py

@@ -3,7 +3,6 @@
 import argparse
 import copy
 import itertools
-import os
 
 import torch
 from weight_shapes import WEIGHT_SHAPES
@@ -24,45 +23,21 @@ PROVIDER_CFGS = {
     "torch-bf16": dict(enabled=True),
     "nvfp4": dict(no_a_quant=False, enabled=True),
     "nvfp4-noquant": dict(no_a_quant=True, enabled=True),
-    "fbgemm-nvfp4": dict(fbgemm=True, no_a_quant=False, enabled=True),
-    "fbgemm-nvfp4-noquant": dict(fbgemm=True, no_a_quant=True, enabled=True),
 }
 
-_needs_fbgemm = any(
-    v.get("fbgemm", False) for v in PROVIDER_CFGS.values() if v.get("enabled", False)
-)
-if _needs_fbgemm:
-    try:
-        from fbgemm_gpu.experimental.gemm.triton_gemm.fp4_quantize import (
-            triton_scale_nvfp4_quant,
-        )
-    except ImportError:
-        print(
-            "WARNING: FBGEMM providers are enabled but fbgemm_gpu is not installed. "
-            "These providers will be skipped. Please install fbgemm_gpu with: "
-            "'pip install fbgemm-gpu-genai' to run them."
-        )
-        # Disable FBGEMM providers so the benchmark can run.
-        for cfg in PROVIDER_CFGS.values():
-            if cfg.get("fbgemm"):
-                cfg["enabled"] = False
-
 _enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
 
 
-def _quant_weight_nvfp4(b: torch.Tensor, device: str, cfg):
+def _quant_weight_nvfp4(b: torch.Tensor, device: str):
     # Compute global scale for weight
     b_amax = torch.abs(b).max().to(torch.float32)
     b_global_scale = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / b_amax
-    if "fbgemm" in cfg and cfg["fbgemm"]:
-        b_fp4, scale_b_fp4 = triton_scale_nvfp4_quant(b, b_global_scale)
-    else:
-        b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
+    b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
     return b_fp4, scale_b_fp4, b_global_scale
 
 
 def build_nvfp4_runner(cfg, a, b, dtype, device):
-    b_fp4, scale_b_fp4, b_global_scale = _quant_weight_nvfp4(b, device, cfg)
+    b_fp4, scale_b_fp4, b_global_scale = _quant_weight_nvfp4(b, device)
 
     # Compute global scale for activation
     # NOTE: This is generally provided ahead-of-time by the model checkpoint.
@@ -71,35 +46,6 @@ def build_nvfp4_runner(cfg, a, b, dtype, device):
 
     # Alpha for the GEMM operation
     alpha = 1.0 / (a_global_scale * b_global_scale)
-    if "fbgemm" in cfg and cfg["fbgemm"]:
-        if cfg["no_a_quant"]:
-            a_fp4, scale_a_fp4 = triton_scale_nvfp4_quant(a, a_global_scale)
-
-            def run():
-                return torch.ops.fbgemm.f4f4bf16(
-                    a_fp4,
-                    b_fp4,
-                    scale_a_fp4,
-                    scale_b_fp4,
-                    global_scale=alpha,
-                    use_mx=False,
-                )
-
-            return run
-        else:
-
-            def run():
-                a_fp4, scale_a_fp4 = triton_scale_nvfp4_quant(a, a_global_scale)
-                return torch.ops.fbgemm.f4f4bf16(
-                    a_fp4,
-                    b_fp4,
-                    scale_a_fp4,
-                    scale_b_fp4,
-                    global_scale=alpha,
-                    use_mx=False,
-                )
-
-            return run
 
     if cfg["no_a_quant"]:
         # Pre-quantize activation
@@ -184,13 +130,10 @@ if __name__ == "__main__":
 
     for K, N, model in prepare_shapes(args):
         print(f"{model}, N={N} K={K}, BF16 vs NVFP4 GEMMs TFLOP/s:")
-        save_dir = f"bench_nvfp4_res_n{N}_k{K}"
-        os.makedirs(save_dir, exist_ok=True)
-
         benchmark.run(
             print_data=True,
             show_plots=True,
-            save_path=save_dir,
+            save_path=f"bench_nvfp4_res_n{N}_k{K}",
             N=N,
             K=K,
         )

benchmarks/kernels/bench_nvfp4_qutlass.py

@@ -1,207 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-#
-# Copyright (C) 2025 Roberto L. Castro (Roberto.LopezCastro@ist.ac.at).
-# All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#       http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-
-import argparse
-import copy
-import itertools
-
-import torch
-from compressed_tensors.transform.utils.hadamard import deterministic_hadamard_matrix
-from weight_shapes import WEIGHT_SHAPES
-
-from vllm import _custom_ops as ops  # use existing nvfp4 gemm in vllm
-from vllm._custom_ops import fusedQuantizeNv
-from vllm.model_executor.layers.quantization.qutlass_utils import to_blocked
-from vllm.triton_utils import triton
-
-PROVIDER_CFGS = {
-    "torch-bf16": dict(enabled=True),
-    "nvfp4": dict(no_a_quant=False, enabled=True),
-    "nvfp4-noquant": dict(no_a_quant=True, enabled=True),
-}
-
-_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
-
-
-def get_hadamard_matrix(group_size: int, dtype: torch.dtype, device: torch.device):
-    return (
-        deterministic_hadamard_matrix(group_size, dtype=dtype, device=device)
-        * group_size**-0.5
-    )
-
-
-def _quant_weight_nvfp4(
-    b: torch.Tensor,
-    forward_hadamard_matrix: torch.Tensor,
-    global_scale: torch.Tensor,
-    device: str,
-    M: int,
-    N: int,
-    K: int,
-):
-    weight_hf_e2m1, weight_hf_e8m0 = fusedQuantizeNv(
-        b, forward_hadamard_matrix, global_scale
-    )
-    weight_hf_scale_block = to_blocked(weight_hf_e8m0, backend="triton").view(
-        -1, K // 16
-    )
-    return weight_hf_e2m1, weight_hf_scale_block
-
-
-def build_nvfp4_runner(cfg, a, b, forward_hadamard_matrix, dtype, device, M, N, K):
-    alpha = torch.tensor([1.0], device="cuda")
-    global_scale = torch.tensor([1.0], device="cuda")
-    weight_hf_e2m1, weight_hf_scale_block = _quant_weight_nvfp4(
-        b, forward_hadamard_matrix, global_scale, device, M, N, K
-    )
-
-    if cfg["no_a_quant"]:
-        # Pre-quantize activation
-        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeNv(
-            a, forward_hadamard_matrix, global_scale
-        )
-        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton").view(
-            -1, K // 16
-        )
-
-        def run():
-            return ops.cutlass_scaled_fp4_mm(
-                input_hf_e2m1,
-                weight_hf_e2m1,
-                input_hf_scale_block,
-                weight_hf_scale_block,
-                alpha,
-                torch.bfloat16,
-            )
-
-        return run
-
-    # Quantize activation on-the-fly
-    def run():
-        input_hf_e2m1, input_hf_e8m0 = fusedQuantizeNv(
-            a, forward_hadamard_matrix, global_scale
-        )
-        input_hf_scale_block = to_blocked(input_hf_e8m0, backend="triton").view(
-            -1, K // 16
-        )
-        return ops.cutlass_scaled_fp4_mm(
-            input_hf_e2m1,
-            weight_hf_e2m1,
-            input_hf_scale_block,
-            weight_hf_scale_block,
-            alpha,
-            torch.bfloat16,
-        )
-
-    return run
-
-
-@triton.testing.perf_report(
-    triton.testing.Benchmark(
-        x_names=["batch_size"],
-        x_vals=[
-            1,
-            4,
-            8,
-            16,
-            32,
-            64,
-            128,
-            256,
-            512,
-            1024,
-            2048,
-            4096,
-            8192,
-            16384,
-            24576,
-            32768,
-        ],
-        x_log=False,
-        line_arg="provider",
-        line_vals=_enabled,
-        line_names=_enabled,
-        ylabel="TFLOP/s (larger is better)",
-        plot_name="BF16 vs NVFP4 GEMMs",
-        args={},
-    )
-)
-def benchmark(batch_size, provider, N, K, had_size):
-    M = batch_size
-    device = "cuda"
-    dtype = torch.bfloat16
-
-    a = torch.randn((M, K), device=device, dtype=dtype)
-    b = torch.randn((N, K), device=device, dtype=dtype)
-    forward_hadamard_matrix = get_hadamard_matrix(had_size, dtype, device)
-
-    quantiles = [0.5, 0.2, 0.8]
-
-    if provider == "torch-bf16":
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: torch.nn.functional.linear(a, b), rep=200, quantiles=quantiles
-        )
-    else:
-        cfg = PROVIDER_CFGS[provider]
-        run_quant = build_nvfp4_runner(
-            cfg, a, b, forward_hadamard_matrix, dtype, device, M, N, K
-        )
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
-            lambda: run_quant(), rep=200, quantiles=quantiles
-        )
-
-    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
-    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
-
-
-def prepare_shapes(args):
-    out = []
-    for model, tp_size in itertools.product(args.models, args.tp_sizes):
-        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
-            KN[tp_dim] //= tp_size
-            KN.append(model)
-            out.append(KN)
-    return out
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--models",
-        nargs="+",
-        type=str,
-        default=["meta-llama/Llama-3.3-70B-Instruct"],
-        choices=list(WEIGHT_SHAPES.keys()),
-    )
-    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
-    args = parser.parse_args()
-
-    for K, N, model in prepare_shapes(args):
-        for had_size in [16, 32, 64, 128]:
-            print(f"{model}, N={N} K={K}, HAD={had_size}, BF16 vs NVFP4 GEMMs TFLOP/s:")
-            benchmark.run(
-                print_data=True,
-                show_plots=True,
-                save_path=f"bench_nvfp4_res_n{N}_k{K}",
-                N=N,
-                K=K,
-                had_size=had_size,
-            )
-
-    print("Benchmark finished!")

benchmarks/kernels/bench_per_token_quant_fp8.py

@@ -51,7 +51,7 @@ def calculate_diff(
 ):
     """Calculate the difference between Inductor and CUDA implementations."""
     device = torch.device("cuda")
-    x = torch.randn((batch_size, hidden_size), dtype=dtype, device=device)
+    x = torch.rand((batch_size * hidden_size, 4096), dtype=dtype, device=device)
 
     quant_fp8 = QuantFP8(False, group_shape, column_major_scales=False)
 
@@ -59,25 +59,23 @@ def calculate_diff(
     torch_eager_out, torch_eager_scale = quant_fp8.forward_native(x)
     cuda_out, cuda_scale = quant_fp8.forward_cuda(x)
 
-    try:
-        torch.testing.assert_close(
-            cuda_out.to(torch.float32),
-            torch_out.to(torch.float32),
-            rtol=1e-3,
-            atol=1e-5,
-        )
-        torch.testing.assert_close(cuda_scale, torch_scale, rtol=1e-3, atol=1e-5)
-        torch.testing.assert_close(
-            cuda_out.to(torch.float32),
-            torch_eager_out.to(torch.float32),
-            rtol=1e-3,
-            atol=1e-5,
-        )
-        torch.testing.assert_close(cuda_scale, torch_eager_scale, rtol=1e-3, atol=1e-5)
+    out_allclose = lambda o1, o2: torch.allclose(
+        o1.to(torch.float32),
+        o2.to(torch.float32),
+        rtol=1e-3,
+        atol=1e-5,
+    )
+    scale_allclose = lambda s1, s2: torch.allclose(s1, s2, rtol=1e-3, atol=1e-5)
+
+    if (
+        out_allclose(cuda_out, torch_out)
+        and scale_allclose(cuda_scale, torch_scale)
+        and out_allclose(cuda_out, torch_eager_out)
+        and scale_allclose(cuda_scale, torch_eager_scale)
+    ):
         print("✅ All implementations match")
-    except AssertionError as e:
+    else:
         print("❌ Implementations differ")
-        print(e)
 
 
 configs = []
@@ -93,7 +91,7 @@ def benchmark_quantization(
 ):
     device = torch.device("cuda")
 
-    x = torch.randn(batch_size, hidden_size, device=device, dtype=dtype)
+    x = torch.randn(batch_size * hidden_size, 4096, device=device, dtype=dtype)
 
     quantiles = [0.5, 0.2, 0.8]
     quant_fp8 = QuantFP8(False, group_shape, column_major_scales=col_major)
@@ -159,21 +157,21 @@ if __name__ == "__main__":
     )
     parser.add_argument("-c", "--check", action="store_true")
     parser.add_argument(
-        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="bfloat16"
+        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="half"
     )
     parser.add_argument(
         "--hidden-sizes",
         type=int,
         nargs="+",
-        default=[896, 1024, 2048, 4096, 7168],
-        help="Hidden sizes to benchmark",
+        default=None,
+        help="Hidden sizes to benchmark (default: 1,16,64,128,256,512,1024,2048,4096)",
     )
     parser.add_argument(
         "--batch-sizes",
         type=int,
         nargs="+",
-        default=[1, 16, 128, 512, 1024],
-        help="Batch sizes to benchmark",
+        default=None,
+        help="Batch sizes to benchmark (default: 1,16,32,64,128)",
     )
     parser.add_argument(
         "--group-sizes",
@@ -194,8 +192,8 @@ if __name__ == "__main__":
 
     dtype = STR_DTYPE_TO_TORCH_DTYPE[args.dtype]
 
-    hidden_sizes = args.hidden_sizes
-    batch_sizes = args.batch_sizes
+    hidden_sizes = args.hidden_sizes or [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
+    batch_sizes = args.batch_sizes or [1, 16, 32, 64, 128]
 
     if args.group_sizes is not None:
         group_shapes = []

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -1,406 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""
-Benchmark the performance of the cutlass_moe_fp8 kernel vs the triton_moe
-kernel. Both kernels take in fp8 quantized weights and 16-bit activations,
-but use different quantization strategies and backends.
-"""
-
-import nvtx
-import torch
-
-from vllm import _custom_ops as ops
-from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
-from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8
-from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
-from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
-
-# Weight shapes for different models: [num_experts, topk, hidden_size,
-# intermediate_size]
-WEIGHT_SHAPES_MOE = {
-    "mixtral-8x7b": [
-        [8, 2, 4096, 14336],
-    ],
-    "deepseek-v2": [
-        [160, 6, 5120, 12288],
-    ],
-    "custom-small": [
-        [8, 2, 2048, 7168],
-    ],
-    "glm45-fp8": [
-        [128, 8, 4096, 1408],
-    ],
-    "Llama-4-Maverick-17B-128E-Instruct-FP8": [
-        [128, 1, 5120, 8192],
-    ],
-}
-
-DEFAULT_MODELS = [
-    "mixtral-8x7b",
-]
-
-DEFAULT_BATCH_SIZES = [4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048]
-DEFAULT_TP_SIZES = [1]
-
-PER_ACT_TOKEN_OPTS = [False, True]
-PER_OUT_CH_OPTS = [False, True]
-
-FP8_DTYPE = current_platform.fp8_dtype()
-
-
-def bench_run(
-    results: list,
-    model: str,
-    num_experts: int,
-    topk: int,
-    per_act_token: bool,
-    per_out_ch: bool,
-    mkn: tuple[int, int, int],
-):
-    (m, k, n) = mkn
-
-    dtype = torch.half
-    device = "cuda"
-
-    # Create input activations
-    a = torch.randn((m, k), device=device, dtype=dtype) / 10
-
-    # Create weights
-    w1 = torch.randn((num_experts, 2 * n, k), device=device, dtype=dtype) / 10
-    w2 = torch.randn((num_experts, k, n), device=device, dtype=dtype) / 10
-
-    # Create FP8 quantized weights and scales for both kernels
-    w1_fp8q = torch.empty((num_experts, 2 * n, k), device=device, dtype=FP8_DTYPE)
-    w2_fp8q = torch.empty((num_experts, k, n), device=device, dtype=FP8_DTYPE)
-
-    # Create scales based on quantization strategy
-    if per_out_ch:
-        # Per-channel quantization
-        w1_scale = torch.empty(
-            (num_experts, 2 * n, 1), device=device, dtype=torch.float32
-        )
-        w2_scale = torch.empty((num_experts, k, 1), device=device, dtype=torch.float32)
-    else:
-        # Per-tensor quantization
-        w1_scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
-        w2_scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
-
-    # Quantize weights
-    for expert in range(num_experts):
-        if per_out_ch:
-            # Per-channel quantization - not yet implemented properly
-            # For now, fall back to per-tensor quantization
-            w1_fp8q[expert], w1_scale_temp = ops.scaled_fp8_quant(w1[expert])
-            w2_fp8q[expert], w2_scale_temp = ops.scaled_fp8_quant(w2[expert])
-            # Expand scalar scales to the expected per-channel shape
-            w1_scale[expert] = w1_scale_temp.expand(2 * n, 1)
-            w2_scale[expert] = w2_scale_temp.expand(k, 1)
-        else:
-            # Per-tensor quantization
-            w1_fp8q[expert], w1_scale_temp = ops.scaled_fp8_quant(w1[expert])
-            w2_fp8q[expert], w2_scale_temp = ops.scaled_fp8_quant(w2[expert])
-            # Store scalar scales in [1, 1] tensors
-            w1_scale[expert, 0, 0] = w1_scale_temp
-            w2_scale[expert, 0, 0] = w2_scale_temp
-
-    # Prepare weights for CUTLASS (no transpose needed)
-    w1_fp8q_cutlass = w1_fp8q  # Keep original [E, 2N, K]
-    w2_fp8q_cutlass = w2_fp8q  # Keep original [E, K, N]
-
-    # Create router scores and get topk
-    score = torch.randn((m, num_experts), device=device, dtype=dtype)
-    topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False)
-
-    # WORKAROUND: CUTLASS MoE FP8 has issues with per-token quantization
-    # Force per-tensor quantization for all cases to match working e2e setup
-    a1_scale = torch.full((), 1e-2, device=device, dtype=torch.float32)
-    a2_scale = torch.full((), 1e-2, device=device, dtype=torch.float32)
-
-    # Force per-tensor quantization for all cases
-    per_act_token = False
-
-    # Create stride tensors for CUTLASS
-    ab_strides1 = torch.full((num_experts,), k, dtype=torch.int64, device=device)
-    ab_strides2 = torch.full((num_experts,), n, dtype=torch.int64, device=device)
-    c_strides1 = torch.full((num_experts,), 2 * n, dtype=torch.int64, device=device)
-    c_strides2 = torch.full((num_experts,), k, dtype=torch.int64, device=device)
-
-    def run_triton_moe(
-        a: torch.Tensor,
-        w1: torch.Tensor,
-        w2: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        w1_scale: torch.Tensor,
-        w2_scale: torch.Tensor,
-        a1_scale: torch.Tensor,
-        a2_scale: torch.Tensor,
-        num_repeats: int,
-    ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            per_act_token_quant=per_act_token,
-            per_out_ch_quant=per_out_ch,
-        )
-
-        for _ in range(num_repeats):
-            fused_experts(
-                a,
-                w1,
-                w2,
-                topk_weights,
-                topk_ids,
-                quant_config=quant_config,
-            )
-
-    def run_cutlass_moe_fp8(
-        a: torch.Tensor,
-        w1: torch.Tensor,
-        w2: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        ab_strides1: torch.Tensor,
-        ab_strides2: torch.Tensor,
-        c_strides1: torch.Tensor,
-        c_strides2: torch.Tensor,
-        w1_scale: torch.Tensor,
-        w2_scale: torch.Tensor,
-        a1_scale: torch.Tensor,
-        a2_scale: torch.Tensor,
-        num_repeats: int,
-    ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            per_act_token_quant=per_act_token,
-            per_out_ch_quant=per_out_ch,
-        )
-
-        for _ in range(num_repeats):
-            with nvtx.annotate("cutlass_moe_fp8", color="blue"):
-                cutlass_moe_fp8(
-                    a=a,
-                    w1_q=w1,
-                    w2_q=w2,
-                    topk_weights=topk_weights,
-                    topk_ids=topk_ids,
-                    ab_strides1=ab_strides1,
-                    ab_strides2=ab_strides2,
-                    c_strides1=c_strides1,
-                    c_strides2=c_strides2,
-                    quant_config=quant_config,
-                    activation="silu",
-                    global_num_experts=num_experts,
-                )
-
-    # Pre-create quantization config to avoid creating it inside CUDA graph
-    quant_config = fp8_w8a8_moe_quant_config(
-        w1_scale=w1_scale,
-        w2_scale=w2_scale,
-        a1_scale=a1_scale,
-        a2_scale=a2_scale,
-        per_act_token_quant=per_act_token,
-        per_out_ch_quant=per_out_ch,
-    )
-
-    # Create CUDA graphs for CUTLASS (match benchmark_moe.py pattern exactly)
-    cutlass_stream = torch.cuda.Stream()
-    cutlass_graph = torch.cuda.CUDAGraph()
-    with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
-        # Capture 10 invocations like benchmark_moe.py
-        for _ in range(10):
-            cutlass_moe_fp8(
-                a=a,
-                w1_q=w1_fp8q_cutlass,
-                w2_q=w2_fp8q_cutlass,
-                topk_weights=topk_weights,
-                topk_ids=topk_ids,
-                ab_strides1=ab_strides1,
-                ab_strides2=ab_strides2,
-                c_strides1=c_strides1,
-                c_strides2=c_strides2,
-                quant_config=quant_config,
-                activation="silu",
-                global_num_experts=num_experts,
-            )
-    torch.cuda.synchronize()
-
-    # Create CUDA graphs for Triton (match benchmark_moe.py pattern exactly)
-    triton_stream = torch.cuda.Stream()
-    triton_graph = torch.cuda.CUDAGraph()
-    with torch.cuda.graph(triton_graph, stream=triton_stream):
-        # Capture 10 invocations like benchmark_moe.py
-        for _ in range(10):
-            fused_experts(
-                a,
-                w1_fp8q,
-                w2_fp8q,
-                topk_weights,
-                topk_ids,
-                quant_config=quant_config,
-            )
-    torch.cuda.synchronize()
-
-    def bench_cuda_graph(graph, num_warmup=5, num_iters=100):
-        """Benchmark CUDA graph using events like benchmark_moe.py"""
-        # Warmup
-        for _ in range(num_warmup):
-            graph.replay()
-        torch.cuda.synchronize()
-
-        # Timing
-        start_event = torch.cuda.Event(enable_timing=True)
-        end_event = torch.cuda.Event(enable_timing=True)
-
-        latencies = []
-        for _ in range(num_iters):
-            torch.cuda.synchronize()
-            start_event.record()
-            graph.replay()
-            end_event.record()
-            end_event.synchronize()
-            latencies.append(start_event.elapsed_time(end_event))
-
-        # Divide by 10 since graph contains 10 calls
-        return sum(latencies) / (num_iters * 10)
-
-    # Benchmark parameters
-    num_warmup = 5
-    num_iters = 100
-
-    # Benchmark only CUDA graphs (more reliable and faster)
-    # Benchmark Triton MoE with CUDA graphs
-    triton_graph_time = bench_cuda_graph(
-        triton_graph, num_warmup=num_warmup, num_iters=num_iters
-    )
-
-    # Benchmark CUTLASS MoE with CUDA graphs
-    cutlass_graph_time = bench_cuda_graph(
-        cutlass_graph, num_warmup=num_warmup, num_iters=num_iters
-    )
-
-    # Convert ms to us and return results
-    triton_time_us = triton_graph_time * 1000
-    cutlass_time_us = cutlass_graph_time * 1000
-
-    return {
-        "batch_size": m,
-        "triton_time_us": triton_time_us,
-        "cutlass_time_us": cutlass_time_us,
-    }
-
-
-def main(args):
-    print("Benchmarking models:")
-    for i, model in enumerate(args.models):
-        print(f"[{i}]  {model}")
-
-    all_results = []
-
-    for model in args.models:
-        for tp in args.tp_sizes:
-            for layer in WEIGHT_SHAPES_MOE[model]:
-                num_experts = layer[0]
-                topk = layer[1]
-                size_k = layer[2]
-                size_n = layer[3] // tp
-
-                if len(args.limit_k) > 0 and size_k not in args.limit_k:
-                    continue
-
-                if len(args.limit_n) > 0 and size_n not in args.limit_n:
-                    continue
-
-                for per_act_token in args.per_act_token_opts:
-                    for per_out_ch in args.per_out_ch_opts:
-                        print(
-                            f"\n=== {model}, experts={num_experts}, topk={topk},"
-                            f"per_act={per_act_token}, per_out_ch={per_out_ch} ==="
-                        )
-
-                        config_results = []
-                        for size_m in args.batch_sizes:
-                            mkn = (size_m, size_k, size_n)
-                            result = bench_run(
-                                [],  # Not used anymore
-                                model,
-                                num_experts,
-                                topk,
-                                per_act_token,
-                                per_out_ch,
-                                mkn,
-                            )
-                            if result:
-                                config_results.append(result)
-
-                        # Print results table for this configuration
-                        if config_results:
-                            print(
-                                f"\n{'Batch Size':<12}"
-                                f"{'Triton (us)':<15}"
-                                f"{'CUTLASS (us)':<15}"
-                            )
-                            print("-" * 45)
-                            for result in config_results:
-                                print(
-                                    f"{result['batch_size']:<12}"
-                                    f"{result['triton_time_us']:<15.2f}"
-                                    f"{result['cutlass_time_us']:<15.2f}"
-                                )
-
-                            all_results.extend(config_results)
-
-    print(f"\nTotal benchmarks completed: {len(all_results)}")
-
-
-if __name__ == "__main__":
-    parser = FlexibleArgumentParser(
-        description="""Benchmark CUTLASS FP8 MOE vs Triton FP8 FUSED MOE
-         across specified models/shapes/batches
-
-        Example usage:
-        python benchmark_cutlass_moe_fp8.py  \
-            --model "Llama-4-Maverick-17B-128E-Instruct-FP8"  \
-            --tp-sizes 8 \
-            --batch-size 2 4 8  \
-            --per-act-token-opts false \
-            --per-out-ch-opts false
-
-        """
-    )
-    parser.add_argument(
-        "--models",
-        nargs="+",
-        type=str,
-        default=DEFAULT_MODELS,
-        choices=WEIGHT_SHAPES_MOE.keys(),
-    )
-    parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES)
-    parser.add_argument(
-        "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES
-    )
-    parser.add_argument("--limit-k", nargs="+", type=int, default=[])
-    parser.add_argument("--limit-n", nargs="+", type=int, default=[])
-    parser.add_argument(
-        "--per-act-token-opts",
-        nargs="+",
-        type=lambda x: x.lower() == "true",
-        default=[False, True],
-        help="Per-activation token quantization options (true/false)",
-    )
-    parser.add_argument(
-        "--per-out-ch-opts",
-        nargs="+",
-        type=lambda x: x.lower() == "true",
-        default=[False, True],
-        help="Per-output channel quantization options (true/false)",
-    )
-
-    args = parser.parse_args()
-    main(args)

benchmarks/kernels/benchmark_device_communicators.py

@@ -7,10 +7,6 @@ Benchmark script for device communicators:
 CustomAllreduce (oneshot, twoshot), PyNcclCommunicator,
 and SymmMemCommunicator (multimem, two-shot).
 
-for NCCL symmetric memory you need to set the environment variables
-NCCL_NVLS_ENABLE=1 NCCL_CUMEM_ENABLE=1 VLLM_USE_NCCL_SYMM_MEM=1, otherwise NCCL does
-not use fast NVLS implementation for all reduce.
-
 Usage:
     torchrun --nproc_per_node=<N> benchmark_device_communicators.py [options]
 
@@ -30,13 +26,7 @@ import torch.distributed as dist
 from torch.distributed import ProcessGroup
 
 from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
-from vllm.distributed.device_communicators.pynccl import (
-    PyNcclCommunicator,
-    register_nccl_symmetric_ops,
-)
-from vllm.distributed.device_communicators.pynccl_allocator import (
-    set_graph_pool_id,
-)
+from vllm.distributed.device_communicators.pynccl import PyNcclCommunicator
 from vllm.distributed.device_communicators.symm_mem import SymmMemCommunicator
 from vllm.logger import init_logger
 from vllm.utils import FlexibleArgumentParser
@@ -108,7 +98,6 @@ class CommunicatorBenchmark:
             )
             if not self.pynccl_comm.disabled:
                 logger.info("Rank %s: PyNcclCommunicator initialized", self.rank)
-                register_nccl_symmetric_ops(self.pynccl_comm)
             else:
                 logger.info("Rank %s: PyNcclCommunicator disabled", self.rank)
                 self.pynccl_comm = None
@@ -205,15 +194,6 @@ class CommunicatorBenchmark:
                     None,  # no env variable needed
                 )
             )
-            communicators.append(
-                (
-                    "pynccl-symm",
-                    lambda t: torch.ops.vllm.all_reduce_symmetric_with_copy(t),
-                    lambda t: True,  # Always available if initialized
-                    nullcontext(),
-                    None,  # no env variable needed
-                )
-            )
 
         if self.symm_mem_comm_multimem is not None:
             comm = self.symm_mem_comm_multimem
@@ -291,9 +271,7 @@ class CommunicatorBenchmark:
                 # Capture the graph using context manager
                 with context:
                     graph = torch.cuda.CUDAGraph()
-                    graph_pool = torch.cuda.graph_pool_handle()
-                    set_graph_pool_id(graph_pool)
-                    with torch.cuda.graph(graph, pool=graph_pool):
+                    with torch.cuda.graph(graph):
                         for _ in range(CUDA_GRAPH_CAPTURE_CYCLES):
                             allreduce_fn(graph_input)
 

benchmarks/kernels/benchmark_lora.py

@@ -79,9 +79,9 @@ def make_rand_lora_weight_tensor(
 
 
 def make_rand_tensors(
-    a_shape: tuple[int, ...],
-    b_shape: tuple[int, ...],
-    c_shape: tuple[int, ...],
+    a_shape: tuple[int],
+    b_shape: tuple[int],
+    c_shape: tuple[int],
     a_dtype: torch.dtype,
     b_dtype: torch.dtype,
     c_dtype: torch.dtype,
@@ -243,7 +243,7 @@ class OpType(Enum):
         lora_rank: int,
         num_loras: int,
         num_slices: int,
-    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
+    ) -> tuple[tuple[int], tuple[int], tuple[int]]:
         """
         Given num_slices, return the shapes of the A, B, and C matrices
         in A x B = C, for the op_type

benchmarks/kernels/benchmark_moe.py

@@ -579,22 +579,18 @@ def main(args: argparse.Namespace):
         E = config.ffn_config.moe_num_experts
         topk = config.ffn_config.moe_top_k
         intermediate_size = config.ffn_config.ffn_hidden_size
-        hidden_size = config.hidden_size
     elif config.architectures[0] == "JambaForCausalLM":
         E = config.num_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
-        hidden_size = config.hidden_size
     elif config.architectures[0] in (
-        "DeepseekV2ForCausalLM",
         "DeepseekV3ForCausalLM",
-        "DeepseekV32ForCausalLM",
+        "DeepseekV2ForCausalLM",
         "Glm4MoeForCausalLM",
     ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
-        hidden_size = config.hidden_size
     elif config.architectures[0] in (
         "Qwen2MoeForCausalLM",
         "Qwen3MoeForCausalLM",
@@ -603,18 +599,10 @@ def main(args: argparse.Namespace):
         E = config.num_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
-        hidden_size = config.hidden_size
-    elif config.architectures[0] == "Qwen3VLMoeForConditionalGeneration":
-        text_config = config.get_text_config()
-        E = text_config.num_experts
-        topk = text_config.num_experts_per_tok
-        intermediate_size = text_config.moe_intermediate_size
-        hidden_size = text_config.hidden_size
     elif config.architectures[0] in ("HunYuanMoEV1ForCausalLM"):
         E = config.num_experts
         topk = config.moe_topk[0]
         intermediate_size = config.moe_intermediate_size[0]
-        hidden_size = config.hidden_size
     else:
         # Support for llama4
         config = config.get_text_config()
@@ -622,7 +610,6 @@ def main(args: argparse.Namespace):
         E = config.num_local_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
-        hidden_size = config.hidden_size
     enable_ep = bool(args.enable_expert_parallel)
     if enable_ep:
         ensure_divisibility(E, args.tp_size, "Number of experts")
@@ -631,6 +618,7 @@ def main(args: argparse.Namespace):
     else:
         ensure_divisibility(intermediate_size, args.tp_size, "intermediate_size")
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
+    hidden_size = config.hidden_size
     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"

benchmarks/kernels/benchmark_reshape_and_cache.py

@@ -1,174 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from __future__ import annotations
-
-import random
-import time
-
-import torch
-from tabulate import tabulate
-
-from vllm import _custom_ops as ops
-from vllm.logger import init_logger
-from vllm.platforms import current_platform
-from vllm.utils import (
-    STR_DTYPE_TO_TORCH_DTYPE,
-    FlexibleArgumentParser,
-    create_kv_caches_with_random,
-)
-
-logger = init_logger(__name__)
-
-
-@torch.inference_mode()
-def run_benchmark(
-    num_tokens: int,
-    num_heads: int,
-    head_size: int,
-    block_size: int,
-    num_blocks: int,
-    dtype: torch.dtype,
-    kv_cache_dtype: str,
-    num_iters: int,
-    benchmark_mode: str,
-    device: str = "cuda",
-) -> float:
-    """Return latency (seconds) for given num_tokens."""
-
-    if kv_cache_dtype == "fp8" and head_size % 16:
-        raise ValueError("fp8 kv-cache requires head_size to be a multiple of 16.")
-
-    current_platform.seed_everything(42)
-    torch.set_default_device(device)
-
-    # create random key / value tensors [T, H, D].
-    key = torch.randn(num_tokens, num_heads, head_size, dtype=dtype, device=device)
-    value = torch.randn_like(key)
-
-    # prepare the slot mapping.
-    # each token is assigned a unique slot in the KV-cache.
-    num_slots = block_size * num_blocks
-    if num_tokens > num_slots:
-        raise ValueError("num_tokens cannot exceed the total number of cache slots")
-    slot_mapping_lst = random.sample(range(num_slots), num_tokens)
-    slot_mapping = torch.tensor(slot_mapping_lst, dtype=torch.long, device=device)
-
-    key_caches, value_caches = create_kv_caches_with_random(
-        num_blocks,
-        block_size,
-        1,  # num_layers
-        num_heads,
-        head_size,
-        kv_cache_dtype,
-        dtype,
-        device=device,
-    )
-    key_cache, value_cache = key_caches[0], value_caches[0]
-    # to free unused memory
-    del key_caches, value_caches
-
-    # compute per-kernel scaling factors for fp8 conversion (if used).
-    k_scale = (key.amax() / 64.0).to(torch.float32)
-    v_scale = (value.amax() / 64.0).to(torch.float32)
-
-    function_under_test = lambda: ops.reshape_and_cache(
-        key,  # noqa: F821
-        value,  # noqa: F821
-        key_cache,  # noqa: F821
-        value_cache,  # noqa: F821
-        slot_mapping,  # noqa: F821
-        kv_cache_dtype,
-        k_scale,
-        v_scale,
-    )
-
-    if benchmark_mode == "cudagraph":
-        g = torch.cuda.CUDAGraph()
-        with torch.cuda.graph(g):
-            function_under_test()
-        torch.cuda.synchronize()
-        function_under_test = lambda: g.replay()
-
-    def run_cuda_benchmark(n_iters: int) -> float:
-        nonlocal key, value, key_cache, value_cache, slot_mapping
-        torch.cuda.synchronize()
-        start = time.perf_counter()
-        for _ in range(n_iters):
-            function_under_test()
-            torch.cuda.synchronize()
-        end = time.perf_counter()
-        return (end - start) / n_iters
-
-    # warm-up
-    run_cuda_benchmark(3)
-
-    lat = run_cuda_benchmark(num_iters)
-
-    # free tensors to mitigate OOM when sweeping
-    del key, value, key_cache, value_cache, slot_mapping
-    torch.cuda.empty_cache()
-
-    return lat
-
-
-def main(args):
-    rows = []
-    for exp in range(1, 17):
-        n_tok = 2**exp
-        lat = run_benchmark(
-            num_tokens=n_tok,
-            num_heads=args.num_heads,
-            head_size=args.head_size,
-            block_size=args.block_size,
-            num_blocks=args.num_blocks,
-            dtype=STR_DTYPE_TO_TORCH_DTYPE[args.dtype],
-            kv_cache_dtype=args.kv_cache_dtype,
-            num_iters=args.iters,
-            benchmark_mode=args.mode,
-            device="cuda",
-        )
-        rows.append([n_tok, lat * 1e6])  # convert to microseconds
-
-    print(f"Benchmark results for implementation cuda (measuring with {args.mode}):")
-    print(tabulate(rows, headers=["num_tokens", "latency (µs)"], floatfmt=".3f"))
-
-
-if __name__ == "__main__":
-    parser = FlexibleArgumentParser()
-
-    parser.add_argument("--num-heads", type=int, default=128)
-    parser.add_argument(
-        "--head-size",
-        type=int,
-        choices=[64, 80, 96, 112, 120, 128, 192, 256],
-        default=128,
-    )
-    parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
-    parser.add_argument("--num-blocks", type=int, default=128 * 128)
-
-    parser.add_argument(
-        "--dtype",
-        type=str,
-        choices=["half", "bfloat16", "float"],
-        default="bfloat16",
-    )
-
-    parser.add_argument(
-        "--kv-cache-dtype",
-        type=str,
-        choices=["auto", "fp8"],
-        default="auto",
-    )
-
-    parser.add_argument("--iters", type=int, default=200)
-
-    parser.add_argument(
-        "--mode",
-        type=str,
-        choices=["cudagraph", "no_graph"],
-        default="cudagraph",
-    )
-
-    args = parser.parse_args()
-
-    main(args)

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

@@ -9,9 +9,6 @@ import torch
 from tabulate import tabulate
 
 from vllm import _custom_ops as ops
-from vllm.attention.ops.triton_reshape_and_cache_flash import (
-    triton_reshape_and_cache_flash,
-)
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.utils import (
@@ -34,8 +31,6 @@ def run_benchmark(
     kv_cache_dtype: str,
     kv_cache_layout: str,
     num_iters: int,
-    implementation: str,
-    benchmark_mode: str,
     device: str = "cuda",
 ) -> float:
     """Return latency (seconds) for given num_tokens."""
@@ -43,14 +38,6 @@ def run_benchmark(
     if kv_cache_dtype == "fp8" and head_size % 16:
         raise ValueError("fp8 kv-cache requires head_size to be a multiple of 16.")
 
-    if implementation not in ("cuda", "triton"):
-        raise ValueError(
-            f"Unsupported implementation: {implementation}. "
-            "Only 'cuda' and 'triton' are supported."
-        )
-    if implementation == "triton" and kv_cache_layout == "HND":
-        return float("nan")  # Triton does not support HND layout yet.
-
     current_platform.seed_everything(42)
     torch.set_default_device(device)
 
@@ -78,49 +65,27 @@ def run_benchmark(
         cache_layout=kv_cache_layout,
     )
     key_cache, value_cache = key_caches[0], value_caches[0]
-    # to free unused memory
-    del key_caches, value_caches
 
     # compute per-kernel scaling factors for fp8 conversion (if used).
     k_scale = (key.amax() / 64.0).to(torch.float32)
     v_scale = (value.amax() / 64.0).to(torch.float32)
 
-    if implementation == "cuda":
-        function_under_test = lambda: ops.reshape_and_cache_flash(
-            key,  # noqa: F821
-            value,  # noqa: F821
-            key_cache,  # noqa: F821
-            value_cache,  # noqa: F821
-            slot_mapping,  # noqa: F821
-            kv_cache_dtype,
-            k_scale,
-            v_scale,
-        )
-    else:
-        function_under_test = lambda: triton_reshape_and_cache_flash(
-            key,  # noqa: F821
-            value,  # noqa: F821
-            key_cache,  # noqa: F821
-            value_cache,  # noqa: F821
-            slot_mapping,  # noqa: F821
-            kv_cache_dtype,
-            k_scale,
-            v_scale,
-        )
-    if benchmark_mode == "cudagraph":
-        g = torch.cuda.CUDAGraph()
-        with torch.cuda.graph(g):
-            function_under_test()
-        torch.cuda.synchronize()
-        function_under_test = lambda: g.replay()
-
     def run_cuda_benchmark(n_iters: int) -> float:
         nonlocal key, value, key_cache, value_cache, slot_mapping
         torch.cuda.synchronize()
         start = time.perf_counter()
         for _ in range(n_iters):
-            function_under_test()
-            torch.cuda.synchronize()
+            ops.reshape_and_cache_flash(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                slot_mapping,
+                kv_cache_dtype,
+                k_scale,
+                v_scale,
+            )
+        torch.cuda.synchronize()
         end = time.perf_counter()
         return (end - start) / n_iters
 
@@ -151,16 +116,10 @@ def main(args):
                 kv_cache_dtype=args.kv_cache_dtype,
                 kv_cache_layout=layout,
                 num_iters=args.iters,
-                implementation=args.implementation,
-                benchmark_mode=args.mode,
                 device="cuda",
             )
             rows.append([n_tok, layout, f"{lat * 1e6:.3f}"])
 
-    print(
-        f"Benchmark results for implementation {args.implementation}"
-        f" (measuring with {args.mode}):"
-    )
     print(tabulate(rows, headers=["num_tokens", "layout", "latency (µs)"]))
 
 
@@ -192,21 +151,6 @@ if __name__ == "__main__":
     )
 
     parser.add_argument("--iters", type=int, default=100)
-
-    parser.add_argument(
-        "--implementation",
-        type=str,
-        choices=["cuda", "triton"],
-        default="cuda",
-    )
-
-    parser.add_argument(
-        "--mode",
-        type=str,
-        choices=["cudagraph", "no_graph"],
-        default="cudagraph",
-    )
-
     args = parser.parse_args()
 
     main(args)

benchmarks/kernels/benchmark_silu_mul_fp8_quant.py

@@ -1,19 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-"""
-Comprehensive 3-way SiLU Benchmark Suite
-
-This benchmark compares three SiLU implementations:
-1. SiLU V2 (CUDA) - Optimized CUDA kernel implementation
-2. Triton Kernel - Triton-based implementation
-
-The suite generates detailed performance comparisons including:
-- Memory bandwidth utilization
-- Speedup ratios (baseline vs optimized implementations)
-- Performance across different expert configurations and token distributions
-"""
-
 from collections.abc import Callable
 
 import matplotlib.pyplot as plt
@@ -21,7 +7,7 @@ import numpy as np
 import torch
 
 from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
-    persistent_masked_m_silu_mul_quant,
+    silu_mul_fp8_quant_deep_gemm_cuda,
 )
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
@@ -108,7 +94,6 @@ def silu_mul_fp8_quant_deep_gemm_triton(
     num_parallel_tokens,
     group_size: int = 128,
     eps: float = 1e-10,
-    expert_offsets: torch.Tensor = None,
 ) -> tuple[torch.Tensor, torch.Tensor]:
     """Quantize silu(y[..., :H]) * y[..., H:] to FP8 with group per-token scales
 
@@ -189,7 +174,7 @@ def silu_mul_fp8_quant_deep_gemm_triton(
 
 
 # Parse generation strategies
-strategies = ["random_imbalanced", "uniform", "max_t"]
+strategies = ["uniform", "max_t", "first_t"]
 
 
 def benchmark(
@@ -210,27 +195,15 @@ def benchmark(
         current_platform.seed_everything(42 + seed_offset)
         y = torch.rand((E, T, 2 * H), dtype=torch.bfloat16, device="cuda").contiguous()
 
-        if gen_strategy == "random_imbalanced":
-
-            def generate_expert_loads(n_e, total_tokens, ratio, device="cuda"):
-                mean = total_tokens // n_e
-                min_max = mean // ratio
-                e = torch.ones(size=(E,), dtype=torch.int64, device=device) * mean
-                e[0] = min_max
-                r = torch.rand(size=(E - 1,))
-                r /= r.sum()
-                r *= total_tokens - min_max
-                r = r.round().long()
-                e[1:] = r.to(device=device)
-                return e
-
-            tokens_per_expert = generate_expert_loads(E, total_tokens, 0.7, "cuda")
-        elif gen_strategy == "uniform":
-            r = torch.rand(size=(E,))
+        if gen_strategy == "uniform":
+            r = torch.rand(size=(E,), device="cuda")
             r /= r.sum()
             r *= total_tokens
-            r = r.round().long()
-            tokens_per_expert = r
+            tokens_per_expert = r.int()
+            tokens_per_expert = torch.minimum(
+                tokens_per_expert,
+                torch.ones((E,), device=r.device, dtype=torch.int) * T,
+            )
         elif gen_strategy == "max_t":
             tokens_per_expert = torch.empty(size=(E,), dtype=torch.int32, device="cuda")
             tokens_per_expert.fill_(total_tokens / E)
@@ -308,34 +281,40 @@ def benchmark(
 
 
 def create_comparison_plot(
-    ratios, silu_v2_times, triton_times, config_labels, strategy_name, id
+    ratio, cuda_times, baseline_times, config_labels, strategy_name, id
 ):
-    fig, ax = plt.subplots(1, 1, figsize=(18, 6))
+    """Create a comparison plot for a specific generation strategy"""
+    fig, ax = plt.subplots(1, 1, figsize=(16, 6))
 
     # Configure x-axis positions
     x = np.arange(len(config_labels))
-    width = 0.25
+    width = 0.35
 
     # Execution Time plot (lower is better)
-    ax.bar(x, silu_v2_times, width, label="SiLU V2 (CUDA)", alpha=0.8, color="blue")
     ax.bar(
-        x + width, triton_times, width, label="Triton Kernel", alpha=0.8, color="green"
+        x - width / 2, cuda_times, width, label="CUDA Kernel", alpha=0.8, color="blue"
+    )
+    ax.bar(
+        x + width / 2,
+        baseline_times,
+        width,
+        label="Baseline",
+        alpha=0.8,
+        color="orange",
     )
 
-    # Add speedup labels over each bar trio
+    # Add speedup labels over each bar pair
     for i in range(len(x)):
-        triton_v2_speedup = ratios[i][1]  # triton/v2
-        max_height = max(silu_v2_times[i], triton_times[i])
-
-        # Triton/V2 speedup
+        speedup = ratio[i]
+        max_height = max(cuda_times[i], baseline_times[i])
         ax.text(
-            x[i] + width / 2,
+            x[i],
             max_height + max_height * 0.02,
-            f"{triton_v2_speedup:.2f}x",
+            f"{speedup:.2f}x",
             ha="center",
             va="bottom",
             fontweight="bold",
-            fontsize=8,
+            fontsize=9,
         )
 
     ax.set_xlabel("Configuration")
@@ -353,75 +332,56 @@ def create_comparison_plot(
 
 
 def create_combined_plot(all_results):
+    """Create a combined plot with all strategies in one PNG"""
     num_strategies = len(all_results)
-    fig, axes = plt.subplots(num_strategies, 1, figsize=(22, 7 * num_strategies))
+    fig, axes = plt.subplots(num_strategies, 1, figsize=(20, 6 * num_strategies))
 
     if num_strategies == 1:
         axes = [axes]
 
     for idx, (
         strategy_name,
-        all_ratios,
-        all_silu_v2_results,
-        all_triton_results,
+        ratio,
+        cuda_times,
+        baseline_times,
         config_labels,
-        config_x_axis,
     ) in enumerate(all_results):
         ax = axes[idx]
 
-        # Flatten the nested results to get bandwidth percentages for plotting
-        silu_v2_bandwidths = []
-        triton_bandwidths = []
-        flat_ratios = []
-
-        for config_results in all_silu_v2_results:
-            for result in config_results:
-                silu_v2_bandwidths.append(result[3])  # bandwidth percentage
-
-        for config_results in all_triton_results:
-            for result in config_results:
-                triton_bandwidths.append(result[3])  # bandwidth percentage
-
-        for config_ratios in all_ratios:
-            for ratio in config_ratios:
-                flat_ratios.append(ratio)
-
         # Configure x-axis positions
         x = np.arange(len(config_labels))
-        width = 0.25
+        width = 0.35
 
-        # Bandwidth utilization plot (higher is better)
+        # Execution Time plot (lower is better)
         ax.bar(
-            x,
-            silu_v2_bandwidths,
+            x - width / 2,
+            cuda_times,
             width,
-            label="SiLU V2 (CUDA)",
+            label="CUDA Kernel",
             alpha=0.8,
             color="blue",
         )
         ax.bar(
-            x + width,
-            triton_bandwidths,
+            x + width / 2,
+            baseline_times,
             width,
-            label="Triton Kernel",
+            label="Baseline",
             alpha=0.8,
-            color="green",
+            color="orange",
         )
 
-        # Add speedup labels over each bar trio
+        # Add speedup labels over each bar pair
         for i in range(len(x)):
-            triton_v2_speedup = flat_ratios[i]  # triton/v2
-            max_height = max(silu_v2_bandwidths[i], triton_bandwidths[i])
-
-            # Triton/V2 speedup
+            speedup = ratio[i]
+            max_height = max(cuda_times[i], baseline_times[i])
             ax.text(
-                x[i] + width / 2,
+                x[i],
                 max_height + max_height * 0.02,
-                f"{triton_v2_speedup:.2f}x",
+                f"{speedup:.2f}x",
                 ha="center",
                 va="bottom",
                 fontweight="bold",
-                fontsize=8,
+                fontsize=9,
             )
 
         ax.set_xlabel("Configuration")
@@ -435,7 +395,7 @@ def create_combined_plot(all_results):
         ax.grid(True, alpha=0.3)
 
     plt.tight_layout()
-    filename = "silu_benchmark_combined_3way.png"
+    filename = "../../silu_bench/silu_benchmark_combined.png"
     plt.savefig(filename, dpi=300, bbox_inches="tight")
     plt.show()
 
@@ -445,9 +405,7 @@ def create_combined_plot(all_results):
 outer_dim = 7168
 configs = [
     # DeepSeekV3 Configs
-    # (1, 56, 7168),
     (8, 1024, 7168),
-    # (32, 56, 7168),
     # DeepSeekV3 Configs
     (32, 1024, 7168),
     # DeepSeekV3 Configs
@@ -459,7 +417,6 @@ num_warmups = 20
 
 strategy_descriptions = {
     "uniform": "Uniform Random",
-    "random_imbalanced": "Imbalanced Random",
     "max_t": "Even Assignment",
     "first_t": "experts[0] = T, experts[1:] = 0",
 }
@@ -476,31 +433,28 @@ for id, strategy in enumerate(strategies):
     print(f"Testing strategy: {strategy_descriptions[strategy]}")
     print(f"{'=' * 60}")
 
-    # Collect benchmark data for all three algorithms
+    # Collect benchmark data for both algorithms
     config_labels = []
     config_x_axis = []
-    all_silu_v2_results = []
-    all_triton_results = []
+    all_cuda_results = []
+    all_baseline_results = []
     all_ratios = []
 
     for E, T, H in configs:
-        total_tokens_config = []
-        for i in [8, 16, 32, 64, 128, 256, 512]:
-            if i <= T:
-                total_tokens_config.append(i * E)
+        total_tokens_config = [8 * E, 16 * E, 32 * E, 64 * E, 128 * E, 256 * E]
         config_x_axis.append(total_tokens_config)
 
-        silu_v2_results = []
-        triton_results = []
+        cuda_results = []
+        baseline_results = []
         ratios = []
 
         for total_tokens in total_tokens_config:
             config_label = f"E={E},T={T},H={H},TT={total_tokens}"
             config_labels.append(config_label)
 
-            # SiLU V2 (CUDA kernel) results
-            time_ms_silu_v2, gflops, gbps, perc = benchmark(
-                persistent_masked_m_silu_mul_quant,
+            # CUDA kernel results
+            time_ms_cuda, gflops, gbps, perc = benchmark(
+                silu_mul_fp8_quant_deep_gemm_cuda,
                 E,
                 T,
                 H,
@@ -509,9 +463,9 @@ for id, strategy in enumerate(strategies):
                 num_warmups=num_warmups,
                 gen_strategy=strategy,
             )
-            silu_v2_results.append((time_ms_silu_v2, gflops, gbps, perc))
+            cuda_results.append((time_ms_cuda, gflops, gbps, perc))
 
-            # Triton kernel results
+            # Baseline results
             time_ms_triton, gflops, gbps, perc = benchmark(
                 silu_mul_fp8_quant_deep_gemm_triton,
                 E,
@@ -522,20 +476,12 @@ for id, strategy in enumerate(strategies):
                 num_warmups=num_warmups,
                 gen_strategy=strategy,
             )
-            triton_results.append((time_ms_triton, gflops, gbps, perc))
+            baseline_results.append((time_ms_triton, gflops, gbps, perc))
+            ratios.append(time_ms_triton / time_ms_cuda)
 
-            # Calculate speedup ratios (triton baseline / implementation)
-            triton_v2_ratio = time_ms_triton / time_ms_silu_v2
-            ratios.append(triton_v2_ratio)
-
-            print(
-                f"Completed: {config_label}:"
-                f" V2: {time_ms_silu_v2:.3f}ms,"
-                f" Triton: {time_ms_triton:.3f}ms"
-            )
-
-        all_silu_v2_results.append(silu_v2_results)
-        all_triton_results.append(triton_results)
+            print(f"Completed: {config_label}")
+        all_cuda_results.append(cuda_results)
+        all_baseline_results.append(baseline_results)
         all_ratios.append(ratios)
 
     # Store results for combined plotting
@@ -543,8 +489,8 @@ for id, strategy in enumerate(strategies):
         (
             strategy_descriptions[strategy],
             all_ratios,
-            all_silu_v2_results,
-            all_triton_results,
+            all_cuda_results,
+            all_baseline_results,
             config_labels,
             config_x_axis,
         )
@@ -552,18 +498,15 @@ for id, strategy in enumerate(strategies):
 
     # Print summary table for this strategy
     print(f"\nSummary Table - {strategy_descriptions[strategy]}:")
-    print(f" {'V2 Time(ms)':<12} {'Triton Time(ms)':<14} {'Triton/V2':<10}")
-    print("-" * 90)
+    print(f"{'Config':<20} {'CUDA Time(ms)':<12} {'Base Time(ms)':<12} {'Speedup':<8}")
+    print("-" * 60)
 
     for i, (E, T, H) in enumerate(configs):
-        # Get the first result for each config (simplifying for summary)
-        v2_time = silu_v2_results[i][0]
-        triton_time = triton_results[i][0]
-        triton_v2_speedup = triton_time / v2_time
+        speedup = baseline_results[i][0] / cuda_results[i][0]
         config_label = f"E={E:3d},T={T:4d},H={H:4d}"
         print(
-            f"{config_label:<20} {v2_time:8.5f} {triton_time:10.5f} "
-            f"{triton_v2_speedup:8.2f}x"
+            f"{config_label:<20} {cuda_results[i][0]:8.5f} "
+            f"{baseline_results[i][0]:8.5f} {speedup:6.2f}x"
         )
 
 
@@ -571,14 +514,15 @@ def create_total_tokens_plot(all_results):
     num_strategies = len(all_results)
     num_configs = len(configs)
 
+    # Create side-by-side subplots: 2 columns for speedup and bandwidth percentage
     fig, axs = plt.subplots(
-        num_strategies, num_configs * 2, figsize=(32, 8 * num_strategies)
+        num_strategies, num_configs * 2, figsize=(28, 6 * num_strategies)
     )
 
     # Add main title to the entire figure
     fig.suptitle(
-        "Performance Analysis: Speedup vs Bandwidth Utilization (SiLU V2, and Triton)",
-        fontsize=18,
+        "Performance Analysis: Speedup vs Bandwidth Utilization (Triton & CUDA)",
+        fontsize=16,
         fontweight="bold",
         y=0.98,
     )
@@ -595,8 +539,8 @@ def create_total_tokens_plot(all_results):
         (
             strategy_name,
             all_ratios,
-            all_silu_v2_results,
-            all_triton_results,
+            all_cuda_results,
+            all_baseline_results,
             config_labels,
             config_x_axis,
         ) = result
@@ -611,54 +555,42 @@ def create_total_tokens_plot(all_results):
             ratios = all_ratios[config_idx]
             total_tokens_values = config_x_axis[config_idx]
 
-            # Extract speedup ratios
-            triton_v2_ratios = [ratio for ratio in ratios]
-
-            # Extract bandwidth percentages for all implementations
-            v2_bandwidth_percentages = [
-                result[3] for result in all_silu_v2_results[config_idx]
+            # Extract CUDA and Triton bandwidth percentages
+            cuda_bandwidth_percentages = [
+                result[3] for result in all_cuda_results[config_idx]
             ]
             triton_bandwidth_percentages = [
-                result[3] for result in all_triton_results[config_idx]
+                result[3] for result in all_baseline_results[config_idx]
             ]
 
             # Plot speedup ratios vs total tokens (left plot)
             ax_speedup.plot(
-                total_tokens_values,
-                triton_v2_ratios,
-                "go-",
-                linewidth=3,
-                markersize=8,
-                label="Triton/V2 Speedup",
+                total_tokens_values, ratios, "bo-", linewidth=3, markersize=8
             )
             ax_speedup.set_title(
-                f"{strategy_name}\nSpeedup vs Baseline (Triton)\nE={E}, T={T}, H={H}",
+                f"{strategy_name}\nSpeedup (CUDA/Triton)\nE={E}, T={T}, H={H}",
                 fontsize=12,
                 fontweight="bold",
             )
             ax_speedup.set_xlabel("Total Tokens", fontweight="bold", fontsize=11)
             ax_speedup.set_ylabel("Speedup Ratio", fontweight="bold", fontsize=11)
-            ax_speedup.legend(prop={"weight": "bold"})
             ax_speedup.grid(True, alpha=0.3)
 
-            # Plot bandwidth utilization (right plot)
             ax_bandwidth.plot(
                 total_tokens_values,
-                v2_bandwidth_percentages,
-                "o-",
+                cuda_bandwidth_percentages,
+                "ro-",
                 linewidth=3,
                 markersize=8,
-                label="SiLU V2",
-                color="blue",
+                label="CUDA",
             )
             ax_bandwidth.plot(
                 total_tokens_values,
                 triton_bandwidth_percentages,
-                "o-",
+                "go-",
                 linewidth=3,
                 markersize=8,
                 label="Triton",
-                color="green",
             )
             ax_bandwidth.set_title(
                 f"{strategy_name}\nBandwidth Utilization (Hopper)\nE={E}, T={T}, H={H}",
@@ -686,12 +618,38 @@ def create_total_tokens_plot(all_results):
                 for label in ax.get_xticklabels() + ax.get_yticklabels():
                     label.set_fontweight("bold")
 
-            # Add value labels on Triton/V2 speedup points
-            for x, y in zip(total_tokens_values, triton_v2_ratios):
+            # Add value labels on speedup points
+            for x, y in zip(total_tokens_values, ratios):
                 ax_speedup.annotate(
                     f"{y:.2f}x",
                     (x, y),
                     textcoords="offset points",
+                    xytext=(0, 12),
+                    ha="center",
+                    fontsize=10,
+                    fontweight="bold",
+                    bbox=dict(boxstyle="round,pad=0.3", facecolor="white", alpha=0.7),
+                )
+
+            # Add value labels on CUDA bandwidth points
+            for x, y in zip(total_tokens_values, cuda_bandwidth_percentages):
+                ax_bandwidth.annotate(
+                    f"{y:.1f}%",
+                    (x, y),
+                    textcoords="offset points",
+                    xytext=(0, 12),
+                    ha="center",
+                    fontsize=9,
+                    fontweight="bold",
+                    bbox=dict(boxstyle="round,pad=0.2", facecolor="red", alpha=0.3),
+                )
+
+            # Add value labels on Triton bandwidth points
+            for x, y in zip(total_tokens_values, triton_bandwidth_percentages):
+                ax_bandwidth.annotate(
+                    f"{y:.1f}%",
+                    (x, y),
+                    textcoords="offset points",
                     xytext=(0, -15),
                     ha="center",
                     fontsize=9,
@@ -701,20 +659,17 @@ def create_total_tokens_plot(all_results):
 
     plt.tight_layout()
     plt.subplots_adjust(top=0.93)  # Make room for main title
-    filename = "silu_benchmark_total_tokens_3way.png"
+    filename = "silu_benchmark_total_tokens.png"
     plt.savefig(filename, dpi=300, bbox_inches="tight")
     plt.show()
 
     return filename
 
 
-# Create comprehensive 3-way comparison plots
-combined_plot_filename = create_combined_plot(all_results)
-total_tokens_plot_filename = create_total_tokens_plot(all_results)
+# Create combined plot with all strategies
+combined_plot_filename = create_total_tokens_plot(all_results)
 
-print(f"\n{'=' * 80}")
-print("3-Way Benchmark Suite Complete!")
-print(f"Generated combined comparison plot: {combined_plot_filename}")
-print(f"Generated total tokens analysis plot: {total_tokens_plot_filename}")
-print("Compared: SiLU V2 (CUDA), and Triton implementations")
-print(f"{'=' * 80}")
+print(f"\n{'=' * 60}")
+print("Benchmark Complete!")
+print(f"Generated combined plot: {combined_plot_filename}")
+print(f"{'=' * 60}")

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -11,13 +11,13 @@ from datetime import datetime
 from typing import Any
 
 import torch
+import triton
 from tqdm import tqdm
 
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    _w8a8_triton_block_scaled_mm,
+    _w8a8_block_fp8_matmul,
 )
 from vllm.platforms import current_platform
-from vllm.triton_utils import triton
 from vllm.utils import FlexibleArgumentParser
 
 mp.set_start_method("spawn", force=True)
@@ -83,7 +83,7 @@ def w8a8_block_matmul(
         )
 
     if A.dtype == torch.float8_e4m3fn:
-        kernel = _w8a8_triton_block_scaled_mm
+        kernel = _w8a8_block_fp8_matmul
     else:
         raise RuntimeError("Currently, only support tune w8a8 block fp8 kernel.")
 

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -1,5 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# fmt: off
 # ruff: noqa: E501
 import time
 
@@ -7,33 +8,27 @@ import torch
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    get_col_major_tma_aligned_tensor,
     per_token_group_quant_fp8,
-    w8a8_triton_block_scaled_mm,
+    w8a8_block_fp8_matmul,
 )
 from vllm.triton_utils import triton
-from vllm.utils.deep_gemm import (
-    calc_diff,
-    fp8_gemm_nt,
-    get_col_major_tma_aligned_tensor,
-    per_block_cast_to_fp8,
-)
+from vllm.utils.deep_gemm import calc_diff, fp8_gemm_nt, per_block_cast_to_fp8
 
 
-def benchmark_shape(
-    m: int,
-    n: int,
-    k: int,
-    warmup: int = 100,
-    repeat: int = 10000,
-    verbose: bool = False,
-) -> dict:
+def benchmark_shape(m: int,
+                    n: int,
+                    k: int,
+                    warmup: int = 100,
+                    repeat: int = 10000,
+                    verbose: bool = False) -> dict:
     """Benchmark all implementations for a specific (m, n, k) shape."""
     if verbose:
         print(f"\n=== Benchmarking shape: m={m}, n={n}, k={k} ===")
 
     # Create test tensors
-    A = torch.randn((m, k), device="cuda", dtype=torch.bfloat16)
-    B = torch.randn((n, k), device="cuda", dtype=torch.bfloat16)
+    A = torch.randn((m, k), device='cuda', dtype=torch.bfloat16)
+    B = torch.randn((n, k), device='cuda', dtype=torch.bfloat16)
 
     # Reference result in BF16
     torch.cuda.synchronize()
@@ -50,39 +45,34 @@ def benchmark_shape(
     # Pre-quantize A for all implementations
     A_deepgemm, A_scale_deepgemm = per_token_group_quant_fp8(A, block_size[1])
     A_scale_deepgemm = get_col_major_tma_aligned_tensor(A_scale_deepgemm)
-    C_deepgemm = torch.empty((m, n), device="cuda", dtype=torch.bfloat16)
+    C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16)
     A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1])
     A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8(
-        A, block_size[1], column_major_scales=True
-    )
+        A, block_size[1], column_major_scales=True)
 
     # === DeepGEMM Implementation ===
     def deepgemm_gemm():
-        fp8_gemm_nt(
-            (A_deepgemm, A_scale_deepgemm), (B_deepgemm, B_scale_deepgemm), C_deepgemm
-        )
+        fp8_gemm_nt((A_deepgemm, A_scale_deepgemm),
+                                       (B_deepgemm, B_scale_deepgemm),
+                                       C_deepgemm)
         return C_deepgemm
 
     # === vLLM Triton Implementation ===
     def vllm_triton_gemm():
-        return w8a8_triton_block_scaled_mm(
-            A_vllm,
-            B_vllm,
-            A_scale_vllm,
-            B_scale_vllm,
-            block_size,
-            output_dtype=torch.bfloat16,
-        )
+        return w8a8_block_fp8_matmul(A_vllm,
+                                     B_vllm,
+                                     A_scale_vllm,
+                                     B_scale_vllm,
+                                     block_size,
+                                     output_dtype=torch.bfloat16)
 
     # === vLLM CUTLASS Implementation ===
     def vllm_cutlass_gemm():
-        return ops.cutlass_scaled_mm(
-            A_vllm_cutlass,
-            B_vllm.T,
-            scale_a=A_scale_vllm_cutlass,
-            scale_b=B_scale_vllm.T,
-            out_dtype=torch.bfloat16,
-        )
+        return ops.cutlass_scaled_mm(A_vllm_cutlass,
+                                     B_vllm.T,
+                                     scale_a=A_scale_vllm_cutlass,
+                                     scale_b=B_scale_vllm.T,
+                                     out_dtype=torch.bfloat16)
 
     # Run correctness check first
     if verbose:
@@ -99,23 +89,26 @@ def benchmark_shape(
         print(f"DeepGEMM vs Reference difference: {deepgemm_diff:.6f}")
         print(f"vLLM Triton vs Reference difference: {vllm_triton_diff:.6f}")
         print(f"vLLM CUTLASS vs Reference difference: {vllm_cutlass_diff:.6f}")
-        print(
-            "vLLM Triton vs DeepGEMM difference: "
-            f"{calc_diff(C_vllm_triton, C_deepgemm):.6f}"
-        )
-        print(
-            "vLLM CUTLASS vs DeepGEMM difference: "
-            f"{calc_diff(C_vllm_cutlass, C_deepgemm):.6f}"
-        )
+        print("vLLM Triton vs DeepGEMM difference: "
+              f"{calc_diff(C_vllm_triton, C_deepgemm):.6f}")
+        print("vLLM CUTLASS vs DeepGEMM difference: "
+              f"{calc_diff(C_vllm_cutlass, C_deepgemm):.6f}")
 
     # Benchmark implementations
     implementations = {
         "DeepGEMM": deepgemm_gemm,
         "vLLM Triton": vllm_triton_gemm,
-        "vLLM CUTLASS": vllm_cutlass_gemm,
+        "vLLM CUTLASS": vllm_cutlass_gemm
     }
 
-    benchmark_results = {"shape": {"m": m, "n": n, "k": k}, "implementations": {}}
+    benchmark_results = {
+        "shape": {
+            "m": m,
+            "n": n,
+            "k": k
+        },
+        "implementations": {}
+    }
 
     for name, func in implementations.items():
         # Warmup
@@ -143,36 +136,38 @@ def benchmark_shape(
             "tflops": tflops,
             "gb_s": gb_s,
             "diff": {
-                "DeepGEMM": 0.0
-                if name == "DeepGEMM"
-                else calc_diff(func(), C_deepgemm),
-                "Reference": deepgemm_diff
-                if name == "DeepGEMM"
-                else (vllm_triton_diff if name == "vLLM Triton" else vllm_cutlass_diff),
-            },
+                "DeepGEMM":
+                0.0 if name == "DeepGEMM" else calc_diff(func(), C_deepgemm),
+                "Reference":
+                deepgemm_diff if name == "DeepGEMM" else
+                (vllm_triton_diff
+                 if name == "vLLM Triton" else vllm_cutlass_diff)
+            }
         }
 
         if verbose:
-            print(f"{name}: {avg_time_ms:.3f} ms, {tflops:.2f} TFLOPS, {gb_s:.2f} GB/s")
+            print(
+                f"{name}: {avg_time_ms:.3f} ms, {tflops:.2f} TFLOPS, {gb_s:.2f} GB/s"
+            )
 
     # Calculate speedups
     baseline = benchmark_results["implementations"]["DeepGEMM"]["time_ms"]
     for name, data in benchmark_results["implementations"].items():
         if name != "DeepGEMM":
             speedup = baseline / data["time_ms"]
-            benchmark_results["implementations"][name]["speedup_vs_deepgemm"] = speedup
+            benchmark_results["implementations"][name][
+                "speedup_vs_deepgemm"] = speedup
             if verbose:
-                print(
-                    f"DeepGEMM is {1 / speedup:.2f}x "
-                    f"{'faster' if 1 / speedup > 1 else 'slower'} than {name}"
-                )
+                print(f"DeepGEMM is {1/speedup:.2f}x "
+                      f"{'faster' if 1/speedup > 1 else 'slower'} than {name}")
 
-    vllm_triton_time = benchmark_results["implementations"]["vLLM Triton"]["time_ms"]
-    vllm_cutlass_time = benchmark_results["implementations"]["vLLM CUTLASS"]["time_ms"]
+    vllm_triton_time = benchmark_results["implementations"]["vLLM Triton"][
+        "time_ms"]
+    vllm_cutlass_time = benchmark_results["implementations"]["vLLM CUTLASS"][
+        "time_ms"]
     cutlass_vs_triton = vllm_triton_time / vllm_cutlass_time
-    benchmark_results["implementations"]["vLLM CUTLASS"]["speedup_vs_triton"] = (
-        cutlass_vs_triton
-    )
+    benchmark_results["implementations"]["vLLM CUTLASS"][
+        "speedup_vs_triton"] = cutlass_vs_triton
     if verbose:
         print(
             f"vLLM CUTLASS is {cutlass_vs_triton:.2f}x "
@@ -184,7 +179,8 @@ def benchmark_shape(
 
 def format_table_row(values, widths):
     """Format a row with specified column widths."""
-    return "| " + " | ".join(f"{val:{w}}" for val, w in zip(values, widths)) + " |"
+    return "| " + " | ".join(f"{val:{w}}"
+                             for val, w in zip(values, widths)) + " |"
 
 
 def print_table(headers, rows, title=None):
@@ -292,50 +288,38 @@ def run_benchmarks(verbose: bool = False):
     for result in all_results:
         shape = result["shape"]
         impl_data = result["implementations"]["DeepGEMM"]
-        deepgemm_rows.append(
-            [
-                shape["m"],
-                shape["n"],
-                shape["k"],
-                f"{impl_data['time_us']:.1f}",
-                f"{impl_data['tflops']:.1f}",
-                f"{impl_data['gb_s']:.1f}",
-            ]
-        )
+        deepgemm_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}"
+        ])
 
-    print_table(deepgemm_headers, deepgemm_rows, title="DeepGEMM Implementation:")
+    print_table(deepgemm_headers,
+                deepgemm_rows,
+                title="DeepGEMM Implementation:")
 
     # Print vLLM Triton table
-    triton_headers = ["m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM"]
+    triton_headers = [
+        "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM"
+    ]
     triton_rows = []
     for result in all_results:
         shape = result["shape"]
         impl_data = result["implementations"]["vLLM Triton"]
         speedup = impl_data.get("speedup_vs_deepgemm", 1.0)
-        triton_rows.append(
-            [
-                shape["m"],
-                shape["n"],
-                shape["k"],
-                f"{impl_data['time_us']:.1f}",
-                f"{impl_data['tflops']:.1f}",
-                f"{impl_data['gb_s']:.1f}",
-                format_speedup(speedup),
-            ]
-        )
+        triton_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}",
+            format_speedup(speedup)
+        ])
 
-    print_table(triton_headers, triton_rows, title="vLLM Triton Implementation:")
+    print_table(triton_headers,
+                triton_rows,
+                title="vLLM Triton Implementation:")
 
     # Print vLLM CUTLASS table
     cutlass_headers = [
-        "m",
-        "n",
-        "k",
-        "Time (μs)",
-        "TFLOPS",
-        "GB/s",
-        "vs DeepGEMM",
-        "vs Triton",
+        "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM",
+        "vs Triton"
     ]
     cutlass_rows = []
     for result in all_results:
@@ -343,27 +327,28 @@ def run_benchmarks(verbose: bool = False):
         impl_data = result["implementations"]["vLLM CUTLASS"]
         vs_deepgemm = impl_data.get("speedup_vs_deepgemm", 1.0)
         vs_triton = impl_data.get("speedup_vs_triton", 1.0)
-        cutlass_rows.append(
-            [
-                shape["m"],
-                shape["n"],
-                shape["k"],
-                f"{impl_data['time_us']:.1f}",
-                f"{impl_data['tflops']:.1f}",
-                f"{impl_data['gb_s']:.1f}",
-                format_speedup(vs_deepgemm),
-                format_speedup(vs_triton),
-            ]
-        )
+        cutlass_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}",
+            format_speedup(vs_deepgemm),
+            format_speedup(vs_triton)
+        ])
 
-    print_table(cutlass_headers, cutlass_rows, title="vLLM CUTLASS Implementation:")
+    print_table(cutlass_headers,
+                cutlass_rows,
+                title="vLLM CUTLASS Implementation:")
 
     # Calculate and print averages
     print("\n===== AVERAGE PERFORMANCE =====")
 
     implementations = ["DeepGEMM", "vLLM Triton", "vLLM CUTLASS"]
     avg_metrics = {
-        impl: {"tflops": 0, "gb_s": 0, "time_ms": 0} for impl in implementations
+        impl: {
+            "tflops": 0,
+            "gb_s": 0,
+            "time_ms": 0
+        }
+        for impl in implementations
     }
 
     for result in all_results:
@@ -381,9 +366,9 @@ def run_benchmarks(verbose: bool = False):
         avg_tflops = avg_metrics[impl]["tflops"] / num_shapes
         avg_mem_bw = avg_metrics[impl]["gb_s"] / num_shapes
         avg_time = avg_metrics[impl]["time_ms"] / num_shapes
-        avg_rows.append(
-            [impl, f"{avg_tflops:.2f}", f"{avg_mem_bw:.2f}", f"{avg_time:.2f}"]
-        )
+        avg_rows.append([
+            impl, f"{avg_tflops:.2f}", f"{avg_mem_bw:.2f}", f"{avg_time:.2f}"
+        ])
 
     print_table(avg_headers, avg_rows)
 
@@ -391,19 +376,21 @@ def run_benchmarks(verbose: bool = False):
     avg_speedups = {
         "DeepGEMM vs vLLM Triton": 0,
         "DeepGEMM vs vLLM CUTLASS": 0,
-        "vLLM CUTLASS vs vLLM Triton": 0,
+        "vLLM CUTLASS vs vLLM Triton": 0
     }
 
     for result in all_results:
         deepgemm_time = result["implementations"]["DeepGEMM"]["time_ms"]
         vllm_triton_time = result["implementations"]["vLLM Triton"]["time_ms"]
-        vllm_cutlass_time = result["implementations"]["vLLM CUTLASS"]["time_ms"]
+        vllm_cutlass_time = result["implementations"]["vLLM CUTLASS"][
+            "time_ms"]
 
-        avg_speedups["DeepGEMM vs vLLM Triton"] += vllm_triton_time / deepgemm_time
-        avg_speedups["DeepGEMM vs vLLM CUTLASS"] += vllm_cutlass_time / deepgemm_time
-        avg_speedups["vLLM CUTLASS vs vLLM Triton"] += (
-            vllm_triton_time / vllm_cutlass_time
-        )
+        avg_speedups[
+            "DeepGEMM vs vLLM Triton"] += vllm_triton_time / deepgemm_time
+        avg_speedups[
+            "DeepGEMM vs vLLM CUTLASS"] += vllm_cutlass_time / deepgemm_time
+        avg_speedups[
+            "vLLM CUTLASS vs vLLM Triton"] += vllm_triton_time / vllm_cutlass_time
 
     print("\n===== AVERAGE SPEEDUPS =====")
     speedup_headers = ["Comparison", "Speedup"]
@@ -421,7 +408,8 @@ def run_benchmarks(verbose: bool = False):
 
     for result in all_results:
         for impl in implementations:
-            avg_diff[impl] += result["implementations"][impl]["diff"]["Reference"]
+            avg_diff[impl] += result["implementations"][impl]["diff"][
+                "Reference"]
 
     diff_headers = ["Implementation", "Avg Diff vs Reference"]
     diff_rows = []

benchmarks/multi_turn/benchmark_serving_multi_turn.py

@@ -13,7 +13,7 @@ from datetime import datetime
 from enum import Enum
 from http import HTTPStatus
 from statistics import mean
-from typing import NamedTuple, Union
+from typing import NamedTuple, Optional, Union
 
 import aiohttp  # type: ignore
 import numpy as np  # type: ignore
@@ -46,9 +46,9 @@ class ConversationSampling(str, Enum):
 
 class ClientArgs(NamedTuple):
     seed: int
-    max_num_requests: int | None
+    max_num_requests: Optional[int]
     skip_first_turn: bool
-    max_turns: int | None
+    max_turns: Optional[int]
     max_active_conversations: int
     verbose: bool
     print_content: bool
@@ -109,9 +109,9 @@ class RequestStats(NamedTuple):
 
 class MetricStats:
     def __init__(self) -> None:
-        self.min: float | None = None
-        self.max: float | None = None
-        self.avg: float | None = None
+        self.min: Optional[float] = None
+        self.max: Optional[float] = None
+        self.avg: Optional[float] = None
         self.sum = 0.0
         self.count = 0
 
@@ -143,7 +143,7 @@ class MovingAverage:
         self.index = 0
         self.sum = 0.0
         self.count = 0
-        self.avg: float | None = None
+        self.avg: Optional[float] = None
 
     def update(self, new_value: float) -> None:
         if self.count < self.window_size:
@@ -198,6 +198,14 @@ class DebugStats:
         self.logger.info("-" * 50)
 
 
+# Must support Python 3.8, we can't use str.removeprefix(prefix)
+# introduced in Python 3.9
+def remove_prefix(text: str, prefix: str) -> str:
+    if text.startswith(prefix):
+        return text[len(prefix) :]
+    return text
+
+
 def nanosec_to_millisec(value: float) -> float:
     return value / 1000000.0
 
@@ -212,8 +220,8 @@ async def send_request(
     chat_url: str,
     model: str,
     stream: bool = True,
-    min_tokens: int | None = None,
-    max_tokens: int | None = None,
+    min_tokens: Optional[int] = None,
+    max_tokens: Optional[int] = None,
 ) -> ServerResponse:
     payload = {
         "model": model,
@@ -242,9 +250,9 @@ async def send_request(
     timeout = aiohttp.ClientTimeout(total=timeout_sec)
 
     valid_response = True
-    ttft: float | None = None
+    ttft: Optional[float] = None
     chunk_delay: list[int] = []
-    latency: float | None = None
+    latency: Optional[float] = None
     first_chunk = ""
     generated_text = ""
 
@@ -261,7 +269,7 @@ async def send_request(
                 if not chunk_bytes:
                     continue
 
-                chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
+                chunk = remove_prefix(chunk_bytes.decode("utf-8"), "data: ")
                 if chunk == "[DONE]":
                     # End of stream
                     latency = time.perf_counter_ns() - start_time
@@ -356,7 +364,7 @@ async def send_turn(
     req_args: RequestArgs,
     verbose: bool,
     verify_output: bool,
-) -> RequestStats | None:
+) -> Optional[RequestStats]:
     assert messages_to_use > 0
     assert messages_to_use <= len(conversation_messages)
 
@@ -761,7 +769,7 @@ def get_client_config(
             "Number of conversations must be equal or larger than the number of clients"
         )
 
-    max_req_per_client: int | None = None
+    max_req_per_client: Optional[int] = None
     if args.max_num_requests is not None:
         # Max number of requests per client
         req_per_client = args.max_num_requests // args.num_clients
@@ -1024,7 +1032,7 @@ def process_statistics(
     warmup_percentages: list[float],
     test_params: dict,
     verbose: bool,
-    gen_conv_args: GenConvArgs | None = None,
+    gen_conv_args: Optional[GenConvArgs] = None,
     excel_output: bool = False,
 ) -> None:
     if len(client_metrics) == 0:

benchmarks/pyproject.toml

@@ -0,0 +1,49 @@
+# This local pyproject file is part of the migration from yapf to ruff format.
+# It uses the same core rules as the main pyproject.toml file, but with the
+# following differences:
+# - ruff line length is overridden to 88
+# - deprecated typing ignores (UP006, UP035) have been removed
+
+[tool.ruff]
+line-length = 88
+
+[tool.ruff.lint.per-file-ignores]
+"vllm/third_party/**" = ["ALL"]
+"vllm/version.py" = ["F401"]
+"vllm/_version.py" = ["ALL"]
+
+[tool.ruff.lint]
+select = [
+    # pycodestyle
+    "E",
+    # Pyflakes
+    "F",
+    # pyupgrade
+    "UP",
+    # flake8-bugbear
+    "B",
+    # flake8-simplify
+    "SIM",
+    # isort
+    "I",
+    # flake8-logging-format
+    "G",
+]
+ignore = [
+    # star imports
+    "F405", "F403",
+    # lambda expression assignment
+    "E731",
+    # Loop control variable not used within loop body
+    "B007",
+    # f-string format
+    "UP032",
+    # Can remove once 3.10+ is the minimum Python version
+    "UP007",
+]
+
+[tool.ruff.lint.isort]
+known-first-party = ["vllm"]
+
+[tool.ruff.format]
+docstring-code-format = true

cmake/cpu_extension.cmake

@@ -101,7 +101,6 @@ else()
     find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
     find_isa(${CPUINFO} "bf16" ARM_BF16_FOUND) # Check for ARM BF16 support
     find_isa(${CPUINFO} "S390" S390_FOUND)
-    find_isa(${CPUINFO} "v" RVV_FOUND) # Check for RISC-V RVV support
 endif()
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
@@ -178,14 +177,8 @@ elseif (S390_FOUND)
         "-mzvector"
         "-march=native"
         "-mtune=native")
-elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
-    if(RVV_FOUND)
-	    message(FAIL_ERROR "Can't support rvv now.")
-    else()
-        list(APPEND CXX_COMPILE_FLAGS "-march=rv64gc")
-    endif()
 else()
-    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA, S390X ISA, ARMv8 or RISC-V support.")
+    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA, S390X ISA or ARMv8 support.")
 endif()
 
 #
@@ -198,24 +191,13 @@ else()
 endif()
 
 if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
-    set(FETCHCONTENT_SOURCE_DIR_ONEDNN "$ENV{FETCHCONTENT_SOURCE_DIR_ONEDNN}" CACHE PATH "Path to a local oneDNN source directory.")
-
-    if(FETCHCONTENT_SOURCE_DIR_ONEDNN)
-        message(STATUS "Using oneDNN from specified source directory: ${FETCHCONTENT_SOURCE_DIR_ONEDNN}")
-        FetchContent_Declare(
-            oneDNN
-            SOURCE_DIR ${FETCHCONTENT_SOURCE_DIR_ONEDNN}
-        )
-    else()
-        message(STATUS "Downloading oneDNN from GitHub")
-        FetchContent_Declare(
-            oneDNN
-            GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-            GIT_TAG v3.9
-            GIT_PROGRESS TRUE
-            GIT_SHALLOW TRUE
-        )
-    endif()
+    FetchContent_Declare(
+        oneDNN
+        GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
+        GIT_TAG v3.9
+        GIT_PROGRESS TRUE
+        GIT_SHALLOW TRUE
+    )
 
     if(USE_ACL)
         find_library(ARM_COMPUTE_LIBRARY NAMES arm_compute PATHS $ENV{ACL_ROOT_DIR}/build/)
@@ -224,7 +206,6 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
         endif()
         set(ONEDNN_AARCH64_USE_ACL "ON")
         set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wl,-rpath,$ENV{ACL_ROOT_DIR}/build/")
-        add_compile_definitions(VLLM_USE_ACL)
     endif()
 
     set(ONEDNN_LIBRARY_TYPE "STATIC")
@@ -277,8 +258,7 @@ set(VLLM_EXT_SRC
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
-    "csrc/cpu/torch_bindings.cpp"
-    "csrc/moe/dynamic_4bit_int_moe_cpu.cpp")
+    "csrc/cpu/torch_bindings.cpp")
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     set(VLLM_EXT_SRC
@@ -320,4 +300,4 @@ define_gpu_extension_target(
     WITH_SOABI
 )
 
-message(STATUS "Enabling C extension.")
+message(STATUS "Enabling C extension.")

cmake/external_projects/flashmla.cmake

@@ -18,8 +18,8 @@ if(FLASH_MLA_SRC_DIR)
 else()
   FetchContent_Declare(
         flashmla
-        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
-        GIT_TAG 5f65b85703c7ed75fda01e06495077caad207c3f
+        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA.git
+        GIT_TAG a757314c04eedd166e329e846c820eb1bdd702de
         GIT_PROGRESS TRUE
         CONFIGURE_COMMAND ""
         BUILD_COMMAND ""
@@ -33,64 +33,23 @@ message(STATUS "FlashMLA is available at ${flashmla_SOURCE_DIR}")
 # The FlashMLA kernels only work on hopper and require CUDA 12.3 or later.
 # Only build FlashMLA kernels if we are building for something compatible with 
 # sm90a
-
-set(SUPPORT_ARCHS)
-if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3)
-    list(APPEND SUPPORT_ARCHS 9.0a)
-endif()
-if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8)
-    list(APPEND SUPPORT_ARCHS 10.0a)
-endif()
-
-
-cuda_archs_loose_intersection(FLASH_MLA_ARCHS "${SUPPORT_ARCHS}" "${CUDA_ARCHS}")
-if(FLASH_MLA_ARCHS)
-    set(VLLM_FLASHMLA_GPU_FLAGS ${VLLM_GPU_FLAGS})
-    list(APPEND VLLM_FLASHMLA_GPU_FLAGS "--expt-relaxed-constexpr" "--expt-extended-lambda" "--use_fast_math")
-
+cuda_archs_loose_intersection(FLASH_MLA_ARCHS "9.0a" "${CUDA_ARCHS}")
+if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3 AND FLASH_MLA_ARCHS)
     set(FlashMLA_SOURCES
-        ${flashmla_SOURCE_DIR}/csrc/torch_api.cpp
-        ${flashmla_SOURCE_DIR}/csrc/pybind.cpp
-        ${flashmla_SOURCE_DIR}/csrc/smxx/get_mla_metadata.cu
-        ${flashmla_SOURCE_DIR}/csrc/smxx/mla_combine.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm90/decode/dense/splitkv_mla.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm90/decode/sparse_fp8/splitkv_mla.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm90/prefill/sparse/fwd.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm100/decode/sparse_fp8/splitkv_mla.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm100/prefill/dense/fmha_cutlass_fwd_sm100.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm100/prefill/dense/fmha_cutlass_bwd_sm100.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm100/prefill/sparse/fwd.cu
-    )
-
-    set(FlashMLA_Extension_SOURCES
-        ${flashmla_SOURCE_DIR}/csrc/extension/torch_api.cpp
-        ${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/pybind.cpp
-        ${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/flash_fwd_mla_fp8_sm90.cu
-    )
+        ${flashmla_SOURCE_DIR}/csrc/flash_api.cpp
+        ${flashmla_SOURCE_DIR}/csrc/kernels/get_mla_metadata.cu
+        ${flashmla_SOURCE_DIR}/csrc/kernels/mla_combine.cu
+        ${flashmla_SOURCE_DIR}/csrc/kernels/splitkv_mla.cu
+        ${flashmla_SOURCE_DIR}/csrc/kernels_fp8/flash_fwd_mla_fp8_sm90.cu)
 
     set(FlashMLA_INCLUDES
-        ${flashmla_SOURCE_DIR}/csrc
-        ${flashmla_SOURCE_DIR}/csrc/sm90
         ${flashmla_SOURCE_DIR}/csrc/cutlass/include
-        ${flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
-    )
-
-    set(FlashMLA_Extension_INCLUDES
-        ${flashmla_SOURCE_DIR}/csrc
-        ${flashmla_SOURCE_DIR}/csrc/sm90
-        ${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/
-        ${flashmla_SOURCE_DIR}/csrc/cutlass/include
-        ${flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
-    )
+        ${flashmla_SOURCE_DIR}/csrc)
 
     set_gencode_flags_for_srcs(
         SRCS "${FlashMLA_SOURCES}"
         CUDA_ARCHS "${FLASH_MLA_ARCHS}")
 
-    set_gencode_flags_for_srcs(
-        SRCS "${FlashMLA_Extension_SOURCES}"
-        CUDA_ARCHS "${FLASH_MLA_ARCHS}")
-
     define_gpu_extension_target(
         _flashmla_C
         DESTINATION vllm
@@ -101,32 +60,8 @@ if(FLASH_MLA_ARCHS)
         INCLUDE_DIRECTORIES ${FlashMLA_INCLUDES}
         USE_SABI 3
         WITH_SOABI)
-
-    # Keep Stable ABI for the module, but *not* for CUDA/C++ files.
-    # This prevents Py_LIMITED_API from affecting nvcc and C++ compiles.
-    target_compile_options(_flashmla_C PRIVATE
-        $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
-        $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
-
-    define_gpu_extension_target(
-        _flashmla_extension_C
-        DESTINATION vllm
-        LANGUAGE ${VLLM_GPU_LANG}
-        SOURCES ${FlashMLA_Extension_SOURCES}
-        COMPILE_FLAGS ${VLLM_FLASHMLA_GPU_FLAGS}
-        ARCHITECTURES ${VLLM_GPU_ARCHES}
-        INCLUDE_DIRECTORIES ${FlashMLA_Extension_INCLUDES}
-        USE_SABI 3
-        WITH_SOABI)
-
-    # Keep Stable ABI for the module, but *not* for CUDA/C++ files.
-    # This prevents Py_LIMITED_API from affecting nvcc and C++ compiles.
-    target_compile_options(_flashmla_extension_C PRIVATE
-        $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
-        $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
 else()
-    # Create empty targets for setup.py when not targeting sm90a systems
+    # Create an empty target for setup.py when not targeting sm90a systems
     add_custom_target(_flashmla_C)
-    add_custom_target(_flashmla_extension_C)
 endif()
 

cmake/external_projects/qutlass.cmake

@@ -1,97 +0,0 @@
-include(FetchContent)
-
-set(CUTLASS_INCLUDE_DIR "${CUTLASS_INCLUDE_DIR}" CACHE PATH "Path to CUTLASS include/ directory")
-
-if(DEFINED ENV{QUTLASS_SRC_DIR})
-  set(QUTLASS_SRC_DIR $ENV{QUTLASS_SRC_DIR})
-endif()
-
-if(QUTLASS_SRC_DIR)
-  FetchContent_Declare(
-    qutlass
-    SOURCE_DIR ${QUTLASS_SRC_DIR}
-    CONFIGURE_COMMAND ""
-    BUILD_COMMAND ""
-  )
-else()
-  FetchContent_Declare(
-    qutlass
-    GIT_REPOSITORY https://github.com/IST-DASLab/qutlass.git
-    GIT_TAG 830d2c4537c7396e14a02a46fbddd18b5d107c65
-    GIT_PROGRESS TRUE
-    CONFIGURE_COMMAND ""
-    BUILD_COMMAND ""
-  )
-  FetchContent_Populate(qutlass)
-  set(qutlass_SOURCE_DIR "${qutlass_SOURCE_DIR}")
-endif()
-
-if(NOT qutlass_SOURCE_DIR)
-  message(FATAL_ERROR "[QUTLASS] source directory could not be resolved.")
-endif()
-message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}")
-
-cuda_archs_loose_intersection(QUTLASS_ARCHS "12.0a;10.0a" "${CUDA_ARCHS}")
-if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND QUTLASS_ARCHS)
-
-  if(QUTLASS_ARCHS MATCHES "10\\.0a")
-    set(QUTLASS_TARGET_CC 100)
-  elseif(QUTLASS_ARCHS MATCHES "12\\.0a")
-    set(QUTLASS_TARGET_CC 120)
-  else()
-    message(FATAL_ERROR "[QUTLASS] internal error parsing CUDA_ARCHS='${QUTLASS_ARCHS}'.")
-  endif()
-
-  set(QUTLASS_SOURCES
-    ${qutlass_SOURCE_DIR}/qutlass/csrc/bindings.cpp
-    ${qutlass_SOURCE_DIR}/qutlass/csrc/gemm.cu
-    ${qutlass_SOURCE_DIR}/qutlass/csrc/gemm_ada.cu
-    ${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_mx.cu
-    ${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_nv.cu
-    ${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_mx_sm100.cu
-    ${qutlass_SOURCE_DIR}/qutlass/csrc/fused_quantize_nv_sm100.cu
-  )
-
-  set(QUTLASS_INCLUDES
-    ${qutlass_SOURCE_DIR}
-    ${qutlass_SOURCE_DIR}/qutlass
-    ${qutlass_SOURCE_DIR}/qutlass/csrc/include
-    ${qutlass_SOURCE_DIR}/qutlass/csrc/include/cutlass_extensions
-  )
-
-  if(CUTLASS_INCLUDE_DIR AND EXISTS "${CUTLASS_INCLUDE_DIR}/cutlass/cutlass.h")
-    list(APPEND QUTLASS_INCLUDES "${CUTLASS_INCLUDE_DIR}")
-  elseif(EXISTS "${qutlass_SOURCE_DIR}/qutlass/third_party/cutlass/include/cutlass/cutlass.h")
-    list(APPEND QUTLASS_INCLUDES "${qutlass_SOURCE_DIR}/qutlass/third_party/cutlass/include")
-    message(STATUS "[QUTLASS] Using QuTLASS vendored CUTLASS headers (no vLLM CUTLASS detected).")
-  else()
-    message(FATAL_ERROR "[QUTLASS] CUTLASS headers not found. "
-                        "Set -DCUTLASS_INCLUDE_DIR=/path/to/cutlass/include")
-  endif()
-
-  set_gencode_flags_for_srcs(
-    SRCS "${QUTLASS_SOURCES}"
-    CUDA_ARCHS "${QUTLASS_ARCHS}"
-  )
-
-  target_sources(_C PRIVATE ${QUTLASS_SOURCES})
-  target_include_directories(_C PRIVATE ${QUTLASS_INCLUDES})
-  target_compile_definitions(_C PRIVATE
-    QUTLASS_DISABLE_PYBIND=1
-    TARGET_CUDA_ARCH=${QUTLASS_TARGET_CC}
-  )
-
-  set_property(SOURCE ${QUTLASS_SOURCES} APPEND PROPERTY COMPILE_OPTIONS
-    $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr --use_fast_math -O3>
-  )
-
-else()
-  if("${CMAKE_CUDA_COMPILER_VERSION}" VERSION_LESS "12.8")
-    message(STATUS
-      "[QUTLASS] Skipping build: CUDA 12.8 or newer is required (found ${CMAKE_CUDA_COMPILER_VERSION}).")
-  else()
-    message(STATUS
-      "[QUTLASS] Skipping build: no supported arch (12.0a / 10.0a) found in "
-      "CUDA_ARCHS='${CUDA_ARCHS}'.")
-  endif()
-endif()

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 8f468e7da54a8e2f98abfa7c38636aac91c0cba1
+          GIT_TAG ee4d25bd84e0cbc7e0b9b9685085fd5db2dcb62a
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/hipify.py

@@ -16,7 +16,7 @@ import shutil
 
 from torch.utils.hipify.hipify_python import hipify
 
-if __name__ == "__main__":
+if __name__ == '__main__':
     parser = argparse.ArgumentParser()
 
     # Project directory where all the source + include files live.
@@ -34,14 +34,15 @@ if __name__ == "__main__":
     )
 
     # Source files to convert.
-    parser.add_argument(
-        "sources", help="Source files to hipify.", nargs="*", default=[]
-    )
+    parser.add_argument("sources",
+                        help="Source files to hipify.",
+                        nargs="*",
+                        default=[])
 
     args = parser.parse_args()
 
     # Limit include scope to project_dir only
-    includes = [os.path.join(args.project_dir, "*")]
+    includes = [os.path.join(args.project_dir, '*')]
 
     # Get absolute path for all source files.
     extra_files = [os.path.abspath(s) for s in args.sources]
@@ -50,31 +51,25 @@ if __name__ == "__main__":
     # The directory might already exist to hold object files so we ignore that.
     shutil.copytree(args.project_dir, args.output_dir, dirs_exist_ok=True)
 
-    hipify_result = hipify(
-        project_directory=args.project_dir,
-        output_directory=args.output_dir,
-        header_include_dirs=[],
-        includes=includes,
-        extra_files=extra_files,
-        show_detailed=True,
-        is_pytorch_extension=True,
-        hipify_extra_files_only=True,
-    )
+    hipify_result = hipify(project_directory=args.project_dir,
+                           output_directory=args.output_dir,
+                           header_include_dirs=[],
+                           includes=includes,
+                           extra_files=extra_files,
+                           show_detailed=True,
+                           is_pytorch_extension=True,
+                           hipify_extra_files_only=True)
 
     hipified_sources = []
     for source in args.sources:
         s_abs = os.path.abspath(source)
-        hipified_s_abs = (
-            hipify_result[s_abs].hipified_path
-            if (
-                s_abs in hipify_result
-                and hipify_result[s_abs].hipified_path is not None
-            )
-            else s_abs
-        )
+        hipified_s_abs = (hipify_result[s_abs].hipified_path if
+                          (s_abs in hipify_result
+                           and hipify_result[s_abs].hipified_path is not None)
+                          else s_abs)
         hipified_sources.append(hipified_s_abs)
 
-    assert len(hipified_sources) == len(args.sources)
+    assert (len(hipified_sources) == len(args.sources))
 
     # Print hipified source files.
     print("\n".join(hipified_sources))

cmake/utils.cmake

@@ -310,13 +310,13 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   list(REMOVE_DUPLICATES _PTX_ARCHS)
   list(REMOVE_DUPLICATES _SRC_CUDA_ARCHS)
 
-  # If x.0a or x.0f is in SRC_CUDA_ARCHS and x.0 is in CUDA_ARCHS then we should
-  # remove x.0a or x.0f from SRC_CUDA_ARCHS and add x.0a or x.0f to _CUDA_ARCHS
+  # if x.0a is in SRC_CUDA_ARCHS and x.0 is in CUDA_ARCHS then we should
+  # remove x.0a from SRC_CUDA_ARCHS and add x.0a to _CUDA_ARCHS
   set(_CUDA_ARCHS)
   foreach(_arch ${_SRC_CUDA_ARCHS})
-    if(_arch MATCHES "[af]$")
+    if(_arch MATCHES "\\a$")
       list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
-      string(REGEX REPLACE "[af]$" "" _base "${_arch}")
+      string(REPLACE "a" "" _base "${_arch}")
       if ("${_base}" IN_LIST TGT_CUDA_ARCHS)
         list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
         list(APPEND _CUDA_ARCHS "${_arch}")

csrc/attention/attention_kernels.cuh

@@ -28,10 +28,10 @@
 
 #ifdef USE_ROCM
   #include <hip/hip_bf16.h>
-  #include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
+  #include "../quantization/fp8/amd/quant_utils.cuh"
 typedef __hip_bfloat16 __nv_bfloat16;
 #else
-  #include "../quantization/w8a8/fp8/nvidia/quant_utils.cuh"
+  #include "../quantization/fp8/nvidia/quant_utils.cuh"
 #endif
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))

csrc/attention/mla/cutlass_sm100_mla/device/sm100_mla.hpp

@@ -135,10 +135,10 @@ public:
     max_splits = min(16, max_splits);
 
     // TODO: This avoids a hang when the batch size larger than 1 and 
-    // there is more than 1 kv_splits. 
+    // there is more than 4 kv_splits. 
     // Discuss with NVIDIA how this can be fixed.
     if (B > 1) {
-      max_splits = min(1, max_splits);
+      max_splits = min(2, max_splits);
     }
     
     // printf("    max_splits = %d\n", max_splits);

csrc/attention/mla/cutlass_sm100_mla/kernel/sm100_fmha_mla_tma_warpspecialized.hpp

@@ -580,22 +580,22 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
       for (; tile_scheduler.is_valid(); ++tile_scheduler) {
         auto blk_coord = tile_scheduler.get_block_coord();
         auto problem_shape = params.problem_shape;
-        auto local_split_kv = params.split_kv;
+	auto local_split_kv = params.split_kv;
         if (params.mainloop.ptr_seq != nullptr) {
           get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
-          if (params.ptr_split_kv != nullptr) {
+	  if (params.ptr_split_kv != nullptr) {
             local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
           }
         }
-        if (local_split_kv <= get<3>(blk_coord))
-          continue;
+	if (local_split_kv <= get<3>(blk_coord))
+	  continue;
         load_page_table(
           blk_coord,
           problem_shape,
           params.mainloop,
           shared_storage.tensors,
           pipeline_page_table, pipeline_pt_producer_state,
-          local_split_kv
+	  local_split_kv
         );
       }
     }
@@ -604,15 +604,15 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
         CUTLASS_PRAGMA_NO_UNROLL
         for (; tile_scheduler.is_valid(); ++tile_scheduler) {
           auto blk_coord = tile_scheduler.get_block_coord();
-          auto problem_shape = params.problem_shape;
-          auto local_split_kv = params.split_kv;
+	  auto problem_shape = params.problem_shape;
+	  auto local_split_kv = params.split_kv;
           if (params.mainloop.ptr_seq != nullptr) {
             get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
-            if (params.ptr_split_kv != nullptr) {
+	    if (params.ptr_split_kv != nullptr) {
               local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
             }
           }
-          if (local_split_kv <= get<3>(blk_coord))
+	  if (local_split_kv <= get<3>(blk_coord))
             continue;
           load_cpasync(
             blk_coord,
@@ -621,7 +621,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
             params.mainloop_params,
             shared_storage.tensors,
             pipeline_load_qk, pipeline_load_qk_producer_state,
-            local_split_kv,
+	    local_split_kv,
             /* must be shared pipe */
             pipeline_page_table, pipeline_pt_consumer_state
           );
@@ -633,15 +633,15 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
           CUTLASS_PRAGMA_NO_UNROLL
           for (; tile_scheduler.is_valid(); ++tile_scheduler) {
             auto blk_coord = tile_scheduler.get_block_coord();
-            auto problem_shape = params.problem_shape;
-            auto local_split_kv = params.split_kv;
+	    auto problem_shape = params.problem_shape;
+	    auto local_split_kv = params.split_kv;
             if (params.mainloop.ptr_seq != nullptr) {
               get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
-              if (params.ptr_split_kv != nullptr) {
-                local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
-              }
+	      if (params.ptr_split_kv != nullptr) {
+	        local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
+	      }
             }
-            if (local_split_kv <= get<3>(blk_coord))
+	    if (local_split_kv <= get<3>(blk_coord))
               continue;
             load_tma</* paged= */ true>(
               blk_coord,
@@ -651,7 +651,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
               shared_storage.tensors,
               pipeline_load_qk, pipeline_load_qk_producer_state,
               pipeline_load_qk, pipeline_load_qk_producer_state,
-              local_split_kv
+	      local_split_kv
             );
             cutlass::arch::NamedBarrier((kNumComputeWarps + kNumLoadWarps) * NumThreadsPerWarp, kNamedBarrierEpilogue).arrive_and_wait();
           }
@@ -660,15 +660,15 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
           CUTLASS_PRAGMA_NO_UNROLL
           for (; tile_scheduler.is_valid(); ++tile_scheduler) {
             auto blk_coord = tile_scheduler.get_block_coord();
-            auto problem_shape = params.problem_shape;
-            auto local_split_kv = params.split_kv;
+	    auto problem_shape = params.problem_shape;
+	    auto local_split_kv = params.split_kv;
             if (params.mainloop.ptr_seq != nullptr) {
               get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
-              if (params.ptr_split_kv != nullptr) {
+	      if (params.ptr_split_kv != nullptr) {
                 local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
-              }
+	      }
             }
-            if (local_split_kv <= get<3>(blk_coord))
+	    if (local_split_kv <= get<3>(blk_coord))
               continue;
             load_tma<false>(
               blk_coord,
@@ -678,7 +678,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
               shared_storage.tensors,
               pipeline_load_qk, pipeline_load_qk_producer_state,
               pipeline_load_qk, pipeline_load_qk_producer_state,
-              local_split_kv
+	      local_split_kv
             );
             cutlass::arch::NamedBarrier((kNumComputeWarps + kNumLoadWarps) * NumThreadsPerWarp, kNamedBarrierEpilogue).arrive_and_wait();
           }
@@ -694,14 +694,14 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
         for (; tile_scheduler.is_valid(); ++tile_scheduler) {
           auto blk_coord = tile_scheduler.get_block_coord();
           auto problem_shape = params.problem_shape;
-          auto local_split_kv = params.split_kv;
+	  auto local_split_kv = params.split_kv;
           if (params.mainloop.ptr_seq != nullptr) {
             get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
             if (params.ptr_split_kv != nullptr) {
                 local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
             }
           }
-          if (local_split_kv <= get<3>(blk_coord))
+	  if (local_split_kv <= get<3>(blk_coord))
             continue;
           mma(blk_coord,
             problem_shape,
@@ -711,7 +711,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
             pipeline_mma_s, pipeline_mma_s_producer_state,
             pipeline_p_mma, pipeline_p_mma_consumer_state,
             pipeline_mma_o, pipeline_mma_o_producer_state,
-            local_split_kv
+	    local_split_kv
           );
         }
       }
@@ -726,15 +726,15 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
       for (; tile_scheduler.is_valid(); ++tile_scheduler) {
         auto blk_coord = tile_scheduler.get_block_coord();
         auto problem_shape = params.problem_shape;
-        auto split_kv = params.split_kv;
-        auto local_split_kv = split_kv;
+	auto split_kv = params.split_kv;
+	auto local_split_kv = split_kv;
         if (params.mainloop.ptr_seq != nullptr) {
           get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
-          if (params.ptr_split_kv != nullptr) {
+	  if (params.ptr_split_kv != nullptr) {
             local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
           }
         }
-        if (local_split_kv <= get<3>(blk_coord))
+	if (local_split_kv <= get<3>(blk_coord))
           continue;
         compute(
           blk_coord,
@@ -745,7 +745,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
           pipeline_mma_s, pipeline_mma_s_consumer_state,
           pipeline_p_mma, pipeline_p_mma_producer_state,
           pipeline_mma_o, pipeline_mma_o_consumer_state,
-          local_split_kv
+	  local_split_kv
         );
       }
 
@@ -1900,7 +1900,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
       cutlass::arch::NamedBarrier(
           (kNumComputeWarps + kNumLoadWarps) * NumThreadsPerWarp,
           kNamedBarrierEpilogue
-      ).arrive_and_wait();
+      ).arrive();
 
       return;
     }

csrc/cache.h

@@ -56,19 +56,3 @@ void cp_gather_cache(
     torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
     torch::Tensor const& cu_seq_lens,  // [BATCH+1]
     int64_t batch_size, std::optional<torch::Tensor> seq_starts = std::nullopt);
-
-// Indexer K quantization and cache function
-void indexer_k_quant_and_cache(
-    torch::Tensor& k,             // [num_tokens, head_dim]
-    torch::Tensor& kv_cache,      // [num_blocks, block_size, cache_stride]
-    torch::Tensor& slot_mapping,  // [num_tokens]
-    int64_t quant_block_size,     // quantization block size
-    const std::string& scale_fmt);
-
-// Extract function to gather quantized K cache
-void cp_gather_indexer_k_quant_cache(
-    const torch::Tensor& kv_cache,  // [num_blocks, block_size, cache_stride]
-    torch::Tensor& dst_k,           // [num_tokens, head_dim]
-    torch::Tensor& dst_scale,  // [num_tokens, head_dim / quant_block_size * 4]
-    const torch::Tensor& block_table,   // [batch_size, num_blocks]
-    const torch::Tensor& cu_seq_lens);  // [batch_size + 1]

csrc/cache_kernels.cu

@@ -9,14 +9,15 @@
 #include "quantization/vectorization_utils.cuh"
 
 #ifdef USE_ROCM
-  #include "quantization/w8a8/fp8/amd/quant_utils.cuh"
+  #include "quantization/fp8/amd/quant_utils.cuh"
 #else
-  #include "quantization/w8a8/fp8/nvidia/quant_utils.cuh"
+  #include "quantization/fp8/nvidia/quant_utils.cuh"
 #endif
 
 #include <algorithm>
 #include <cassert>
-#include <cfloat>
+#include <map>
+#include <vector>
 
 #ifdef USE_ROCM
   #include <hip/hip_bf16.h>
@@ -208,20 +209,6 @@ void copy_blocks_mla(std::vector<torch::Tensor> const& kv_caches,
 
 namespace vllm {
 
-// Used to copy/convert one element
-template <typename OutT, typename InT, Fp8KVCacheDataType kv_dt>
-struct CopyWithScaleOp {
-  float scale;
-
-  __device__ __forceinline__ void operator()(OutT& dst, const InT src) const {
-    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
-      dst = static_cast<OutT>(src);
-    } else {
-      dst = fp8::scaled_convert<OutT, InT, kv_dt>(src, scale);
-    }
-  }
-};
-
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
 __global__ void reshape_and_cache_kernel(
     const scalar_t* __restrict__ key,    // [num_tokens, num_heads, head_size]
@@ -237,51 +224,59 @@ __global__ void reshape_and_cache_kernel(
   const int64_t token_idx = blockIdx.x;
   const int64_t slot_idx = slot_mapping[token_idx];
   if (slot_idx < 0) {
+    // Padding token that should be ignored.
     return;
   }
 
   const int64_t block_idx = slot_idx / block_size;
   const int64_t block_offset = slot_idx % block_size;
-  const int h_block_count = head_size / x;  // head_size//x
 
-  const int h_block_idx = threadIdx.x;
-  if (h_block_idx >= num_heads * h_block_count) {
-    return;
-  }
+  const int n = num_heads * head_size;
+  for (int i = threadIdx.x; i < n; i += blockDim.x) {
+    const int64_t src_key_idx = token_idx * key_stride + i;
+    const int64_t src_value_idx = token_idx * value_stride + i;
 
-  const int head_idx = h_block_idx / h_block_count;
-  const int h_block = h_block_idx % h_block_count;
+    const int head_idx = i / head_size;
+    const int head_offset = i % head_size;
+    const int x_idx = head_offset / x;
+    const int x_offset = head_offset % x;
 
-  const scalar_t* __restrict__ key_src =
-      key + token_idx * key_stride + head_idx * head_size + h_block * x;
-  const int64_t src_value_start =
-      token_idx * value_stride + head_idx * head_size + h_block * x;
-
-  cache_t* __restrict__ key_dst =
-      key_cache + block_idx * num_heads * h_block_count * block_size * x +
-      head_idx * h_block_count * block_size * x + h_block * block_size * x +
-      block_offset * x;
-  const int64_t tgt_value_start =
-      block_idx * num_heads * h_block_count * x * block_size +
-      head_idx * h_block_count * x * block_size + h_block * x * block_size +
-      block_offset;
-
-  constexpr int VEC_SIZE = (sizeof(scalar_t) == 2) ? 8 : 4;
-  float k_scale_val = (kv_dt == Fp8KVCacheDataType::kAuto) ? 0.f : *k_scale;
-  CopyWithScaleOp<cache_t, scalar_t, kv_dt> k_op{k_scale_val};
-  float v_scale_val = (kv_dt == Fp8KVCacheDataType::kAuto) ? 0.f : *v_scale;
-  CopyWithScaleOp<cache_t, scalar_t, kv_dt> v_op{v_scale_val};
-
-  vectorize_with_alignment<VEC_SIZE>(key_src, key_dst, x, 0, 1, k_op);
-
-  const scalar_t* __restrict__ value_src = value + src_value_start;
-  cache_t* __restrict__ value_dst = value_cache + tgt_value_start;
-#pragma unroll
-  for (int i = 0; i < x; i++) {
-    v_op(value_dst[i * block_size], value_src[i]);
+    const int64_t tgt_key_idx =
+        block_idx * num_heads * (head_size / x) * block_size * x +
+        head_idx * (head_size / x) * block_size * x + x_idx * block_size * x +
+        block_offset * x + x_offset;
+    const int64_t tgt_value_idx =
+        block_idx * num_heads * head_size * block_size +
+        head_idx * head_size * block_size + head_offset * block_size +
+        block_offset;
+    scalar_t tgt_key = key[src_key_idx];
+    scalar_t tgt_value = value[src_value_idx];
+    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
+      key_cache[tgt_key_idx] = tgt_key;
+      value_cache[tgt_value_idx] = tgt_value;
+    } else {
+      key_cache[tgt_key_idx] =
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, *k_scale);
+      value_cache[tgt_value_idx] =
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, *v_scale);
+    }
   }
 }
 
+// Used by vectorization_utils to copy/convert one element
+template <typename OutT, typename InT, Fp8KVCacheDataType kv_dt>
+struct CopyWithScaleOp {
+  float scale;
+
+  __device__ __forceinline__ void operator()(OutT& dst, const InT src) const {
+    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
+      dst = static_cast<OutT>(src);
+    } else {
+      dst = fp8::scaled_convert<OutT, InT, kv_dt>(src, scale);
+    }
+  }
+};
+
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
 __global__ void reshape_and_cache_flash_kernel(
     const scalar_t* __restrict__ key,    // [num_tokens, num_heads, head_size]
@@ -401,241 +396,6 @@ __global__ void concat_and_cache_mla_kernel(
   copy(k_pe, kv_cache, k_pe_stride, block_stride, pe_dim, kv_lora_rank);
 }
 
-template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
-__global__ void concat_and_cache_ds_mla_kernel(
-    const scalar_t* __restrict__ kv_c,  // [num_tokens, kv_lora_rank]
-    const scalar_t* __restrict__ k_pe,  // [num_tokens, pe_dim]
-    cache_t* __restrict__ kv_cache,  // [num_blocks, block_size, (kv_lora_rank
-                                     // + pe_dim)]
-    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
-    const int block_stride,                    //
-    const int entry_stride,                    //
-    const int kv_c_stride,                     //
-    const int k_pe_stride,                     //
-    const int kv_lora_rank,                    //
-    const int pe_dim,                          //
-    const int block_size,                      //
-    const float* scale                         //
-) {
-  const int64_t token_idx = blockIdx.x;
-  const int64_t slot_idx = slot_mapping[token_idx];
-  // NOTE: slot_idx can be -1 if the token is padded
-  if (slot_idx < 0) {
-    return;
-  }
-  const int64_t block_idx = slot_idx / block_size;
-  const int64_t block_offset = slot_idx % block_size;
-  const int64_t dst_idx_start =
-      block_idx * block_stride + block_offset * entry_stride;
-
-  // For the NoPE part, each tile of 128 elements is handled by half of one warp
-  // (16 threads). There are 4 total tiles, so 2 warps (64 threads).
-  // Lanes 0 and 16 of each warp write the scale values for that warp's tiles.
-  // The RoPE part (last 64 elements) is handled by another 1 warp (32 threads).
-  // So in total, we use 3 warps (96 threads) per block.
-
-  // Cast kv_cache to 16_bit for RoPE values
-  scalar_t* kv_cache_16bit =
-      reinterpret_cast<scalar_t*>(&kv_cache[dst_idx_start]);
-
-  // The last warp handles the RoPE part
-  if (threadIdx.x >= 64) {
-    // Each thread handles two elements of RoPE
-    const int8_t pe_idx_start = (threadIdx.x - 64) * 2;
-    const int64_t src_idx = token_idx * k_pe_stride + pe_idx_start;
-    // Vectorized load of two 16-bit values, performed as one 32-bit load
-    const int32_t vals = *reinterpret_cast<const int32_t*>(&k_pe[src_idx]);
-    // RoPE values start after the packed 8-bit NoPE values and the
-    // 32-bit scales
-    const int64_t dst_idx = kv_lora_rank / 2 + 8 + pe_idx_start;
-    // Vectorized store of two 16-bit values, performed as one 32-bit store
-    *reinterpret_cast<int32_t*>(&kv_cache_16bit[dst_idx]) = vals;
-    return;
-  }
-
-  // The first two warps handle the NoPE part
-  const int8_t warp_idx = threadIdx.x >> 5;
-  const int8_t lane_idx = threadIdx.x & 31;
-  const int8_t tile_idx = warp_idx * 2 + (lane_idx >> 4);
-
-  // Each thread handles 8 elements of NoPE
-  // Load the NoPE elements for this thread into registers
-  const int64_t src_idx_start = token_idx * kv_c_stride + (threadIdx.x * 8);
-  // Vectorized load of eight 16-bit values, performed as an int4 load
-  const int4 vals_i4 = *reinterpret_cast<const int4*>(&kv_c[src_idx_start]);
-  const scalar_t* vals = reinterpret_cast<const scalar_t*>(&vals_i4);
-
-  // Max absolute value of this thread's elements
-  float max_abs = fmaxf(fmaxf(fmaxf(fabsf(vals[0]), fabsf(vals[1])),
-                              fmaxf(fabsf(vals[2]), fabsf(vals[3]))),
-                        fmaxf(fmaxf(fabsf(vals[4]), fabsf(vals[5])),
-                              fmaxf(fabsf(vals[6]), fabsf(vals[7]))));
-
-  // Warp-level reduction to find the max absolute value in each half-warp
-#pragma unroll
-  for (int offset = 8; offset > 0; offset /= 2) {
-    max_abs = fmaxf(max_abs, VLLM_SHFL_XOR_SYNC_WIDTH(max_abs, offset, 16));
-  }
-
-  // Compute the scale for the tile
-  float tile_scale = max_abs / 448.f;
-  tile_scale = fmaxf(tile_scale, FLT_MIN);
-
-  // The first lane of each half-warp writes the scale to kv_cache
-  if ((lane_idx == 0) || (lane_idx == 16)) {
-    float* kv_cache_32bit = reinterpret_cast<float*>(&kv_cache[dst_idx_start]);
-    const uint64_t dst_idx = kv_lora_rank / 4 + tile_idx;
-    kv_cache_32bit[dst_idx] = tile_scale;
-  }
-
-  // Now all threads in the block scale and write their elements
-  // NoPE data is packed in the first kv_lora_rank/2 bytes (first 256 bytes)
-  const int64_t dst_idx_base = dst_idx_start + (threadIdx.x * 8);
-
-  uint8_t result[8];
-#pragma unroll
-  for (int i = 0; i < 8; i++) {
-    result[i] =
-        fp8::scaled_convert<uint8_t, scalar_t, Fp8KVCacheDataType::kFp8E4M3>(
-            vals[i], tile_scale);
-  }
-
-  // Store as aligned 64-bit writes
-  *reinterpret_cast<uint64_t*>(&kv_cache[dst_idx_base]) =
-      *reinterpret_cast<const uint64_t*>(result);
-}
-
-template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
-__global__ void indexer_k_quant_and_cache_kernel(
-    const scalar_t* __restrict__ k,  // [num_tokens, head_dim]
-    cache_t* __restrict__ kv_cache,  // [num_blocks, block_size, cache_stride]
-    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
-    const int head_dim,                        // dimension of each head
-    const int quant_block_size,                // quantization block size
-    const int cache_block_size,                // cache block size
-    const int cache_stride,  // stride for each token in kv_cache
-    const bool use_ue8m0     // use ue8m0 scale format
-) {
-  constexpr int VEC_SIZE = 4;
-  const int64_t token_idx = blockIdx.x;
-  const int64_t head_dim_idx = (blockIdx.y * blockDim.y * blockDim.x +
-                                threadIdx.y * blockDim.x + threadIdx.x) *
-                               VEC_SIZE;
-  const int64_t slot_idx = slot_mapping[token_idx];
-  const int64_t block_idx = slot_idx / cache_block_size;
-  const int64_t block_offset = slot_idx % cache_block_size;
-
-  // NOTE: slot_idx can be -1 if the token is padded
-  if (slot_idx < 0 || (head_dim_idx >= head_dim)) {
-    return;
-  }
-
-  float2 k_val = (reinterpret_cast<const float2*>(
-      k))[(token_idx * head_dim + head_dim_idx) / VEC_SIZE];
-  scalar_t* k_val_ptr = reinterpret_cast<scalar_t*>(&k_val);
-  float amax = 0.0f;
-  for (int i = 0; i < VEC_SIZE; i++) {
-    amax = fmaxf(amax, fabsf(float(k_val_ptr[i])));
-  }
-#ifndef USE_ROCM
-  __syncwarp();
-#endif
-
-  // Reduced amax
-  for (int mask = 16; mask > 0; mask /= 2) {
-#ifdef USE_ROCM
-    amax = fmaxf(amax, __shfl_xor_sync(uint64_t(-1), amax, mask));
-#else
-    amax = fmaxf(amax, __shfl_xor_sync(unsigned(-1), amax, mask));
-#endif
-  }
-#ifndef USE_ROCM
-  __syncwarp();
-#endif
-  float scale = fmaxf(amax, 1e-4) / 448.0f;
-  if (use_ue8m0) {
-    scale = exp2f(ceilf(log2f(scale)));
-  }
-
-  const int64_t dst_offset = block_idx * cache_block_size * cache_stride +
-                             block_offset * head_dim + head_dim_idx;
-  for (int i = 0; i < VEC_SIZE; i++) {
-    kv_cache[dst_offset + i] =
-        fp8::scaled_convert<cache_t, scalar_t, kv_dt>(k_val_ptr[i], scale);
-  }
-  if (threadIdx.x == 0) {
-    const int64_t dst_scale_idx =
-        block_idx * cache_block_size * cache_stride +
-        cache_block_size * head_dim +
-        (block_offset * head_dim + head_dim_idx) * 4 / quant_block_size;
-    reinterpret_cast<float*>(kv_cache)[dst_scale_idx / 4] = scale;
-  }
-}
-
-template <int BLOCK_Y_SIZE>
-__global__ void cp_gather_indexer_k_quant_cache_kernel(
-    const char* __restrict__ kv_cache,  // [num_blocks, block_size,
-                                        // cache_stride]
-    char* __restrict__ dst_k,           // [num_tokens, head_dim]
-    char* __restrict__ dst_scale,  // [num_tokens, head_dim / quant_block_size *
-                                   // 4]
-    const int* __restrict__ block_table,  // [batch_size, num_blocks]
-    const int* __restrict__ cu_seq_lens,  // [batch_size + 1]
-    const int batch_size,                 // batch size
-    const int64_t token_stride,           // stride for each token in dst_k
-    const int64_t head_dim,               // dimension of each head
-    const int64_t block_stride,           // stride for each block in kv_cache
-    const int64_t cache_token_stride,     // stride for each token in kv_cache
-    const int64_t cache_block_size,  // num_tokens for each block in kv_cache
-    const int num_blocks,            // number of blocks
-    const int num_tokens,            // number of tokens
-    const int quant_block_size       // quantization block size
-) {
-  constexpr int VEC_SIZE = sizeof(float4) / sizeof(char);
-  const int token_idx = blockIdx.x * blockDim.y + threadIdx.y;
-  const int head_idx = (blockIdx.y * blockDim.x + threadIdx.x) * VEC_SIZE;
-  // Find batch index within a block
-  __shared__ int batch_idx[BLOCK_Y_SIZE];
-  for (int iter = 0; iter < cuda_utils::ceil_div(batch_size, int(blockDim.x));
-       iter++) {
-    int tid = iter * blockDim.x + threadIdx.x;
-    if (tid < batch_size) {
-      const int seq_start = cu_seq_lens[tid];
-      const int seq_end = cu_seq_lens[tid + 1];
-      if (token_idx >= seq_start && token_idx < seq_end) {
-        batch_idx[threadIdx.y] = tid;
-      }
-    }
-  }
-
-#ifndef USE_ROCM
-  __syncwarp();
-#endif
-
-  if (head_idx >= head_dim || token_idx >= num_tokens) {
-    return;
-  }
-  const int inbatch_seq_idx = token_idx - cu_seq_lens[batch_idx[threadIdx.y]];
-  const int block_idx = block_table[batch_idx[threadIdx.y] * num_blocks +
-                                    inbatch_seq_idx / cache_block_size];
-  const int64_t src_block_offset = block_idx * block_stride;
-  const int64_t cache_inblock_offset =
-      (inbatch_seq_idx % cache_block_size) * head_dim + head_idx;
-  const int64_t src_inblock_offset = src_block_offset + cache_inblock_offset;
-  const int64_t dst_inblock_offset = token_idx * token_stride + head_idx;
-
-  reinterpret_cast<float4*>(dst_k)[dst_inblock_offset / VEC_SIZE] =
-      reinterpret_cast<const float4*>(kv_cache)[src_inblock_offset / VEC_SIZE];
-  ;
-  if (threadIdx.x == 0) {
-    const int64_t src_scale_offset =
-        src_block_offset + cache_block_size * head_dim +
-        cache_inblock_offset * 4 / quant_block_size;
-    reinterpret_cast<float*>(dst_scale)[dst_inblock_offset / quant_block_size] =
-        reinterpret_cast<const float*>(kv_cache)[src_scale_offset / 4];
-  }
-}
-
 }  // namespace vllm
 
 // KV_T is the data type of key and value tensors.
@@ -671,15 +431,14 @@ void reshape_and_cache(
 
   int key_stride = key.stride(0);
   int value_stride = value.stride(0);
-  int head_div_x = head_size / x;
 
   dim3 grid(num_tokens);
-  dim3 block(std::min(num_heads * head_div_x, 512));
+  dim3 block(std::min(num_heads * head_size, 512));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(key));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
   DISPATCH_BY_KV_CACHE_DTYPE(key.dtype(), kv_cache_dtype,
-                             CALL_RESHAPE_AND_CACHE);
+                             CALL_RESHAPE_AND_CACHE)
 }
 
 // KV_T is the data type of key and value tensors.
@@ -750,18 +509,6 @@ void reshape_and_cache_flash(
           kv_c_stride, k_pe_stride, kv_lora_rank, pe_dim, block_size,   \
           reinterpret_cast<const float*>(scale.data_ptr()));
 
-// KV_T is the data type of key and value tensors.
-// CACHE_T is the stored data type of kv-cache.
-#define CALL_CONCAT_AND_CACHE_DS_MLA(KV_T, CACHE_T, KV_DTYPE)           \
-  vllm::concat_and_cache_ds_mla_kernel<KV_T, CACHE_T, KV_DTYPE>         \
-      <<<grid, block, 0, stream>>>(                                     \
-          reinterpret_cast<KV_T*>(kv_c.data_ptr()),                     \
-          reinterpret_cast<KV_T*>(k_pe.data_ptr()),                     \
-          reinterpret_cast<CACHE_T*>(kv_cache.data_ptr()),              \
-          slot_mapping.data_ptr<int64_t>(), block_stride, entry_stride, \
-          kv_c_stride, k_pe_stride, kv_lora_rank, pe_dim, block_size,   \
-          reinterpret_cast<const float*>(scale.data_ptr()));
-
 void concat_and_cache_mla(
     torch::Tensor& kv_c,          // [num_tokens, kv_lora_rank]
     torch::Tensor& k_pe,          // [num_tokens, pe_dim]
@@ -784,43 +531,20 @@ void concat_and_cache_mla(
   int pe_dim = k_pe.size(1);
   int block_size = kv_cache.size(1);
 
-  if (kv_cache_dtype == "fp8_ds_mla") {
-    TORCH_CHECK(kv_lora_rank == 512, "kv_lora_rank must be 512 for fp8_ds_mla");
-    TORCH_CHECK(pe_dim == 64, "pe_dim must be 64 for fp8_ds_mla");
-    TORCH_CHECK(kv_cache.size(2) == 656 / kv_cache.itemsize(),
-                "kv_cache.size(2) must be 656 bytes for fp8_ds_mla");
-    TORCH_CHECK(kv_c.itemsize() == 2,
-                "kv_c.itemsize() must be 2 for fp8_ds_mla");
-    TORCH_CHECK(k_pe.itemsize() == 2,
-                "k_pe.itemsize() must be 2 for fp8_ds_mla");
-  } else {
-    TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
-  }
+  TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
 
   int kv_c_stride = kv_c.stride(0);
   int k_pe_stride = k_pe.stride(0);
   int block_stride = kv_cache.stride(0);
   int entry_stride = kv_cache.stride(1);
 
+  dim3 grid(num_tokens);
+  dim3 block(std::min(kv_lora_rank, 512));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(kv_c));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  if (kv_cache_dtype == "fp8_ds_mla") {
-    dim3 grid(num_tokens);
-    // For the NoPE part, each tile of 128 elements is handled by half of one
-    // warp (16 threads). There are 4 total tiles, so 2 warps (64 threads).
-    // Lanes 0 and 16 of each warp write the scale values for that warp's tiles.
-    // The RoPE part (last 64 elements) is handled by another 1 warp (32
-    // threads). So in total, we use 3 warps (96 threads) per block.
-    dim3 block(96);
-    DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
-                               CALL_CONCAT_AND_CACHE_DS_MLA);
-  } else {
-    dim3 grid(num_tokens);
-    dim3 block(std::min(kv_lora_rank, 512));
-    DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
-                               CALL_CONCAT_AND_CACHE_MLA);
-  }
+  DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
+                             CALL_CONCAT_AND_CACHE_MLA);
 }
 
 namespace vllm {
@@ -1198,98 +922,3 @@ void cp_gather_cache(
     TORCH_CHECK(false, "Unsupported data type width: ", dtype_bits);
   }
 }
-
-// Macro to dispatch the kernel based on the data type.
-#define CALL_INDEXER_K_QUANT_AND_CACHE(KV_T, CACHE_T, KV_DTYPE)         \
-  vllm::indexer_k_quant_and_cache_kernel<KV_T, CACHE_T, KV_DTYPE>       \
-      <<<grid, block, 0, stream>>>(                                     \
-          reinterpret_cast<KV_T*>(k.data_ptr()),                        \
-          reinterpret_cast<CACHE_T*>(kv_cache.data_ptr()),              \
-          slot_mapping.data_ptr<int64_t>(), head_dim, quant_block_size, \
-          cache_block_size, cache_stride, use_ue8m0);
-
-void indexer_k_quant_and_cache(
-    torch::Tensor& k,             // [num_tokens, head_dim]
-    torch::Tensor& kv_cache,      // [num_blocks, block_size, cache_stride]
-    torch::Tensor& slot_mapping,  // [num_tokens]
-    int64_t quant_block_size,     // quantization block size
-    const std::string& scale_fmt) {
-  int num_tokens = k.size(0);
-  int head_dim = k.size(1);
-  int cache_block_size = kv_cache.size(1);
-  int cache_stride = kv_cache.size(2);
-  bool use_ue8m0 = scale_fmt == "ue8m0";
-
-  TORCH_CHECK(k.device() == kv_cache.device(),
-              "k and kv_cache must be on the same device");
-  TORCH_CHECK(k.device() == slot_mapping.device(),
-              "k and slot_mapping must be on the same device");
-  TORCH_CHECK(head_dim % quant_block_size == 0,
-              "head_dim must be divisible by quant_block_size");
-
-  constexpr int vec_size = 4;
-  dim3 grid(num_tokens, (head_dim + quant_block_size * vec_size - 1) /
-                            (quant_block_size * vec_size));
-  dim3 block(32, vec_size);
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(k));
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-
-  DISPATCH_BY_KV_CACHE_DTYPE(k.dtype(), "fp8_e4m3",
-                             CALL_INDEXER_K_QUANT_AND_CACHE);
-}
-
-// Macro to dispatch the kernel based on the data amount.
-#define CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(BLOCK_Y_SIZE)                  \
-  vllm::cp_gather_indexer_k_quant_cache_kernel<BLOCK_Y_SIZE>                \
-      <<<dim3((num_tokens + BLOCK_Y_SIZE - 1) / BLOCK_Y_SIZE,               \
-              (head_dim + 8 * vec_size - 1) / (8 * vec_size)),              \
-         dim3(8, BLOCK_Y_SIZE), 0, stream>>>(                               \
-          reinterpret_cast<char*>(kv_cache.data_ptr()),                     \
-          reinterpret_cast<char*>(dst_k.data_ptr()),                        \
-          reinterpret_cast<char*>(dst_scale.data_ptr()),                    \
-          block_table.data_ptr<int32_t>(), cu_seq_lens.data_ptr<int32_t>(), \
-          batch_size, dst_k.stride(0), dst_k.size(1), kv_cache.stride(0),   \
-          kv_cache.stride(1), kv_cache.size(1), block_table.size(1),        \
-          num_tokens, quant_block_size);
-
-void cp_gather_indexer_k_quant_cache(
-    const torch::Tensor& kv_cache,  // [num_blocks, block_size, cache_stride]
-    torch::Tensor& dst_k,           // [num_tokens, head_dim]
-    torch::Tensor& dst_scale,  // [num_tokens, head_dim / quant_block_size * 4]
-    const torch::Tensor& block_table,  // [batch_size, num_blocks]
-    const torch::Tensor& cu_seq_lens   // [batch_size + 1]
-) {
-  int batch_size = block_table.size(0);
-  int num_tokens = dst_k.size(0);
-  int head_dim = dst_k.size(1);
-  int quant_block_size = head_dim * 4 / dst_scale.size(1);
-
-  TORCH_CHECK(kv_cache.device() == dst_k.device(),
-              "kv_cache and dst_k must be on the same device");
-  TORCH_CHECK(kv_cache.device() == dst_scale.device(),
-              "kv_cache and dst_scale must be on the same device");
-  TORCH_CHECK(kv_cache.device() == block_table.device(),
-              "kv_cache and block_table must be on the same device");
-  TORCH_CHECK(kv_cache.device() == cu_seq_lens.device(),
-              "kv_cache and cu_seq_lens must be on the same device");
-  TORCH_CHECK(head_dim % quant_block_size == 0,
-              "head_dim must be divisible by quant_block_size");
-
-  constexpr int vec_size = 16;
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(kv_cache));
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-
-  if (num_tokens < 32) {
-    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(1);
-  } else if (num_tokens < 64) {
-    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(2);
-  } else if (num_tokens < 128) {
-    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(4);
-  } else if (num_tokens < 256) {
-    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(8);
-  } else if (num_tokens < 512) {
-    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(16);
-  } else {
-    CALL_CP_GATHER_INDEXER_K_QUANT_CACHE(32);
-  }
-}

csrc/core/batch_invariant.hpp

@@ -1,19 +0,0 @@
-#pragma once
-#include <cstdlib>
-#include <string>
-#include <cctype>
-
-namespace vllm {
-
-// vllm_kernel_override_batch_invariant(); returns true
-// if env VLLM_KERNEL_OVERRIDE_BATCH_INVARIANT=1
-inline bool vllm_kernel_override_batch_invariant() {
-  static bool cached = []() {
-    std::string env_key = "VLLM_KERNEL_OVERRIDE_BATCH_INVARIANT";
-    const char* val = std::getenv(env_key.c_str());
-    return (val && std::atoi(val) != 0) ? 1 : 0;
-  }();
-  return cached;
-}
-
-}  // namespace vllm

csrc/cpu/cpu_types.hpp

@@ -14,8 +14,7 @@
   // arm implementation
   #include "cpu_types_arm.hpp"
 #else
-  #warning "unsupported vLLM cpu implementation, vLLM will compile with scalar"
-  #include "cpu_types_scalar.hpp"
+  #warning "unsupported vLLM cpu implementation"
 #endif
 
 #ifdef _OPENMP

csrc/cpu/cpu_types_scalar.hpp

@@ -1,513 +0,0 @@
-#include <cmath>
-#include <cstdint>
-#include <cstring>
-#include <torch/all.h>
-#include "float_convert.hpp"
-
-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__) \
-  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
-
-#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)
-#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;
-#endif
-
-#define FORCE_INLINE __attribute__((always_inline)) inline
-
-#define __max(a, b) ((a) > (b) ? (a) : (b))
-#define __min(a, b) ((a) < (b) ? (a) : (b))
-#define __abs(a) ((a) < (0) ? (0 - a) : (a))
-
-typedef struct f16x8_t {
-  uint16_t val[8];
-} f16x8_t;
-
-typedef struct f16x16_t {
-  uint16_t val[16];
-} f16x16_t;
-
-typedef struct f16x32_t {
-  uint16_t val[32];
-} f16x32_t;
-
-typedef struct f32x4_t {
-  float val[4];
-} f32x4_t;
-
-typedef struct f32x8_t {
-  float val[8];
-} f32x8_t;
-
-typedef struct f32x16_t {
-  float val[16];
-} f32x16_t;
-
-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>{}), ...);
-};
-};  // 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) {
-  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
-}
-
-template <typename T>
-struct Vec {
-  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; }
-};
-
-struct FP32Vec8;
-struct FP32Vec16;
-
-struct FP16Vec8 : public Vec<FP16Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-  f16x8_t reg;
-
-  explicit FP16Vec8(const void* ptr)
-      : reg(*reinterpret_cast<const f16x8_t*>(ptr)) {};
-
-  explicit FP16Vec8(const FP32Vec8&);
-
-  void save(void* ptr) const { *reinterpret_cast<f16x8_t*>(ptr) = reg; }
-};
-
-struct FP16Vec16 : public Vec<FP16Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  f16x16_t reg;
-
-  explicit FP16Vec16(const void* ptr)
-      : reg(*reinterpret_cast<const f16x16_t*>(ptr)) {};
-
-  explicit FP16Vec16(const FP32Vec16&);
-
-  void save(void* ptr) const { *reinterpret_cast<f16x16_t*>(ptr) = reg; }
-
-  void save(void* ptr, const int elem_num) const {
-    int num = __min(elem_num, VEC_ELEM_NUM);
-    std::memcpy(ptr, &(reg.val[0]), num * sizeof(uint16_t));
-  }
-};
-
-struct BF16Vec8 : public Vec<BF16Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-  f16x8_t reg;
-
-  explicit BF16Vec8(const void* ptr)
-      : reg(*reinterpret_cast<const f16x8_t*>(ptr)) {};
-
-  explicit BF16Vec8(const FP32Vec8&);
-
-  void save(void* ptr) const { *reinterpret_cast<f16x8_t*>(ptr) = reg; }
-};
-
-struct BF16Vec16 : public Vec<BF16Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  f16x16_t reg;
-
-  explicit BF16Vec16(const void* ptr)
-      : reg(*reinterpret_cast<const f16x16_t*>(ptr)) {};
-
-  explicit BF16Vec16(const FP32Vec16&);
-
-  void save(void* ptr) const { *reinterpret_cast<f16x16_t*>(ptr) = reg; }
-
-  void save(void* ptr, const int elem_num) const {
-    int num = __min(elem_num, VEC_ELEM_NUM);
-    std::memcpy(ptr, &(reg.val[0]), num * sizeof(uint16_t));
-  }
-};
-
-struct BF16Vec32 : public Vec<BF16Vec32> {
-  constexpr static int VEC_ELEM_NUM = 32;
-  f16x32_t reg;
-
-  explicit BF16Vec32(const void* ptr)
-      : reg(*reinterpret_cast<const f16x32_t*>(ptr)) {};
-
-  explicit BF16Vec32(f16x32_t data) : reg(data) {};
-
-  explicit BF16Vec32(BF16Vec8& vec8_data) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = vec8_data.reg.val[i % BF16Vec8::VEC_ELEM_NUM];
-    }
-  }
-
-  void save(void* ptr) const { *reinterpret_cast<f16x32_t*>(ptr) = reg; }
-};
-
-struct FP32Vec4 : public Vec<FP32Vec4> {
-  constexpr static int VEC_ELEM_NUM = 4;
-
-  f32x4_t reg;
-
-  explicit FP32Vec4(float v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = v;
-    }
-  }
-
-  explicit FP32Vec4() {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = 0.0f;
-    }
-  }
-
-  explicit FP32Vec4(const float* ptr)
-      : reg(*reinterpret_cast<const f32x4_t*>(ptr)) {};
-
-  explicit FP32Vec4(f32x4_t data) : reg(data) {};
-
-  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {};
-};
-
-struct FP32Vec8 : public Vec<FP32Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-
-  f32x8_t reg;
-
-  explicit FP32Vec8(float v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = v;
-    }
-  }
-
-  explicit FP32Vec8() {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = 0.0f;
-    }
-  }
-
-  explicit FP32Vec8(const float* ptr)
-      : reg(*reinterpret_cast<const f32x8_t*>(ptr)) {};
-
-  explicit FP32Vec8(f32x8_t data) : reg(data) {};
-
-  explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {};
-
-  explicit FP32Vec8(const FP16Vec8& v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = fp16_to_float(v.reg.val[i]);
-    }
-  }
-
-  FP32Vec8(const BF16Vec8& v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = bf16_to_float(v.reg.val[i]);
-    }
-  }
-
-  float reduce_sum() const {
-    float result = 0;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result += reg.val[i];
-    }
-    return result;
-  }
-
-  FP32Vec8 exp() const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = expf(reg.val[i]);
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 tanh() const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = tanhf(reg.val[i]);
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 er() const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = erf(reg.val[i]);
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 operator*(const FP32Vec8& b) const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = reg.val[i] * b.reg.val[i];
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 operator+(const FP32Vec8& b) const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = reg.val[i] + b.reg.val[i];
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 operator-(const FP32Vec8& b) const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = reg.val[i] - b.reg.val[i];
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 operator/(const FP32Vec8& b) const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = reg.val[i] / b.reg.val[i];
-    }
-    return FP32Vec8(ret);
-  }
-
-  void save(void* ptr) const { *reinterpret_cast<f32x8_t*>(ptr) = reg; }
-};
-
-struct FP32Vec16 : public Vec<FP32Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  f32x16_t reg;
-
-  explicit FP32Vec16(float v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = v;
-    }
-  }
-
-  explicit FP32Vec16() {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = 0.0f;
-    }
-  }
-
-  explicit FP32Vec16(const float* ptr)
-      : reg(*reinterpret_cast<const f32x16_t*>(ptr)) {};
-
-  explicit FP32Vec16(f32x16_t data) : reg(data) {};
-
-  FP32Vec16(const FP32Vec4& data) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = data.reg.val[i % FP32Vec4::VEC_ELEM_NUM];
-    }
-  }
-
-  FP32Vec16(const FP32Vec8& data) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = data.reg.val[i % FP32Vec8::VEC_ELEM_NUM];
-    }
-  }
-
-  FP32Vec16(const FP32Vec16& data) : reg(data.reg) {};
-
-  explicit FP32Vec16(const FP16Vec16& v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = fp16_to_float(v.reg.val[i]);
-    }
-  }
-
-  explicit FP32Vec16(const BF16Vec16& v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = bf16_to_float(v.reg.val[i]);
-    }
-  }
-
-  explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
-
-  FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
-
-  FP32Vec16 operator*(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = reg.val[i] * b.reg.val[i];
-    }
-    return result;
-  }
-
-  FP32Vec16 operator+(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = reg.val[i] + b.reg.val[i];
-    }
-    return result;
-  }
-
-  FP32Vec16 operator-(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = reg.val[i] - b.reg.val[i];
-    }
-    return result;
-  }
-
-  FP32Vec16 operator/(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = reg.val[i] / b.reg.val[i];
-    }
-    return result;
-  }
-
-  FP32Vec16 max(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = __max(reg.val[i], b.reg.val[i]);
-    }
-    return result;
-  }
-
-  FP32Vec16 min(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = __min(reg.val[i], b.reg.val[i]);
-    }
-    return result;
-  }
-
-  FP32Vec16 abs() const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = __abs(reg.val[i]);
-    }
-    return result;
-  }
-
-  float reduce_sum() const {
-    float result = 0.0f;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result += reg.val[i];
-    }
-    return result;
-  }
-
-  float reduce_max() const {
-    float result = reg.val[0];
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result = __max(reg.val[i], result);
-    }
-    return result;
-  }
-
-  float reduce_min() const {
-    float result = reg.val[0];
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result = __min(reg.val[i], result);
-    }
-    return result;
-  }
-
-  template <int group_size>
-  float reduce_sub_sum(int idx) {
-    static_assert(VEC_ELEM_NUM % group_size == 0);
-    float sum = 0.0;
-    int start = idx * group_size;
-    int end = (idx + 1) * group_size;
-
-    for (; (start < VEC_ELEM_NUM) && (start < end); ++start) {
-      sum += reg.val[start];
-    }
-
-    return sum;
-  }
-
-  void save(void* ptr) const { *reinterpret_cast<f32x16_t*>(ptr) = reg; }
-};
-
-template <typename T>
-struct VecType {
-  using vec_type = void;
-};
-
-template <typename T>
-using vec_t = typename VecType<T>::vec_type;
-
-template <>
-struct VecType<float> {
-  using vec_type = FP32Vec8;
-};
-
-template <>
-struct VecType<c10::Half> {
-  using vec_type = FP16Vec8;
-};
-
-template <>
-struct VecType<c10::BFloat16> {
-  using vec_type = BF16Vec8;
-};
-
-template <typename T>
-void storeFP32(float v, T* ptr) {
-  *ptr = v;
-}
-
-/*
-template <> inline void storeFP32<c10::Half>(float v, c10::Half *ptr) {
-  c10::Half __attribute__((__may_alias__)) *v_ptr =
-      reinterpret_cast<c10::Half *>(&v);
-  *ptr = *(v_ptr + 1);
-}
-*/
-
-template <>
-inline void storeFP32<c10::Half>(float v, c10::Half* ptr) {
-  uint16_t fp16 = float_to_fp16(v);
-  *reinterpret_cast<uint16_t*>(ptr) = fp16;
-}
-
-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);
-}
-
-inline FP16Vec16::FP16Vec16(const FP32Vec16& v) {
-  int i = 0;
-  for (i = 0; i < FP16Vec16::VEC_ELEM_NUM; ++i) {
-    reg.val[i] = float_to_fp16(v.reg.val[i]);
-  }
-}
-
-inline FP16Vec8 ::FP16Vec8(const FP32Vec8& v) {
-  int i = 0;
-  for (i = 0; i < FP16Vec8::VEC_ELEM_NUM; ++i) {
-    reg.val[i] = float_to_fp16(v.reg.val[i]);
-  }
-}
-
-inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) {
-  acc = acc + a * b;
-}
-
-inline BF16Vec8::BF16Vec8(const FP32Vec8& v) {
-  int i = 0;
-  for (i = 0; i < BF16Vec8::VEC_ELEM_NUM; ++i) {
-    reg.val[i] = float_to_bf16(v.reg.val[i]);
-  }
-}
-
-inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
-  int i = 0;
-  for (i = 0; i < BF16Vec16::VEC_ELEM_NUM; ++i) {
-    reg.val[i] = float_to_bf16(v.reg.val[i]);
-  }
-}
-
-inline void prefetch(const void* addr) { __builtin_prefetch(addr, 0, 3); }
-
-};  // namespace vec_op

csrc/cpu/dnnl_helper.cpp

@@ -137,8 +137,9 @@ DNNLMatMulPrimitiveHandler::DNNLMatMulPrimitiveHandler(
 }
 
 void DNNLMatMulPrimitiveHandler::prepack_weight(
-    void* original_b_ptr, dnnl::memory::desc original_b_md,
-    dnnl::memory::desc b_target_mem_desc) {
+    void* original_b_ptr, dnnl::memory::desc b_target_mem_desc) {
+  dnnl::memory::desc original_b_md({b_k_size_, b_n_size_}, b_type_,
+                                   {b_k_stride_, b_n_stride_});
   dnnl::memory original_weight(original_b_md, default_engine(), original_b_ptr);
   dnnl::memory packed_weight(b_target_mem_desc, default_engine());
   {
@@ -249,9 +250,7 @@ W8A8MatMulPrimitiveHandler::W8A8MatMulPrimitiveHandler(const Args& args)
   if (a_qs_ == QuantizationStrategy::PER_TOKEN) {
     assert(!use_azp_);
   };
-  dnnl::memory::desc original_b_md({b_k_size_, b_n_size_}, b_type_,
-                                   {b_k_stride_, b_n_stride_});
-  prepack_weight(args.b_ptr, original_b_md,
+  prepack_weight(args.b_ptr,
                  create_primitive_desc(
                      MSizeCacheKey{.a_m_size = DNNL_RUNTIME_DIM_VAL,
                                    .use_bias = false,
@@ -413,25 +412,12 @@ MatMulPrimitiveHandler::MatMulPrimitiveHandler(const Args& args)
   assert(ab_type_ == dnnl::memory::data_type::f32 ||
          ab_type_ == dnnl::memory::data_type::bf16 ||
          ab_type_ == dnnl::memory::data_type::f16);
-
-  dnnl::memory::desc original_b_md({b_k_size_, b_n_size_}, b_type_,
-                                   {b_k_stride_, b_n_stride_});
-
-  prepack_weight(args.b_ptr, original_b_md,
+  prepack_weight(args.b_ptr,
                  create_primitive_desc(
-                     MSizeCacheKey{
-#ifdef VLLM_USE_ACL
-                         // Arm Compute Library (ACL) backend for oneDNN does
-                         // not support runtime
-                         // dimensions, so we set M to a default value
-                         .a_m_size = 128,
-                         .a_m_stride = b_k_size_,
-#else
-                         .a_m_size = DNNL_RUNTIME_DIM_VAL,
-                         .a_m_stride = DNNL_RUNTIME_DIM_VAL,
-#endif
-                         .use_bias = false,
-                         .bias_type = dnnl::memory::data_type::undef},
+                     MSizeCacheKey{.a_m_size = DNNL_RUNTIME_DIM_VAL,
+                                   .a_m_stride = DNNL_RUNTIME_DIM_VAL,
+                                   .use_bias = false,
+                                   .bias_type = dnnl::memory::data_type::undef},
                      true)
                      .weights_desc());
   init_runtime_memory_cache(args);
@@ -457,30 +443,12 @@ void MatMulPrimitiveHandler::execute(ExecArgs& args) {
   c_storage->set_data_handle((void*)args.c_ptr);
   c_mem_desc->dims[0] = args.a_m_size;
 
-#ifndef VLLM_USE_ACL
-  // We do not support in ACL backend of oneDNN, we handle bias by:
-  // 1. copying it into the result tensor
-  // 2. attaching a fused-sum post-op to the matmul primitive
   if (args.use_bias) {
     auto&& [bias_storage, bias_mem_desc] = get_runtime_memory_ptr(2);
     bias_storage->set_data_handle((void*)args.bias_ptr);
   }
-#endif
-  dnnl::matmul matmul = get_matmul_cache(args);
 
-// With ACL backend of oneDNN, the required memory format might change when the
-// source tensor dims change. This does not really happen in practice, so isn't
-// a performance hit, but we need to support it because the API allows for it.
-#ifdef VLLM_USE_ACL
-  auto new_expected_wei_desc =
-      dnnl::matmul::primitive_desc(
-          const_cast<dnnl_primitive_desc_t>(matmul.get_primitive_desc()))
-          .weights_desc();
-  if (new_expected_wei_desc != b_target_mem_desc_) {
-    prepack_weight(memory_cache_[DNNL_ARG_WEIGHTS].get_data_handle(),
-                   b_target_mem_desc_, new_expected_wei_desc);
-  }
-#endif
+  dnnl::matmul matmul = get_matmul_cache(args);
 
   auto&& [scratchpad_storage, scratchpad_mem_desc] = get_runtime_memory_ptr(3);
   scratchpad_storage->set_data_handle(
@@ -516,13 +484,7 @@ dnnl::matmul::primitive_desc MatMulPrimitiveHandler::create_primitive_desc(
   } else {
     a_md = dnnl::memory::desc({key.a_m_size, b_k_size_}, b_type_,
                               {key.a_m_stride, 1});
-#ifdef VLLM_USE_ACL
-    // ACL's backend of oneDNN always expects the weight format to be "any"
-    b_md = dnnl::memory::desc({b_k_size_, b_n_size_}, b_type_,
-                              dnnl::memory::format_tag::any);
-#else
     b_md = b_target_mem_desc_;
-#endif
   }
   dnnl::memory::desc c_md({key.a_m_size, b_n_size_}, c_type_,
                           dnnl::memory::format_tag::ab);
@@ -532,18 +494,8 @@ dnnl::matmul::primitive_desc MatMulPrimitiveHandler::create_primitive_desc(
 
   if (key.use_bias) {
     dnnl::memory::desc bias_md({1, b_n_size_}, key.bias_type, {b_n_size_, 1});
-// Since ACL's matmuls don't support passing a bias_md, we apply the bias
-// through a fused-sum post-op
-#ifdef VLLM_USE_ACL
-    dnnl::post_ops post_ops;
-    post_ops.append_sum();
-    attr.set_post_ops(post_ops);
-    return dnnl::matmul::primitive_desc(default_engine(), a_md, b_md, c_md,
-                                        attr);
-#else
     return dnnl::matmul::primitive_desc(default_engine(), a_md, b_md, bias_md,
                                         c_md, attr);
-#endif
   } else {
     return dnnl::matmul::primitive_desc(default_engine(), a_md, b_md, c_md,
                                         attr);
@@ -559,23 +511,13 @@ void MatMulPrimitiveHandler::init_runtime_memory_cache(const Args& args) {
                    default_engine(), nullptr);
   set_runtime_memory_ptr(1, memory_cache_[DNNL_ARG_DST].get());
 
-// ACL matmuls don't support bias_md, so we don't need these
-#ifndef VLLM_USE_ACL
   memory_cache_[DNNL_ARG_BIAS] =
       dnnl::memory({{b_n_size_}, dnnl::memory::data_type::f32, {1}},
                    default_engine(), nullptr);
   set_runtime_memory_ptr(2, memory_cache_[DNNL_ARG_BIAS].get());
-#endif
+
   memory_cache_[DNNL_ARG_SCRATCHPAD] =
       dnnl::memory({{b_n_size_}, dnnl::memory::data_type::f32, {1}},
                    default_engine(), nullptr);
   set_runtime_memory_ptr(3, memory_cache_[DNNL_ARG_SCRATCHPAD].get());
 }
-
-bool is_onednn_acl_supported() {
-#ifdef VLLM_USE_ACL
-  return true;
-#else
-  return false;
-#endif
-}

csrc/cpu/dnnl_helper.h

@@ -101,7 +101,7 @@ class DNNLMatMulPrimitiveHandler {
  protected:
   DNNLMatMulPrimitiveHandler(const Args& args, dnnl::memory::data_type b_type);
 
-  void prepack_weight(void* original_b_ptr, dnnl::memory::desc original_b_md,
+  void prepack_weight(void* original_b_ptr,
                       dnnl::memory::desc b_target_mem_desc);
 
   void set_runtime_memory_ptr(size_t index, dnnl_memory* memory_ptr);

csrc/cpu/dnnl_kernels.cpp

@@ -527,42 +527,21 @@ void onednn_mm(torch::Tensor& c,        // [M, OC], row-major
   MatMulPrimitiveHandler* ptr =
       reinterpret_cast<MatMulPrimitiveHandler*>(handler);
 
-// ACL matmuls expect contiguous source tensors
-#ifdef VLLM_USE_ACL
-  torch::Tensor a_contig = a.contiguous();
-#endif
-
   MatMulPrimitiveHandler::ExecArgs exec_args;
-
-#ifdef VLLM_USE_ACL
-  exec_args.a_m_size = a_contig.size(0);
-  exec_args.a_m_stride = a_contig.stride(0);
-#else
   exec_args.a_m_size = a.size(0);
   exec_args.a_m_stride = a.stride(0);
-#endif
+
   VLLM_DISPATCH_FLOATING_TYPES(a.scalar_type(), "onednn_mm", [&] {
     if (bias.has_value()) {
       exec_args.use_bias = true;
       exec_args.bias_type = get_dnnl_type<scalar_t>();
-#ifdef VLLM_USE_ACL
-      // ACL matmuls in oneDNN do not support a bias.
-      // We handle a matmul with bias by doing: c = bias; c += matmul(a, b)
-      c.copy_(bias.value());
-#else
       exec_args.bias_ptr = bias->data_ptr<scalar_t>();
-#endif
     } else {
       exec_args.use_bias = false;
       exec_args.bias_type = get_dnnl_type<void>();
       exec_args.bias_ptr = nullptr;
     }
-#ifdef VLLM_USE_ACL
-    exec_args.a_ptr = a_contig.data_ptr<scalar_t>();
-#else
     exec_args.a_ptr = a.data_ptr<scalar_t>();
-
-#endif
     exec_args.c_ptr = c.data_ptr<scalar_t>();
 
     ptr->execute(exec_args);

csrc/cpu/float_convert.hpp

@@ -1,106 +0,0 @@
-
-static float bf16_to_float(uint16_t bf16) {
-  uint32_t bits = static_cast<uint32_t>(bf16) << 16;
-  float fp32;
-  std::memcpy(&fp32, &bits, sizeof(fp32));
-  return fp32;
-}
-
-static uint16_t float_to_bf16(float fp32) {
-  uint32_t bits;
-  std::memcpy(&bits, &fp32, sizeof(fp32));
-  return static_cast<uint16_t>(bits >> 16);
-}
-
-/************************************************
- * Copyright (c) 2015 Princeton Vision Group
- * Licensed under the MIT license.
- * Codes below copied from
- * https://github.com/PrincetonVision/marvin/tree/master/tools/tensorIO_matlab
- *************************************************/
-static uint16_t float_to_fp16(float fp32) {
-  uint16_t fp16;
-
-  unsigned x;
-  unsigned u, remainder, shift, lsb, lsb_s1, lsb_m1;
-  unsigned sign, exponent, mantissa;
-
-  std::memcpy(&x, &fp32, sizeof(fp32));
-  u = (x & 0x7fffffff);
-
-  // Get rid of +NaN/-NaN case first.
-  if (u > 0x7f800000) {
-    fp16 = 0x7fffU;
-    return fp16;
-  }
-
-  sign = ((x >> 16) & 0x8000);
-
-  // Get rid of +Inf/-Inf, +0/-0.
-  if (u > 0x477fefff) {
-    fp16 = sign | 0x7c00U;
-    return fp16;
-  }
-  if (u < 0x33000001) {
-    fp16 = (sign | 0x0000);
-    return fp16;
-  }
-
-  exponent = ((u >> 23) & 0xff);
-  mantissa = (u & 0x7fffff);
-
-  if (exponent > 0x70) {
-    shift = 13;
-    exponent -= 0x70;
-  } else {
-    shift = 0x7e - exponent;
-    exponent = 0;
-    mantissa |= 0x800000;
-  }
-  lsb = (1 << shift);
-  lsb_s1 = (lsb >> 1);
-  lsb_m1 = (lsb - 1);
-
-  // Round to nearest even.
-  remainder = (mantissa & lsb_m1);
-  mantissa >>= shift;
-  if (remainder > lsb_s1 || (remainder == lsb_s1 && (mantissa & 0x1))) {
-    ++mantissa;
-    if (!(mantissa & 0x3ff)) {
-      ++exponent;
-      mantissa = 0;
-    }
-  }
-
-  fp16 = (sign | (exponent << 10) | mantissa);
-
-  return fp16;
-}
-
-static float fp16_to_float(uint16_t fp16) {
-  unsigned sign = ((fp16 >> 15) & 1);
-  unsigned exponent = ((fp16 >> 10) & 0x1f);
-  unsigned mantissa = ((fp16 & 0x3ff) << 13);
-  int temp;
-  float fp32;
-  if (exponent == 0x1f) { /* NaN or Inf */
-    mantissa = (mantissa ? (sign = 0, 0x7fffff) : 0);
-    exponent = 0xff;
-  } else if (!exponent) { /* Denorm or Zero */
-    if (mantissa) {
-      unsigned int msb;
-      exponent = 0x71;
-      do {
-        msb = (mantissa & 0x400000);
-        mantissa <<= 1; /* normalize */
-        --exponent;
-      } while (!msb);
-      mantissa &= 0x7fffff; /* 1.mantissa is implicit */
-    }
-  } else {
-    exponent += 0x70;
-  }
-  temp = ((sign << 31) | (exponent << 23) | mantissa);
-  std::memcpy(&fp32, &temp, sizeof(temp));
-  return fp32;
-}

csrc/cpu/torch_bindings.cpp

@@ -27,8 +27,6 @@ int64_t create_onednn_mm_handler(const torch::Tensor& b,
 void onednn_mm(torch::Tensor& c, const torch::Tensor& a,
                const std::optional<torch::Tensor>& bias, int64_t handler);
 
-bool is_onednn_acl_supported();
-
 void mla_decode_kvcache(torch::Tensor& out, torch::Tensor& query,
                         torch::Tensor& kv_cache, double scale,
                         torch::Tensor& block_tables, torch::Tensor& seq_lens);
@@ -90,18 +88,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "    int tp_rank, int blocksparse_local_blocks,"
       "    int blocksparse_vert_stride, int blocksparse_block_size,"
       "    int blocksparse_head_sliding_step) -> ()");
-
   ops.impl("paged_attention_v1", torch::kCPU, &paged_attention_v1);
 
-  ops.def(
-      "dynamic_4bit_int_moe("
-      "Tensor x, Tensor topk_ids, Tensor topk_weights,"
-      "Tensor w13_packed, Tensor w2_packed, int H, int I, int I2,"
-      "int group_size, bool apply_router_weight_on_input, int activation_kind"
-      ") -> Tensor");
-
-  ops.impl("dynamic_4bit_int_moe", torch::kCPU, &dynamic_4bit_int_moe_cpu);
-
   // PagedAttention V2.
   ops.def(
       "paged_attention_v2("
@@ -183,9 +171,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "int handler) -> ()");
   ops.impl("onednn_mm", torch::kCPU, &onednn_mm);
 
-  // Check if oneDNN was built with ACL backend
-  ops.def("is_onednn_acl_supported() -> bool", &is_onednn_acl_supported);
-
   // Create oneDNN W8A8 handler
   ops.def(
       "create_onednn_scaled_mm_handler(Tensor b, Tensor b_scales, ScalarType "

csrc/cub_helpers.h

@@ -12,7 +12,6 @@ using CubMaxOp = cub::Max;
   #endif  // CUB_VERSION
 #else
   #include <hipcub/hipcub.hpp>
-namespace cub = hipcub;
-using CubAddOp = hipcub::Sum;
-using CubMaxOp = hipcub::Max;
+using CubAddOp = cub::Sum;
+using CubMaxOp = cub::Max;
 #endif  // USE_ROCM

csrc/cutlass_extensions/vllm_cutlass_library_extension.py

@@ -27,7 +27,7 @@ VLLMDataTypeNames: dict[Union[VLLMDataType, DataType], str] = {
     **{
         VLLMDataType.u4b8: "u4b8",
         VLLMDataType.u8b128: "u8b128",
-    },
+    }
 }
 
 VLLMDataTypeTag: dict[Union[VLLMDataType, DataType], str] = {
@@ -35,7 +35,7 @@ VLLMDataTypeTag: dict[Union[VLLMDataType, DataType], str] = {
     **{
         VLLMDataType.u4b8: "cutlass::vllm_uint4b8_t",
         VLLMDataType.u8b128: "cutlass::vllm_uint8b128_t",
-    },
+    }
 }
 
 VLLMDataTypeSize: dict[Union[VLLMDataType, DataType], int] = {
@@ -43,7 +43,7 @@ VLLMDataTypeSize: dict[Union[VLLMDataType, DataType], int] = {
     **{
         VLLMDataType.u4b8: 4,
         VLLMDataType.u8b128: 8,
-    },
+    }
 }
 
 VLLMDataTypeVLLMScalarTypeTag: dict[Union[VLLMDataType, DataType], str] = {
@@ -67,13 +67,15 @@ VLLMDataTypeTorchDataTypeTag: dict[Union[VLLMDataType, DataType], str] = {
     DataType.f32: "at::ScalarType::Float",
 }
 
-VLLMKernelScheduleTag: dict[
-    Union[MixedInputKernelScheduleType, KernelScheduleType], str
-] = {
-    **KernelScheduleTag,  # type: ignore
-    **{
-        MixedInputKernelScheduleType.TmaWarpSpecialized: "cutlass::gemm::KernelTmaWarpSpecialized",  # noqa: E501
-        MixedInputKernelScheduleType.TmaWarpSpecializedPingpong: "cutlass::gemm::KernelTmaWarpSpecializedPingpong",  # noqa: E501
-        MixedInputKernelScheduleType.TmaWarpSpecializedCooperative: "cutlass::gemm::KernelTmaWarpSpecializedCooperative",  # noqa: E501
-    },
-}
+VLLMKernelScheduleTag: dict[Union[
+    MixedInputKernelScheduleType, KernelScheduleType], str] = {
+        **KernelScheduleTag,  # type: ignore
+        **{
+            MixedInputKernelScheduleType.TmaWarpSpecialized:
+            "cutlass::gemm::KernelTmaWarpSpecialized",
+            MixedInputKernelScheduleType.TmaWarpSpecializedPingpong:
+            "cutlass::gemm::KernelTmaWarpSpecializedPingpong",
+            MixedInputKernelScheduleType.TmaWarpSpecializedCooperative:
+            "cutlass::gemm::KernelTmaWarpSpecializedCooperative",
+        }
+    }

csrc/launch_bounds_utils.h

@@ -1,64 +0,0 @@
-#pragma once
-
-#include <cuda_runtime_api.h>
-#include <algorithm>
-
-// maximum blocks per SM cap
-#ifndef VLLM_LAUNCH_BLOCKS_CAP
-  #define VLLM_LAUNCH_BLOCKS_CAP 4
-#endif
-
-// Compile-time estimate of max threads per SM for launch bounds.
-// Families: 1024, 1536, 2048 threads/SM.
-#ifndef VLLM_MAX_THREADS_PER_SM
-  #ifdef __CUDA_ARCH__
-
-    /* 1024 thr/SM: Turing (sm_75) */
-    #if (__CUDA_ARCH__ == 750)
-      #define VLLM_MAX_THREADS_PER_SM 1024
-
-    /* 1536 thr/SM: Ampere GA10x (sm_86/87), Ada (sm_89),
-        GB20x consumer (sm_120/121), Thor (sm_101 or sm_110) */
-    #elif (__CUDA_ARCH__ == 860) || (__CUDA_ARCH__ == 870) || \
-        (__CUDA_ARCH__ == 890) || (__CUDA_ARCH__ == 1010) ||  \
-        (__CUDA_ARCH__ == 1100) || (__CUDA_ARCH__ == 1200) || \
-        (__CUDA_ARCH__ == 1210)
-      #define VLLM_MAX_THREADS_PER_SM 1536
-
-    /* 2048 thr/SM: Volta (sm_70/72), Ampere GA100 (sm_80),
-        Hopper (sm_90), Blackwell (sm_100/103) */
-    #elif (__CUDA_ARCH__ == 700) || (__CUDA_ARCH__ == 720) || \
-        (__CUDA_ARCH__ == 800) || (__CUDA_ARCH__ == 900) ||   \
-        (__CUDA_ARCH__ == 1000) || (__CUDA_ARCH__ == 1030)
-      #define VLLM_MAX_THREADS_PER_SM 2048
-
-    /* Fallback: use 2048 for unknown future CCs */
-    #else
-      #define VLLM_MAX_THREADS_PER_SM 2048
-    #endif
-
-  #else
-  /* Host pass (no __CUDA_ARCH__): neutral default */
-    #define VLLM_MAX_THREADS_PER_SM 2048
-  #endif
-#endif
-
-// compute the number of blocks per SM to request in __launch_bounds__
-#define VLLM_BLOCKS_DIV(VAL) (VLLM_MAX_THREADS_PER_SM / (VAL))
-#define VLLM_CLAMP_BLOCKS_PER_SM(VAL) \
-  (((VAL) <= 0)                       \
-       ? 1                            \
-       : (((VAL) < VLLM_LAUNCH_BLOCKS_CAP) ? (VAL) : VLLM_LAUNCH_BLOCKS_CAP))
-#define VLLM_BLOCKS_PER_SM(BLOCK_THREADS) \
-  VLLM_CLAMP_BLOCKS_PER_SM(VLLM_BLOCKS_DIV(BLOCK_THREADS))
-
-// runtime-time helper to compute blocks/SM
-static inline int vllm_runtime_blocks_per_sm(int block_threads) {
-  int device = -1;
-  cudaGetDevice(&device);
-  int max_threads_per_sm = VLLM_MAX_THREADS_PER_SM;
-  cudaDeviceGetAttribute(&max_threads_per_sm,
-                         cudaDevAttrMaxThreadsPerMultiProcessor, device);
-  int blocks = (block_threads > 0) ? (max_threads_per_sm / block_threads) : 1;
-  return VLLM_CLAMP_BLOCKS_PER_SM(blocks);
-}

csrc/layernorm_kernels.cu

@@ -1,7 +1,6 @@
 #include "type_convert.cuh"
 #include "dispatch_utils.h"
 #include "cub_helpers.h"
-#include "core/batch_invariant.hpp"
 
 #include <torch/cuda.h>
 #include <c10/cuda/CUDAGuard.h>
@@ -414,9 +413,7 @@ void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
                           wt_ptr % req_alignment_bytes == 0;
   bool offsets_are_multiple_of_vector_width =
       hidden_size % vector_width == 0 && input_stride % vector_width == 0;
-  bool batch_invariant_launch = vllm::vllm_kernel_override_batch_invariant();
-  if (ptrs_are_aligned && offsets_are_multiple_of_vector_width &&
-      !batch_invariant_launch) {
+  if (ptrs_are_aligned && offsets_are_multiple_of_vector_width) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
     LAUNCH_FUSED_ADD_RMS_NORM(0);
@@ -462,8 +459,7 @@ void poly_norm(torch::Tensor& out,     // [..., hidden_size]
   auto inp_ptr = reinterpret_cast<std::uintptr_t>(input.data_ptr());
   auto out_ptr = reinterpret_cast<std::uintptr_t>(out.data_ptr());
   bool ptrs_are_aligned = inp_ptr % 16 == 0 && out_ptr % 16 == 0;
-  bool batch_invariant_launch = vllm::vllm_kernel_override_batch_invariant();
-  if (ptrs_are_aligned && hidden_size % 8 == 0 && !batch_invariant_launch) {
+  if (ptrs_are_aligned && hidden_size % 8 == 0) {
     LAUNCH_FUSED_POLY_NORM(8);
   } else {
     LAUNCH_FUSED_POLY_NORM(0);

csrc/layernorm_quant_kernels.cu

@@ -6,10 +6,9 @@
  */
 
 #include "type_convert.cuh"
-#include "quantization/w8a8/fp8/common.cuh"
+#include "quantization/fp8/common.cuh"
 #include "dispatch_utils.h"
 #include "cub_helpers.h"
-#include "core/batch_invariant.hpp"
 
 #include <torch/cuda.h>
 #include <c10/cuda/CUDAGuard.h>
@@ -241,9 +240,7 @@ void fused_add_rms_norm_static_fp8_quant(
   auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight.data_ptr());
   bool ptrs_are_aligned =
       inp_ptr % 16 == 0 && res_ptr % 16 == 0 && wt_ptr % 16 == 0;
-  bool batch_invariant_launch = vllm::vllm_kernel_override_batch_invariant();
-  if (ptrs_are_aligned && hidden_size % 8 == 0 && input_stride % 8 == 0 &&
-      !batch_invariant_launch) {
+  if (ptrs_are_aligned && hidden_size % 8 == 0 && input_stride % 8 == 0) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
     LAUNCH_FUSED_ADD_RMS_NORM(0);

csrc/moe/dynamic_4bit_int_moe_cpu.cpp

@@ -1,156 +0,0 @@
-#include <ATen/ATen.h>
-#include <ATen/Parallel.h>
-#include <torch/all.h>
-
-// _dyn_quant_matmul_4bit is only available on AArch64.
-#if defined(__aarch64__)
-  #include <ATen/ops/_dyn_quant_matmul_4bit.h>
-#endif
-
-inline torch::Tensor mm(const torch::Tensor& a, const torch::Tensor& packed_w,
-                        int64_t group_size_eff, int64_t in_features,
-                        int64_t out_features) {
-#if defined(__aarch64__)
-  return at::_ops::_dyn_quant_matmul_4bit::call(a, packed_w, group_size_eff,
-                                                in_features, out_features);
-#else
-  TORCH_CHECK(false,
-              "dynamic 4-bit int MoE path requires AArch64 (ARM64); "
-              "_dyn_quant_matmul_4bit is unavailable on this architecture");
-  return {};
-#endif
-}
-
-enum ActivationKind : int64_t {
-  SwiGLU_Gu = 0,  // act = SiLU(g) * u
-  SwiGLUOAI = 1,  // act = SiLU(u) * g
-  SiLU = 2        // SiLU
-};
-
-torch::Tensor dynamic_4bit_int_moe_cpu(
-    torch::Tensor x, torch::Tensor topk_ids, torch::Tensor topk_weights,
-    torch::Tensor w13_packed, torch::Tensor w2_packed, int64_t H, int64_t I,
-    int64_t I2, int64_t group_size, bool apply_router_weight_on_input,
-    int64_t activation_kind) {
-  TORCH_CHECK(x.dim() == 2, "x must be 2D");
-  TORCH_CHECK(topk_ids.dim() == 2 && topk_weights.dim() == 2,
-              "topk tensors must be [T, K]");
-  TORCH_CHECK(
-      w13_packed.size(0) == w2_packed.size(0),
-      "w13_packed and w2_packed must have same number of experts in dim 0");
-  TORCH_CHECK(I2 == 2 * I, "I2 must equal 2*I");
-
-  const int64_t T = x.size(0);
-  const int64_t K = topk_ids.size(1);
-  const int64_t E = w13_packed.size(0);
-  const int64_t N = T * K;
-
-  auto x_c = x.contiguous();
-  auto ids_c = topk_ids.contiguous();
-  auto gates_c = topk_weights.to(at::kFloat).contiguous();
-
-  // bucketing tokens -> experts
-  c10::SmallVector<int64_t, 64> counts(
-      E, 0);  // Small vector uses stack allocation
-  {
-    const auto* ids_ptr = ids_c.data_ptr<int64_t>();
-    for (int64_t i = 0; i < N; ++i) {
-      const int64_t e_id = ids_ptr[i];
-      TORCH_CHECK(0 <= e_id && e_id < E, "expert id out of range");
-      counts[e_id]++;
-    }
-  }
-  c10::SmallVector<int64_t, 65> offsets(E + 1, 0);  // ( E +1 )
-  for (int64_t e = 0; e < E; ++e) offsets[e + 1] = offsets[e] + counts[e];
-
-  auto expert_tokens = at::empty({offsets[E]}, ids_c.options());
-  auto expert_gates = at::empty({offsets[E]}, gates_c.options());
-  {
-    c10::SmallVector<int64_t, 64> cursor(E, 0);
-    const auto* ids_ptr = ids_c.data_ptr<int64_t>();
-    const auto* gts_ptr = gates_c.data_ptr<float>();
-    auto* tok_ptr = expert_tokens.data_ptr<int64_t>();
-    auto* gate_ptr = expert_gates.data_ptr<float>();
-
-    for (int64_t t = 0; t < T; ++t) {
-      const int64_t base = t * K;
-      for (int64_t k = 0; k < K; ++k) {
-        const int64_t idx = base + k;
-        const int64_t e = ids_ptr[idx];
-        const int64_t p = offsets[e] + (cursor[e]++);
-        tok_ptr[p] = t;
-        gate_ptr[p] = gts_ptr[idx];
-      }
-    }
-  }
-
-  const int64_t g_eff_13 = (group_size != -1) ? group_size : H;
-  const int64_t g_eff_2 = (group_size != -1) ? group_size : I;
-
-  // Per-expert outputs filled in parallel
-  std::vector<torch::Tensor> y_list(E);
-  y_list.resize(E);
-
-  at::parallel_for(0, E, 1, [&](int64_t e_begin, int64_t e_end) {
-    for (int64_t e = e_begin; e < e_end; ++e) {
-      const int64_t te = counts[e];
-      if (te == 0) {
-        y_list[e] = at::empty({0, H}, x_c.options());
-        continue;
-      }
-
-      const int64_t start = offsets[e];
-
-      auto sel_tokens =
-          expert_tokens.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);
-      auto gates_e =
-          expert_gates.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);
-
-      auto x_e = x_c.index_select(/*dim=*/0, sel_tokens);
-
-      if (apply_router_weight_on_input) {
-        x_e = x_e.mul(gates_e.unsqueeze(1));
-      }
-
-      auto w13_e = w13_packed.select(/*dim=*/0, e);
-      auto w2_e = w2_packed.select(/*dim=*/0, e);
-
-      // W13
-      auto y13 =
-          mm(x_e, w13_e, g_eff_13, /*in_features=*/H, /*out_features=*/I2);
-
-      auto g_part = y13.narrow(/*dim=*/1, /*start=*/0, /*length=*/I);
-      auto u_part = y13.narrow(/*dim=*/1, /*start=*/I, /*length=*/I);
-
-      torch::Tensor act;
-      if (activation_kind == ActivationKind::SwiGLUOAI) {  // SwiGLUOAI
-        constexpr double kAlpha = 1.702;                   // GPT-OSS default
-        constexpr double kLimit = 7.0;                     // GPT-OSS default
-        auto gate_c = at::clamp_max(g_part, kLimit);
-        auto up_c = at::clamp(u_part, -kLimit, kLimit);
-        auto glu = gate_c.mul(at::sigmoid(gate_c.mul(kAlpha)));
-        act = up_c.add(1.0).mul(glu);
-      } else {  // SiLU , SwiGLU_GU, vLLM maps silu to SiluAndMul()
-        act = at::silu(g_part).mul(u_part);
-      }
-
-      // W2
-      auto y = mm(act, w2_e, g_eff_2, /*in_features=*/I, /*out_features=*/H);
-
-      if (!apply_router_weight_on_input) {
-        y = y.mul(gates_e.unsqueeze(1));
-      }
-
-      // Store per-expert result
-      y_list[e] = y;
-    }
-  });
-
-  // Concatenate all expert outputs to match expert_tokens order
-  auto Y_all = at::cat(y_list, /*dim=*/0);
-  auto out = at::zeros({T, H}, x.options());
-  out =
-      at::index_add(out, /*dim=*/0, /*index=*/expert_tokens, /*source=*/Y_all);
-
-  return out;
-}

csrc/moe/grouped_topk_kernels.cu

@@ -418,15 +418,6 @@ __device__ inline T neg_inf() {
   return cuda_cast<T, float>(-cuda::std::numeric_limits<float>::infinity());
 }
 
-template <typename T>
-__device__ inline bool is_finite(const T val) {
-#if (__CUDACC_VER_MAJOR__ * 10000 + __CUDACC_VER_MINOR__ * 100 >= 120800)
-  return cuda::std::isfinite(val);
-#else
-  return isfinite(cuda_cast<float, T>(val));
-#endif
-}
-
 template <typename T>
 __device__ void topk_with_k2(T* output, T const* input,
                              cg::thread_block_tile<32> const& tile,
@@ -542,7 +533,7 @@ __global__ void group_idx_and_topk_idx_kernel(
     // calculate group_idx
     int32_t target_num_min = WARP_SIZE - n_group + topk_group;
     // The check is necessary to avoid abnormal input
-    if (lane_id < n_group && is_finite(group_scores[lane_id])) {
+    if (lane_id < n_group && cuda::std::isfinite(group_scores[lane_id])) {
       value = group_scores[lane_id];
     }
 
@@ -577,10 +568,11 @@ __global__ void group_idx_and_topk_idx_kernel(
         int32_t offset = i_group * num_experts_per_group;
         for (int32_t i = lane_id; i < align_num_experts_per_group;
              i += WARP_SIZE) {
-          T candidates = (i < num_experts_per_group) &&
-                                 is_finite(scores_with_bias[offset + i])
-                             ? scores_with_bias[offset + i]
-                             : neg_inf<T>();
+          T candidates =
+              (i < num_experts_per_group) &&
+                      cuda::std::isfinite(scores_with_bias[offset + i])
+                  ? scores_with_bias[offset + i]
+                  : neg_inf<T>();
           queue.add(candidates, offset + i);
         }
         if (group_scores[i_group] == topk_group_value) {

csrc/moe/marlin_moe_wna16/generate_kernels.py

@@ -17,30 +17,25 @@ FILE_HEAD = """
 namespace MARLIN_NAMESPACE_NAME {
 """.strip()
 
-TEMPLATE = (
-    "template __global__ void Marlin<"
-    "{{scalar_t}}, "
-    "{{w_type_id}}, "
-    "{{s_type_id}}, "
-    "{{threads}}, "
-    "{{thread_m_blocks}}, "
-    "{{thread_n_blocks}}, "
-    "{{thread_k_blocks}}, "
-    "{{'true' if m_block_size_8 else 'false'}}, "
-    "{{stages}}, "
-    "{{group_blocks}}, "
-    "{{'true' if is_zp_float else 'false'}}>"
-    "( MARLIN_KERNEL_PARAMS );"
-)
+TEMPLATE = ("template __global__ void Marlin<"
+            "{{scalar_t}}, "
+            "{{w_type_id}}, "
+            "{{s_type_id}}, "
+            "{{threads}}, "
+            "{{thread_m_blocks}}, "
+            "{{thread_n_blocks}}, "
+            "{{thread_k_blocks}}, "
+            "{{'true' if m_block_size_8 else 'false'}}, "
+            "{{stages}}, "
+            "{{group_blocks}}, "
+            "{{'true' if is_zp_float else 'false'}}>"
+            "( MARLIN_KERNEL_PARAMS );")
 
 # int8 with zero point case (vllm::kU8) is also supported,
 # we don't add it to reduce wheel size.
 SCALAR_TYPES = [
-    "vllm::kU4",
-    "vllm::kU4B8",
-    "vllm::kU8B128",
-    "vllm::kFE4M3fn",
-    "vllm::kFE2M1f",
+    "vllm::kU4", "vllm::kU4B8", "vllm::kU8B128", "vllm::kFE4M3fn",
+    "vllm::kFE2M1f"
 ]
 THREAD_CONFIGS = [(128, 128, 256), (64, 256, 256), (64, 128, 128)]
 
@@ -63,12 +58,11 @@ def generate_new_kernels():
         all_template_str_list = []
 
         for group_blocks, m_blocks, thread_configs in itertools.product(
-            GROUP_BLOCKS, THREAD_M_BLOCKS, THREAD_CONFIGS
-        ):
+                GROUP_BLOCKS, THREAD_M_BLOCKS, THREAD_CONFIGS):
+
             # act order case only support gptq-int4 and gptq-int8
             if group_blocks == 0 and scalar_type not in [
-                "vllm::kU4B8",
-                "vllm::kU8B128",
+                    "vllm::kU4B8", "vllm::kU8B128"
             ]:
                 continue
             if thread_configs[2] == 256:

csrc/moe/moe_align_sum_kernels.cu

@@ -44,9 +44,6 @@ __global__ void moe_align_block_size_kernel(
 
   for (size_t i = tid; i < numel; i += stride) {
     int expert_id = topk_ids[i];
-    if (expert_id >= num_experts) {
-      continue;
-    }
     int warp_idx = expert_id / experts_per_warp;
     int expert_offset = expert_id % experts_per_warp;
     atomicAdd(&shared_counts[warp_idx * experts_per_warp + expert_offset], 1);
@@ -98,15 +95,12 @@ template <typename scalar_t>
 __global__ void count_and_sort_expert_tokens_kernel(
     const scalar_t* __restrict__ topk_ids,
     int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ cumsum_buffer,
-    size_t numel, int32_t num_experts) {
+    size_t numel) {
   const size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
   const size_t stride = blockDim.x * gridDim.x;
 
   for (size_t i = tid; i < numel; i += stride) {
     int32_t expert_id = topk_ids[i];
-    if (expert_id >= num_experts) {
-      continue;
-    }
     int32_t rank_post_pad = atomicAdd(&cumsum_buffer[expert_id], 1);
     sorted_token_ids[rank_post_pad] = i;
   }
@@ -275,7 +269,7 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
           sort_kernel<<<actual_blocks, block_threads, 0, stream>>>(
               topk_ids.data_ptr<scalar_t>(),
               sorted_token_ids.data_ptr<int32_t>(),
-              cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel(), num_experts);
+              cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel());
         }
       });
 }

csrc/moe/topk_softmax_kernels.cu

@@ -21,7 +21,6 @@
 #include <c10/cuda/CUDAGuard.h>
 #include "../cuda_compat.h"
 #include "../cub_helpers.h"
-#include "../core/batch_invariant.hpp"
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -406,8 +405,7 @@ void topkGatingSoftmaxLauncherHelper(const float* input, const bool* finished, f
     using Constants = detail::TopkConstants<EXPERTS, BYTES_PER_LDG, WARP_SIZE_PARAM>;
     static constexpr int VPT = Constants::VPT;
     static constexpr int ROWS_PER_WARP = Constants::ROWS_PER_WARP;
-    const bool batch_invariant_launch = vllm::vllm_kernel_override_batch_invariant();
-    const int num_warps = batch_invariant_launch ? 32 : (num_rows + ROWS_PER_WARP - 1) / ROWS_PER_WARP;
+    const int num_warps = (num_rows + ROWS_PER_WARP - 1) / ROWS_PER_WARP;
     const int num_blocks = (num_warps + WARPS_PER_TB - 1) / WARPS_PER_TB;
 
     dim3 block_dim(WARP_SIZE_PARAM, WARPS_PER_TB);

csrc/ops.h

@@ -100,11 +100,6 @@ void apply_repetition_penalties_(torch::Tensor& logits,
                                  const torch::Tensor& output_mask,
                                  const torch::Tensor& repetition_penalties);
 
-void top_k_per_row(const torch::Tensor& logits, const torch::Tensor& rowStarts,
-                   const torch::Tensor& rowEnds, torch::Tensor& indices,
-                   torch::Tensor& values, int64_t numRows, int64_t stride0,
-                   int64_t stride1);
-
 void rms_norm_static_fp8_quant(torch::Tensor& out, torch::Tensor& input,
                                torch::Tensor& weight, torch::Tensor& scale,
                                double epsilon);
@@ -138,12 +133,12 @@ void silu_and_mul_nvfp4_quant(torch::Tensor& out,
                               torch::Tensor& input,
                               torch::Tensor& input_global_scale);
 #endif
-void persistent_masked_m_silu_mul_quant(
+void silu_mul_fp8_quant_deep_gemm_cuda(
     const at::Tensor& input,   // (E, T, 2*H)
     const at::Tensor& counts,  // (E)
     at::Tensor& y_q,           // (E, T, H) [OUT]
     at::Tensor& y_s,           // (E, T, H//group_size) [OUT]
-    bool use_ue8m0);
+    int64_t group_size, bool use_ue8m0, int64_t num_parallel_tokens);
 
 void mul_and_silu(torch::Tensor& out, torch::Tensor& input);
 
@@ -333,12 +328,6 @@ void selective_scan_fwd(const torch::Tensor& u, const torch::Tensor& delta,
                         const std::optional<torch::Tensor>& has_initial_state,
                         const torch::Tensor& ssm_states, int64_t pad_slot_id);
 
-torch::Tensor dynamic_4bit_int_moe_cpu(
-    torch::Tensor x, torch::Tensor topk_ids, torch::Tensor topk_weights,
-    torch::Tensor w13_packed, torch::Tensor w2_packed, int64_t H, int64_t I,
-    int64_t I2, int64_t group_size, bool apply_router_weight_on_input,
-    int64_t activation_kind);
-
 using fptr_t = int64_t;
 fptr_t init_custom_ar(const std::vector<int64_t>& fake_ipc_ptrs,
                       torch::Tensor& rank_data, int64_t rank,

csrc/quantization/activation_kernels.cu

@@ -7,7 +7,7 @@
 #include "../cuda_compat.h"
 #include "dispatch_utils.h"
 
-#include "quantization/w8a8/fp8/common.cuh"
+#include "quantization/fp8/common.cuh"
 
 #include <c10/util/Float8_e4m3fn.h>
 
@@ -23,14 +23,9 @@
 typedef __hip_bfloat162 __nv_bfloat162;
 typedef __hip_bfloat16 __nv_bfloat16;
 typedef __hip_bfloat16_raw __nv_bfloat16_raw;
-  #if defined(HIP_FP8_TYPE_OCP)
+
 typedef __hip_fp8_e4m3 __nv_fp8_e4m3;
 typedef __hip_fp8x4_e4m3 __nv_fp8x4_e4m3;
-  #else
-// ROCm 6.2 fallback: only *_fnuz types exist
-typedef __hip_fp8_e4m3_fnuz __nv_fp8_e4m3;
-typedef __hip_fp8x4_e4m3_fnuz __nv_fp8x4_e4m3;
-  #endif
 #endif
 
 #include "core/registration.h"
@@ -114,22 +109,13 @@ __global__ void act_and_mul_quant_kernel(
 }
 
 __device__ __forceinline__ float silu(float x) {
-  return __fdividef(x, (1.f + expf(-x)));
+  return (__fdividef(x, (1.f + expf(-x))));
 }
 
 __device__ __forceinline__ float2 silu2(float2 x) {
   return make_float2(silu(x.x), silu(x.y));
 }
 
-__device__ __forceinline__ __nv_bfloat162 silu2_v2(float2 x) {
-#ifndef USE_ROCM
-  return make_bfloat162(__float2bfloat16_rn(silu(x.x)),
-                        __float2bfloat16_rn(silu(x.y)));
-#else
-  return __float22bfloat162_rn(make_float2(silu(x.x), silu(x.y)));
-#endif
-}
-
 #ifndef USE_ROCM
 __device__ __forceinline__ float warp_max(float v) {
   static constexpr unsigned FULL_MASK = 0xffffffffu;
@@ -232,308 +218,224 @@ constexpr __nv_bfloat16 get_fp8_min() {
     return __nv_bfloat16(__nv_bfloat16_raw{.x = 50032});
   }
 }
-
-template <typename Idx_t>
-__device__ __forceinline__ int warp_expert_search(
-    int idx, int n, const Idx_t* __restrict__ input, Idx_t val) {
-  const Idx_t* input_ptr = input + idx;
-  int base_offset = 0;
-
-  for (;;) {
-    bool move_on = (idx < n && *input_ptr <= val);
-
-    unsigned mask = __ballot_sync(0xffffffff, move_on);
-
-    if (mask != 0xffffffffu) {
-      int last_lane = 31 - __clz(mask);
-      return base_offset + last_lane;
-    }
-
-    input_ptr += 32;
-    base_offset += 32;
-    idx += 32;
-  }
-}
-
-template <int num_parallel_tokens>
-__device__ __forceinline__ void token_bounds(int32_t n_tokens,
-                                             int32_t worker_id,
-                                             int32_t& n_tokens_lower,
-                                             int32_t& n_tokens_upper) {
-  if (n_tokens < num_parallel_tokens && worker_id < n_tokens) {
-    if (worker_id >= num_parallel_tokens) return;
-    n_tokens_lower = worker_id;
-    n_tokens_upper = worker_id + 1;
-  } else {
-    int32_t chunk_size = n_tokens / num_parallel_tokens;
-    int32_t residual = n_tokens - chunk_size * num_parallel_tokens;
-    auto calc_id = [&](int32_t id) {
-      if (id < residual)
-        return min(n_tokens, id * (chunk_size + 1));
-      else
-        return min(n_tokens, id * chunk_size + residual);
-    };
-    n_tokens_lower = calc_id(worker_id);
-    n_tokens_upper = calc_id(worker_id + 1);
-  }
-}
-
-template <int BLOCK_COUNT, int SMEM_SIZE_BYTES_Y, typename fp8_type,
-          int THREADS, typename Idx_t, bool USE_UE8M0, int GROUP_SIZE = 128,
+#ifndef USE_ROCM
+template <typename fp8_type, int32_t NUM_WARPS, typename Idx_t,
+          int NUM_PARALLEL_TOKENS, bool USE_UE8M0, int GROUP_SIZE = 128,
           int NUM_STAGES = 3>
 __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
     const __nv_bfloat16* __restrict__ _input, fp8_type* __restrict__ _y_q,
-    float* __restrict__ _y_s, const int32_t* __restrict__ tokens_per_expert,
+    float* __restrict__ _y_s, const int32_t* __restrict__ counts,
+
     // sizes
-    Idx_t E, Idx_t T, Idx_t H,
+    int H, int G,
+
     // strides (in elements)
     Idx_t stride_i_e, Idx_t stride_i_t, Idx_t stride_i_h, Idx_t stride_yq_e,
     Idx_t stride_yq_t, Idx_t stride_yq_h, Idx_t stride_ys_e, Idx_t stride_ys_t,
     Idx_t stride_ys_g, Idx_t stride_counts_e) {
-#ifndef USE_ROCM
-  static constexpr int NUM_WARPS = THREADS / WARP_SIZE;
-
-  static constexpr int LOAD_STAGE_SIZE = 2 * GROUP_SIZE / 8;
-  static constexpr int LOAD_STAGE_MOD = NUM_STAGES * LOAD_STAGE_SIZE;
-
-  static constexpr int COMPUTE_STAGE_SIZE = 2 * GROUP_SIZE / 4;
-  static constexpr int COMPUTE_STAGE_MOD = COMPUTE_STAGE_SIZE * NUM_STAGES;
-
-  extern __shared__ __align__(16) __int128_t smem_128[];
-
-  int* s_expert_offsets =
-      reinterpret_cast<int*>(smem_128 + (SMEM_SIZE_BYTES_Y / 16));
-
   static constexpr __nv_bfloat16 fp8_min = get_fp8_min<fp8_type>();
   static constexpr __nv_bfloat16 fp8_max = get_fp8_max<fp8_type>();
-  // We assign EPS with it's 16-bit unsigned counterpart to allow constexpr.
+  // We assign EPS with its 16-bit unsigned counterpart to allow constexpr.
   static constexpr __nv_bfloat16 EPS = (__nv_bfloat16_raw{.x = 11996});
-  int tid = threadIdx.x;
-  int warp_id = tid >> 5;
-  int lane_id = tid & 0x1f;
 
-  int running_sum{};
-  if (!warp_id) {
-    for (int i = 0; i < E; i += WARP_SIZE) {
-      bool valid = (i + threadIdx.x) < E;
-      int value =
-          (valid ? tokens_per_expert[i + threadIdx.x * stride_counts_e] : 0) +
-          (!lane_id ? running_sum : 0);
+  // We pack 8 16-bit bfloat16 values into a 128-bit __int128_t.
+  static constexpr int32_t BFLOAT16_PER_GROUP = 8;
 
-      for (int offset = 1; offset < 32; offset *= 2) {
-        int n = __shfl_up_sync(0xFFFFFFFFu, value, offset);
-        if (lane_id >= offset) value += n;
-      }
+  // We split the shared memory in half, corresponding to gate and up matrices:
+  // [...gate_i, ...up_i]  where 0 <= i < stages.
+  static constexpr int32_t S_NUM_128 =
+      2u * (GROUP_SIZE / BFLOAT16_PER_GROUP) * NUM_WARPS * NUM_STAGES;
+  static constexpr auto THREAD_COUNT = NUM_WARPS * WARP_SIZE;
+  static constexpr int HALF_THREAD_COUNT = THREAD_COUNT / 2;
+  static constexpr int32_t S_NUM_64 = S_NUM_128 * 2;
+  __shared__ __int128_t __align__(16) s_buff_128[S_NUM_128];
 
-      if (valid) {
-        s_expert_offsets[i + threadIdx.x + 1] = value;
-      }
+  const int32_t tid = threadIdx.x;
+  const int32_t warp_id = tid / WARP_SIZE;
+  const int32_t lane_id = tid % WARP_SIZE;
 
-      running_sum = __shfl_sync(0xFFFFFFFFu, value, WARP_SIZE - 1);
-    }
+  auto s_buff_compute_32 = reinterpret_cast<__nv_bfloat162*>(s_buff_128);
 
-    if (!lane_id) {
-      s_expert_offsets[0] = 0;
-    }
+  // block handles one (expert e, group g)
+  int32_t pid = blockIdx.x;
+  int32_t e = pid / G;
+  int32_t g = pid % G;
+
+  const int32_t n_tokens = counts[e * stride_counts_e];
+
+  if (!n_tokens) {
+    return;  // Exit ASAP.
   }
 
-  __syncthreads();
+  const Idx_t stride_i_t_128 = stride_i_t / 8u;
 
-  int32_t total_tokens = s_expert_offsets[E];
+  int32_t n_tokens_lower, n_tokens_upper;
 
-  const int warp_position_yq = warp_id * (H / NUM_WARPS);
-  const int warp_position_scales = warp_id * (H / (GROUP_SIZE * NUM_WARPS));
-
-  // A single block will handle tokens_per_block tokens.
   // Each block i iterates over tokens of a slice of n_tokens =
   // expert_counts[i], with the size of chunk being
   // (n_tokens / NUM_PARALLEL_TOKENS) + residual, instead of
   // updiv(n_tokens, NUM_PARALLEL_TOKENS) for better scheduling.
-
-  // Each warp will get space to store its hidden dim for gate and up.
-  __int128_t* s_hidden_load = smem_128 + warp_id * ((2 * 128 / 8) * NUM_STAGES);
-  __int128_t* smem_load_ptr = s_hidden_load + lane_id;
-
-  const __nv_bfloat16 fp8_inv = __hdiv(__float2bfloat16(1.f), fp8_max);
-
-  int32_t compute_pipeline_offset_64 = 0;
-  int32_t load_stage_offset{};
-  const __nv_bfloat16 one_bf16 = __float2bfloat16_rn(1.f);
-
-  __int64_t* smem_compute_ptr = reinterpret_cast<__int64_t*>(smem_128) +
-                                warp_id * (2 * (GROUP_SIZE / 4) * NUM_STAGES) +
-                                lane_id;
-  __int64_t* s_gate64_ptr = smem_compute_ptr;
-  __int64_t* s_up64_ptr = smem_compute_ptr + GROUP_SIZE / 4;
-
-  int tokens_lower, tokens_upper;
-
-  token_bounds<BLOCK_COUNT>(total_tokens, blockIdx.x, tokens_lower,
-                            tokens_upper);
-
-  Idx_t expert_id{}, expert_offset{}, next_expert_offset{};
-  int token_id = tokens_lower;
-  int32_t t_load{};
-
-  if (token_id < tokens_upper) {
-    expert_id = warp_expert_search<int>(lane_id, E, s_expert_offsets, token_id);
-    expert_offset = s_expert_offsets[expert_id];
-    next_expert_offset = s_expert_offsets[expert_id + 1];
+  if (n_tokens < NUM_PARALLEL_TOKENS && blockIdx.y < n_tokens) {
+    // Specialize this, but can be likely fused.
+    if (blockIdx.y >= NUM_PARALLEL_TOKENS) {
+      return;
+    }
+    n_tokens_lower = blockIdx.y;
+    n_tokens_upper = blockIdx.y + 1;
   } else {
-    // This thread block has no work to do.
+    auto chunk_size = n_tokens / NUM_PARALLEL_TOKENS;
+    auto residual = n_tokens - chunk_size * NUM_PARALLEL_TOKENS;
+    auto calc_id = [&](int32_t id) {
+      if (id < residual) {
+        return min(n_tokens, id * (chunk_size + 1));
+      } else {
+        return min(n_tokens, id * chunk_size + residual);
+      }
+    };
+    n_tokens_lower = calc_id(blockIdx.y);
+    n_tokens_upper = calc_id(blockIdx.y + 1);
+  }
+
+  if (n_tokens_lower >= n_tokens_upper) {
     return;
   }
 
-  int t_load_bound = H / (GROUP_SIZE * NUM_WARPS);
+  // We do calculations here, using constexpr wherever possible.
+  const Idx_t base_i = e * stride_i_e + NUM_WARPS * g * GROUP_SIZE * stride_i_h;
+  const Idx_t base_ys = e * stride_ys_e + NUM_WARPS * g * stride_ys_g;
+  const Idx_t base_yq =
+      e * stride_yq_e + NUM_WARPS * g * GROUP_SIZE * stride_yq_h;
+  Idx_t gate_off_128 = (base_i / static_cast<Idx_t>(8u));
+  auto input_128_ptr = reinterpret_cast<const __int128_t*>(_input);
+  auto gate_128_ptr = input_128_ptr + gate_off_128 + (tid % HALF_THREAD_COUNT) +
+                      stride_i_t_128 * n_tokens_lower;
+  auto up_128_ptr = gate_128_ptr + (H * stride_i_h) / 8u;
+  auto y_s_ptr =
+      _y_s + base_ys + warp_id * stride_ys_g + n_tokens_lower * stride_ys_t;
+  auto y_q_ptr = _y_q + base_yq + warp_id * GROUP_SIZE +
+                 stride_yq_t * n_tokens_lower + 4 * lane_id;
+  int32_t t_load = n_tokens_lower, load_stage_id = 0;
+  auto s_buff_gate_load_128 = s_buff_128 + (tid % HALF_THREAD_COUNT);
+  auto s_buff_up_load_128 = s_buff_gate_load_128 + S_NUM_128 / 2u;
+  int32_t stage_offset{};
 
-  Idx_t base_i = ((expert_id * stride_i_e) / 8) +
-                 (token_id - expert_offset) * stride_i_t / 8;
-  const Idx_t gate_warp_offset =
-      warp_id * ((stride_i_h * H) / (8 * NUM_WARPS)) + (lane_id & 0b1111);
-
-  const __int128_t* input_128_ptr =
-      reinterpret_cast<const __int128_t*>(_input) + gate_warp_offset +
-      ((lane_id < 16) ? 0 : ((H * stride_i_h) / 8));
-  __int128_t* load_ptr = const_cast<__int128_t*>(input_128_ptr + base_i);
-
-  auto token_offset = token_id - expert_offset;
+  static constexpr int32_t LOAD_STAGE_SIZE = (NUM_WARPS * WARP_SIZE / 2);
+  static constexpr int32_t LOAD_STAGE_MOD =
+      NUM_STAGES * (NUM_WARPS * WARP_SIZE / 2);
 
+  // Two halves of all threads in a block conduct global loads for gate and up,
+  // repsectively.
   auto load_and_advance_y_pred = [&] {
-    if (t_load < t_load_bound) {
-      // Here we are simply continuing to load data
-      // from the current token.
-      auto smem_load_ptr_staged = smem_load_ptr + load_stage_offset;
+    if (t_load < n_tokens_upper) {
+      auto s_gate_stage_128_staged_ptr = s_buff_gate_load_128 + stage_offset;
+      auto s_up_stage_128_staged_ptr = s_buff_up_load_128 + stage_offset;
 
       // It is very important that LOAD_STAGE_SIZE is constexpr to avoid
       // unnecessary ALU ops.
-      load_stage_offset += LOAD_STAGE_SIZE;
-      load_stage_offset %= LOAD_STAGE_MOD;
+      stage_offset += LOAD_STAGE_SIZE;
+      stage_offset %= LOAD_STAGE_MOD;
 
-      cp_async4(smem_load_ptr_staged, load_ptr);
-      load_ptr += GROUP_SIZE / 8;
-      ++t_load;
-    } else if (token_id + 1 < tokens_upper) {
-      // We loaded everything from the current token, let's move on
-      // to the next one, and we checked that we have more tokens to load.
-      ++token_id;
-      t_load = 0;
-      if (token_id >= next_expert_offset) {
-        // We need to find the next expert.
-        do {
-          // This is a loop because it's possible
-          // that some experts are assigned 0 tokens.
-          // NOTE: We are guaranteed that there's at least
-          // one more token left so we don't have to check for
-          // expert_id bounds.
-          ++expert_id;
-          // This skips 1 memory read.
-          expert_offset = next_expert_offset;
-          next_expert_offset = s_expert_offsets[expert_id + 1];
-        } while (next_expert_offset == expert_offset);
-
-        base_i = expert_id * (stride_i_e / 8);
-        token_offset = 0;
-        load_ptr = const_cast<__int128_t*>(input_128_ptr + base_i);
+      if (tid < HALF_THREAD_COUNT) {
+        cp_async4(s_gate_stage_128_staged_ptr, gate_128_ptr);
+        gate_128_ptr += stride_i_t_128;
       } else {
-        // We remain within the same expert, so just
-        // move by H/4 __int128_t (2 * H/8).
-        base_i += stride_yq_t / 4;
-        token_offset++;
+        cp_async4(s_up_stage_128_staged_ptr, up_128_ptr);
+        up_128_ptr += stride_i_t_128;
       }
-
-      load_ptr = const_cast<__int128_t*>(input_128_ptr + base_i);
-
-      auto smem_load_ptr_staged = smem_load_ptr + load_stage_offset;
-
-      // It is very important that LOAD_STAGE_SIZE is constexpr to avoid
-      // unnecessary ALU ops.
-      load_stage_offset += LOAD_STAGE_SIZE;
-      load_stage_offset %= LOAD_STAGE_MOD;
-
-      cp_async4(smem_load_ptr_staged, load_ptr);
-      load_ptr += GROUP_SIZE / 8;
       ++t_load;
+      ++load_stage_id;
     }
     // We fence even if there is nothing to load to simplify pipelining.
     cp_async_fence();
   };
 
-  // We need to warm-up the pipeline.
   #pragma unroll
   for (int i = 0; i < NUM_STAGES - 1; i++) {
     load_and_advance_y_pred();
   }
 
-  __nv_fp8x4_e4m3* y_q_base_ptr =
-      reinterpret_cast<__nv_fp8x4_e4m3*>(_y_q) + lane_id;
-  auto y_scale_base_ptr = _y_s + warp_position_scales * stride_ys_g;
+  __int64_t* s_gate_ptr = reinterpret_cast<__int64_t*>(
+                              s_buff_compute_32 + warp_id * (GROUP_SIZE / 2)) +
+                          lane_id;
+  __int64_t* s_up_ptr = s_gate_ptr + S_NUM_64 / 2;
 
-  for (auto j = tokens_lower; j < tokens_upper; j++) {
-    const Idx_t base_ys = expert_id * stride_ys_e;
-    auto y_s_ptr = y_scale_base_ptr + base_ys + token_offset * stride_ys_t;
-    __nv_fp8x4_e4m3* y_q_ptr =
-        y_q_base_ptr + (expert_id * stride_yq_e + token_offset * stride_yq_t +
-                        warp_position_yq * stride_yq_h) /
-                           4;
-    const int COMPUTE_LIMIT = H / (GROUP_SIZE * NUM_WARPS);
+  static constexpr int32_t STAGE_SIZE = (GROUP_SIZE * NUM_WARPS) / 4u;
+  static constexpr int32_t STAGE_MOD = STAGE_SIZE * NUM_STAGES;
 
-    for (int i = 0; i < COMPUTE_LIMIT; i++) {
-      cp_async_wait<NUM_STAGES - 2>();
-      __syncthreads();
-      load_and_advance_y_pred();
+  int32_t compute_pipeline_offset_64 = 0;
 
-      __int64_t* gate64_ptr = s_gate64_ptr + compute_pipeline_offset_64;
-      __int64_t* up64_ptr = s_up64_ptr + compute_pipeline_offset_64;
+  for (int32_t t = n_tokens_lower; t < n_tokens_upper; ++t) {
+    __nv_bfloat162 results_bf162[2];
 
-      // COMPUTE_STAGE_SIZE/MOD must also be constexpr!
-      compute_pipeline_offset_64 += COMPUTE_STAGE_SIZE;
-      compute_pipeline_offset_64 %= COMPUTE_STAGE_MOD;
+    cp_async_wait<NUM_STAGES - 2>();
+    __syncthreads();
 
-      __int64_t gate64 = *gate64_ptr;
-      __int64_t up64 = *up64_ptr;
+    // We double-buffer pipelined loads so that the next load will
+    // concurrently run with compute without overwrites.
+    load_and_advance_y_pred();
 
-      // Compute
-      __nv_bfloat162 res[2];
-      __nv_bfloat162* s_up_comp = reinterpret_cast<__nv_bfloat162*>(&up64);
-      __nv_bfloat162* s_gate_comp = reinterpret_cast<__nv_bfloat162*>(&gate64);
+    auto s_gate_compute_64 = s_gate_ptr + compute_pipeline_offset_64;
+    auto s_up_compute_64 = s_up_ptr + compute_pipeline_offset_64;
+
+    // STAGE_SIZE must also be constexpr!
+    compute_pipeline_offset_64 += STAGE_SIZE;
+    compute_pipeline_offset_64 %= STAGE_MOD;
+
+    // Each thread loads (gate/up) 2X 4X bfloat16 values into registers.
+    __int64_t gate64 = *s_gate_compute_64;
+    __nv_bfloat162* s_gate_compute_32 =
+        reinterpret_cast<__nv_bfloat162*>(&gate64);
+
+    __int64_t up64 = *s_up_compute_64;
+    __nv_bfloat162* s_up_compute_32 = reinterpret_cast<__nv_bfloat162*>(&up64);
 
   #pragma unroll
-      for (int32_t k = 0; k < 2; ++k) {
-        __nv_bfloat162 gate = silu2_v2(__bfloat1622float2(s_gate_comp[k]));
-        res[k] = __hmul2(gate, s_up_comp[k]);
-      }
-
-      auto _y_max2 = __hmax2(__habs2(res[0]), __habs2(res[1]));
-
-      _y_max2.x = __hmax(__hmax(_y_max2.x, _y_max2.y), EPS);
-
-      __nv_bfloat16 y_s = __hmul(warp_max(_y_max2.x), fp8_inv);
-
-      if constexpr (USE_UE8M0) {
-        y_s = hexp2(hceil(hlog2(y_s)));
-      }
-
-      __nv_bfloat16 inv_y = __hdiv(one_bf16, y_s);
-
-      auto y_s2 = make_bfloat162(inv_y, inv_y);
+    for (int i = 0; i < 2; i++) {
+      // For silu, we make sure that div is emitted.
+      float2 gate = silu2(__bfloat1622float2(s_gate_compute_32[i]));
+      results_bf162[i] = __float22bfloat162_rn(gate);
+    }
 
   #pragma unroll
-      for (int32_t k = 0; k < 2; ++k) {
-        res[k] = clip(__hmul2(res[k], y_s2), __bfloat162bfloat162(fp8_min),
-                      __bfloat162bfloat162(fp8_max));
-      }
+    for (int i = 0; i < 2; i++) {
+      results_bf162[i] = __hmul2(results_bf162[i], s_up_compute_32[i]);
+    }
 
-      *y_q_ptr = __nv_fp8x4_e4m3(res[0], res[1]);
-      y_q_ptr += WARP_SIZE * stride_yq_h;
+    auto _y_max2 =
+        __hmax2(__habs2(results_bf162[0]), __habs2(results_bf162[1]));
 
-      if (!lane_id) {
-        *y_s_ptr = y_s;
-        y_s_ptr += stride_ys_g;
-      }
+    __nv_bfloat16 y_max_bf16 = __hmax(EPS, __hmax(_y_max2.x, _y_max2.y));
+
+    // An entire group is assigned to a single warp, so a simple warp reduce
+    // is used.
+    __nv_bfloat16 y_s = warp_max(y_max_bf16) / fp8_max;
+
+    if constexpr (USE_UE8M0) {
+      y_s = hexp2(hceil(hlog2(y_s)));
+    }
+
+    auto inv_y = __float2bfloat16_rn(1.f) / y_s;
+
+    auto y_s2 = make_bfloat162(inv_y, inv_y);
+
+  #pragma unroll
+    for (int32_t i = 0; i < 2; ++i) {
+      results_bf162[i] =
+          clip(__hmul2(results_bf162[i], y_s2), __bfloat162bfloat162(fp8_min),
+               __bfloat162bfloat162(fp8_max));
+    }
+
+    auto fp8x4 = __nv_fp8x4_e4m3(results_bf162[0], results_bf162[1]);
+    *reinterpret_cast<__nv_fp8x4_e4m3*>(y_q_ptr) = fp8x4;
+    y_q_ptr += stride_yq_t;
+
+    if (lane_id == 0) {
+      *y_s_ptr = y_s;
+      y_s_ptr += stride_ys_t;
     }
   }
-#endif
 }
+#endif
 
 }  // namespace vllm
 
@@ -568,14 +470,14 @@ void silu_and_mul_quant(torch::Tensor& out,    // [..., d]
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel);
 }
 
-void persistent_masked_m_silu_mul_quant(
-    const at::Tensor& input,              // (E, T, 2*H)
-    const at::Tensor& tokens_per_expert,  // (E)
-    at::Tensor& y_q,                      // (E, T, H) [OUT]
-    at::Tensor& y_s,                      // (E, T, H//group_size) [OUT]
-    bool use_ue8m0) {
+void silu_mul_fp8_quant_deep_gemm_cuda(
+    const at::Tensor& input,   // (E, T, 2*H)
+    const at::Tensor& counts,  // (E)
+    at::Tensor& y_q,           // (E, T, H) [OUT]
+    at::Tensor& y_s,           // (E, T, H//group_size) [OUT]
+    int64_t group_size, bool use_ue8m0, int64_t num_parallel_tokens) {
 #ifndef USE_ROCM
-
+  // This kernel relies heavily on cp.async and fp8 support.
   // This kernel currently only supports H % 128 == 0 and assumes a
   // fixed GROUP_SIZE of 128.
   TORCH_CHECK(input.dtype() == torch::kBFloat16);
@@ -584,6 +486,10 @@ void persistent_masked_m_silu_mul_quant(
   TORCH_CHECK(y_s.dtype() == torch::kFloat32);
   TORCH_CHECK(input.size(-1) % 256 == 0);
 
+  // Check that num_parallel_tokens is of power of 2 and between 1 and 64.
+  TORCH_CHECK(1 <= num_parallel_tokens && num_parallel_tokens <= 64);
+  TORCH_CHECK(!(num_parallel_tokens & (num_parallel_tokens - 1)));
+
   using Idx_t = int64_t;
 
   Idx_t E = input.size(0);
@@ -599,54 +505,81 @@ void persistent_masked_m_silu_mul_quant(
   Idx_t stride_ys_t = y_s.stride(1);
   Idx_t stride_ys_g = y_s.stride(2);
 
-  Idx_t stride_counts_e = tokens_per_expert.stride(0);
+  Idx_t stride_counts_e = counts.stride(0);
 
   static constexpr int GROUP_SIZE = 128;
 
+  #define KERNEL_FN                                                         \
+    if (use_ue8m0) {                                                        \
+      vllm::silu_mul_fp8_quant_deep_gemm_kernel<fp8_t, NUM_WARPS, Idx_t,    \
+                                                NUM_PARALLEL_TOKENS, true>  \
+          <<<grid, block, 0, stream>>>(                                     \
+              reinterpret_cast<__nv_bfloat16*>(input.data_ptr()),           \
+              (fp8_t*)y_q.data_ptr(), y_s.data_ptr<float>(),                \
+              reinterpret_cast<int32_t*>(counts.data_ptr<int>()), H, G,     \
+              stride_i_e, stride_i_t, stride_i_h, stride_yq_e, stride_yq_t, \
+              stride_yq_h, stride_ys_e, stride_ys_t, stride_ys_g,           \
+              stride_counts_e);                                             \
+    } else {                                                                \
+      vllm::silu_mul_fp8_quant_deep_gemm_kernel<fp8_t, NUM_WARPS, Idx_t,    \
+                                                NUM_PARALLEL_TOKENS, false> \
+          <<<grid, block, 0, stream>>>(                                     \
+              reinterpret_cast<__nv_bfloat16*>(input.data_ptr()),           \
+              (fp8_t*)y_q.data_ptr(), y_s.data_ptr<float>(),                \
+              reinterpret_cast<int32_t*>(counts.data_ptr<int>()), H, G,     \
+              stride_i_e, stride_i_t, stride_i_h, stride_yq_e, stride_yq_t, \
+              stride_yq_h, stride_ys_e, stride_ys_t, stride_ys_g,           \
+              stride_counts_e);                                             \
+    }
+
+  #define KERNEL_CALL_H                                       \
+    if (H % (4 * GROUP_SIZE) == 0) {                          \
+      static constexpr int NUM_WARPS = 4;                     \
+      populate_launch_params(NUM_WARPS, NUM_PARALLEL_TOKENS); \
+      KERNEL_FN                                               \
+    } else {                                                  \
+      static constexpr int NUM_WARPS = 1;                     \
+      populate_launch_params(NUM_WARPS, NUM_PARALLEL_TOKENS); \
+      KERNEL_FN                                               \
+    }
+
+  #define KERNEL_CALL_TOP_LEVEL                      \
+    if (num_parallel_tokens == 1) {                  \
+      static constexpr int NUM_PARALLEL_TOKENS = 1;  \
+      KERNEL_CALL_H                                  \
+    } else if (num_parallel_tokens == 2) {           \
+      static constexpr int NUM_PARALLEL_TOKENS = 2;  \
+      KERNEL_CALL_H                                  \
+    } else if (num_parallel_tokens == 4) {           \
+      static constexpr int NUM_PARALLEL_TOKENS = 4;  \
+      KERNEL_CALL_H                                  \
+    } else if (num_parallel_tokens == 8) {           \
+      static constexpr int NUM_PARALLEL_TOKENS = 8;  \
+      KERNEL_CALL_H                                  \
+    } else if (num_parallel_tokens == 16) {          \
+      static constexpr int NUM_PARALLEL_TOKENS = 16; \
+      KERNEL_CALL_H                                  \
+    } else if (num_parallel_tokens == 32) {          \
+      static constexpr int NUM_PARALLEL_TOKENS = 32; \
+      KERNEL_CALL_H                                  \
+    } else if (num_parallel_tokens == 64) {          \
+      static constexpr int NUM_PARALLEL_TOKENS = 64; \
+      KERNEL_CALL_H                                  \
+    }
+
+  Idx_t G;
+  dim3 block, grid;
+  auto populate_launch_params = [&](int num_warps, int _num_parallel_tokens) {
+    G = H / Idx_t(group_size * num_warps);
+    grid = dim3(E * G, _num_parallel_tokens);
+    block = dim3(num_warps * WARP_SIZE);
+  };
+
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-
-  #define KERNEL(BLOCK_COUNT, USE_UE8M0, THREAD_COUNT, STAGES)                 \
-    static constexpr int NUM_WARPS = THREAD_COUNT / WARP_SIZE;                 \
-    int sms = SILU_V2_BLOCK_COUNT;                                             \
-    static constexpr int max_shared_mem_bytes =                                \
-        GROUP_SIZE * 2 * STAGES * NUM_WARPS * 2;                               \
-    dim3 grid(sms), block(THREAD_COUNT);                                       \
-    const at::cuda::OptionalCUDAGuard device_guard(device_of(input));          \
-    VLLM_DISPATCH_FP8_TYPES(                                                   \
-        y_q.scalar_type(), "silu_mul_fp8_quant_deep_gemm_kernel", [&] {        \
-          vllm::silu_mul_fp8_quant_deep_gemm_kernel<                           \
-              BLOCK_COUNT, max_shared_mem_bytes, fp8_t, THREAD_COUNT, Idx_t,   \
-              USE_UE8M0, GROUP_SIZE, STAGES>                                   \
-              <<<grid, block, max_shared_mem_bytes + (E + 1) * 16, stream>>>(  \
-                  reinterpret_cast<__nv_bfloat16*>(input.data_ptr()),          \
-                  (fp8_t*)y_q.data_ptr(), y_s.data_ptr<float>(),               \
-                  reinterpret_cast<int32_t*>(tokens_per_expert.data_ptr()), E, \
-                  T, H, stride_i_e, stride_i_t, stride_i_h, stride_yq_e,       \
-                  stride_yq_t, stride_yq_h, stride_ys_e, stride_ys_t,          \
-                  stride_ys_g, stride_counts_e);                               \
-        });
-
-  static constexpr int SILU_V2_BLOCK_COUNT = 132 * 32;
-
-  if (!use_ue8m0) {
-    if (H >= 4096) {
-      static constexpr int NUM_STAGES = 4;
-      static constexpr int THREAD_COUNT = 256;
-      KERNEL(SILU_V2_BLOCK_COUNT, false, THREAD_COUNT, NUM_STAGES);
-    } else {
-      static constexpr int THREAD_COUNT = 32;
-      KERNEL(SILU_V2_BLOCK_COUNT, false, THREAD_COUNT, 2);
-    }
-  } else {
-    if (H >= 4096) {
-      static constexpr int NUM_STAGES = 4;
-      static constexpr int THREAD_COUNT = 256;
-      KERNEL(SILU_V2_BLOCK_COUNT, true, THREAD_COUNT, NUM_STAGES);
-    } else {
-      static constexpr int THREAD_COUNT = 32;
-      KERNEL(SILU_V2_BLOCK_COUNT, true, THREAD_COUNT, 2);
-    }
-  }
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
+  VLLM_DISPATCH_FP8_TYPES(y_q.scalar_type(),
+                          "silu_mul_fp8_quant_deep_gemm_kernel",
+                          [&] { KERNEL_CALL_TOP_LEVEL });
 
 #endif
 }

csrc/quantization/compressed_tensors/int8_quant_kernels.cu

@@ -1,11 +1,15 @@
 #include <ATen/cuda/CUDAContext.h>
 #include <torch/all.h>
 
+#ifndef USE_ROCM
+  #include "../per_token_group_quant_8bit.h"
+#endif
+
 #include <cmath>
 
-#include "dispatch_utils.h"
-#include "quantization/vectorization_utils.cuh"
-#include "cub_helpers.h"
+#include "../../cub_helpers.h"
+#include "../../dispatch_utils.h"
+#include "../vectorization_utils.cuh"
 
 static inline __device__ int8_t float_to_int8_rn(float x) {
 #ifdef USE_ROCM
@@ -21,6 +25,7 @@ static inline __device__ int8_t float_to_int8_rn(float x) {
   float dst = std::nearbyint(x);
 
   // saturate
+
   // See https://github.com/pytorch/pytorch/issues/127666
   // See https://github.com/llvm/llvm-project/issues/95183
   // hip-clang std::clamp __glibcxx_assert_fail host function when building on
@@ -79,6 +84,7 @@ static inline __device__ int8_t int32_to_int8(int32_t x) {
       static_cast<int32_t>(std::numeric_limits<int8_t>::max());
 
   // saturate
+
   // See https://github.com/pytorch/pytorch/issues/127666
   // See https://github.com/llvm/llvm-project/issues/95183
   // hip-clang std::clamp __glibcxx_assert_fail host function when building on
@@ -170,6 +176,7 @@ __global__ void dynamic_scaled_int8_quant_kernel(
 
   float inv_s = (absmax == 0.f) ? 0.f : 127.f / absmax;
 
+  // 2. quantize
   vectorize_with_alignment<16>(
       row_in, row_out, hidden_size, tid, stride,
       [=] __device__(int8_t& dst, const scalar_t& src) {
@@ -187,6 +194,7 @@ struct MinMax {
 
   __host__ __device__ explicit MinMax(float v) : min(v), max(v) {}
 
+  // add a value to the MinMax
   __host__ __device__ MinMax& operator+=(float v) {
     min = fminf(min, v);
     max = fmaxf(max, v);
@@ -220,6 +228,7 @@ __global__ void dynamic_scaled_int8_azp_quant_kernel(
   const scalar_t* row_in = input + token_idx * hidden_size;
   int8_t* row_out = output + token_idx * hidden_size;
 
+  // 1. calculate min & max
   MinMax thread_mm;
   vectorize_read_with_alignment<16>(row_in, hidden_size, tid, stride,
                                     [&] __device__(const scalar_t& src) {
@@ -252,6 +261,7 @@ __global__ void dynamic_scaled_int8_azp_quant_kernel(
   const float inv_s = 1.f / scale_sh;
   const azp_t azp = azp_sh;
 
+  // 2. quantize
   vectorize_with_alignment<16>(
       row_in, row_out, hidden_size, tid, stride,
       [=] __device__(int8_t& dst, const scalar_t& src) {
@@ -322,4 +332,14 @@ void dynamic_scaled_int8_quant(
                   hidden_size);
         }
       });
-}
+}
+
+#ifndef USE_ROCM
+void per_token_group_quant_int8(const torch::Tensor& input,
+                                torch::Tensor& output_q,
+                                torch::Tensor& output_s, int64_t group_size,
+                                double eps, double int8_min, double int8_max) {
+  per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
+                             int8_min, int8_max);
+}
+#endif

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8_dispatch.cuh

@@ -231,7 +231,7 @@ void cutlass_gemm_blockwise_sm100_fp8_dispatch(torch::Tensor& out,
       } else {
         cutlass_gemm_caller_blockwise<cutlass_3x_gemm_fp8_blockwise<
             OutType, 1, TILE_N, TILE_K, Shape<_64, Int<TILE_N>, Int<TILE_K>>,
-            Shape<_1, _1, _1>, cutlass::epilogue::BlockwiseNoSmemWarpSpecialized1Sm,
+            Shape<_1, _1, _1>, cutlass::epilogue::NoSmemWarpSpecialized1Sm,
             cutlass::gemm::KernelTmaWarpSpecializedBlockwise1SmSm100>>(
             out, a, b, a_scales, b_scales);
       }
@@ -245,7 +245,7 @@ void cutlass_gemm_blockwise_sm100_fp8_dispatch(torch::Tensor& out,
       } else {
         cutlass_gemm_caller_blockwise<cutlass_3x_gemm_fp8_blockwise<
             OutType, 1, TILE_N, TILE_K, Shape<_128, Int<TILE_N>, Int<TILE_K>>,
-            Shape<_1, _1, _1>, cutlass::epilogue::BlockwiseNoSmemWarpSpecialized1Sm,
+            Shape<_1, _1, _1>, cutlass::epilogue::NoSmemWarpSpecialized1Sm,
             cutlass::gemm::KernelTmaWarpSpecializedBlockwise1SmSm100>>(
             out, a, b, a_scales, b_scales);
       }
@@ -259,7 +259,7 @@ void cutlass_gemm_blockwise_sm100_fp8_dispatch(torch::Tensor& out,
       } else {
           cutlass_gemm_caller_blockwise<cutlass_3x_gemm_fp8_blockwise<
               OutType, 1, TILE_N, TILE_K, Shape<_256, Int<TILE_N>, Int<TILE_K>>,
-            Shape<_2, _1, _1>, cutlass::epilogue::BlockwiseNoSmemWarpSpecialized2Sm,
+            Shape<_2, _1, _1>, cutlass::epilogue::NoSmemWarpSpecialized2Sm,
             cutlass::gemm::KernelTmaWarpSpecializedBlockwise2SmSm100>>(
             out, a, b, a_scales, b_scales);
       }
@@ -271,10 +271,10 @@ void cutlass_gemm_blockwise_sm100_fp8_dispatch(torch::Tensor& out,
     // TMA epilogue isn't compatible with Swap A/B
     cutlass_gemm_caller_blockwise<cutlass_3x_gemm_fp8_blockwise<
         OutType, TILE_M, 1, TILE_K, Shape<Int<TILE_M>, Int<TILE_N>, Int<TILE_K>>,
-        Shape<_1, _1, _1>, cutlass::epilogue::BlockwiseNoSmemWarpSpecialized1Sm,
+        Shape<_1, _1, _1>, cutlass::epilogue::NoSmemWarpSpecialized1Sm,
         cutlass::gemm::KernelTmaWarpSpecializedBlockwise1SmSm100, true>>(
         out, a, b, a_scales, b_scales);
   }
 }
 
-}  // namespace vllm
+}  // namespace vllm

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_helper.hpp

@@ -25,10 +25,7 @@ void dispatch_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
       if constexpr (!std::is_same_v<Int8Func, std::nullptr_t>) {
         int8_func(c, a, b, a_scales, b_scales, bias);
       } else {
-        int32_t version_num = get_sm_version_num();
-        TORCH_CHECK(
-            false, "Int8 not supported on SM", version_num,
-            ". Use FP8 quantization instead, or run on older arch (SM < 100).");
+        TORCH_CHECK(false, "Int8 not supported for this architecture");
       }
     }
   } else {

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8_dispatch.cuh

@@ -133,4 +133,4 @@ void cutlass_scaled_mm_sm100_fp8_epilogue(torch::Tensor& out,
   }
 }
 
-}  // namespace vllm
+}  // namespace vllm