[Quantization/NVFP4] Speed up TRTLLM NVFP4 MOE weight loading and fix K/V scale loading for MLA Attn (#25968)

Signed-off-by: Pavani Majety <pmajety@nvidia.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com>

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

@@ -181,18 +181,14 @@ 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="python3 \
-        -m vllm.entrypoints.openai.api_server \
+    server_command="vllm serve $model \
         -tp $tp \
-        --model $model \
         --port $port \
         $server_args"
   else
     echo "Key 'fp8' does not exist in common params."
-    server_command="python3 \
-        -m vllm.entrypoints.openai.api_server \
+    server_command="vllm serve $model \
         -tp $tp \
-        --model $model \
         --port $port \
         $server_args"
   fi

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

@@ -365,8 +365,7 @@ run_serving_tests() {
       continue
     fi
 
-    server_command="$server_envs python3 \
-      -m vllm.entrypoints.openai.api_server \
+    server_command="$server_envs vllm serve \
       $server_args"
 
     # run the server

.buildkite/release-pipeline.yaml

@@ -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 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 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 push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
 
   # Add job to create multi-arch manifest

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

@@ -86,10 +86,6 @@ 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.sh

@@ -58,11 +58,8 @@ 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
 
-    # 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/generation -m cpu_model
+    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model
 
     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

@@ -0,0 +1,191 @@
+#!/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

@@ -62,7 +62,7 @@ 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
+    && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
 echo "--- Python dependencies installed ---"
 export VLLM_USE_V1=1
 export VLLM_XLA_CHECK_RECOMPILATION=1

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

@@ -62,7 +62,7 @@ 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
+    && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
 echo "--- Python dependencies installed ---"
 export VLLM_USE_V1=1
 export VLLM_XLA_CHECK_RECOMPILATION=1

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

@@ -35,16 +35,15 @@ 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_VLLM_V1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    VLLM_ATTENTION_BACKEND=TRITON_ATTN 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_eagle.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_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_metrics
     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
-python3 -m vllm.entrypoints.openai.api_server --model meta-llama/Llama-2-7b-chat-hf &
+vllm serve 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

@@ -0,0 +1,59 @@
+#!/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/test-pipeline.yaml

@@ -6,24 +6,28 @@
 # to generate the final pipeline yaml file.
 
 # Documentation
-# 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.
+# 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).
 # command(str): the single command to run for tests. incompatible with commands.
-# 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
+# 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
 #     command runs on the second host.
-# 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.
+# 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.
 
 # When adding a test
-# - If the test belong to an existing group, add it there
+# - If the test belongs 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.
@@ -46,19 +50,28 @@ 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 multimodal
-  - pytest -v -s utils_ # Utils
-  - pytest -v -s transformers_utils # transformers_utils
+  - pytest -v -s -m 'cpu_test' multimodal
+  - pytest -v -s transformers_utils
 
 - label: Python-only Installation Test # 10min
   timeout_in_minutes: 20
@@ -110,7 +123,7 @@ 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
-  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+  - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Entrypoints Integration Test (API Server) # 100min
   timeout_in_minutes: 130
@@ -148,7 +161,6 @@ steps:
   num_gpus: 4
   source_file_dependencies:
   - vllm/distributed/
-  - vllm/core/
   - tests/distributed/test_utils
   - tests/distributed/test_pynccl
   - tests/distributed/test_events
@@ -156,28 +168,34 @@ steps:
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
-  - 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/distributed
   - tests/v1/engine/test_engine_core_client.py
+  - tests/distributed/test_symm_mem_allreduce.py
   commands:
-  # test with tp=2 and external_dp=2
-  - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  # test with torchrun tp=2 and external_dp=2
   - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
-  # test with tp=2 and pp=2
+  # test with torchrun 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/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
+  - 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
   - 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
@@ -278,24 +296,34 @@ steps:
     - tests/v1
   commands:
     # split the test to avoid interference
-    - 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 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 -m 'not cpu_test' v1/kv_connector/unit
+    - pytest -v -s -m 'not cpu_test' v1/metrics
     - pytest -v -s v1/test_oracle.py
-    - pytest -v -s v1/test_metrics_reader.py
+    - pytest -v -s v1/test_request.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 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]
@@ -314,12 +342,13 @@ 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
-    - 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 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
-    - VLLM_USE_V1=0 python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
+    - 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
 
 - label: Platform Tests (CUDA) # 4min
   timeout_in_minutes: 15
@@ -450,29 +479,18 @@ steps:
   commands:
     - pytest -v -s kernels/mamba
 
-- label: Tensorizer Test # 14min
-  timeout_in_minutes: 25
-  mirror_hardwares: [amdexperimental]
-  source_file_dependencies:
-  - 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 tensorizer_loader
-    - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py
-
-- label: Model Executor Test # 7min
-  timeout_in_minutes: 20
+- label: Model Executor Test # 23min
+  timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/model_executor
   - tests/model_executor
+  - 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 entrypoints/openai/test_tensorizer_entrypoint.py
 
 - label: Benchmarks # 11min
   timeout_in_minutes: 20
@@ -507,7 +525,7 @@ 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
@@ -535,10 +553,17 @@ steps:
   source_file_dependencies:
     - vllm/
     - tests/tool_use
-    - tests/mistral_tool_use
   commands:
-    - pytest -v -s tool_use
-    - pytest -v -s mistral_tool_use
+    - 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
 
 #####  models test  #####
 
@@ -578,13 +603,19 @@ 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_transformers.py \
-                   models/test_registry.py \
-                   models/test_utils.py \
-                   models/test_vision.py
+    - pytest -v -s models/test_utils.py models/test_vision.py
 
 - label: Language Models Tests (Standard)
   timeout_in_minutes: 25
@@ -754,11 +785,13 @@ 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
@@ -812,7 +845,24 @@ 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 --server-args '--tensor-parallel-size 2'
+    - 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
 
 #####  1 GPU test  #####
 #####  multi gpus test  #####
@@ -856,47 +906,58 @@ 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) # 110min
-  timeout_in_minutes: 150
+- label: Distributed Tests (2 GPUs) # 68min
+  timeout_in_minutes: 90
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
+  - vllm/compilation/
   - vllm/distributed/
   - vllm/engine/
   - vllm/executor/
-  - vllm/model_executor/models/
-  - tests/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
   - vllm/v1/engine/
+  - vllm/v1/worker/
+  - tests/compile/test_basic_correctness.py
+  - tests/compile/test_wrapper.py
+  - tests/distributed/
+  - tests/entrypoints/llm/test_collective_rpc.py
+  - tests/v1/distributed
+  - tests/v1/entrypoints/openai/test_multi_api_servers.py
+  - tests/v1/shutdown
+  - tests/v1/worker/test_worker_memory_snapshot.py
   commands:
-  - 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
+  - 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
   - 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
@@ -1030,3 +1091,16 @@ 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

@@ -4,19 +4,14 @@
 # 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
@@ -31,11 +26,13 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # vLLM V1
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
-/vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
-/vllm/v1/spec_decode @benchislett @luccafong
+/vllm/v1/attention @LucasWilkinson
 /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
 
@@ -57,7 +54,7 @@ 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
 
@@ -75,6 +72,7 @@ mkdocs.yaml @hmellor
 # Linting
 .markdownlint.yaml @hmellor
 .pre-commit-config.yaml @hmellor
+/tools/pre_commit @hmellor
 
 # CPU
 /vllm/v1/worker/cpu* @bigPYJ1151

.github/ISSUE_TEMPLATE/750-RFC.yml

@@ -43,10 +43,6 @@ 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

@@ -2,6 +2,7 @@ pull_request_rules:
 - name: label-documentation
   description: Automatically apply documentation label
   conditions:
+    - label != stale
     - or:
       - files~=^[^/]+\.md$
       - files~=^docs/
@@ -14,6 +15,7 @@ pull_request_rules:
 - name: label-ci-build
   description: Automatically apply ci/build label
   conditions:
+    - label != stale
     - or:
       - files~=^\.github/
       - files~=\.buildkite/
@@ -30,6 +32,7 @@ pull_request_rules:
 - name: label-deepseek
   description: Automatically apply deepseek label
   conditions:
+    - label != stale
     - or:
       - files~=^examples/.*deepseek.*\.py
       - files~=^tests/.*deepseek.*\.py
@@ -46,6 +49,7 @@ pull_request_rules:
 - name: label-frontend
   description: Automatically apply frontend label
   conditions:
+    - label != stale
     - files~=^vllm/entrypoints/
   actions:
     label:
@@ -55,6 +59,7 @@ pull_request_rules:
 - name: label-llama
   description: Automatically apply llama label
   conditions:
+    - label != stale
     - or:
       - files~=^examples/.*llama.*\.py
       - files~=^tests/.*llama.*\.py
@@ -70,6 +75,7 @@ pull_request_rules:
 - name: label-multi-modality
   description: Automatically apply multi-modality label
   conditions:
+    - label != stale
     - or:
       - files~=^vllm/multimodal/
       - files~=^tests/multimodal/
@@ -83,6 +89,7 @@ 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
@@ -94,6 +101,7 @@ pull_request_rules:
 - name: label-performance
   description: Automatically apply performance label
   conditions:
+    - label != stale
     - or:
       - files~=^benchmarks/
       - files~=^vllm/benchmarks/
@@ -107,6 +115,7 @@ pull_request_rules:
 - name: label-qwen
   description: Automatically apply qwen label
   conditions:
+    - label != stale
     - or:
       - files~=^examples/.*qwen.*\.py
       - files~=^tests/.*qwen.*\.py
@@ -121,6 +130,7 @@ 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
@@ -142,6 +152,7 @@ pull_request_rules:
 - name: label-rocm
   description: Automatically apply rocm label
   conditions:
+    - label != stale
     - or:
       - files~=^csrc/rocm/
       - files~=^docker/Dockerfile.rocm
@@ -162,6 +173,7 @@ 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
@@ -181,6 +193,7 @@ 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/
@@ -196,6 +209,7 @@ pull_request_rules:
 - name: label-v1
   description: Automatically apply v1 label
   conditions:
+    - label != stale
     - or:
       - files~=^vllm/v1/
       - files~=^tests/v1/
@@ -208,6 +222,7 @@ 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
@@ -223,6 +238,7 @@ 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
@@ -237,9 +253,9 @@ 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/
@@ -256,8 +272,9 @@ pull_request_rules:
 
 - name: ping author on conflicts and add 'needs-rebase' label
   conditions:
-      - conflict
-      - -closed
+    - label != stale
+    - conflict
+    - -closed
   actions:
     label:
       add:
@@ -271,10 +288,12 @@ 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/tensorizer_loader/
+      - files~=^tests/model_executor/model_loader/tensorizer_loader/
   actions:
     assign:
       users:
@@ -282,6 +301,7 @@ 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$
@@ -296,8 +316,8 @@ pull_request_rules:
 
 - name: remove 'needs-rebase' label when conflict is resolved
   conditions:
-      - -conflict
-      - -closed
+    - -conflict
+    - -closed
   actions:
     label:
       remove:
@@ -306,6 +326,7 @@ 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[^/]*/.*

.pre-commit-config.yaml

@@ -49,7 +49,7 @@ repos:
   rev: 0.6.17
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128, --python-platform, x86_64-manylinux_2_28]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
@@ -60,38 +60,32 @@ repos:
     files: ^requirements/test\.(in|txt)$
   - id: mypy-local
     name: Run mypy for local Python installation
-    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]
+    entry: python tools/pre_commit/mypy.py 0 "local"
     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
+    entry: python tools/pre_commit/mypy.py 1 "3.9"
+    <<: *mypy_common
     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: tools/mypy.sh 1 "3.10"
-    language: python
-    types: [python]
-    additional_dependencies: *mypy_deps
+    entry: python tools/pre_commit/mypy.py 1 "3.10"
+    <<: *mypy_common
     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: tools/mypy.sh 1 "3.11"
-    language: python
-    types: [python]
-    additional_dependencies: *mypy_deps
+    entry: python tools/pre_commit/mypy.py 1 "3.11"
+    <<: *mypy_common
     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: tools/mypy.sh 1 "3.12"
-    language: python
-    types: [python]
-    additional_dependencies: *mypy_deps
+    entry: python tools/pre_commit/mypy.py 1 "3.12"
+    <<: *mypy_common
     stages: [manual] # Only run in CI
   - id: shellcheck
     name: Lint shell scripts
@@ -155,11 +149,10 @@ repos:
     additional_dependencies: [regex]
   - id: check-pickle-imports
     name: Prevent new pickle/cloudpickle imports
-    entry: python tools/check_pickle_imports.py
+    entry: python tools/pre_commit/check_pickle_imports.py
     language: python
     types: [python]
-    pass_filenames: false
-    additional_dependencies: [pathspec, regex]
+    additional_dependencies: [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

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

CMakeLists.txt

@@ -37,7 +37,7 @@ install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
 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")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -86,6 +86,9 @@ 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()
@@ -175,6 +178,15 @@ 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.
 #
@@ -257,8 +269,8 @@ set(VLLM_EXT_SRC
   "csrc/sampler.cu"
   "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
-  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
-  "csrc/quantization/fp8/common.cu"
+  "csrc/quantization/w8a8/int8/scaled_quant.cu"
+  "csrc/quantization/w8a8/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"
@@ -270,7 +282,7 @@ 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 "v4.0.0" CACHE STRING "CUTLASS revision to use")
+  set(CUTLASS_REVISION "v4.2.1" 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})
@@ -302,13 +314,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_EXT_SRC
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/permute_cols.cu"
-    "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
+    "csrc/quantization/w8a8/cutlass/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/fp8/per_token_group_quant.cu")
+    "csrc/quantization/w8a8/fp8/per_token_group_quant.cu"
+    "csrc/quantization/w8a8/int8/per_token_group_quant.cu")
 
   set_gencode_flags_for_srcs(
     SRCS "${VLLM_EXT_SRC}"
@@ -412,11 +424,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND SCALED_MM_ARCHS)
     set(SRCS
-       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
-       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
+       "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm90.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_fp8.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm90_int8.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_azp_sm90_int8.cu"
+       "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm90_fp8.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -440,12 +452,16 @@ 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
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
+  if(${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()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
-      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm120_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm120.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm120_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm120_fp8.cu"
     )
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -470,12 +486,16 @@ 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
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
+  if(${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()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
-      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/scaled_mm_c3x_sm100.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_sm100_fp8.cu"
+      "csrc/quantization/w8a8/cutlass/c3x/scaled_mm_blockwise_sm100_fp8.cu"
     )
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -506,7 +526,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # subtract out the archs that are already built for 3x
   list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
   if (SCALED_MM_2X_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/scaled_mm_c2x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_2X_ARCHS}")
@@ -550,7 +570,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # The nvfp4_scaled_mm_sm120 kernels for Geforce Blackwell SM120 require
   # CUDA 12.8 or later
-  cuda_archs_loose_intersection(FP4_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
+  if(${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()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
@@ -569,7 +593,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # FP4 Archs and flags
-  cuda_archs_loose_intersection(FP4_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${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()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
@@ -591,7 +619,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # CUTLASS MLA Archs and flags
-  cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${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()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
     set(SRCS
       "csrc/attention/mla/sm100_cutlass_mla_kernel.cu")
@@ -617,7 +649,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # if it's possible to compile MoE kernels that use its output.
   cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm90.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm90.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -635,9 +667,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${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()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x_sm100.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -656,9 +692,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # moe_data.cu is used by all CUTLASS MoE kernels.
-  cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
+  if(${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()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/moe_data.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
@@ -675,9 +715,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${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()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu")
+    set(SRCS "csrc/quantization/w8a8/cutlass/moe/blockwise_scaled_group_mm_sm100.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")

README.md

@@ -21,6 +21,7 @@ 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).

benchmarks/auto_tune/auto_tune.sh

@@ -103,10 +103,15 @@ start_server() {
         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
@@ -118,7 +123,7 @@ start_server() {
     done
 
     if (( ! server_started )); then
-        echo "server did not start within 10 minutes. Please check server log at $vllm_log".
+        echo "server did not start within 10 minutes or crashed. Please check server log at $vllm_log".
         return 1
     else
         return 0

benchmarks/benchmark_ngram_proposer.py

@@ -1,17 +1,31 @@
 # 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 tabulate import tabulate
 
 from benchmark_utils import TimeCollector
-from vllm.config import ModelConfig, SpeculativeConfig, VllmConfig
+from vllm.config import (
+    CacheConfig,
+    DeviceConfig,
+    LoadConfig,
+    ModelConfig,
+    ParallelConfig,
+    SchedulerConfig,
+    SpeculativeConfig,
+    VllmConfig,
+)
+from vllm.platforms import current_platform
 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 main(args):
+def benchmark_propose(args):
     rows = []
     for max_ngram in args.max_ngram:
         collector = TimeCollector(TimeCollector.US)
@@ -69,10 +83,88 @@ def main(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"
+    )  # noqa: E501
     parser.add_argument(
         "--num-iteration",
         type=int,
@@ -105,8 +197,17 @@ def invoke_main() -> None:
         help="Number of speculative tokens to generate",
     )
     args = parser.parse_args()
-    main(args)
+
+    if not args.batched:
+        benchmark_propose(args)
+    else:
+        benchmark_batched_propose(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

@@ -449,7 +449,8 @@ async def benchmark(
     def prepare_extra_body(request) -> dict:
         extra_body = {}
         # Add the schema to the extra_body
-        extra_body[request.structure_type] = request.schema
+        extra_body["structured_outputs"] = {}
+        extra_body["structured_outputs"][request.structure_type] = request.schema
         return extra_body
 
     print("Starting initial single prompt test run...")

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_block_fp8_matmul,
+    w8a8_triton_block_scaled_mm,
 )
 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_block_fp8_matmul(
+        "triton_fp8_fp8_fp16_scaled_mm_blockwise": lambda: w8a8_triton_block_scaled_mm(
             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

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

benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh

@@ -38,16 +38,12 @@ wait_for_server() {
 launch_chunked_prefill() {
   model="meta-llama/Meta-Llama-3.1-8B-Instruct"
   # disagg prefill
-  CUDA_VISIBLE_DEVICES=0 python3 \
-    -m vllm.entrypoints.openai.api_server \
-    --model $model \
+  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
     --port 8100 \
     --max-model-len 10000 \
     --enable-chunked-prefill \
     --gpu-memory-utilization 0.6 &
-  CUDA_VISIBLE_DEVICES=1 python3 \
-    -m vllm.entrypoints.openai.api_server \
-    --model $model \
+  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
     --port 8200 \
     --max-model-len 10000 \
     --enable-chunked-prefill \
@@ -62,18 +58,14 @@ launch_chunked_prefill() {
 launch_disagg_prefill() {
   model="meta-llama/Meta-Llama-3.1-8B-Instruct"
   # disagg prefill
-  CUDA_VISIBLE_DEVICES=0 python3 \
-    -m vllm.entrypoints.openai.api_server \
-    --model $model \
+  CUDA_VISIBLE_DEVICES=0 vllm serve $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 python3 \
-    -m vllm.entrypoints.openai.api_server \
-    --model $model \
+  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
     --port 8200 \
     --max-model-len 10000 \
     --gpu-memory-utilization 0.6 \

benchmarks/kernels/bench_nvfp4_gemm.py

@@ -3,6 +3,7 @@
 import argparse
 import copy
 import itertools
+import os
 
 import torch
 from weight_shapes import WEIGHT_SHAPES
@@ -23,21 +24,45 @@ 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):
+def _quant_weight_nvfp4(b: torch.Tensor, device: str, cfg):
     # 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
-    b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
+    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)
     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)
+    b_fp4, scale_b_fp4, b_global_scale = _quant_weight_nvfp4(b, device, cfg)
 
     # Compute global scale for activation
     # NOTE: This is generally provided ahead-of-time by the model checkpoint.
@@ -46,6 +71,35 @@ 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
@@ -130,10 +184,13 @@ 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=f"bench_nvfp4_res_n{N}_k{K}",
+            save_path=save_dir,
             N=N,
             K=K,
         )

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

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -0,0 +1,406 @@
+# 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

@@ -7,6 +7,10 @@ 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]
 
@@ -26,7 +30,13 @@ 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
+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.symm_mem import SymmMemCommunicator
 from vllm.logger import init_logger
 from vllm.utils import FlexibleArgumentParser
@@ -98,6 +108,7 @@ 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
@@ -194,6 +205,15 @@ 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
@@ -271,7 +291,9 @@ class CommunicatorBenchmark:
                 # Capture the graph using context manager
                 with context:
                     graph = torch.cuda.CUDAGraph()
-                    with torch.cuda.graph(graph):
+                    graph_pool = torch.cuda.graph_pool_handle()
+                    set_graph_pool_id(graph_pool)
+                    with torch.cuda.graph(graph, pool=graph_pool):
                         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

@@ -584,8 +584,9 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
     elif config.architectures[0] in (
-        "DeepseekV3ForCausalLM",
         "DeepseekV2ForCausalLM",
+        "DeepseekV3ForCausalLM",
+        "DeepseekV32ForCausalLM",
         "Glm4MoeForCausalLM",
     ):
         E = config.n_routed_experts

benchmarks/kernels/benchmark_reshape_and_cache.py

@@ -0,0 +1,174 @@
+# 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

@@ -9,6 +9,9 @@ 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 (
@@ -31,6 +34,8 @@ 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."""
@@ -38,6 +43,14 @@ 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)
 
@@ -65,27 +78,49 @@ 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):
-            ops.reshape_and_cache_flash(
-                key,
-                value,
-                key_cache,
-                value_cache,
-                slot_mapping,
-                kv_cache_dtype,
-                k_scale,
-                v_scale,
-            )
-        torch.cuda.synchronize()
+            function_under_test()
+            torch.cuda.synchronize()
         end = time.perf_counter()
         return (end - start) / n_iters
 
@@ -116,10 +151,16 @@ 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)"]))
 
 
@@ -151,6 +192,21 @@ 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/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -8,12 +8,16 @@ 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_block_fp8_matmul,
+    w8a8_triton_block_scaled_mm,
 )
 from vllm.triton_utils import triton
-from vllm.utils.deep_gemm import calc_diff, fp8_gemm_nt, per_block_cast_to_fp8
+from vllm.utils.deep_gemm import (
+    calc_diff,
+    fp8_gemm_nt,
+    get_col_major_tma_aligned_tensor,
+    per_block_cast_to_fp8,
+)
 
 
 def benchmark_shape(m: int,
@@ -59,7 +63,7 @@ def benchmark_shape(m: int,
 
     # === vLLM Triton Implementation ===
     def vllm_triton_gemm():
-        return w8a8_block_fp8_matmul(A_vllm,
+        return w8a8_triton_block_scaled_mm(A_vllm,
                                      B_vllm,
                                      A_scale_vllm,
                                      B_scale_vllm,

cmake/cpu_extension.cmake

@@ -101,6 +101,7 @@ 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)
@@ -177,8 +178,14 @@ 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 or ARMv8 support.")
+    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA, S390X ISA, ARMv8 or RISC-V support.")
 endif()
 
 #
@@ -258,7 +265,8 @@ set(VLLM_EXT_SRC
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
-    "csrc/cpu/torch_bindings.cpp")
+    "csrc/cpu/torch_bindings.cpp"
+    "csrc/moe/dynamic_4bit_int_moe_cpu.cpp")
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     set(VLLM_EXT_SRC

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
-        GIT_TAG a757314c04eedd166e329e846c820eb1bdd702de
+        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
+        GIT_TAG 5f65b85703c7ed75fda01e06495077caad207c3f
         GIT_PROGRESS TRUE
         CONFIGURE_COMMAND ""
         BUILD_COMMAND ""
@@ -33,23 +33,64 @@ message(STATUS "FlashMLA is available at ${flashmla_SOURCE_DIR}")
 # The FlashMLA kernels only work on hopper and require CUDA 12.3 or later.
 # Only build FlashMLA kernels if we are building for something compatible with 
 # sm90a
-cuda_archs_loose_intersection(FLASH_MLA_ARCHS "9.0a" "${CUDA_ARCHS}")
-if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3 AND FLASH_MLA_ARCHS)
+
+set(SUPPORT_ARCHS)
+if(${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")
+
     set(FlashMLA_SOURCES
-        ${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)
+        ${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
+    )
 
     set(FlashMLA_INCLUDES
+        ${flashmla_SOURCE_DIR}/csrc
+        ${flashmla_SOURCE_DIR}/csrc/sm90
         ${flashmla_SOURCE_DIR}/csrc/cutlass/include
-        ${flashmla_SOURCE_DIR}/csrc)
+        ${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
+    )
 
     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
@@ -60,8 +101,32 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3 AND 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 an empty target for setup.py when not targeting sm90a systems
+    # Create empty targets for setup.py when not targeting sm90a systems
     add_custom_target(_flashmla_C)
+    add_custom_target(_flashmla_extension_C)
 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 ee4d25bd84e0cbc7e0b9b9685085fd5db2dcb62a
+          GIT_TAG 4695e6bed5366c41e28c06cd86170166e4f43d00
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

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 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
+  # 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
   set(_CUDA_ARCHS)
   foreach(_arch ${_SRC_CUDA_ARCHS})
-    if(_arch MATCHES "\\a$")
+    if(_arch MATCHES "[af]$")
       list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
-      string(REPLACE "a" "" _base "${_arch}")
+      string(REGEX REPLACE "[af]$" "" _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/fp8/amd/quant_utils.cuh"
+  #include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
 typedef __hip_bfloat16 __nv_bfloat16;
 #else
-  #include "../quantization/fp8/nvidia/quant_utils.cuh"
+  #include "../quantization/w8a8/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 4 kv_splits. 
+    // there is more than 1 kv_splits. 
     // Discuss with NVIDIA how this can be fixed.
     if (B > 1) {
-      max_splits = min(2, max_splits);
+      max_splits = min(1, 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();
+      ).arrive_and_wait();
 
       return;
     }

csrc/cache.h

@@ -56,3 +56,11 @@ 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);

csrc/cache_kernels.cu

@@ -9,15 +9,14 @@
 #include "quantization/vectorization_utils.cuh"
 
 #ifdef USE_ROCM
-  #include "quantization/fp8/amd/quant_utils.cuh"
+  #include "quantization/w8a8/fp8/amd/quant_utils.cuh"
 #else
-  #include "quantization/fp8/nvidia/quant_utils.cuh"
+  #include "quantization/w8a8/fp8/nvidia/quant_utils.cuh"
 #endif
 
 #include <algorithm>
 #include <cassert>
-#include <map>
-#include <vector>
+#include <cfloat>
 
 #ifdef USE_ROCM
   #include <hip/hip_bf16.h>
@@ -209,6 +208,20 @@ 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]
@@ -224,59 +237,51 @@ __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 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 h_block_idx = threadIdx.x;
+  if (h_block_idx >= num_heads * h_block_count) {
+    return;
+  }
 
-    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 int head_idx = h_block_idx / h_block_count;
+  const int h_block = h_block_idx % h_block_count;
 
-    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);
-    }
+  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]);
   }
 }
 
-// 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]
@@ -396,6 +401,177 @@ __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;
+  }
+}
+
 }  // namespace vllm
 
 // KV_T is the data type of key and value tensors.
@@ -431,14 +607,15 @@ 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_size, 512));
+  dim3 block(std::min(num_heads * head_div_x, 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.
@@ -509,6 +686,18 @@ 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]
@@ -531,20 +720,43 @@ void concat_and_cache_mla(
   int pe_dim = k_pe.size(1);
   int block_size = kv_cache.size(1);
 
-  TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
+  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);
+  }
 
   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();
 
-  DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
-                             CALL_CONCAT_AND_CACHE_MLA);
+  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);
+  }
 }
 
 namespace vllm {
@@ -922,3 +1134,42 @@ 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);
+}

csrc/core/batch_invariant.hpp

@@ -0,0 +1,16 @@
+#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() {
+  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;
+}
+
+}  // namespace vllm

csrc/cpu/cpu_types.hpp

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

csrc/cpu/cpu_types_scalar.hpp

@@ -0,0 +1,513 @@
+#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/float_convert.hpp

@@ -0,0 +1,106 @@
+
+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

@@ -88,8 +88,18 @@ 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("

csrc/cub_helpers.h

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

csrc/launch_bounds_utils.h

@@ -0,0 +1,64 @@
+#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

@@ -1,6 +1,7 @@
 #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>
@@ -413,7 +414,9 @@ 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;
-  if (ptrs_are_aligned && offsets_are_multiple_of_vector_width) {
+  bool batch_invariant_launch = vllm::vllm_kernel_override_batch_invariant();
+  if (ptrs_are_aligned && offsets_are_multiple_of_vector_width &&
+      !batch_invariant_launch) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
     LAUNCH_FUSED_ADD_RMS_NORM(0);
@@ -459,7 +462,8 @@ 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;
-  if (ptrs_are_aligned && hidden_size % 8 == 0) {
+  bool batch_invariant_launch = vllm::vllm_kernel_override_batch_invariant();
+  if (ptrs_are_aligned && hidden_size % 8 == 0 && !batch_invariant_launch) {
     LAUNCH_FUSED_POLY_NORM(8);
   } else {
     LAUNCH_FUSED_POLY_NORM(0);

csrc/layernorm_quant_kernels.cu

@@ -6,9 +6,10 @@
  */
 
 #include "type_convert.cuh"
-#include "quantization/fp8/common.cuh"
+#include "quantization/w8a8/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>
@@ -240,7 +241,9 @@ 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;
-  if (ptrs_are_aligned && hidden_size % 8 == 0 && input_stride % 8 == 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) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
     LAUNCH_FUSED_ADD_RMS_NORM(0);

csrc/moe/dynamic_4bit_int_moe_cpu.cpp

@@ -0,0 +1,156 @@
+#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

@@ -418,6 +418,15 @@ __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,
@@ -533,7 +542,7 @@ __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 && cuda::std::isfinite(group_scores[lane_id])) {
+    if (lane_id < n_group && is_finite(group_scores[lane_id])) {
       value = group_scores[lane_id];
     }
 
@@ -568,11 +577,10 @@ __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) &&
-                      cuda::std::isfinite(scores_with_bias[offset + i])
-                  ? scores_with_bias[offset + i]
-                  : neg_inf<T>();
+          T candidates = (i < num_experts_per_group) &&
+                                 is_finite(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/moe_align_sum_kernels.cu

@@ -44,6 +44,9 @@ __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);
@@ -95,12 +98,15 @@ 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) {
+    size_t numel, int32_t num_experts) {
   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;
   }
@@ -269,7 +275,7 @@ 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());
+              cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel(), num_experts);
         }
       });
 }

csrc/moe/topk_softmax_kernels.cu

@@ -21,6 +21,7 @@
 #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))
@@ -405,7 +406,8 @@ 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 int num_warps = (num_rows + ROWS_PER_WARP - 1) / 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_blocks = (num_warps + WARPS_PER_TB - 1) / WARPS_PER_TB;
 
     dim3 block_dim(WARP_SIZE_PARAM, WARPS_PER_TB);

csrc/ops.h

@@ -328,6 +328,12 @@ 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/fp8/common.cuh"
+#include "quantization/w8a8/fp8/common.cuh"
 
 #include <c10/util/Float8_e4m3fn.h>
 
@@ -23,9 +23,14 @@
 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"

csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu

@@ -26,6 +26,7 @@
 #include "dispatch_utils.h"
 
 #include "cuda_utils.h"
+#include "launch_bounds_utils.h"
 #include "nvfp4_utils.cuh"
 
 namespace vllm {
@@ -63,7 +64,7 @@ __inline__ __device__ PackedVec<Type> compute_silu_mul(PackedVec<Type>& vec,
 
 // Use UE4M3 by default.
 template <class Type, bool UE8M0_SF = false>
-__global__ void __launch_bounds__(1024, 4)
+__global__ void __launch_bounds__(1024, VLLM_BLOCKS_PER_SM(1024))
     silu_mul_cvt_fp16_to_fp4(int32_t numRows, int32_t numCols, Type const* in,
                              float const* SFScale, uint32_t* out,
                              uint32_t* SFout) {
@@ -131,7 +132,8 @@ void silu_and_mul_nvfp4_quant_sm1xxa(torch::Tensor& output,  // [..., d]
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   auto stream = at::cuda::getCurrentCUDAStream(input.get_device());
   dim3 block(std::min(int(n / ELTS_PER_THREAD), 1024));
-  int const numBlocksPerSM = 2048 / block.x;
+  int const numBlocksPerSM =
+      vllm_runtime_blocks_per_sm(static_cast<int>(block.x));
   dim3 grid(std::min(int(m), multiProcessorCount * numBlocksPerSM));
 
   VLLM_DISPATCH_HALF_TYPES(

csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu

@@ -14,6 +14,8 @@
  * limitations under the License.
  */
 
+#include "core/registration.h"
+
 #include <torch/all.h>
 #include <cutlass/arch/arch.h>
 
@@ -418,3 +420,7 @@ void cutlass_fp4_group_mm(
       "12.8 or above.");
 #endif
 }
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("cutlass_fp4_group_mm", &cutlass_fp4_group_mm);
+}

csrc/quantization/fp4/nvfp4_experts_quant.cu

@@ -26,12 +26,13 @@
 #include "dispatch_utils.h"
 
 #include "nvfp4_utils.cuh"
+#include "launch_bounds_utils.h"
 
 namespace vllm {
 
 // Use UE4M3 by default.
 template <class Type, bool UE8M0_SF = false, bool SMALL_NUM_EXPERTS = false>
-__global__ void __launch_bounds__(512, 4)
+__global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
     cvt_fp16_to_fp4(int32_t numRows, int32_t numCols, Type const* in,
                     float const* SFScale, uint32_t* out, uint32_t* SFout,
                     uint32_t* input_offset_by_experts,
@@ -129,7 +130,7 @@ __global__ void __launch_bounds__(512, 4)
 
 // Kernel for LARGE_M_TOPK = true (large m_topk optimized version)
 template <class Type, bool UE8M0_SF = false, bool SMALL_NUM_EXPERTS = false>
-__global__ void __launch_bounds__(1024, 4)
+__global__ void __launch_bounds__(1024, VLLM_BLOCKS_PER_SM(1024))
     cvt_fp16_to_fp4(int32_t numRows, int32_t numCols, Type const* in,
                     float const* SFScale, uint32_t* out, uint32_t* SFout,
                     uint32_t* input_offset_by_experts,
@@ -233,8 +234,9 @@ void quant_impl(void* output, void* output_scale, void* input,
   int const workSizePerRow = k / ELTS_PER_THREAD;
   int const totalWorkSize = m_topk * workSizePerRow;
   dim3 block(std::min(workSizePerRow, 512));
-  // Get number of blocks per SM (assume we can fully utilize the SM).
-  int const numBlocksPerSM = 2048 / block.x;
+  // Get number of blocks per SM
+  int const numBlocksPerSM =
+      vllm_runtime_blocks_per_sm(static_cast<int>(block.x));
   dim3 grid(std::min(static_cast<int>((totalWorkSize + block.x - 1) / block.x),
                      multiProcessorCount * numBlocksPerSM));
   while (grid.x <= multiProcessorCount && block.x > 64) {

csrc/quantization/fp4/nvfp4_quant_kernels.cu

@@ -26,13 +26,14 @@
 #include "dispatch_utils.h"
 
 #include "cuda_utils.h"
+#include "launch_bounds_utils.h"
 #include "nvfp4_utils.cuh"
 
 namespace vllm {
 
 // Use UE4M3 by default.
 template <class Type, bool UE8M0_SF = false>
-__global__ void __launch_bounds__(512, 4)
+__global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
     cvt_fp16_to_fp4(int32_t numRows, int32_t numCols, Type const* in,
                     float const* SFScale, uint32_t* out, uint32_t* SFout) {
   using PackedVec = PackedVec<Type>;
@@ -75,8 +76,9 @@ void invokeFP4Quantization(int m, int n, T const* input, float const* SFScale,
   // Grid, Block size.
   // Each thread converts 8 values.
   dim3 block(std::min(int(n / ELTS_PER_THREAD), 512));
-  // Get number of blocks per SM (assume we can fully utilize the SM).
-  int const numBlocksPerSM = 2048 / block.x;
+  // Get number of blocks per SM
+  int const numBlocksPerSM =
+      vllm_runtime_blocks_per_sm(static_cast<int>(block.x));
   dim3 grid(std::min(int(m), multiProcessorCount * numBlocksPerSM));
 
   // Launch the cvt kernel.

csrc/quantization/fused_kernels/quant_conversions.cuh

@@ -6,7 +6,7 @@
 
 #include "quantization/vectorization.cuh"
 // TODO(luka/varun):refactor common.cuh to use this file instead
-#include "quantization/fp8/common.cuh"
+#include "quantization/w8a8/fp8/common.cuh"
 
 namespace vllm {
 

csrc/quantization/w8a8/cutlass/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::NoSmemWarpSpecialized1Sm,
+            Shape<_1, _1, _1>, cutlass::epilogue::BlockwiseNoSmemWarpSpecialized1Sm,
             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::NoSmemWarpSpecialized1Sm,
+            Shape<_1, _1, _1>, cutlass::epilogue::BlockwiseNoSmemWarpSpecialized1Sm,
             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::NoSmemWarpSpecialized2Sm,
+            Shape<_2, _1, _1>, cutlass::epilogue::BlockwiseNoSmemWarpSpecialized2Sm,
             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::NoSmemWarpSpecialized1Sm,
+        Shape<_1, _1, _1>, cutlass::epilogue::BlockwiseNoSmemWarpSpecialized1Sm,
         cutlass::gemm::KernelTmaWarpSpecializedBlockwise1SmSm100, true>>(
         out, a, b, a_scales, b_scales);
   }
 }
 
-}  // namespace vllm
+}  // namespace vllm

csrc/quantization/w8a8/cutlass/c3x/scaled_mm_helper.hpp

@@ -25,7 +25,10 @@ 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 {
-        TORCH_CHECK(false, "Int8 not supported for this architecture");
+        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).");
       }
     }
   } else {

csrc/quantization/w8a8/cutlass/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

csrc/quantization/w8a8/cutlass/scaled_mm_entry.cu

@@ -67,8 +67,9 @@ void cutlass_scaled_mm_sm100(torch::Tensor& c, torch::Tensor const& a,
                              std::optional<torch::Tensor> const& bias);
 #endif
 
-#if defined(ENABLE_SCALED_MM_SM90) && ENABLE_SCALED_MM_SM90 || \
-    defined(ENABLE_SCALED_MM_SM100) && ENABLE_SCALED_MM_SM100
+#if defined(ENABLE_SCALED_MM_SM90) && ENABLE_SCALED_MM_SM90 ||   \
+    defined(ENABLE_SCALED_MM_SM100) && ENABLE_SCALED_MM_SM100 || \
+    defined(ENABLE_SCALED_MM_SM120) && ENABLE_SCALED_MM_SM120
 void get_cutlass_moe_mm_data_caller(
     const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
     torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
@@ -253,7 +254,7 @@ void cutlass_moe_mm(
     bool per_act_token, bool per_out_ch) {
   int32_t version_num = get_sm_version_num();
 #if defined ENABLE_CUTLASS_MOE_SM100 && ENABLE_CUTLASS_MOE_SM100
-  if (version_num >= 100) {
+  if (version_num >= 100 && version_num < 110) {
     cutlass_moe_mm_sm100(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
                          expert_offsets, problem_sizes, a_strides, b_strides,
                          c_strides, per_act_token, per_out_ch);
@@ -261,7 +262,7 @@ void cutlass_moe_mm(
   }
 #endif
 #if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
-  if (version_num >= 90) {
+  if (version_num >= 90 && version_num < 100) {
     cutlass_moe_mm_sm90(out_tensors, a_tensors, b_tensors, a_scales, b_scales,
                         expert_offsets, problem_sizes, a_strides, b_strides,
                         c_strides, per_act_token, per_out_ch);

csrc/quantization/w8a8/fp8/amd/quant_utils.cuh

@@ -5,7 +5,7 @@
 #include <hip/hip_bf16.h>
 #include <hip/hip_bfloat16.h>
 
-#include "../../../attention/attention_dtypes.h"
+#include "../../../../attention/attention_dtypes.h"
 
 namespace vllm {
 #ifdef USE_ROCM