Merge branch 'main' into fix_hang

Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>

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

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

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

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

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

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

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

@@ -42,7 +42,7 @@ docker run \
     pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
     pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
     pytest -v -s v1/structured_output
-    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py
+    pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_eagle.py --ignore=v1/spec_decode/test_tree_attention.py
     pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_shared_storage_connector.py
     pytest -v -s v1/test_metrics
     pytest -v -s v1/test_serial_utils.py

.buildkite/scripts/run-benchmarks.sh

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

.buildkite/test-pipeline.yaml

@@ -50,28 +50,19 @@ steps:
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
-  - tests/multimodal
-  - tests/utils_
-  commands:
-  - pytest -v -s -m 'not cpu_test' multimodal
-  - pytest -v -s utils_
-
-- label: Async Engine, Inputs, Utils, Worker Test (CPU) # 4 mins
-  timeout_in_minutes: 10
-  source_file_dependencies:
-  - vllm/
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/multimodal
+  - tests/utils_
   - tests/standalone_tests/lazy_imports.py
   - tests/transformers_utils
-  no_gpu: true
   commands:
   - python3 standalone_tests/lazy_imports.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
-  - pytest -v -s -m 'cpu_test' multimodal
-  - pytest -v -s transformers_utils
+  - pytest -v -s multimodal
+  - pytest -v -s utils_ # Utils
+  - pytest -v -s transformers_utils # transformers_utils
 
 - label: Python-only Installation Test # 10min
   timeout_in_minutes: 20
@@ -296,34 +287,23 @@ 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 -m 'not cpu_test' v1/kv_connector/unit
-    - pytest -v -s -m 'not cpu_test' v1/metrics
+    - pytest -v -s v1/kv_connector/unit
+    - pytest -v -s v1/metrics
     - pytest -v -s v1/test_oracle.py
     - pytest -v -s v1/test_request.py
+    - pytest -v -s v1/test_serial_utils.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]
@@ -553,17 +533,10 @@ steps:
   source_file_dependencies:
     - vllm/
     - tests/tool_use
+    - tests/mistral_tool_use
   commands:
-    - pytest -v -s -m 'not cpu_test' tool_use
-
-- label: OpenAI-Compatible Tool Use (CPU) # 5 mins
-  timeout_in_minutes: 10
-  source_file_dependencies:
-    - vllm/
-    - tests/tool_use
-  no_gpu: true
-  commands:
-    - pytest -v -s -m 'cpu_test' tool_use
+    - pytest -v -s tool_use
+    - pytest -v -s mistral_tool_use
 
 #####  models test  #####
 
@@ -603,19 +576,13 @@ steps:
   - vllm/
   - tests/models/test_transformers.py
   - tests/models/test_registry.py
-  commands:
-    - pytest -v -s models/test_transformers.py models/test_registry.py
-
-- label: Basic Models Test (Other CPU) # 5min
-  timeout_in_minutes: 10
-  torch_nightly: true
-  source_file_dependencies:
-  - vllm/
   - tests/models/test_utils.py
   - tests/models/test_vision.py
-  no_gpu: true
   commands:
-    - pytest -v -s models/test_utils.py models/test_vision.py
+    - pytest -v -s models/test_transformers.py \
+                   models/test_registry.py \
+                   models/test_utils.py \
+                   models/test_vision.py
 
 - label: Language Models Tests (Standard)
   timeout_in_minutes: 25
@@ -785,7 +752,6 @@ 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
@@ -845,7 +811,7 @@ steps:
   - vllm/v1/attention/backends/flashinfer.py
   commands:
     - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
+    - 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'
 
 - label: Blackwell Quantized MoE Test
   timeout_in_minutes: 60

.github/mergify.yml

@@ -2,7 +2,6 @@ pull_request_rules:
 - name: label-documentation
   description: Automatically apply documentation label
   conditions:
-    - label != stale
     - or:
       - files~=^[^/]+\.md$
       - files~=^docs/
@@ -15,7 +14,6 @@ pull_request_rules:
 - name: label-ci-build
   description: Automatically apply ci/build label
   conditions:
-    - label != stale
     - or:
       - files~=^\.github/
       - files~=\.buildkite/
@@ -32,7 +30,6 @@ pull_request_rules:
 - name: label-deepseek
   description: Automatically apply deepseek label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*deepseek.*\.py
       - files~=^tests/.*deepseek.*\.py
@@ -49,7 +46,6 @@ pull_request_rules:
 - name: label-frontend
   description: Automatically apply frontend label
   conditions:
-    - label != stale
     - files~=^vllm/entrypoints/
   actions:
     label:
@@ -59,7 +55,6 @@ pull_request_rules:
 - name: label-llama
   description: Automatically apply llama label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*llama.*\.py
       - files~=^tests/.*llama.*\.py
@@ -75,7 +70,6 @@ pull_request_rules:
 - name: label-multi-modality
   description: Automatically apply multi-modality label
   conditions:
-    - label != stale
     - or:
       - files~=^vllm/multimodal/
       - files~=^tests/multimodal/
@@ -89,7 +83,6 @@ pull_request_rules:
 - name: label-new-model
   description: Automatically apply new-model label
   conditions:
-    - label != stale
     - and:
       - files~=^vllm/model_executor/models/
       - files=vllm/model_executor/models/registry.py
@@ -101,7 +94,6 @@ pull_request_rules:
 - name: label-performance
   description: Automatically apply performance label
   conditions:
-    - label != stale
     - or:
       - files~=^benchmarks/
       - files~=^vllm/benchmarks/
@@ -115,7 +107,6 @@ pull_request_rules:
 - name: label-qwen
   description: Automatically apply qwen label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*qwen.*\.py
       - files~=^tests/.*qwen.*\.py
@@ -130,7 +121,6 @@ pull_request_rules:
 - name: label-gpt-oss
   description: Automatically apply gpt-oss label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/.*gpt[-_]?oss.*\.py
       - files~=^tests/.*gpt[-_]?oss.*\.py
@@ -152,7 +142,6 @@ pull_request_rules:
 - name: label-rocm
   description: Automatically apply rocm label
   conditions:
-    - label != stale
     - or:
       - files~=^csrc/rocm/
       - files~=^docker/Dockerfile.rocm
@@ -173,7 +162,6 @@ pull_request_rules:
 - name: label-structured-output
   description: Automatically apply structured-output label
   conditions:
-    - label != stale
     - or:
       - files~=^benchmarks/structured_schemas/
       - files=benchmarks/benchmark_serving_structured_output.py
@@ -193,7 +181,6 @@ pull_request_rules:
 - name: label-speculative-decoding
   description: Automatically apply speculative-decoding label
   conditions:
-    - label != stale
     - or:
       - files~=^vllm/v1/spec_decode/
       - files~=^tests/v1/spec_decode/
@@ -209,7 +196,6 @@ pull_request_rules:
 - name: label-v1
   description: Automatically apply v1 label
   conditions:
-    - label != stale
     - or:
       - files~=^vllm/v1/
       - files~=^tests/v1/
@@ -222,7 +208,6 @@ pull_request_rules:
   description: Automatically apply tpu label
   # Keep this list in sync with `label-tpu-remove` conditions
   conditions:
-    - label != stale
     - or:
       - files~=tpu.py
       - files~=_tpu
@@ -238,7 +223,6 @@ pull_request_rules:
   description: Automatically remove tpu label
   # Keep this list in sync with `label-tpu` conditions
   conditions:
-    - label != stale
     - and:
       - -files~=tpu.py
       - -files~=_tpu
@@ -253,9 +237,9 @@ pull_request_rules:
 - name: label-tool-calling
   description: Automatically add tool-calling label
   conditions:
-    - label != stale
     - or:
       - files~=^tests/tool_use/
+      - files~=^tests/mistral_tool_use/
       - files~=^tests/entrypoints/openai/tool_parsers/
       - files=tests/entrypoints/openai/test_chat_with_tool_reasoning.py
       - files~=^vllm/entrypoints/openai/tool_parsers/
@@ -272,9 +256,8 @@ pull_request_rules:
 
 - name: ping author on conflicts and add 'needs-rebase' label
   conditions:
-    - label != stale
-    - conflict
-    - -closed
+      - conflict
+      - -closed
   actions:
     label:
       add:
@@ -288,8 +271,6 @@ 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
@@ -301,7 +282,6 @@ pull_request_rules:
 
 - name: assign reviewer for modelopt changes
   conditions:
-    - label != stale
     - or:
         - files~=^vllm/model_executor/layers/quantization/modelopt\.py$
         - files~=^vllm/model_executor/layers/quantization/__init__\.py$
@@ -316,8 +296,8 @@ pull_request_rules:
 
 - name: remove 'needs-rebase' label when conflict is resolved
   conditions:
-    - -conflict
-    - -closed
+      - -conflict
+      - -closed
   actions:
     label:
       remove:
@@ -326,7 +306,6 @@ pull_request_rules:
 - name: label-kv-connector
   description: Automatically apply kv-connector label
   conditions:
-    - label != stale
     - or:
       - files~=^examples/online_serving/disaggregated[^/]*/.*
       - files~=^examples/offline_inference/disaggregated[^/]*/.*

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

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

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

benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh

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

benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh

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

benchmarks/kernels/benchmark_moe.py

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

benchmarks/kernels/benchmark_reshape_and_cache.py

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

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -9,7 +9,7 @@ import torch
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8,
-    w8a8_triton_block_scaled_mm,
+    w8a8_block_fp8_matmul,
 )
 from vllm.triton_utils import triton
 from vllm.utils.deep_gemm import (
@@ -63,7 +63,7 @@ def benchmark_shape(m: int,
 
     # === vLLM Triton Implementation ===
     def vllm_triton_gemm():
-        return w8a8_triton_block_scaled_mm(A_vllm,
+        return w8a8_block_fp8_matmul(A_vllm,
                                      B_vllm,
                                      A_scale_vllm,
                                      B_scale_vllm,

cmake/external_projects/flashmla.cmake

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

cmake/external_projects/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 4695e6bed5366c41e28c06cd86170166e4f43d00
+          GIT_TAG ee4d25bd84e0cbc7e0b9b9685085fd5db2dcb62a
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

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

csrc/attention/attention_kernels.cuh

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

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

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

csrc/cache.h

@@ -56,11 +56,3 @@ void cp_gather_cache(
     torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
     torch::Tensor const& cu_seq_lens,  // [BATCH+1]
     int64_t batch_size, std::optional<torch::Tensor> seq_starts = std::nullopt);
-
-// Indexer K quantization and cache function
-void indexer_k_quant_and_cache(
-    torch::Tensor& k,             // [num_tokens, head_dim]
-    torch::Tensor& kv_cache,      // [num_blocks, block_size, cache_stride]
-    torch::Tensor& slot_mapping,  // [num_tokens]
-    int64_t quant_block_size,     // quantization block size
-    const std::string& scale_fmt);

csrc/cache_kernels.cu

@@ -9,14 +9,15 @@
 #include "quantization/vectorization_utils.cuh"
 
 #ifdef USE_ROCM
-  #include "quantization/w8a8/fp8/amd/quant_utils.cuh"
+  #include "quantization/fp8/amd/quant_utils.cuh"
 #else
-  #include "quantization/w8a8/fp8/nvidia/quant_utils.cuh"
+  #include "quantization/fp8/nvidia/quant_utils.cuh"
 #endif
 
 #include <algorithm>
 #include <cassert>
-#include <cfloat>
+#include <map>
+#include <vector>
 
 #ifdef USE_ROCM
   #include <hip/hip_bf16.h>
@@ -208,20 +209,6 @@ void copy_blocks_mla(std::vector<torch::Tensor> const& kv_caches,
 
 namespace vllm {
 
-// Used to copy/convert one element
-template <typename OutT, typename InT, Fp8KVCacheDataType kv_dt>
-struct CopyWithScaleOp {
-  float scale;
-
-  __device__ __forceinline__ void operator()(OutT& dst, const InT src) const {
-    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
-      dst = static_cast<OutT>(src);
-    } else {
-      dst = fp8::scaled_convert<OutT, InT, kv_dt>(src, scale);
-    }
-  }
-};
-
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
 __global__ void reshape_and_cache_kernel(
     const scalar_t* __restrict__ key,    // [num_tokens, num_heads, head_size]
@@ -237,51 +224,59 @@ __global__ void reshape_and_cache_kernel(
   const int64_t token_idx = blockIdx.x;
   const int64_t slot_idx = slot_mapping[token_idx];
   if (slot_idx < 0) {
+    // Padding token that should be ignored.
     return;
   }
 
   const int64_t block_idx = slot_idx / block_size;
   const int64_t block_offset = slot_idx % block_size;
-  const int h_block_count = head_size / x;  // head_size//x
 
-  const int h_block_idx = threadIdx.x;
-  if (h_block_idx >= num_heads * h_block_count) {
-    return;
-  }
+  const int n = num_heads * head_size;
+  for (int i = threadIdx.x; i < n; i += blockDim.x) {
+    const int64_t src_key_idx = token_idx * key_stride + i;
+    const int64_t src_value_idx = token_idx * value_stride + i;
 
-  const int head_idx = h_block_idx / h_block_count;
-  const int h_block = h_block_idx % h_block_count;
+    const int head_idx = i / head_size;
+    const int head_offset = i % head_size;
+    const int x_idx = head_offset / x;
+    const int x_offset = head_offset % x;
 
-  const scalar_t* __restrict__ key_src =
-      key + token_idx * key_stride + head_idx * head_size + h_block * x;
-  const int64_t src_value_start =
-      token_idx * value_stride + head_idx * head_size + h_block * x;
-
-  cache_t* __restrict__ key_dst =
-      key_cache + block_idx * num_heads * h_block_count * block_size * x +
-      head_idx * h_block_count * block_size * x + h_block * block_size * x +
-      block_offset * x;
-  const int64_t tgt_value_start =
-      block_idx * num_heads * h_block_count * x * block_size +
-      head_idx * h_block_count * x * block_size + h_block * x * block_size +
-      block_offset;
-
-  constexpr int VEC_SIZE = (sizeof(scalar_t) == 2) ? 8 : 4;
-  float k_scale_val = (kv_dt == Fp8KVCacheDataType::kAuto) ? 0.f : *k_scale;
-  CopyWithScaleOp<cache_t, scalar_t, kv_dt> k_op{k_scale_val};
-  float v_scale_val = (kv_dt == Fp8KVCacheDataType::kAuto) ? 0.f : *v_scale;
-  CopyWithScaleOp<cache_t, scalar_t, kv_dt> v_op{v_scale_val};
-
-  vectorize_with_alignment<VEC_SIZE>(key_src, key_dst, x, 0, 1, k_op);
-
-  const scalar_t* __restrict__ value_src = value + src_value_start;
-  cache_t* __restrict__ value_dst = value_cache + tgt_value_start;
-#pragma unroll
-  for (int i = 0; i < x; i++) {
-    v_op(value_dst[i * block_size], value_src[i]);
+    const int64_t tgt_key_idx =
+        block_idx * num_heads * (head_size / x) * block_size * x +
+        head_idx * (head_size / x) * block_size * x + x_idx * block_size * x +
+        block_offset * x + x_offset;
+    const int64_t tgt_value_idx =
+        block_idx * num_heads * head_size * block_size +
+        head_idx * head_size * block_size + head_offset * block_size +
+        block_offset;
+    scalar_t tgt_key = key[src_key_idx];
+    scalar_t tgt_value = value[src_value_idx];
+    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
+      key_cache[tgt_key_idx] = tgt_key;
+      value_cache[tgt_value_idx] = tgt_value;
+    } else {
+      key_cache[tgt_key_idx] =
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, *k_scale);
+      value_cache[tgt_value_idx] =
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, *v_scale);
+    }
   }
 }
 
+// Used by vectorization_utils to copy/convert one element
+template <typename OutT, typename InT, Fp8KVCacheDataType kv_dt>
+struct CopyWithScaleOp {
+  float scale;
+
+  __device__ __forceinline__ void operator()(OutT& dst, const InT src) const {
+    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
+      dst = static_cast<OutT>(src);
+    } else {
+      dst = fp8::scaled_convert<OutT, InT, kv_dt>(src, scale);
+    }
+  }
+};
+
 template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
 __global__ void reshape_and_cache_flash_kernel(
     const scalar_t* __restrict__ key,    // [num_tokens, num_heads, head_size]
@@ -401,177 +396,6 @@ __global__ void concat_and_cache_mla_kernel(
   copy(k_pe, kv_cache, k_pe_stride, block_stride, pe_dim, kv_lora_rank);
 }
 
-template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
-__global__ void concat_and_cache_ds_mla_kernel(
-    const scalar_t* __restrict__ kv_c,  // [num_tokens, kv_lora_rank]
-    const scalar_t* __restrict__ k_pe,  // [num_tokens, pe_dim]
-    cache_t* __restrict__ kv_cache,  // [num_blocks, block_size, (kv_lora_rank
-                                     // + pe_dim)]
-    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
-    const int block_stride,                    //
-    const int entry_stride,                    //
-    const int kv_c_stride,                     //
-    const int k_pe_stride,                     //
-    const int kv_lora_rank,                    //
-    const int pe_dim,                          //
-    const int block_size,                      //
-    const float* scale                         //
-) {
-  const int64_t token_idx = blockIdx.x;
-  const int64_t slot_idx = slot_mapping[token_idx];
-  // NOTE: slot_idx can be -1 if the token is padded
-  if (slot_idx < 0) {
-    return;
-  }
-  const int64_t block_idx = slot_idx / block_size;
-  const int64_t block_offset = slot_idx % block_size;
-  const int64_t dst_idx_start =
-      block_idx * block_stride + block_offset * entry_stride;
-
-  // For the NoPE part, each tile of 128 elements is handled by half of one warp
-  // (16 threads). There are 4 total tiles, so 2 warps (64 threads).
-  // Lanes 0 and 16 of each warp write the scale values for that warp's tiles.
-  // The RoPE part (last 64 elements) is handled by another 1 warp (32 threads).
-  // So in total, we use 3 warps (96 threads) per block.
-
-  // Cast kv_cache to 16_bit for RoPE values
-  scalar_t* kv_cache_16bit =
-      reinterpret_cast<scalar_t*>(&kv_cache[dst_idx_start]);
-
-  // The last warp handles the RoPE part
-  if (threadIdx.x >= 64) {
-    // Each thread handles two elements of RoPE
-    const int8_t pe_idx_start = (threadIdx.x - 64) * 2;
-    const int64_t src_idx = token_idx * k_pe_stride + pe_idx_start;
-    // Vectorized load of two 16-bit values, performed as one 32-bit load
-    const int32_t vals = *reinterpret_cast<const int32_t*>(&k_pe[src_idx]);
-    // RoPE values start after the packed 8-bit NoPE values and the
-    // 32-bit scales
-    const int64_t dst_idx = kv_lora_rank / 2 + 8 + pe_idx_start;
-    // Vectorized store of two 16-bit values, performed as one 32-bit store
-    *reinterpret_cast<int32_t*>(&kv_cache_16bit[dst_idx]) = vals;
-    return;
-  }
-
-  // The first two warps handle the NoPE part
-  const int8_t warp_idx = threadIdx.x >> 5;
-  const int8_t lane_idx = threadIdx.x & 31;
-  const int8_t tile_idx = warp_idx * 2 + (lane_idx >> 4);
-
-  // Each thread handles 8 elements of NoPE
-  // Load the NoPE elements for this thread into registers
-  const int64_t src_idx_start = token_idx * kv_c_stride + (threadIdx.x * 8);
-  // Vectorized load of eight 16-bit values, performed as an int4 load
-  const int4 vals_i4 = *reinterpret_cast<const int4*>(&kv_c[src_idx_start]);
-  const scalar_t* vals = reinterpret_cast<const scalar_t*>(&vals_i4);
-
-  // Max absolute value of this thread's elements
-  float max_abs = fmaxf(fmaxf(fmaxf(fabsf(vals[0]), fabsf(vals[1])),
-                              fmaxf(fabsf(vals[2]), fabsf(vals[3]))),
-                        fmaxf(fmaxf(fabsf(vals[4]), fabsf(vals[5])),
-                              fmaxf(fabsf(vals[6]), fabsf(vals[7]))));
-
-  // Warp-level reduction to find the max absolute value in each half-warp
-#pragma unroll
-  for (int offset = 8; offset > 0; offset /= 2) {
-    max_abs = fmaxf(max_abs, VLLM_SHFL_XOR_SYNC_WIDTH(max_abs, offset, 16));
-  }
-
-  // Compute the scale for the tile
-  float tile_scale = max_abs / 448.f;
-  tile_scale = fmaxf(tile_scale, FLT_MIN);
-
-  // The first lane of each half-warp writes the scale to kv_cache
-  if ((lane_idx == 0) || (lane_idx == 16)) {
-    float* kv_cache_32bit = reinterpret_cast<float*>(&kv_cache[dst_idx_start]);
-    const uint64_t dst_idx = kv_lora_rank / 4 + tile_idx;
-    kv_cache_32bit[dst_idx] = tile_scale;
-  }
-
-  // Now all threads in the block scale and write their elements
-  // NoPE data is packed in the first kv_lora_rank/2 bytes (first 256 bytes)
-  const int64_t dst_idx_base = dst_idx_start + (threadIdx.x * 8);
-
-  uint8_t result[8];
-#pragma unroll
-  for (int i = 0; i < 8; i++) {
-    result[i] =
-        fp8::scaled_convert<uint8_t, scalar_t, Fp8KVCacheDataType::kFp8E4M3>(
-            vals[i], tile_scale);
-  }
-
-  // Store as aligned 64-bit writes
-  *reinterpret_cast<uint64_t*>(&kv_cache[dst_idx_base]) =
-      *reinterpret_cast<const uint64_t*>(result);
-}
-
-template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
-__global__ void indexer_k_quant_and_cache_kernel(
-    const scalar_t* __restrict__ k,  // [num_tokens, head_dim]
-    cache_t* __restrict__ kv_cache,  // [num_blocks, block_size, cache_stride]
-    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
-    const int head_dim,                        // dimension of each head
-    const int quant_block_size,                // quantization block size
-    const int cache_block_size,                // cache block size
-    const int cache_stride,  // stride for each token in kv_cache
-    const bool use_ue8m0     // use ue8m0 scale format
-) {
-  constexpr int VEC_SIZE = 4;
-  const int64_t token_idx = blockIdx.x;
-  const int64_t head_dim_idx = (blockIdx.y * blockDim.y * blockDim.x +
-                                threadIdx.y * blockDim.x + threadIdx.x) *
-                               VEC_SIZE;
-  const int64_t slot_idx = slot_mapping[token_idx];
-  const int64_t block_idx = slot_idx / cache_block_size;
-  const int64_t block_offset = slot_idx % cache_block_size;
-
-  // NOTE: slot_idx can be -1 if the token is padded
-  if (slot_idx < 0 || (head_dim_idx >= head_dim)) {
-    return;
-  }
-
-  float2 k_val = (reinterpret_cast<const float2*>(
-      k))[(token_idx * head_dim + head_dim_idx) / VEC_SIZE];
-  scalar_t* k_val_ptr = reinterpret_cast<scalar_t*>(&k_val);
-  float amax = 0.0f;
-  for (int i = 0; i < VEC_SIZE; i++) {
-    amax = fmaxf(amax, fabsf(float(k_val_ptr[i])));
-  }
-#ifndef USE_ROCM
-  __syncwarp();
-#endif
-
-  // Reduced amax
-  for (int mask = 16; mask > 0; mask /= 2) {
-#ifdef USE_ROCM
-    amax = fmaxf(amax, __shfl_xor_sync(uint64_t(-1), amax, mask));
-#else
-    amax = fmaxf(amax, __shfl_xor_sync(unsigned(-1), amax, mask));
-#endif
-  }
-#ifndef USE_ROCM
-  __syncwarp();
-#endif
-  float scale = fmaxf(amax, 1e-4) / 448.0f;
-  if (use_ue8m0) {
-    scale = exp2f(ceilf(log2f(scale)));
-  }
-
-  const int64_t dst_offset = block_idx * cache_block_size * cache_stride +
-                             block_offset * head_dim + head_dim_idx;
-  for (int i = 0; i < VEC_SIZE; i++) {
-    kv_cache[dst_offset + i] =
-        fp8::scaled_convert<cache_t, scalar_t, kv_dt>(k_val_ptr[i], scale);
-  }
-  if (threadIdx.x == 0) {
-    const int64_t dst_scale_idx =
-        block_idx * cache_block_size * cache_stride +
-        cache_block_size * head_dim +
-        (block_offset * head_dim + head_dim_idx) * 4 / quant_block_size;
-    reinterpret_cast<float*>(kv_cache)[dst_scale_idx / 4] = scale;
-  }
-}
-
 }  // namespace vllm
 
 // KV_T is the data type of key and value tensors.
@@ -607,15 +431,14 @@ void reshape_and_cache(
 
   int key_stride = key.stride(0);
   int value_stride = value.stride(0);
-  int head_div_x = head_size / x;
 
   dim3 grid(num_tokens);
-  dim3 block(std::min(num_heads * head_div_x, 512));
+  dim3 block(std::min(num_heads * head_size, 512));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(key));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
   DISPATCH_BY_KV_CACHE_DTYPE(key.dtype(), kv_cache_dtype,
-                             CALL_RESHAPE_AND_CACHE);
+                             CALL_RESHAPE_AND_CACHE)
 }
 
 // KV_T is the data type of key and value tensors.
@@ -686,18 +509,6 @@ void reshape_and_cache_flash(
           kv_c_stride, k_pe_stride, kv_lora_rank, pe_dim, block_size,   \
           reinterpret_cast<const float*>(scale.data_ptr()));
 
-// KV_T is the data type of key and value tensors.
-// CACHE_T is the stored data type of kv-cache.
-#define CALL_CONCAT_AND_CACHE_DS_MLA(KV_T, CACHE_T, KV_DTYPE)           \
-  vllm::concat_and_cache_ds_mla_kernel<KV_T, CACHE_T, KV_DTYPE>         \
-      <<<grid, block, 0, stream>>>(                                     \
-          reinterpret_cast<KV_T*>(kv_c.data_ptr()),                     \
-          reinterpret_cast<KV_T*>(k_pe.data_ptr()),                     \
-          reinterpret_cast<CACHE_T*>(kv_cache.data_ptr()),              \
-          slot_mapping.data_ptr<int64_t>(), block_stride, entry_stride, \
-          kv_c_stride, k_pe_stride, kv_lora_rank, pe_dim, block_size,   \
-          reinterpret_cast<const float*>(scale.data_ptr()));
-
 void concat_and_cache_mla(
     torch::Tensor& kv_c,          // [num_tokens, kv_lora_rank]
     torch::Tensor& k_pe,          // [num_tokens, pe_dim]
@@ -720,43 +531,20 @@ void concat_and_cache_mla(
   int pe_dim = k_pe.size(1);
   int block_size = kv_cache.size(1);
 
-  if (kv_cache_dtype == "fp8_ds_mla") {
-    TORCH_CHECK(kv_lora_rank == 512, "kv_lora_rank must be 512 for fp8_ds_mla");
-    TORCH_CHECK(pe_dim == 64, "pe_dim must be 64 for fp8_ds_mla");
-    TORCH_CHECK(kv_cache.size(2) == 656 / kv_cache.itemsize(),
-                "kv_cache.size(2) must be 656 bytes for fp8_ds_mla");
-    TORCH_CHECK(kv_c.itemsize() == 2,
-                "kv_c.itemsize() must be 2 for fp8_ds_mla");
-    TORCH_CHECK(k_pe.itemsize() == 2,
-                "k_pe.itemsize() must be 2 for fp8_ds_mla");
-  } else {
-    TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
-  }
+  TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
 
   int kv_c_stride = kv_c.stride(0);
   int k_pe_stride = k_pe.stride(0);
   int block_stride = kv_cache.stride(0);
   int entry_stride = kv_cache.stride(1);
 
+  dim3 grid(num_tokens);
+  dim3 block(std::min(kv_lora_rank, 512));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(kv_c));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  if (kv_cache_dtype == "fp8_ds_mla") {
-    dim3 grid(num_tokens);
-    // For the NoPE part, each tile of 128 elements is handled by half of one
-    // warp (16 threads). There are 4 total tiles, so 2 warps (64 threads).
-    // Lanes 0 and 16 of each warp write the scale values for that warp's tiles.
-    // The RoPE part (last 64 elements) is handled by another 1 warp (32
-    // threads). So in total, we use 3 warps (96 threads) per block.
-    dim3 block(96);
-    DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
-                               CALL_CONCAT_AND_CACHE_DS_MLA);
-  } else {
-    dim3 grid(num_tokens);
-    dim3 block(std::min(kv_lora_rank, 512));
-    DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
-                               CALL_CONCAT_AND_CACHE_MLA);
-  }
+  DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
+                             CALL_CONCAT_AND_CACHE_MLA);
 }
 
 namespace vllm {
@@ -1134,42 +922,3 @@ void cp_gather_cache(
     TORCH_CHECK(false, "Unsupported data type width: ", dtype_bits);
   }
 }
-
-// Macro to dispatch the kernel based on the data type.
-#define CALL_INDEXER_K_QUANT_AND_CACHE(KV_T, CACHE_T, KV_DTYPE)         \
-  vllm::indexer_k_quant_and_cache_kernel<KV_T, CACHE_T, KV_DTYPE>       \
-      <<<grid, block, 0, stream>>>(                                     \
-          reinterpret_cast<KV_T*>(k.data_ptr()),                        \
-          reinterpret_cast<CACHE_T*>(kv_cache.data_ptr()),              \
-          slot_mapping.data_ptr<int64_t>(), head_dim, quant_block_size, \
-          cache_block_size, cache_stride, use_ue8m0);
-
-void indexer_k_quant_and_cache(
-    torch::Tensor& k,             // [num_tokens, head_dim]
-    torch::Tensor& kv_cache,      // [num_blocks, block_size, cache_stride]
-    torch::Tensor& slot_mapping,  // [num_tokens]
-    int64_t quant_block_size,     // quantization block size
-    const std::string& scale_fmt) {
-  int num_tokens = k.size(0);
-  int head_dim = k.size(1);
-  int cache_block_size = kv_cache.size(1);
-  int cache_stride = kv_cache.size(2);
-  bool use_ue8m0 = scale_fmt == "ue8m0";
-
-  TORCH_CHECK(k.device() == kv_cache.device(),
-              "k and kv_cache must be on the same device");
-  TORCH_CHECK(k.device() == slot_mapping.device(),
-              "k and slot_mapping must be on the same device");
-  TORCH_CHECK(head_dim % quant_block_size == 0,
-              "head_dim must be divisible by quant_block_size");
-
-  constexpr int vec_size = 4;
-  dim3 grid(num_tokens, (head_dim + quant_block_size * vec_size - 1) /
-                            (quant_block_size * vec_size));
-  dim3 block(32, vec_size);
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(k));
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-
-  DISPATCH_BY_KV_CACHE_DTYPE(k.dtype(), "fp8_e4m3",
-                             CALL_INDEXER_K_QUANT_AND_CACHE);
-}

csrc/cub_helpers.h

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

csrc/launch_bounds_utils.h

@@ -8,37 +8,11 @@
   #define VLLM_LAUNCH_BLOCKS_CAP 4
 #endif
 
-// Compile-time estimate of max threads per SM for launch bounds.
-// Families: 1024, 1536, 2048 threads/SM.
+// compile-time estimate of max threads per SM for launch bounds.
 #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
-
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 300
+    #define VLLM_MAX_THREADS_PER_SM 1536
   #else
-  /* Host pass (no __CUDA_ARCH__): neutral default */
     #define VLLM_MAX_THREADS_PER_SM 2048
   #endif
 #endif

csrc/layernorm_quant_kernels.cu

@@ -6,7 +6,7 @@
  */
 
 #include "type_convert.cuh"
-#include "quantization/w8a8/fp8/common.cuh"
+#include "quantization/fp8/common.cuh"
 #include "dispatch_utils.h"
 #include "cub_helpers.h"
 #include "core/batch_invariant.hpp"

csrc/quantization/activation_kernels.cu

@@ -7,7 +7,7 @@
 #include "../cuda_compat.h"
 #include "dispatch_utils.h"
 
-#include "quantization/w8a8/fp8/common.cuh"
+#include "quantization/fp8/common.cuh"
 
 #include <c10/util/Float8_e4m3fn.h>
 

csrc/quantization/compressed_tensors/int8_quant_kernels.cu

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

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8_dispatch.cuh

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

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_helper.hpp

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

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8_dispatch.cuh

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

csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu

@@ -67,9 +67,8 @@ 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 || \
-    defined(ENABLE_SCALED_MM_SM120) && ENABLE_SCALED_MM_SM120
+#if defined(ENABLE_SCALED_MM_SM90) && ENABLE_SCALED_MM_SM90 || \
+    defined(ENABLE_SCALED_MM_SM100) && ENABLE_SCALED_MM_SM100
 void get_cutlass_moe_mm_data_caller(
     const torch::Tensor& topk_ids, torch::Tensor& expert_offsets,
     torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
@@ -254,7 +253,7 @@ 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 && version_num < 110) {
+  if (version_num >= 100) {
     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);
@@ -262,7 +261,7 @@ void cutlass_moe_mm(
   }
 #endif
 #if defined ENABLE_CUTLASS_MOE_SM90 && ENABLE_CUTLASS_MOE_SM90
-  if (version_num >= 90 && version_num < 100) {
+  if (version_num >= 90) {
     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/fp4/nvfp4_blockwise_moe_kernel.cu

@@ -14,8 +14,6 @@
  * limitations under the License.
  */
 
-#include "core/registration.h"
-
 #include <torch/all.h>
 #include <cutlass/arch/arch.h>
 
@@ -420,7 +418,3 @@ 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/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

csrc/quantization/fp8/common.cu

@@ -1,7 +1,7 @@
 #include "common.cuh"
 #include "dispatch_utils.h"
-#include "cub_helpers.h"
-#include "quantization/vectorization_utils.cuh"
+#include "../../cub_helpers.h"
+#include "../vectorization_utils.cuh"
 #include <c10/cuda/CUDAGuard.h>
 #include <ATen/cuda/Exceptions.h>
 

csrc/quantization/fp8/nvidia/quant_utils.cuh

@@ -1,6 +1,6 @@
 #pragma once
 
-#include "../../../../attention/attention_dtypes.h"
+#include "../../../attention/attention_dtypes.h"
 #include <assert.h>
 #include <float.h>
 #include <stdint.h>
@@ -576,17 +576,6 @@ __inline__ __device__ Tout scaled_convert(const Tin& x, const float scale) {
           TORCH_CHECK(false,                                                   \
                       "Unsupported input type of kv cache: ", SRC_DTYPE);      \
         }                                                                      \
-      } else if (KV_DTYPE == "fp8_ds_mla") {                                   \
-        if (SRC_DTYPE == at::ScalarType::Float) {                              \
-          FN(float, uint8_t, vllm::Fp8KVCacheDataType::kFp8E4M3);              \
-        } else if (SRC_DTYPE == at::ScalarType::Half) {                        \
-          FN(uint16_t, uint8_t, vllm::Fp8KVCacheDataType::kFp8E4M3);           \
-        } else if (SRC_DTYPE == at::ScalarType::BFloat16) {                    \
-          FN(__nv_bfloat16, uint8_t, vllm::Fp8KVCacheDataType::kFp8E4M3);      \
-        } else {                                                               \
-          TORCH_CHECK(false,                                                   \
-                      "Unsupported input type of kv cache: ", SRC_DTYPE);      \
-        }                                                                      \
       } else {                                                                 \
         TORCH_CHECK(false, "Unsupported data type of kv cache: ", KV_DTYPE);   \
       }                                                                        \

csrc/quantization/fp8/per_token_group_quant.cu

@@ -1,6 +1,6 @@
 #include <ATen/cuda/CUDAContext.h>
 
-#include "quantization/w8a8/per_token_group_quant_8bit.h"
+#include "../per_token_group_quant_8bit.h"
 
 #include <cmath>
 
@@ -8,9 +8,9 @@
 
 #include <torch/all.h>
 
-#include "quantization/vectorization.cuh"
-#include "quantization/vectorization_utils.cuh"
-#include "dispatch_utils.h"
+#include "../vectorization.cuh"
+#include "../vectorization_utils.cuh"
+#include "../../dispatch_utils.h"
 
 __device__ __forceinline__ float GroupReduceMax(float val) {
   unsigned mask = threadIdx.x % 32 >= 16 ? 0xffff0000 : 0x0000ffff;
@@ -212,4 +212,4 @@ void per_token_group_quant_fp8(const torch::Tensor& input,
                                double fp8_max, bool scale_ue8m0) {
   per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
                              fp8_min, fp8_max, scale_ue8m0);
-}
+}

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/w8a8/fp8/common.cuh"
+#include "quantization/fp8/common.cuh"
 
 namespace vllm {
 

csrc/quantization/per_token_group_quant_8bit.h

@@ -1,6 +1,7 @@
 #pragma once
 #include <torch/all.h>
 
+// TODO(wentao): refactor the folder to 8bit, then includes fp8 and int8 folders
 // 8-bit per-token-group quantization helper used by both FP8 and INT8
 void per_token_group_quant_8bit(const torch::Tensor& input,
                                 torch::Tensor& output_q,

csrc/quantization/w8a8/int8/per_token_group_quant.cu

@@ -1,12 +0,0 @@
-#include <ATen/cuda/CUDAContext.h>
-#include <torch/all.h>
-
-#include "quantization/w8a8/per_token_group_quant_8bit.h"
-
-void per_token_group_quant_int8(const torch::Tensor& input,
-                                torch::Tensor& output_q,
-                                torch::Tensor& output_s, int64_t group_size,
-                                double eps, double int8_min, double int8_max) {
-  per_token_group_quant_8bit(input, output_q, output_s, group_size, eps,
-                             int8_min, int8_max);
-}

csrc/rocm/attention.cu

@@ -23,7 +23,7 @@
 
 #include <algorithm>
 #include "../attention/dtype_fp8.cuh"
-#include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
+#include "../quantization/fp8/amd/quant_utils.cuh"
 
 // ROCm 6.2 compatibility: map OCP fp8 types to FNUZ variants if OCP is absent
 #if !defined(HIP_FP8_TYPE_OCP)
@@ -40,8 +40,7 @@ using __hip_fp8_e5m2 = __hip_fp8_e5m2_fnuz;
   #define __HIP__FP8MFMA__
 #endif
 
-#if defined(__HIPCC__) && (defined(__gfx1100__) || defined(__gfx1101__) || \
-                           defined(__gfx1150__) || defined(__gfx1151__))
+#if defined(__HIPCC__) && (defined(__gfx1100__) || defined(__gfx1101__))
   #define __HIP__GFX11__
 #endif
 

csrc/rocm/skinny_gemms.cu

@@ -11,7 +11,7 @@
 
 #include "../cuda_compat.h"
 #include "dispatch_utils.h"
-#include "quantization/w8a8/fp8/common.cuh"
+#include "quantization/fp8/common.cuh"
 
 #if defined(__HIPCC__) && \
     (defined(__gfx90a__) || defined(__gfx942__) || defined(__gfx950__))

csrc/torch_bindings.cpp

@@ -397,7 +397,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       " Tensor a_blockscale, Tensor b_blockscales, Tensor alphas,"
       " Tensor problem_sizes, Tensor expert_offsets, Tensor sf_offsets) -> ()",
       {stride_tag});
-  // conditionally compiled so impl registration is in source file
+  ops.impl("cutlass_fp4_group_mm", torch::kCUDA, &cutlass_fp4_group_mm);
 
   // CUTLASS w8a8 GEMM, supporting symmetric per-tensor or per-row/column
   // quantization, as well as bias
@@ -713,13 +713,6 @@ TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
       "cp_gather_cache(Tensor src_cache, Tensor! dst, Tensor block_table, "
       "Tensor cu_seq_lens, int batch_size, Tensor? seq_starts) -> ()");
   cache_ops.impl("cp_gather_cache", torch::kCUDA, &cp_gather_cache);
-
-  cache_ops.def(
-      "indexer_k_quant_and_cache(Tensor k, Tensor! kv_cache, Tensor "
-      "slot_mapping, "
-      "int quant_block_size, str kv_cache_dtype) -> ()");
-  cache_ops.impl("indexer_k_quant_and_cache", torch::kCUDA,
-                 &indexer_k_quant_and_cache);
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cuda_utils), cuda_utils) {

docker/Dockerfile

@@ -14,11 +14,6 @@ ARG PYTHON_VERSION=3.12
 #
 # Example:
 # docker build --build-arg BUILD_BASE_IMAGE=registry.acme.org/mirror/nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
-
-# Important: We build with an old version of Ubuntu to maintain broad 
-# compatibility with other Linux OSes. The main reason for this is that the
-# glibc version is baked into the distro, and binaries built with one glibc
-# version are not backwards compatible with OSes that use an earlier version.
 ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
 # TODO: Restore to base image after FlashInfer AOT wheel fixed
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
@@ -80,19 +75,34 @@ ARG TARGETPLATFORM
 ARG INSTALL_KV_CONNECTORS=false
 ENV DEBIAN_FRONTEND=noninteractive
 
+ARG DEADSNAKES_MIRROR_URL
+ARG DEADSNAKES_GPGKEY_URL
 ARG GET_PIP_URL
 
-# Install system dependencies and uv, then create Python virtual environment
+# Install Python and other dependencies
 RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
     && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
     && apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl sudo python3-pip \
-    && curl -LsSf https://astral.sh/uv/install.sh | sh \
-    && $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
-    && rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
-    && ln -s /opt/venv/bin/python3 /usr/bin/python3 \
-    && ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
-    && ln -s /opt/venv/bin/pip /usr/bin/pip \
+    && apt-get install -y ccache software-properties-common git curl sudo \
+    && if [ ! -z ${DEADSNAKES_MIRROR_URL} ] ; then \
+        if [ ! -z "${DEADSNAKES_GPGKEY_URL}" ] ; then \
+            mkdir -p -m 0755 /etc/apt/keyrings ; \
+            curl -L ${DEADSNAKES_GPGKEY_URL} | gpg --dearmor > /etc/apt/keyrings/deadsnakes.gpg ; \
+            sudo chmod 644 /etc/apt/keyrings/deadsnakes.gpg ; \
+            echo "deb [signed-by=/etc/apt/keyrings/deadsnakes.gpg] ${DEADSNAKES_MIRROR_URL} $(lsb_release -cs) main" > /etc/apt/sources.list.d/deadsnakes.list ; \
+        fi ; \
+    else \
+        for i in 1 2 3; do \
+            add-apt-repository -y ppa:deadsnakes/ppa && break || \
+            { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
+        done ; \
+    fi \
+    && apt-get update -y \
+    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
+    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
+    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
+    && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
     && python3 --version && python3 -m pip --version
 
 ARG PIP_INDEX_URL UV_INDEX_URL
@@ -101,9 +111,9 @@ ARG PYTORCH_CUDA_INDEX_BASE_URL
 ARG PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL
 ARG PIP_KEYRING_PROVIDER UV_KEYRING_PROVIDER
 
-# Activate virtual environment and add uv to PATH
-ENV PATH="/opt/venv/bin:/root/.local/bin:$PATH"
-ENV VIRTUAL_ENV="/opt/venv"
+# Install uv for faster pip installs
+RUN --mount=type=cache,target=/root/.cache/uv \
+    python3 -m pip install uv
 
 # This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
 # Reference: https://github.com/astral-sh/uv/pull/1694
@@ -132,7 +142,7 @@ WORKDIR /workspace
 COPY requirements/common.txt requirements/common.txt
 COPY requirements/cuda.txt requirements/cuda.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --python /opt/venv/bin/python3 -r requirements/cuda.txt \
+    uv pip install --system -r requirements/cuda.txt \
     --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
 
 # cuda arch list used by torch
@@ -162,7 +172,7 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE=copy
 
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --python /opt/venv/bin/python3 -r requirements/build.txt \
+    uv pip install --system -r requirements/build.txt \
     --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
 
 COPY . .
@@ -259,7 +269,7 @@ COPY requirements/lint.txt requirements/lint.txt
 COPY requirements/test.txt requirements/test.txt
 COPY requirements/dev.txt requirements/dev.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --python /opt/venv/bin/python3 -r requirements/dev.txt \
+    uv pip install --system -r requirements/dev.txt \
     --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
 #################### DEV IMAGE ####################
 
@@ -446,7 +456,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 ARG DEEPGEMM_GIT_REF
 COPY tools/install_deepgemm.sh /tmp/install_deepgemm.sh
 RUN --mount=type=cache,target=/root/.cache/uv \
-    VLLM_DOCKER_BUILD_CONTEXT=1 TORCH_CUDA_ARCH_LIST="9.0a 10.0a" /tmp/install_deepgemm.sh --cuda-version "${CUDA_VERSION}" ${DEEPGEMM_GIT_REF:+--ref "$DEEPGEMM_GIT_REF"}
+    VLLM_DOCKER_BUILD_CONTEXT=1 /tmp/install_deepgemm.sh --cuda-version "${CUDA_VERSION}" ${DEEPGEMM_GIT_REF:+--ref "$DEEPGEMM_GIT_REF"} 
 
 COPY tools/install_gdrcopy.sh install_gdrcopy.sh
 RUN set -eux; \
@@ -464,12 +474,6 @@ ENV CUDA_HOME=/usr/local/cuda
 RUN export TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-9.0a+PTX}" \
     && bash install_python_libraries.sh
 
-# CUDA image changed from /usr/local/nvidia to /usr/local/cuda in 12.8 but will
-# return to /usr/local/nvidia in 13.0 to allow container providers to mount drivers
-# consistently from the host (see https://github.com/vllm-project/vllm/issues/18859).
-# Until then, add /usr/local/nvidia/lib64 before the image cuda path to allow override.
-ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib64:${LD_LIBRARY_PATH}
-
 #################### vLLM installation IMAGE ####################
 
 #################### TEST IMAGE ####################
@@ -555,5 +559,5 @@ ENTRYPOINT ["./sagemaker-entrypoint.sh"]
 
 FROM vllm-openai-base AS vllm-openai
 
-ENTRYPOINT ["vllm", "serve"]
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
 #################### OPENAI API SERVER ####################

docker/Dockerfile.cpu

@@ -47,7 +47,7 @@ ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 
 ENV UV_HTTP_TIMEOUT=500
 
-# Install Python dependencies
+# Install Python dependencies 
 ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 ENV UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 ENV UV_INDEX_STRATEGY="unsafe-best-match"
@@ -104,7 +104,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=cache,target=/workspace/vllm/.deps,sharing=locked \
     --mount=type=bind,source=.git,target=.git \
-    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel
+    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel 
 
 ######################### TEST DEPS #########################
 FROM base AS vllm-test-deps
@@ -117,7 +117,7 @@ RUN --mount=type=bind,src=requirements/test.in,target=requirements/test.in \
     uv pip compile requirements/cpu-test.in -o requirements/cpu-test.txt --index-strategy unsafe-best-match --torch-backend cpu
 
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install -r requirements/cpu-test.txt
+    uv pip install -r requirements/cpu-test.txt 
 
 ######################### DEV IMAGE #########################
 FROM vllm-build AS vllm-dev
@@ -130,12 +130,12 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
 
 # install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install -e tests/vllm_test_utils
+    uv pip install -e tests/vllm_test_utils 
 
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=bind,source=.git,target=.git \
-    VLLM_TARGET_DEVICE=cpu python3 setup.py develop
+    VLLM_TARGET_DEVICE=cpu python3 setup.py develop 
 
 COPY --from=vllm-test-deps /workspace/vllm/requirements/cpu-test.txt requirements/test.txt
 
@@ -160,12 +160,11 @@ ADD ./benchmarks/ ./benchmarks/
 ADD ./vllm/collect_env.py .
 ADD ./.buildkite/ ./.buildkite/
 
-# Create symlink for vllm-workspace to maintain CI compatibility
-RUN ln -sf /workspace /vllm-workspace
-
 # install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install -e tests/vllm_test_utils
+    uv pip install -e tests/vllm_test_utils 
+
+ENTRYPOINT ["bash"]
 
 ######################### RELEASE IMAGE #########################
 FROM base AS vllm-openai
@@ -177,4 +176,4 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,from=vllm-build,src=/workspace/vllm/dist,target=dist \
     uv pip install dist/*.whl
 
-ENTRYPOINT ["vllm", "serve"]
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

docker/Dockerfile.nightly_torch

@@ -6,7 +6,7 @@ ARG CUDA_VERSION=12.8.0
 #
 #################### BASE BUILD IMAGE ####################
 # prepare basic build environment
-FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS base
+FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04 AS base
 ARG CUDA_VERSION=12.8.0
 ARG PYTHON_VERSION=3.12
 ARG TARGETPLATFORM

docker/Dockerfile.ppc64le

@@ -314,4 +314,4 @@ WORKDIR /workspace/
 
 RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
 
-ENTRYPOINT ["vllm", "serve"]
+ENTRYPOINT ["python", "-m", "vllm.entrypoints.openai.api_server"]

docker/Dockerfile.rocm_base

@@ -15,7 +15,7 @@ FROM ${BASE_IMAGE} AS base
 ENV PATH=/opt/rocm/llvm/bin:/opt/rocm/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin
 ENV ROCM_PATH=/opt/rocm
 ENV LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib:
-ARG PYTORCH_ROCM_ARCH=gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151
+ARG PYTORCH_ROCM_ARCH=gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1200;gfx1201
 ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}
 ENV AITER_ROCM_ARCH=gfx942;gfx950
 
@@ -141,4 +141,4 @@ RUN echo "BASE_IMAGE: ${BASE_IMAGE}" > /app/versions.txt \
     && echo "FA_BRANCH: ${FA_BRANCH}" >> /app/versions.txt \
     && echo "FA_REPO: ${FA_REPO}" >> /app/versions.txt \
     && echo "AITER_BRANCH: ${AITER_BRANCH}" >> /app/versions.txt \
-    && echo "AITER_REPO: ${AITER_REPO}" >> /app/versions.txt
+    && echo "AITER_REPO: ${AITER_REPO}" >> /app/versions.txt

docker/Dockerfile.s390x

@@ -309,4 +309,4 @@ USER 2000
 WORKDIR /home/vllm
 
 # Set the default entrypoint
-ENTRYPOINT ["vllm", "serve"]
+ENTRYPOINT ["python", "-m", "vllm.entrypoints.openai.api_server"]

docker/Dockerfile.xpu

@@ -69,4 +69,4 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 
 # install development dependencies (for testing)
 RUN python3 -m pip install -e tests/vllm_test_utils
-ENTRYPOINT ["vllm", "serve"]
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

docs/api/vllm/.meta.yml

@@ -1,2 +1,2 @@
 search:
-  exclude: true
+  boost: 0.5

docs/contributing/benchmarks.md

@@ -661,7 +661,8 @@ Benchmark the performance of multi-modal requests in vLLM.
 Start vLLM:
 
 ```bash
-vllm serve Qwen/Qwen2.5-VL-7B-Instruct \
+python -m vllm.entrypoints.openai.api_server \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
   --dtype bfloat16 \
   --limit-mm-per-prompt '{"image": 1}' \
   --allowed-local-media-path /path/to/sharegpt4v/images
@@ -687,7 +688,8 @@ vllm bench serve \
 Start vLLM:
 
 ```bash
-vllm serve Qwen/Qwen2.5-VL-7B-Instruct \
+python -m vllm.entrypoints.openai.api_server \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
   --dtype bfloat16 \
   --limit-mm-per-prompt '{"video": 1}' \
   --allowed-local-media-path /path/to/sharegpt4video/videos
@@ -821,30 +823,6 @@ The latest performance results are hosted on the public [vLLM Performance Dashbo
 
 More information on the performance benchmarks and their parameters can be found in [Benchmark README](https://github.com/intel-ai-tce/vllm/blob/more_cpu_models/.buildkite/nightly-benchmarks/README.md) and [performance benchmark description](gh-file:.buildkite/nightly-benchmarks/performance-benchmarks-descriptions.md).
 
-### Continuous Benchmarking
-
-The continuous benchmarking provides automated performance monitoring for vLLM across different models and GPU devices. This helps track vLLM's performance characteristics over time and identify any performance regressions or improvements.
-
-#### How It Works
-
-The continuous benchmarking is triggered via a [GitHub workflow CI](https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-benchmark.yml) in the PyTorch infrastructure repository, which runs automatically every 4 hours. The workflow executes three types of performance tests:
-
-- **Serving tests**: Measure request handling and API performance
-- **Throughput tests**: Evaluate token generation rates
-- **Latency tests**: Assess response time characteristics
-
-#### Benchmark Configuration
-
-The benchmarking currently runs on a predefined set of models configured in the [vllm-benchmarks directory](https://github.com/pytorch/pytorch-integration-testing/tree/main/vllm-benchmarks/benchmarks). To add new models for benchmarking:
-
-1. Navigate to the appropriate GPU directory in the benchmarks configuration
-2. Add your model specifications to the corresponding configuration files
-3. The new models will be included in the next scheduled benchmark run
-
-#### Viewing Results
-
-All continuous benchmarking results are automatically published to the public [vLLM Performance Dashboard](https://hud.pytorch.org/benchmark/llms?repoName=vllm-project%2Fvllm).
-
 [](){ #nightly-benchmarks }
 
 ## Nightly Benchmarks

docs/contributing/model/multimodal.md

@@ -258,21 +258,17 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
             self,
             seq_len: int,
             mm_counts: Mapping[str, int],
-            mm_options: Optional[Mapping[str, BaseDummyOptions]] = None,
         ) -> MultiModalDataDict:
             num_images = mm_counts.get("image", 0)
 
             target_width, target_height = \
                 self.info.get_image_size_with_most_features()
 
-            image_overrides = mm_options.get("image") if mm_options else None
-
             return {
                 "image":
                 self._get_dummy_images(width=target_width,
                                     height=target_height,
-                                    num_images=num_images,
-                                    overrides=image_overrides)
+                                    num_images=num_images)
             }
         ```
 
@@ -442,20 +438,16 @@ Assuming that the memory usage increases with the number of tokens, the dummy in
             self,
             seq_len: int,
             mm_counts: Mapping[str, int],
-            mm_options: Optional[Mapping[str, BaseDummyOptions]] = None,
         ) -> MultiModalDataDict:
             target_width, target_height = \
                 self.info.get_image_size_with_most_features()
             num_images = mm_counts.get("image", 0)
 
-            image_overrides = mm_options.get("image") if mm_options else None
-
             return {
                 "image":
                 self._get_dummy_images(width=target_width,
                                     height=target_height,
-                                    num_images=num_images,
-                                    overrides=image_overrides)
+                                    num_images=num_images)
             }
         ```
 

docs/contributing/profiling.md

@@ -39,7 +39,8 @@ Refer to <gh-file:examples/offline_inference/simple_profiling.py> for an example
 
 ```bash
 VLLM_TORCH_PROFILER_DIR=./vllm_profile \
-    vllm serve meta-llama/Meta-Llama-3-70B
+    python -m vllm.entrypoints.openai.api_server \
+    --model meta-llama/Meta-Llama-3-70B
 ```
 
 vllm bench command:
@@ -159,22 +160,6 @@ GUI example:
 
 <img width="1799" alt="Screenshot 2025-03-05 at 11 48 42 AM" src="https://github.com/user-attachments/assets/c7cff1ae-6d6f-477d-a342-bd13c4fc424c" />
 
-## Continuous Profiling
-
-There is a [GitHub CI workflow](https://github.com/pytorch/pytorch-integration-testing/actions/workflows/vllm-profiling.yml) in the PyTorch infrastructure repository that provides continuous profiling for different models on vLLM. This automated profiling helps track performance characteristics over time and across different model configurations.
-
-### How It Works
-
-The workflow currently runs weekly profiling sessions for selected models, generating detailed performance traces that can be analyzed using different tools to identify performance regressions or optimization opportunities. But, it can be triggered manually as well, using the Github Action tool.
-
-### Adding New Models
-
-To extend the continuous profiling to additional models, you can modify the [profiling-tests.json](https://github.com/pytorch/pytorch-integration-testing/blob/main/vllm-profiling/cuda/profiling-tests.json) configuration file in the PyTorch integration testing repository. Simply add your model specifications to this file to include them in the automated profiling runs.
-
-### Viewing Profiling Results
-
-The profiling traces generated by the continuous profiling workflow are publicly available on the [vLLM Performance Dashboard](https://hud.pytorch.org/benchmark/llms?repoName=vllm-project%2Fvllm). Look for the **Profiling traces** table to access and download the traces for different models and runs.
-
 ## Profiling vLLM Python Code
 
 The Python standard library includes
@@ -223,11 +208,3 @@ One example is [snakeviz](https://jiffyclub.github.io/snakeviz/).
 pip install snakeviz
 snakeviz expensive_function.prof
 ```
-
-### Analyzing Garbage Collection Costs
-
-Leverage VLLM_GC_DEBUG environment variable to debug GC costs.
-
-- VLLM_GC_DEBUG=1: enable GC debugger with gc.collect elpased times
-- VLLM_GC_DEBUG='{"top_objects":5}': enable GC debugger to log top 5
-  collected objects for each gc.collect

docs/deployment/frameworks/autogen.md

@@ -19,7 +19,8 @@ pip install -U "autogen-agentchat" "autogen-ext[openai]"
 1. Start the vLLM server with the supported chat completion model, e.g.
 
     ```bash
-    vllm serve mistralai/Mistral-7B-Instruct-v0.2
+    python -m vllm.entrypoints.openai.api_server \
+        --model mistralai/Mistral-7B-Instruct-v0.2
     ```
 
 1. Call it with AutoGen:

docs/deployment/frameworks/hf_inference_endpoints.md

@@ -1,170 +0,0 @@
-# Hugging Face Inference Endpoints
-
-## Overview
-
-Models compatible with vLLM can be deployed on Hugging Face Inference Endpoints, either starting from the [Hugging Face Hub](https://huggingface.co) or directly from the [Inference Endpoints](https://endpoints.huggingface.co/) interface. This allows you to serve models in a fully managed environment with GPU acceleration, auto-scaling, and monitoring, without managing the infrastructure manually.
-
-For advanced details on vLLM integration and deployment options, see [Advanced Deployment Details](#advanced-deployment-details).
-
-## Deployment Methods
-
-- [**Method 1: Deploy from the Catalog.**](#method-1-deploy-from-the-catalog) One-click deploy models from the Hugging Face Hub with ready-made optimized configurations.
-- [**Method 2: Guided Deployment (Transformers Models).**](#method-2-guided-deployment-transformers-models) Instantly deploy models tagged with `transformers` from the Hub UI using the **Deploy** button.
-- [**Method 3: Manual Deployment (Advanced Models).**](#method-3-manual-deployment-advanced-models) For models that either use custom code with the `transformers` tag, or don’t run with standard `transformers` but are supported by vLLM. This method requires manual configuration.
-
-### Method 1: Deploy from the Catalog
-
-This is the easiest way to get started with vLLM on Hugging Face Inference Endpoints. You can browse a catalog of models with verified and optimized deployment configuration at [Inference Endpoints](https://endpoints.huggingface.co/catalog) to maximize performance.
-
-1. Go to [Endpoints Catalog](https://endpoints.huggingface.co/catalog) and in the **Inference Server** options, select `vLLM`.This will display the current list of models with optimized preconfigured options.
-
-    ![Endpoints Catalog](../../assets/deployment/hf-inference-endpoints-catalog.png)
-
-1. Select the desired model and click **Create Endpoint**.
-
-    ![Create Endpoint](../../assets/deployment/hf-inference-endpoints-create-endpoint.png)
-
-1. Once the deployment is ready, you can use the endpoint. Update the `DEPLOYMENT_URL` with the URL provided in the console, remembering to append `/v1` as required.
-
-    ```python
-    # pip install openai
-    from openai import OpenAI
-    import os
-
-    client = OpenAI(
-        base_url = DEPLOYMENT_URL,
-        api_key = os.environ["HF_TOKEN"] # https://huggingface.co/settings/tokens
-    )
-
-    chat_completion = client.chat.completions.create(
-        model = "HuggingFaceTB/SmolLM3-3B",
-        messages = [
-            {
-                "role": "user",
-                "content": [
-                    {
-                        "type": "text",
-                        "text": "Give me a brief explanation of gravity in simple terms."
-                    }
-                ]
-            }
-        ],
-        stream = True
-    )
-
-    for message in chat_completion:
-        print(message.choices[0].delta.content, end = "")
-    ```
-
-!!! note
-    The catalog provides models optimized for vLLM, including GPU settings and inference engine configurations. You can monitor the endpoint and update the **container or its configuration** from the Inference Endpoints UI.
-
-### Method 2: Guided Deployment (Transformers Models)
-
-This method applies to models with the `transformers` library tag in their metadata. It allows you to deploy a model directly from the Hub UI without manual configuration.
-
-1. Navigate to a model on [Hugging Face Hub](https://huggingface.co/models).  
-   For this example we will use the [`ibm-granite/granite-docling-258M`](https://huggingface.co/ibm-granite/granite-docling-258M) model. You can verify that the model is compatible by checking the front matter in the [README](https://huggingface.co/ibm-granite/granite-docling-258M/blob/main/README.md), where the library is tagged as `library: transformers`.
-
-2. Locate the **Deploy** button. The button appears for models tagged with `transformers` at the top right of the [model card](https://huggingface.co/ibm-granite/granite-docling-258M).
-
-    ![Locate deploy button](../../assets/deployment/hf-inference-endpoints-locate-deploy-button.png)
-
-3. Click to **Deploy** button > **HF Inference Endpoints**. You will be taken to the Inference Endpoints interface to configure the deployment.
-
-    ![Click deploy button](../../assets/deployment/hf-inference-endpoints-click-deploy-button.png)
-
-4. Select the Hardware (we choose AWS>GPU>T4 for the example) and Container Configuration. Choose `vLLM` as the container type and finalize the deployment pressing **Create Endpoint**.
-
-    ![Select Hardware](../../assets/deployment/hf-inference-endpoints-select-hardware.png)
-
-5. Use the deployed endpoint. Update the `DEPLOYMENT_URL` with the URL provided in the console (remember to add `/v1` needed). You can then use your endpoint programmatically or via the SDK.
-
-    ```python
-    # pip install openai
-    from openai import OpenAI
-    import os
-
-    client = OpenAI(
-        base_url = DEPLOYMENT_URL,
-        api_key = os.environ["HF_TOKEN"] # https://huggingface.co/settings/tokens
-    )
-
-    chat_completion = client.chat.completions.create(
-        model = "ibm-granite/granite-docling-258M",
-        messages = [
-            {
-                "role": "user",
-                "content": [
-                    {
-                        "type": "image_url",
-                        "image_url": {
-                            "url": "https://huggingface.co/ibm-granite/granite-docling-258M/resolve/main/assets/new_arxiv.png"
-                        }
-                    },
-                    {
-                        "type": "text",
-                        "text": "Convert this page to docling."
-                    }
-                ]
-            }
-        ],
-        stream = True
-    )
-
-    for message in chat_completion:
-        print(message.choices[0].delta.content, end = "")
-    ```
-
-!!! note
-    This method uses best-guess defaults. You may need to adjust the configuration to fit your specific requirements.
-
-### Method 3: Manual Deployment (Advanced Models)
-
-Some models require manual deployment because they:
-
-- Use custom code with the `transformers` tag
-- Don't run with standard `transformers` but are supported by `vLLM`
-
-These models cannot be deployed using the **Deploy** button on the model card.
-
-In this guide, we demonstrate manual deployment using the [rednote-hilab/dots.ocr](https://huggingface.co/rednote-hilab/dots.ocr) model, an OCR model integrated with vLLM (see vLLM [PR](https://github.com/vllm-project/vllm/pull/24645)).
-
-1. Start a new deployment. Go to [Inference Endpoints](https://endpoints.huggingface.co/) and click `New`.
-
-    ![New Endpoint](../../assets/deployment/hf-inference-endpoints-new-endpoint.png)
-
-2. Search the model in the Hub. In the dialog, switch to **Hub** and search for the desired model.
-
-    ![Select model](../../assets/deployment/hf-inference-endpoints-select-model.png)
-
-3. Choosing infrastructure. On the configuration page, select the cloud provider and hardware from the available options.  
-   For this demo, we choose AWS and L4 GPU. Adjust according to your hardware needs.
-
-    ![Choose Infra](../../assets/deployment/hf-inference-endpoints-choose-infra.png)
-
-4. Configure the container. Scroll to the **Container Configuration** and select `vLLM` as the container type.
-
-    ![Configure Container](../../assets/deployment/hf-inference-endpoints-configure-container.png)
-
-5. Create the endpoint. Click **Create Endpoint** to deploy the model.
-
-    Once the endpoint is ready, you can use it with the OpenAI Completion API, cURL, or other SDKs. Remember to append `/v1` to the deployment URL if needed.
-
-!!! note
-    You can adjust the **container settings** (Container URI, Container Arguments) from the Inference Endpoints UI and press **Update Endpoint**. This redeploys the endpoint with the updated container configuration. Changes to the model itself require creating a new endpoint or redeploying with a different model. For example, for this demo, you may need to update the Container URI to the nightly image (`vllm/vllm-openai:nightly`) and add the `--trust-remote-code` flag in the container arguments.
-
-## Advanced Deployment Details
-
-With the [transformers backend integration](https://blog.vllm.ai/2025/04/11/transformers-backend.html), vLLM now offers Day 0 support for any model compatible with `transformers`. This means you can deploy such models immediately, leveraging vLLM’s optimized inference without additional backend modifications.
-
-Hugging Face Inference Endpoints provides a fully managed environment for serving models via vLLM. You can deploy models without configuring servers, installing dependencies, or managing clusters. Endpoints also support deployment across multiple cloud providers (AWS, Azure, GCP) without the need for separate accounts.
-
-The platform integrates seamlessly with the Hugging Face Hub, allowing you to deploy any vLLM- or `transformers`-compatible model, track usage, and update the inference engine directly. The vLLM engine comes preconfigured, enabling optimized inference and easy switching between models or engines without modifying your code. This setup simplifies production deployment: endpoints are ready in minutes, include monitoring and logging, and let you focus on serving models rather than maintaining infrastructure.
-
-## Next Steps
-
-- Explore the [Inference Endpoints](https://endpoints.huggingface.co/catalog) model catalog
-- Read the Inference Endpoints [documentation](https://huggingface.co/docs/inference-endpoints/en/index)
-- Learn about [Inference Endpoints engines](https://huggingface.co/docs/inference-endpoints/en/engines/vllm)
-- Understand the [transformers backend integration](https://blog.vllm.ai/2025/04/11/transformers-backend.html)

docs/deployment/frameworks/open-webui.md

@@ -20,7 +20,7 @@ To get started with Open WebUI using vLLM, follow these steps:
         For example:
 
         ```console
-        vllm serve <model> --host 0.0.0.0 --port 8000
+        python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port 8000
         ```
 
 3. Start the Open WebUI Docker container:

docs/deployment/frameworks/skypilot.md

@@ -32,7 +32,6 @@ See the vLLM SkyPilot YAML for serving, [serving.yaml](https://github.com/skypil
       ports: 8081  # Expose to internet traffic.
 
     envs:
-      PYTHONUNBUFFERED: 1
       MODEL_NAME: meta-llama/Meta-Llama-3-8B-Instruct
       HF_TOKEN: <your-huggingface-token>  # Change to your own huggingface token, or use --env to pass.
 
@@ -48,8 +47,9 @@ See the vLLM SkyPilot YAML for serving, [serving.yaml](https://github.com/skypil
     run: |
       conda activate vllm
       echo 'Starting vllm api server...'
-      vllm serve $MODEL_NAME \
+      python -u -m vllm.entrypoints.openai.api_server \
         --port 8081 \
+        --model $MODEL_NAME \
         --trust-remote-code \
         --tensor-parallel-size $SKYPILOT_NUM_GPUS_PER_NODE \
         2>&1 | tee api_server.log &
@@ -131,7 +131,6 @@ SkyPilot can scale up the service to multiple service replicas with built-in aut
       ports: 8081  # Expose to internet traffic.
 
     envs:
-      PYTHONUNBUFFERED: 1
       MODEL_NAME: meta-llama/Meta-Llama-3-8B-Instruct
       HF_TOKEN: <your-huggingface-token>  # Change to your own huggingface token, or use --env to pass.
 
@@ -147,8 +146,9 @@ SkyPilot can scale up the service to multiple service replicas with built-in aut
     run: |
       conda activate vllm
       echo 'Starting vllm api server...'
-      vllm serve $MODEL_NAME \
+      python -u -m vllm.entrypoints.openai.api_server \
         --port 8081 \
+        --model $MODEL_NAME \
         --trust-remote-code \
         --tensor-parallel-size $SKYPILOT_NUM_GPUS_PER_NODE \
         2>&1 | tee api_server.log
@@ -243,7 +243,6 @@ This will scale the service up to when the QPS exceeds 2 for each replica.
       ports: 8081  # Expose to internet traffic.
 
     envs:
-      PYTHONUNBUFFERED: 1
       MODEL_NAME: meta-llama/Meta-Llama-3-8B-Instruct
       HF_TOKEN: <your-huggingface-token>  # Change to your own huggingface token, or use --env to pass.
 
@@ -259,8 +258,9 @@ This will scale the service up to when the QPS exceeds 2 for each replica.
     run: |
       conda activate vllm
       echo 'Starting vllm api server...'
-      vllm serve $MODEL_NAME \
+      python -u -m vllm.entrypoints.openai.api_server \
         --port 8081 \
+        --model $MODEL_NAME \
         --trust-remote-code \
         --tensor-parallel-size $SKYPILOT_NUM_GPUS_PER_NODE \
         2>&1 | tee api_server.log

docs/design/arch_overview.md

@@ -69,11 +69,6 @@ Sometimes you may see the API server entrypoint used directly instead of via the
 python -m vllm.entrypoints.openai.api_server --model <model>
 ```
 
-!!! warning
-
-    `python -m vllm.entrypoints.openai.api_server` is deprecated
-    and may become unsupported in a future release.
-
 That code can be found in <gh-file:vllm/entrypoints/openai/api_server.py>.
 
 More details on the API server can be found in the [OpenAI-Compatible Server](../serving/openai_compatible_server.md) document.

docs/design/fused_moe_modular_kernel.md

@@ -242,8 +242,30 @@ Example: `python3 -m tests.kernels.moe.modular_kernel_tools.profile_modular_kern
 
 ## FusedMoEPrepareAndFinalize Implementations
 
-See [Fused MoE Kernel features](./moe_kernel_features.md#fused-moe-modular-all2all-backends) for a list of all the available modular prepare and finalize subclasses.
+The following table lists the `FusedMoEPrepareAndFinalize` implementations at the time of writing,
+
+| Implementation | Type | Comments |
+| :--- | :--- | :--- |
+| DeepEPHTPrepareAndFinalize | Contiguous / Non-Batched | Uses the DeepEP High-Throughput all2all kernels. |
+| DeepEPLLPrepareAndFinalize | Batched | Uses the DeepEP Low-Latency all2all kernels. |
+| PplxPrepareAndFinalize | Batched | Uses the Perplexity all2all kernels. |
+| FlashInferCutlassMoEPrepareAndFinalize | Contiguous | |
+| MoEPrepareAndFinalizeNoEP | Contiguous | This implementation is used when there is no EP. i.e. no all2all kernels are invoked. |
+| BatchedPrepareAndFinalize | Batched | A reference prepare/finalize class that reorganizes the tokens into expert batched format, i.e. E x max_num_tokens x K. (Doesn’t use any all2all kernels. This is primarily used in unit testing) |
 
 ## FusedMoEPermuteExpertsUnpermute
 
-See [Fused MoE Kernel features](./moe_kernel_features.md#fused-moe-experts-kernels) for a list of all the available modular experts.
+The following table lists the `FusedMoEPermuteExpertsUnpermute` implementations at the time of writing,
+
+| Implementation | Type | Comment |
+| :--- | :--- | :--- |
+| BatchedDeepGemmExperts | Batched | Uses the DeepGemm’s Masked Grouped Gemm kernels for the fused_moe operation. |
+| BatchedTritonExperts | Batched | Uses a Triton Kernel for the Batched matmuls. |
+| BatchedTritonOrDeepGemmExperts | Batched | Chooses either the `BatchedDeepGemmExperts` or `BatchedTritonExperts` based on environment settings. |
+| DeepGemmExperts | Contiguous / Non-Batched | Uses DeepGemm’s Grouped Gemm kernels for fused_moe operation. |
+| TritonExperts | Contiguous / Non-Batched | Uses a Triton Kernel for fused_moe matmuls. |
+| TritonOrDeepGemmExperts | Contiguous / Non-Batched | Chooses either the `DeepGemmExperts` or `TritonExperts` based on fused_moe inputs. |
+| CutlassExpertsFP8 | Supports both Batched and Contiguous formats | Uses Cutlass Grouped Gemm implementations for the fp8 matmuls. |
+| CutlassExpertsFP4 | Supports both Batched and Contiguous formats | Uses Cutlass Grouped Gemm implementations for the fp4 matmuls. |
+| FlashInferExperts | Contiguous | Uses fused_moe operation from FlashInfer |
+| NaiveBatchedExperts | Batched | Reference Batched Experts implementation. Primarily used in unit tests. |