Merge tag 'v0.9.1' into v0.9.1-dev

.github/workflows/cleanup_pr_body.yml

@@ -20,7 +20,12 @@ jobs:
         with:
           python-version: '3.12'
 
+      - name: Install Python dependencies
+        run: |
+          python3 -m pip install --upgrade pip
+          python3 -m pip install regex
+
       - name: Update PR description
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: .github/scripts/cleanup_pr_body.sh "${{ github.event.number }}"
+        run: bash .github/scripts/cleanup_pr_body.sh "${{ github.event.number }}"

.github/workflows/lint-and-deploy.yaml

@@ -2,6 +2,9 @@ name: Lint and Deploy Charts
 
 on: pull_request
 
+permissions:
+  contents: read
+
 jobs:
   lint-and-deploy:
     runs-on: ubuntu-latest
@@ -66,7 +69,7 @@ jobs:
           export AWS_SECRET_ACCESS_KEY=minioadmin
           sleep 30 && kubectl -n ns-vllm logs -f "$(kubectl -n ns-vllm get pods | awk '/deployment/ {print $1;exit}')" &
           helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/online_serving/chart-helm -f examples/online_serving/chart-helm/values.yaml --set secrets.s3endpoint=http://minio:9000 --set secrets.s3bucketname=testbucket --set secrets.s3accesskeyid=$AWS_ACCESS_KEY_ID --set secrets.s3accesskey=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
-    
+
       - name: curl test
         run: |
           kubectl -n ns-vllm port-forward service/test-vllm-service 8001:80 &
@@ -79,4 +82,4 @@ jobs:
                           "max_tokens": 7,
                           "temperature": 0
                   }'):$CODE"
-          echo "$CODE"
\ No newline at end of file
+          echo "$CODE"

.github/workflows/pre-commit.yml

@@ -5,6 +5,9 @@ on:
   push:
     branches: [main]
 
+permissions:
+  contents: read
+
 jobs:
   pre-commit:
     runs-on: ubuntu-latest

.github/workflows/reminder_comment.yml

 name: PR Reminder Comment Bot
+permissions:
+  pull-requests: write
 on:
   pull_request_target:
     types: [opened]

.gitignore

@@ -77,10 +77,6 @@ instance/
 # Scrapy stuff:
 .scrapy
 
-# Sphinx documentation
-docs/_build/
-docs/source/getting_started/examples/
-
 # PyBuilder
 .pybuilder/
 target/
@@ -150,6 +146,7 @@ venv.bak/
 
 # mkdocs documentation
 /site
+docs/examples
 
 # mypy
 .mypy_cache/

.pre-commit-config.yaml

@@ -11,42 +11,47 @@ repos:
   hooks:
   - id: yapf
     args: [--in-place, --verbose]
+    # Keep the same list from yapfignore here to avoid yapf failing without any inputs
+    exclude: '(.buildkite|benchmarks|build|examples)/.*'
 - repo: https://github.com/astral-sh/ruff-pre-commit
-  rev: v0.9.3
+  rev: v0.11.7
   hooks:
   - id: ruff
     args: [--output-format, github, --fix]
+  - id: ruff-format
+    files: ^(.buildkite|benchmarks|examples)/.*
 - repo: https://github.com/codespell-project/codespell
-  rev: v2.4.0
+  rev: v2.4.1
   hooks:
   - id: codespell
     additional_dependencies: ['tomli']
     args: ['--toml', 'pyproject.toml']
 - repo: https://github.com/PyCQA/isort
-  rev: 0a0b7a830386ba6a31c2ec8316849ae4d1b8240d # 6.0.0
+  rev: 6.0.1
   hooks:
   - id: isort
 - repo: https://github.com/pre-commit/mirrors-clang-format
-  rev: v19.1.7
+  rev: v20.1.3
   hooks:
   - id: clang-format
     exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
     types_or: [c++, cuda]
     args: [--style=file, --verbose]
 - repo: https://github.com/jackdewinter/pymarkdown
-  rev: v0.9.27
+  rev: v0.9.29
   hooks:
   - id: pymarkdown
+    exclude: '.*\.inc\.md'
     args: [fix]
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:
   - id: actionlint
 - repo: https://github.com/astral-sh/uv-pre-commit
-  rev: 0.6.2
+  rev: 0.6.17
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
@@ -55,7 +60,7 @@ repos:
     entry: tools/mypy.sh 0 "local"
     language: python
     types: [python]
-    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, types-setuptools, types-PyYAML, types-requests]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, types-setuptools, types-PyYAML, types-requests, pydantic]
     stages: [pre-commit] # Don't run in CI
   - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.9
@@ -101,8 +106,8 @@ repos:
     args:
       - -c
       - |
-        if ! grep -q "^Signed-off-by: $(git config user.name) <$(git config user.email)>" .git/COMMIT_EDITMSG; then
-          printf "\nSigned-off-by: $(git config user.name) <$(git config user.email)>\n" >> .git/COMMIT_EDITMSG
+        if ! grep -q "^Signed-off-by: $(git config user.name) <$(git config user.email)>" "$(git rev-parse --git-path COMMIT_EDITMSG)"; then
+          printf "\nSigned-off-by: $(git config user.name) <$(git config user.email)>\n" >> "$(git rev-parse --git-path COMMIT_EDITMSG)"
         fi
     language: system
     verbose: true
@@ -125,8 +130,21 @@ repos:
     name: Update Dockerfile dependency graph
     entry: tools/update-dockerfile-graph.sh
     language: script
-    files: ^docker/Dockerfile$
+  - id: enforce-import-regex-instead-of-re
+    name: Enforce import regex as re
+    entry: python tools/enforce_regex_import.py
+    language: python
+    types: [python]
+    pass_filenames: false
+    additional_dependencies: [regex]
+  # forbid directly import triton
+  - id: forbid-direct-triton-import
+    name: "Forbid direct 'import triton'"
+    entry: python tools/check_triton_import.py
+    language: python
+    types: [python]
     pass_filenames: false
+    additional_dependencies: [regex]
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

.readthedocs.yaml

@@ -8,12 +8,8 @@ build:
   tools:
     python: "3.12"
 
-sphinx:
-  configuration: docs/source/conf.py
-  fail_on_warning: true
-
-# If using Sphinx, optionally build your docs in additional formats such as PDF
-formats: []
+mkdocs:
+  configuration: mkdocs.yaml
 
 # Optionally declare the Python requirements required to build your docs
 python:

CMakeLists.txt

@@ -28,15 +28,15 @@ add_compile_options(-w)
 # Suppress potential warnings about unused manually-specified variables
 set(ignoreMe "${VLLM_PYTHON_PATH}")
 
+# Prevent installation of dependencies (cutlass) by default.
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
+
 #
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
 #
 set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
 
-# Supported NVIDIA architectures.
-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")
-
 # Supported AMD GPU architectures.
 set(HIP_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx906;gfx926;gfx928;gfx936")
 
@@ -50,8 +50,8 @@ set(HIP_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.6.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.6.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.7.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.7.0")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -84,6 +84,15 @@ endif()
 #
 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 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()
+  set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
+endif()
+
 #
 # Forward the non-CUDA device extensions to external CMake scripts.
 #
@@ -178,9 +187,6 @@ include(FetchContent)
 file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
 message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
 
-#
-# Set rocm version dev int.
-#
 if(VLLM_GPU_LANG STREQUAL "HIP")
   #
   # Overriding the default -O set up by cmake, adding ggdb3 for the most verbose devug info
@@ -188,7 +194,6 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
   set(CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG "${CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG} -O0 -ggdb3")
   set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O0 -ggdb3")
 
-
   #
   # Certain HIP functions are marked as [[nodiscard]], yet vllm ignores the result which generates
   # a lot of warnings that always mask real issues. Suppressing until this is properly addressed.
@@ -231,10 +236,13 @@ endif()
 #
 
 set(VLLM_EXT_SRC
+  "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
+  "csrc/mamba/causal_conv1d/causal_conv1d.cu"
   "csrc/cache_kernels.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
   "csrc/attention/merge_attn_states.cu"
+  "csrc/attention/vertical_slash_index.cu"
   "csrc/pos_encoding_kernels.cu"
   "csrc/pos_encoding_tgi_kernels.cu"
   "csrc/activation_kernels.cu"
@@ -248,12 +256,14 @@ set(VLLM_EXT_SRC
   "csrc/attention/attention_with_mask_kernels_opt_tc.cu"
   "csrc/opt/layernorm_kernels_opt.cu"
   # "csrc/layernorm_quant_kernels.cu"
+  "csrc/sampler.cu"
   "csrc/cuda_view.cu"
   # "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
   # "csrc/quantization/fp8/common.cu"
   # "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
   "csrc/quantization/gguf/gguf_kernel.cu"
+  "csrc/quantization/activation_kernels.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/prepare_inputs/advance_step.cu"
   "csrc/custom_all_reduce.cu"
@@ -262,9 +272,8 @@ set(VLLM_EXT_SRC
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
-  # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
-  # Please keep this in sync with FetchContent_Declare line below.
-  set(CUTLASS_REVISION "v3.9.0" CACHE STRING "CUTLASS revision to use")
+  # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
+  set(CUTLASS_REVISION "v3.9.2" 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})
@@ -282,7 +291,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
         cutlass
         GIT_REPOSITORY https://github.com/nvidia/cutlass.git
         # Please keep this in sync with CUTLASS_REVISION line above.
-        GIT_TAG v3.9.0
+        GIT_TAG ${CUTLASS_REVISION}
         GIT_PROGRESS TRUE
 
         # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
@@ -294,14 +303,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   FetchContent_MakeAvailable(cutlass)
 
   list(APPEND VLLM_EXT_SRC
-    "csrc/mamba/mamba_ssm/selective_scan_fwd.cu"
-    "csrc/mamba/causal_conv1d/causal_conv1d.cu"
     "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
     "csrc/permute_cols.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/attention/mla/cutlass_mla_entry.cu")
@@ -313,10 +321,55 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # Only build Marlin kernels if we are building for at least some compatible archs.
   # Keep building Marlin for 9.0 as there are some group sizes and shapes that
   # are not supported by Machete yet.
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
+  # 9.0 for latest bf16 atomicAdd PTX
+  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
+
+    #
+    # For the Marlin kernels we automatically generate sources for various
+    # preselected input type pairs and schedules.
+    # Generate sources:
+    set(MARLIN_GEN_SCRIPT
+      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/gptq_marlin/generate_kernels.py)
+    file(MD5 ${MARLIN_GEN_SCRIPT} MARLIN_GEN_SCRIPT_HASH)
+
+    message(STATUS "Marlin generation script hash: ${MARLIN_GEN_SCRIPT_HASH}")
+    message(STATUS "Last run Marlin generate script hash: $CACHE{MARLIN_GEN_SCRIPT_HASH}")
+
+    if (NOT DEFINED CACHE{MARLIN_GEN_SCRIPT_HASH}
+        OR NOT $CACHE{MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MARLIN_GEN_SCRIPT_HASH})
+      execute_process(
+        COMMAND ${CMAKE_COMMAND} -E env
+        PYTHONPATH=$PYTHONPATH
+          ${Python_EXECUTABLE} ${MARLIN_GEN_SCRIPT}
+        RESULT_VARIABLE marlin_generation_result
+        OUTPUT_VARIABLE marlin_generation_result
+        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
+        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
+      )
+
+      if (NOT marlin_generation_result EQUAL 0)
+        message(FATAL_ERROR "Marlin generation failed."
+                            " Result: \"${marlin_generation_result}\""
+                            "\nCheck the log for details: "
+                            "${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log")
+      else()
+        set(MARLIN_GEN_SCRIPT_HASH ${MARLIN_GEN_SCRIPT_HASH}
+            CACHE STRING "Last run Marlin generate script hash" FORCE)
+        message(STATUS "Marlin generation completed successfully.")
+      endif()
+    else()
+      message(STATUS "Marlin generation script has not changed, skipping generation.")
+    endif()
+
+    file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/gptq_marlin/kernel_*.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
+      CUDA_ARCHS "${MARLIN_ARCHS}")
+
+    list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})
+
     set(MARLIN_SRCS
-       "csrc/quantization/fp8/fp8_marlin.cu"
        "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
        "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
        "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
@@ -388,6 +441,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set(SRCS
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
       "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
     )
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -412,8 +466,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   #
   # For the cutlass_scaled_mm kernels we want to build the c2x (CUTLASS 2.x)
   # kernels for the remaining archs that are not already built for 3x.
+  # (Build 8.9 for FP8)
   cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
-    "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
+    "7.5;8.0;8.7;8.9+PTX" "${CUDA_ARCHS}")
   # subtract out the archs that are already built for 3x
   list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
   if (SCALED_MM_2X_ARCHS)
@@ -464,7 +519,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
-      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu")
+      "csrc/quantization/fp4/nvfp4_experts_quant.cu"
+      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu"
+      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${FP4_ARCHS}")
@@ -500,9 +557,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # CUTLASS MoE kernels
 
   # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and only works
-  # on Hopper). get_cutlass_moe_mm_data should only be compiled if it's possible
-  # to compile MoE kernels that use its output.
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  # on Hopper). get_cutlass_(pplx_)moe_mm_data should only be compiled
+  # if it's possible to compile MoE kernels that use its output.
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
              "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
@@ -641,7 +698,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     CUDA_ARCHS "${CUDA_ARCHS}")
 
   list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
+  # 9.0 for latest bf16 atomicAdd PTX
+  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_MOE_ARCHS)
 
     #
@@ -659,7 +717,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
         OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH})
       execute_process(
         COMMAND ${CMAKE_COMMAND} -E env
-        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
+        PYTHONPATH=$PYTHONPATH
           ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT}
         RESULT_VARIABLE moe_marlin_generation_result
         OUTPUT_VARIABLE moe_marlin_generation_output
@@ -695,6 +753,17 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 endif()
 
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  set(MOE_PERMUTE_SRC
+      "csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.cu"
+      "csrc/moe/moe_permute_unpermute_op.cu")
+
+  set_gencode_flags_for_srcs(
+    SRCS "${MARLIN_PERMUTE_SRC}"
+    CUDA_ARCHS "${MOE_PERMUTE_ARCHS}")
+
+  list(APPEND VLLM_MOE_EXT_SRC "${MOE_PERMUTE_SRC}")
+endif()
 message(STATUS "Enabling moe extension.")
 define_gpu_extension_target(
   _moe_C
@@ -703,6 +772,8 @@ define_gpu_extension_target(
   SOURCES ${VLLM_MOE_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
   USE_SABI 3
   WITH_SOABI)
 
@@ -731,5 +802,7 @@ endif()
 # For CUDA we also build and ship some external projects.
 if (VLLM_GPU_LANG STREQUAL "CUDA")
     include(cmake/external_projects/flashmla.cmake)
+
+    # vllm-flash-attn should be last as it overwrites some CMake functions
     include(cmake/external_projects/vllm_flash_attn.cmake)
 endif ()
\ No newline at end of file

CONTRIBUTING.md

 # Contributing to vLLM
 
-You may find information about contributing to vLLM on [docs.vllm.ai](https://docs.vllm.ai/en/latest/contributing/overview.html).
+You may find information about contributing to vLLM on [docs.vllm.ai](https://docs.vllm.ai/en/latest/contributing).

README.md

@@ -97,7 +97,7 @@ python3 setup.py install （若调试，可使用python3 setup.py develop）
 + 若使用 pip install 下载安装过慢,可添加源:-i https://pypi.tuna.tsinghua.edu.cn/simple/
 
 ## 验证
-- python -c "import vllm; print(vllm.\_\_version__)",版本号与官方版本同步,查询该软件的版本号,例如0.8.5.post1;
+- python -c "import vllm; print(vllm.\_\_version__)",版本号与官方版本同步,查询该软件的版本号,例如0.9.1;
 
 ## Known Issue
 - 无

README_ORIGIN.md

 <p align="center">
   <picture>
-    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/logos/vllm-logo-text-dark.png">
-    <img alt="vLLM" src="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/source/assets/logos/vllm-logo-text-light.png" width=55%>
+    <source media="(prefers-color-scheme: dark)" srcset="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/assets/logos/vllm-logo-text-dark.png">
+    <img alt="vLLM" src="https://raw.githubusercontent.com/vllm-project/vllm/main/docs/assets/logos/vllm-logo-text-light.png" width=55%>
   </picture>
 </p>
 
@@ -16,18 +16,20 @@ Easy, fast, and cheap LLM serving for everyone
 ---
 
 *Latest News* 🔥
+- [2025/05] We hosted [NYC vLLM Meetup](https://lu.ma/c1rqyf1f)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1_q_aW_ioMJWUImf1s1YM-ZhjXz8cUeL0IJvaquOYBeA/edit?usp=sharing).
+- [2025/05] vLLM is now a hosted project under PyTorch Foundation! Please find the announcement [here](https://pytorch.org/blog/pytorch-foundation-welcomes-vllm/).
 - [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
+- [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
+
+<details>
+<summary>Previous News</summary>
+
 - [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
 - [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
 - [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).
 - [2025/02] We hosted [the ninth vLLM meetup](https://lu.ma/h7g3kuj9) with Meta! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1jzC_PZVXrVNSFVCW-V4cFXb6pn7zZ2CyP_Flwo05aqg/edit?usp=sharing) and AMD [here](https://drive.google.com/file/d/1Zk5qEJIkTmlQ2eQcXQZlljAx3m9s7nwn/view?usp=sharing). The slides from Meta will not be posted.
-- [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
 - [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing), and Google Cloud team [here](https://drive.google.com/file/d/1h24pHewANyRL11xy5dXUbvRC9F9Kkjix/view?usp=sharing).
 - [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone!
-
-<details>
-<summary>Previous News</summary>
-
 - [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing).
 - [2024/10] We have just created a developer slack ([slack.vllm.ai](https://slack.vllm.ai)) focusing on coordinating contributions and discussing features. Please feel free to join us there!
 - [2024/10] Ray Summit 2024 held a special track for vLLM! Please find the opening talk slides from the vLLM team [here](https://docs.google.com/presentation/d/1B_KQxpHBTRa_mDF-tR6i8rWdOU5QoTZNcEg2MKZxEHM/edit?usp=sharing). Learn more from the [talks](https://www.youtube.com/playlist?list=PLzTswPQNepXl6AQwifuwUImLPFRVpksjR) from other vLLM contributors and users!
@@ -56,8 +58,8 @@ vLLM is fast with:
 - Efficient management of attention key and value memory with [**PagedAttention**](https://blog.vllm.ai/2023/06/20/vllm.html)
 - Continuous batching of incoming requests
 - Fast model execution with CUDA/HIP graph
-- Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), INT4, INT8, and FP8.
-- Optimized CUDA kernels, including integration with FlashAttention and FlashInfer.
+- Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [AutoRound](https://arxiv.org/abs/2309.05516), INT4, INT8, and FP8
+- Optimized CUDA kernels, including integration with FlashAttention and FlashInfer
 - Speculative decoding
 - Chunked prefill
 
@@ -70,14 +72,14 @@ vLLM is flexible and easy to use with:
 - Tensor parallelism and pipeline parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron.
+- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron
 - Prefix caching support
-- Multi-lora support
+- Multi-LoRA support
 
 vLLM seamlessly supports most popular open-source models on HuggingFace, including:
 - Transformer-like LLMs (e.g., Llama)
 - Mixture-of-Expert LLMs (e.g., Mixtral, Deepseek-V2 and V3)
-- Embedding Models (e.g. E5-Mistral)
+- Embedding Models (e.g., E5-Mistral)
 - Multi-modal LLMs (e.g., LLaVA)
 
 Find the full list of supported models [here](https://docs.vllm.ai/en/latest/models/supported_models.html).
@@ -98,14 +100,14 @@ Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
 ## Contributing
 
 We welcome and value any contributions and collaborations.
-Please check out [Contributing to vLLM](https://docs.vllm.ai/en/stable/contributing/overview.html) for how to get involved.
+Please check out [Contributing to vLLM](https://docs.vllm.ai/en/latest/contributing/index.html) for how to get involved.
 
 ## Sponsors
 
 vLLM is a community project. Our compute resources for development and testing are supported by the following organizations. Thank you for your support!
 
 <!-- Note: Please sort them in alphabetical order. -->
-<!-- Note: Please keep these consistent with docs/source/community/sponsors.md -->
+<!-- Note: Please keep these consistent with docs/community/sponsors.md -->
 Cash Donations:
 - a16z
 - Dropbox
@@ -160,4 +162,4 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 
 ## Media Kit
 
-- If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit).
\ No newline at end of file
+- If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit)
\ No newline at end of file

SECURITY.md

@@ -8,4 +8,6 @@ Please report security issues privately using [the vulnerability submission form
 
 ---
 
+Please see the [Security Guide in the vLLM documentation](https://docs.vllm.ai/en/latest/usage/security.html) for more information on vLLM's security assumptions and recommendations.
+
 Please see [PyTorch's Security Policy](https://github.com/pytorch/pytorch/blob/main/SECURITY.md) for more information and recommendations on how to securely interact with models.

benchmarks/README.md

@@ -64,6 +64,12 @@ become available.
       <td style="text-align: center;">✅</td>
       <td><code>lmms-lab/LLaVA-OneVision-Data</code>, <code>Aeala/ShareGPT_Vicuna_unfiltered</code></td>
     </tr>
+    <tr>
+      <td><strong>Custom</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>Local file: <code>data.jsonl</code></td>
+    </tr>
   </tbody>
 </table>
 
@@ -124,6 +130,38 @@ P99 ITL (ms):                            8.39
 ==================================================
 ```
 
+### Custom Dataset
+If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl
+
+```
+{"prompt": "What is the capital of India?"}
+{"prompt": "What is the capital of Iran?"}
+{"prompt": "What is the capital of China?"}
+``` 
+
+```bash
+# start server
+VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct --disable-log-requests
+```
+
+```bash
+# run benchmarking script
+python3 benchmarks/benchmark_serving.py --port 9001 --save-result --save-detailed \
+  --backend vllm \
+  --model meta-llama/Llama-3.1-8B-Instruct \
+  --endpoint /v1/completions \
+  --dataset-name custom \
+  --dataset-path <path-to-your-data-jsonl> \
+  --custom-skip-chat-template \
+  --num-prompts 80 \
+  --max-concurrency 1 \
+  --temperature=0.3 \
+  --top-p=0.75 \
+  --result-dir "./log/"
+```
+
+You can skip applying chat template if your data already has it by using `--custom-skip-chat-template`.
+
 ### VisionArena Benchmark for Vision Language Models
 
 ```bash
@@ -146,10 +184,9 @@ python3 vllm/benchmarks/benchmark_serving.py \
 
 ``` bash
 VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
-    --speculative-model "[ngram]" \
-    --ngram_prompt_lookup_min 2 \
-    --ngram-prompt-lookup-max 5 \
-    --num_speculative_tokens 5
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
 ```
 
 ``` bash
@@ -204,6 +241,16 @@ python3 vllm/benchmarks/benchmark_serving.py \
     --seed 42
 ```
 
+**`philschmid/mt-bench`**
+
+``` bash
+python3 vllm/benchmarks/benchmark_serving.py \
+    --model Qwen/QwQ-32B \
+    --dataset-name hf \
+    --dataset-path philschmid/mt-bench \
+    --num-prompts 80
+```
+
 ### Running With Sampling Parameters
 
 When using OpenAI-compatible backends such as `vllm`, optional sampling
@@ -274,10 +321,9 @@ python3 vllm/benchmarks/benchmark_throughput.py \
     --output-len=100 \
     --num-prompts=2048 \
     --async-engine \
-    --speculative-model="[ngram]" \
-    --ngram_prompt_lookup_min=2 \
-    --ngram-prompt-lookup-max=5 \
-    --num_speculative_tokens=5
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
 ```
 
 ```

benchmarks/auto_tune.sh

+#!/bin/bash
+
+# This script aims to tune the best server parameter combinations to maximize throughput for given requirement. 
+# The current server parameter combination is  max_num_seqs and max_num_batched_tokens
+# It also supports additional requirement: e2e latency and prefix cache. 
+
+# Pre-requisite:
+# 1. Checkout to your branch, install/ update the correct running env. For TPU, activate conda env and install the corresponding torch, xla version. 
+# 2. If the model is customized, replace the MODEL's config with the customized config.
+# 3. Set variables (ALL REQUIRED)
+#   BASE: your directory for vllm repo
+#   MODEL: the model served by vllm
+#   TP: ways of tensor parallelism
+#   DOWNLOAD_DIR: directory to download and load model weights.
+#   INPUT_LEN: request input len
+#   OUTPUT_LEN: request output len
+#   MIN_CACHE_HIT_PCT: prefix cache rate
+#   MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000
+#   NUM_SEQS_LIST: a list of `max-num-seqs` you want to loop with.
+#   NUM_BATCHED_TOKENS_LIST: a list of `max-num-batched-tokens` you want to loop with.
+#   Note that the default NUM_SEQS_LIST and NUM_BATCHED_TOKENS_LIST are set for medium size input/output len, for extra short context (such as 20:20), you might need to include larger numbers in NUM_SEQS_LIST.
+# 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens.
+# 5. The final result will be saved in RESULT file. 
+
+
+# Example use cases 
+# 1. Given input_len=1800, output_len=20, what's the best max_num_seqs and max_num_batched_tokens to get highest throughput?
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=100000000000
+# 2. If we have latency requirement to be lower than 500ms, what's the best server parameter?
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=500
+# 3. If we want to reach 60% prefix cache, what's the best server parameter? 
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=60, MAX_LATENCY_ALLOWED_MS=500
+
+TAG=$(date +"%Y_%m_%d_%H_%M")
+BASE=""
+MODEL="meta-llama/Llama-3.1-8B-Instruct"
+TP=1
+DOWNLOAD_DIR=""
+INPUT_LEN=4000
+OUTPUT_LEN=16
+MIN_CACHE_HIT_PCT=0
+MAX_LATENCY_ALLOWED_MS=100000000000
+NUM_SEQS_LIST="128 256"
+NUM_BATCHED_TOKENS_LIST="512 1024 2048 4096"
+
+LOG_FOLDER="$BASE/auto-benchmark/$TAG"
+RESULT="$LOG_FOLDER/result.txt"
+
+echo "result file: $RESULT"
+echo "model: $MODEL"
+
+rm -rf $LOG_FOLDER
+mkdir -p $LOG_FOLDER
+
+cd "$BASE/vllm"
+
+pip install -q datasets
+
+current_hash=$(git rev-parse HEAD)
+echo "hash:$current_hash" >> "$RESULT"
+echo "current_hash: $current_hash"
+
+best_throughput=0
+best_max_num_seqs=0
+best_num_batched_tokens=0
+best_goodput=0
+
+start_server() {
+    local gpu_memory_utilization=$1
+    local max_num_seqs=$2
+    local max_num_batched_tokens=$3
+    local vllm_log=$4
+    
+    pkill -f vllm
+
+    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 vllm serve $MODEL \
+        --disable-log-requests \
+        --port 8004 \
+        --gpu-memory-utilization $gpu_memory_utilization \
+        --max-num-seqs $max_num_seqs \
+        --max-num-batched-tokens $max_num_batched_tokens \
+        --tensor-parallel-size $TP \
+        --enable-prefix-caching \
+        --load-format dummy \
+        --download-dir "$DOWNLOAD_DIR" \
+        --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 &
+
+    # wait for 10 minutes...
+    server_started=0
+    for i in {1..60}; do  
+        RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout)
+        STATUS_CODE=$(echo "$RESPONSE" | tail -n 1) 
+        if [[ "$STATUS_CODE" -eq 200 ]]; then
+            server_started=1
+            break
+        else
+            sleep 10
+        fi
+    done
+    if (( ! server_started )); then
+        echo "server did not start within 10 minutes. Please check server log at $vllm_log".
+        return 1
+    else
+        return 0
+    fi
+}
+
+run_benchmark() {
+    local max_num_seqs=$1
+    local max_num_batched_tokens=$2
+    local gpu_memory_utilization=$3
+    echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
+    local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
+    echo "vllm_log: $vllm_log"
+    echo
+    rm -f $vllm_log
+    pkill -f vllm
+
+    echo "starting server..."
+    start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log
+    result=$?
+    if [[ "$result" -eq 1 ]]; then
+        echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
+    else
+        echo "server started."
+    fi
+    echo
+    
+    echo "run benchmark test..."
+    meet_latency_requirement=0
+    # get a basic qps by using request-rate inf
+    bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
+    prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
+    python benchmarks/benchmark_serving.py \
+        --backend vllm \
+        --model $MODEL  \
+        --dataset-name random \
+        --random-input-len $INPUT_LEN \
+        --random-output-len $OUTPUT_LEN \
+        --ignore-eos \
+        --disable-tqdm \
+        --request-rate inf \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+        --num-prompts 1000 \
+        --random-prefix-len $prefix_len \
+        --port 8004 &> "$bm_log"
+    throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+    e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
+    goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+
+    if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
+        meet_latency_requirement=1
+        request_rate=inf
+    fi
+
+    if (( ! meet_latency_requirement )); then
+    # start from request-rate as int(throughput) + 1
+        request_rate=$((${throughput%.*} + 1))
+        while ((request_rate > 0)); do
+            # clear prefix cache
+            curl -X POST http://0.0.0.0:8004/reset_prefix_cache
+            sleep 5
+            bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt"
+            python benchmarks/benchmark_serving.py \
+                --backend vllm \
+                --model $MODEL  \
+                --dataset-name random \
+                --random-input-len $INPUT_LEN \
+                --random-output-len $OUTPUT_LEN \
+                --ignore-eos \
+                --disable-tqdm \
+                --request-rate $request_rate \
+                --percentile-metrics ttft,tpot,itl,e2el \
+                --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+                --num-prompts 100 \
+                --random-prefix-len $prefix_len \
+                --port 8004 &> "$bm_log"
+            throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+            e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
+            goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+            if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
+                meet_latency_requirement=1
+                break
+            fi
+            request_rate=$((request_rate-1))
+        done
+    fi
+    # write the results and update the best result.
+    if ((meet_latency_requirement)); then
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput"
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" >> "$RESULT"
+        if (( $(echo "$throughput > $best_throughput" | bc -l) )); then
+            best_throughput=$throughput
+            best_max_num_seqs=$max_num_seqs
+            best_num_batched_tokens=$max_num_batched_tokens
+            best_goodput=$goodput
+        fi
+    else
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}"
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" >> "$RESULT"
+    fi
+
+    echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
+
+    pkill vllm
+    sleep 10
+    printf '=%.0s' $(seq 1 20)
+    return 0
+}
+
+read -r -a num_seqs_list <<< "$NUM_SEQS_LIST"
+read -r -a num_batched_tokens_list <<< "$NUM_BATCHED_TOKENS_LIST"
+
+# first find out the max gpu-memory-utilization without HBM OOM.
+gpu_memory_utilization=0.98
+find_gpu_memory_utilization=0
+while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do
+    start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log"
+    result=$?
+    if [[ "$result" -eq 0 ]]; then
+        find_gpu_memory_utilization=1
+        break
+    else
+        gpu_memory_utilization=$(echo "$gpu_memory_utilization - 0.01" | bc)
+    fi
+done
+
+if [[ "$find_gpu_memory_utilization" -eq 1 ]]; then
+    echo "Using gpu_memory_utilization=$gpu_memory_utilization to serve model."
+else
+    echo "Cannot find a proper gpu_memory_utilization over 0.9 to serve the model, please check logs in $LOG_FOLDER."
+    exit 1
+fi
+
+for num_seqs in "${num_seqs_list[@]}"; do
+    for num_batched_tokens in "${num_batched_tokens_list[@]}"; do
+        run_benchmark $num_seqs $num_batched_tokens $gpu_memory_utilization
+    done
+done
+echo "finish permutations"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput" >> "$RESULT"
+

benchmarks/benchmark_latency.py

 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark the latency of processing a single batch of requests."""
 
 import argparse
@@ -6,14 +7,13 @@ import dataclasses
 import json
 import os
 import time
-from pathlib import Path
 from typing import Any, Optional
 
 import numpy as np
-import torch
-from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from tqdm import tqdm
 
+import vllm.envs as envs
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
 from vllm.inputs import PromptType
@@ -21,13 +21,14 @@ from vllm.sampling_params import BeamSearchParams
 from vllm.utils import FlexibleArgumentParser
 
 
-def save_to_pytorch_benchmark_format(args: argparse.Namespace,
-                                     results: dict[str, Any]) -> None:
+def save_to_pytorch_benchmark_format(
+    args: argparse.Namespace, results: dict[str, Any]
+) -> None:
     pt_records = convert_to_pytorch_benchmark_format(
         args=args,
         metrics={"latency": results["latencies"]},
-        extra_info={k: results[k]
-                    for k in ["avg_latency", "percentiles"]})
+        extra_info={k: results[k] for k in ["avg_latency", "percentiles"]},
+    )
     if pt_records:
         pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
         write_to_json(pt_file, pt_records)
@@ -42,9 +43,11 @@ def main(args: argparse.Namespace):
     # the engine will automatically process the request in multiple batches.
     llm = LLM(**dataclasses.asdict(engine_args))
     assert llm.llm_engine.model_config.max_model_len >= (
-        args.input_len +
-        args.output_len), ("Please ensure that max_model_len is greater than"
-                           " the sum of input_len and output_len.")
+        args.input_len + args.output_len
+    ), (
+        "Please ensure that max_model_len is greater than"
+        " the sum of input_len and output_len."
+    )
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -55,18 +58,16 @@ def main(args: argparse.Namespace):
         detokenize=not args.disable_detokenize,
     )
     print(sampling_params)
-    dummy_prompt_token_ids = np.random.randint(10000,
-                                               size=(args.batch_size,
-                                                     args.input_len))
-    dummy_prompts: list[PromptType] = [{
-        "prompt_token_ids": batch
-    } for batch in dummy_prompt_token_ids.tolist()]
+    dummy_prompt_token_ids = np.random.randint(
+        10000, size=(args.batch_size, args.input_len)
+    )
+    dummy_prompts: list[PromptType] = [
+        {"prompt_token_ids": batch} for batch in dummy_prompt_token_ids.tolist()
+    ]
 
     def llm_generate():
         if not args.use_beam_search:
-            llm.generate(dummy_prompts,
-                         sampling_params=sampling_params,
-                         use_tqdm=False)
+            llm.generate(dummy_prompts, sampling_params=sampling_params, use_tqdm=False)
         else:
             llm.beam_search(
                 dummy_prompts,
@@ -79,16 +80,9 @@ def main(args: argparse.Namespace):
 
     def run_to_completion(profile_dir: Optional[str] = None):
         if profile_dir:
-            with torch.profiler.profile(
-                    activities=[
-                        torch.profiler.ProfilerActivity.CPU,
-                        torch.profiler.ProfilerActivity.CUDA,
-                    ],
-                    on_trace_ready=torch.profiler.tensorboard_trace_handler(
-                        str(profile_dir)),
-            ) as p:
-                llm_generate()
-            print(p.key_averages().table(sort_by="self_cuda_time_total"))
+            llm.start_profile()
+            llm_generate()
+            llm.stop_profile()
         else:
             start_time = time.perf_counter()
             llm_generate()
@@ -101,10 +95,7 @@ def main(args: argparse.Namespace):
         run_to_completion(profile_dir=None)
 
     if args.profile:
-        profile_dir = args.profile_result_dir
-        if not profile_dir:
-            profile_dir = (Path(".") / "vllm_benchmark_result" /
-                           f"latency_result_{time.time()}")
+        profile_dir = envs.VLLM_TORCH_PROFILER_DIR
         print(f"Profiling (results will be saved to '{profile_dir}')...")
         run_to_completion(profile_dir=profile_dir)
         return
@@ -135,7 +126,8 @@ def main(args: argparse.Namespace):
 if __name__ == "__main__":
     parser = FlexibleArgumentParser(
         description="Benchmark the latency of processing a single batch of "
-        "requests till completion.")
+        "requests till completion."
+    )
     parser.add_argument("--input-len", type=int, default=32)
     parser.add_argument("--output-len", type=int, default=128)
     parser.add_argument("--batch-size", type=int, default=8)
@@ -152,22 +144,14 @@ if __name__ == "__main__":
         default=10,
         help="Number of iterations to run for warmup.",
     )
-    parser.add_argument("--num-iters",
-                        type=int,
-                        default=30,
-                        help="Number of iterations to run.")
+    parser.add_argument(
+        "--num-iters", type=int, default=30, help="Number of iterations to run."
+    )
     parser.add_argument(
         "--profile",
         action="store_true",
         help="profile the generation process of a single batch",
     )
-    parser.add_argument(
-        "--profile-result-dir",
-        type=str,
-        default=None,
-        help=("path to save the pytorch profiler output. Can be visualized "
-              "with ui.perfetto.dev or Tensorboard."),
-    )
     parser.add_argument(
         "--output-json",
         type=str,
@@ -177,10 +161,20 @@ if __name__ == "__main__":
     parser.add_argument(
         "--disable-detokenize",
         action="store_true",
-        help=("Do not detokenize responses (i.e. do not include "
-              "detokenization time in the latency measurement)"),
+        help=(
+            "Do not detokenize responses (i.e. do not include "
+            "detokenization time in the latency measurement)"
+        ),
     )
 
     parser = EngineArgs.add_cli_args(parser)
+    # V1 enables prefix caching by default which skews the latency
+    # numbers. We need to disable prefix caching by default.
+    parser.set_defaults(enable_prefix_caching=False)
     args = parser.parse_args()
-    main(args)
\ No newline at end of file
+    if args.profile and not envs.VLLM_TORCH_PROFILER_DIR:
+        raise OSError(
+            "The environment variable 'VLLM_TORCH_PROFILER_DIR' is not set. "
+            "Please set it to a valid path to use torch profiler."
+        )
+    main(args)