Merging in the latest merge from vllm-project to ROCm

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

@@ -426,7 +426,7 @@ main() {
 
   pip install -U transformers
 
-  pip install -r requirements-dev.txt
+  pip install -r requirements/dev.txt
   which genai-perf
 
   # check storage

.buildkite/run-amd-test.sh

@@ -93,7 +93,12 @@ if [[ $commands == *" kernels "* ]]; then
   --ignore=kernels/test_rand.py \
   --ignore=kernels/test_sampler.py \
   --ignore=kernels/test_cascade_flash_attn.py \
-  --ignore=kernels/test_mamba_mixer2.py"
+  --ignore=kernels/test_mamba_mixer2.py \
+  --ignore=kernels/test_aqlm.py \
+  --ignore=kernels/test_machete_mm.py \
+  --ignore=kernels/test_mha_attn.py \
+  --ignore=kernels/test_block_fp8.py \
+  --ignore=kernels/test_permute_cols.py"
 fi
 
 #ignore certain Entrypoints tests

.buildkite/run-cpu-test.sh

@@ -35,7 +35,7 @@ function cpu_tests() {
   # Run basic model test
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
-    pip install -r vllm/requirements-test.txt
+    pip install -r vllm/requirements/test.txt
     pytest -v -s tests/models/decoder_only/language -m cpu_model
     pytest -v -s tests/models/embedding/language -m cpu_model
     pytest -v -s tests/models/encoder_decoder/language -m cpu_model

.buildkite/test-pipeline.yaml

@@ -35,7 +35,7 @@ steps:
   fast_check: true
   no_gpu: True
   commands:
-  - pip install -r requirements-docs.txt
+  - pip install -r ../../requirements/docs.txt
   - SPHINXOPTS=\"-W\" make html
   # Check API reference (if it fails, you may have missing mock imports)
   - grep \"sig sig-object py\" build/html/api/inference_params.html
@@ -78,6 +78,7 @@ steps:
   - tests/basic_correctness/test_preemption
   - tests/basic_correctness/test_cumem.py
   commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s basic_correctness/test_cumem.py
   - pytest -v -s basic_correctness/test_basic_correctness.py
   - pytest -v -s basic_correctness/test_cpu_offload.py
@@ -115,6 +116,7 @@ steps:
   - tests/entrypoints/test_chat_utils
   - tests/entrypoints/offline_mode
   commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
   - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
@@ -146,8 +148,10 @@ steps:
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
-  - python3 ../examples/offline_inference/rlhf.py
-  - RAY_DEDUP_LOGS=0 python3 ../examples/offline_inference/rlhf_colocate.py
+  - pushd ../examples/offline_inference
+  - python3 rlhf.py
+  - RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
+  - popd
 
 - label: Metrics, Tracing Test # 10min
   num_gpus: 2
@@ -204,6 +208,7 @@ steps:
     - VLLM_USE_V1=1 pytest -v -s v1/engine
     - VLLM_USE_V1=1 pytest -v -s v1/sample
     - VLLM_USE_V1=1 pytest -v -s v1/worker
+    - VLLM_USE_V1=1 pytest -v -s v1/structured_output
     - VLLM_USE_V1=1 pytest -v -s v1/test_stats.py
     - VLLM_USE_V1=1 pytest -v -s v1/test_utils.py
     # TODO: accuracy does not match, whether setting

.github/mergify.yml

@@ -36,6 +36,21 @@ pull_request_rules:
       add:
         - frontend
 
+- name: label-multi-modality
+  description: Automatically apply multi-modality label
+  conditions:
+    - or:
+      - files~=^vllm/multimodal/
+      - files~=^tests/multimodal/
+      - files~=^tests/models/multimodal/
+      - files~=^tests/models/*/audio_language/
+      - files~=^tests/models/*/vision_language/
+      - files=tests/models/test_vision.py
+  actions:
+    label:
+      add:
+        - multi-modality
+
 - name: label-structured-output
   description: Automatically apply structured-output label
   conditions:

.github/workflows/scripts/build.sh

@@ -5,7 +5,7 @@ python_executable=python3
 
 # Update paths
 # Install requirements
-$python_executable -m pip install -r requirements-rocm.txt
+$python_executable -m pip install -r requirements/rocm.txt
 
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1

.gitignore

@@ -197,7 +197,7 @@ _build/
 hip_compat.h
 
 # Benchmark dataset
-benchmarks/*.json
+benchmarks/**/*.json
 
 # Linting
 actionlint

.pre-commit-config.yaml

@@ -44,8 +44,8 @@ repos:
   rev: 0.6.2
   hooks:
     - id: pip-compile
-      args: [requirements-test.in, -o, requirements-test.txt]
-      files: ^requirements-test\.(in|txt)$
+      args: [requirements/test.in, -o, requirements/test.txt]
+      files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
   - id: mypy-local

.readthedocs.yaml

@@ -18,4 +18,4 @@ formats: []
 # Optionally declare the Python requirements required to build your docs
 python:
   install:
-    - requirements: docs/requirements-docs.txt
+    - requirements: requirements/docs.txt

CMakeLists.txt

@@ -31,7 +31,7 @@ set(ignoreMe "${VLLM_PYTHON_PATH}")
 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")
+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 "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
@@ -312,7 +312,7 @@ 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" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
     set(MARLIN_SRCS
        "csrc/quantization/fp8/fp8_marlin.cu"
@@ -334,7 +334,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # Only build AllSpark kernels if we are building for at least some compatible archs.
   cuda_archs_loose_intersection(ALLSPARK_ARCHS "8.0;8.6;8.7;8.9" "${CUDA_ARCHS}")
-  if (ALLSPARK_ARCHS)
+  if (${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND ALLSPARK_ARCHS)
     set(ALLSPARK_SRCS
        "csrc/quantization/gptq_allspark/allspark_repack.cu"
        "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
@@ -345,46 +345,74 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     message(STATUS "Building AllSpark kernels for archs: ${ALLSPARK_ARCHS}")
   else()
     message(STATUS "Not building AllSpark kernels as no compatible archs found"
-                   " in CUDA target architectures")
+                   " in CUDA target architectures, or CUDA not >= 12.0")
   endif()
 
+
+  set(SCALED_MM_3X_ARCHS)
   # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
-  # CUDA 12.0 or later (and only work on Hopper, 9.0a for now).
-  cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
+  # CUDA 12.0 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
     set(SRCS
-       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.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_3X_ARCHS}")
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_C3X=1")
-    message(STATUS "Building scaled_mm_c3x for archs: ${SCALED_MM_3X_ARCHS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM90=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
-      message(STATUS "Not building scaled_mm_c3x as CUDA Compiler version is "
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
                      "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                      "later if you intend on running FP8 quantized models on "
                      "Hopper.")
     else()
-      message(STATUS "Not building scaled_mm_c3x as no compatible archs found "
+      message(STATUS "Not building scaled_mm_c3x_sm90 as no compatible archs found "
                      "in CUDA target architectures")
     endif()
+  endif()
 
-    # clear SCALED_MM_3X_ARCHS so the scaled_mm_c2x kernels know we didn't
-    # build any 3x kernels
-    set(SCALED_MM_3X_ARCHS)
+  # The cutlass_scaled_mm kernels for Blackwell (c3x, i.e. CUTLASS 3.x) require
+  # CUDA 12.8 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;12.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
+    )
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM100=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm100 for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm100 as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
+                     "later if you intend on running FP8 quantized models on "
+                     "Blackwell.")
+    else()
+      message(STATUS "Not building scaled_mm_c3x_100 as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
   endif()
 
   #
   # 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.
   cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
-    "7.5;8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
+    "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${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)
@@ -409,17 +437,17 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # 2:4 Sparse Kernels
 
   # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
-  # require CUDA 12.2 or later (and only work on Hopper, 9.0a for now).
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
+  # require CUDA 12.2 or later (and only work on Hopper and Blackwell).
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
-    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_3X_ARCHS}")
+    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
       message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
                      "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
                      "if you intend on running FP8 sparse quantized models on Hopper.")
@@ -434,8 +462,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   if(${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_scaled_mm_kernels.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${FP4_ARCHS}")
@@ -534,6 +561,7 @@ define_gpu_extension_target(
   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)
 
@@ -557,7 +585,7 @@ set_gencode_flags_for_srcs(
   CUDA_ARCHS "${CUDA_ARCHS}")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
   if (MARLIN_MOE_ARCHS)
     set(MARLIN_MOE_SRC
         "csrc/moe/marlin_kernels/marlin_moe_kernel.h"