Merge branch 'vllm-project:main' into enable_bitsandbytes_quant_rocm

Author: sstamenk

.buildkite/test-pipeline.yaml

@@ -172,6 +172,8 @@ steps:
   - tests/v1/engine/test_engine_core_client.py
   - tests/distributed/test_symm_mem_allreduce.py
   commands:
+  # https://github.com/NVIDIA/nccl/issues/1838
+  - export NCCL_CUMEM_HOST_ENABLE=0
   # test with torchrun tp=2 and external_dp=2
   - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
   # test with torchrun tp=2 and pp=2
@@ -349,7 +351,8 @@ steps:
     - python3 offline_inference/basic/embed.py
     - python3 offline_inference/basic/score.py
     - python3 offline_inference/spec_decode.py --test --method eagle --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
-    - python3 offline_inference/spec_decode.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 2048
+    # https://github.com/vllm-project/vllm/pull/26682 uses slightly more memory in PyTorch 2.9+ causing this test to OOM in 1xL4 GPU
+    - python3 offline_inference/spec_decode.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
 
 - label: Platform Tests (CUDA) # 4min
   timeout_in_minutes: 15
@@ -534,7 +537,7 @@ steps:
   # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now
   # we can only upgrade after this is resolved
   # TODO(jerryzh168): resolve the above comment
-  - uv pip install --system torchao==0.13.0
+  - uv pip install --system torchao==0.13.0 --index-url https://download.pytorch.org/whl/cu129
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
 - label: LM Eval Small Models # 53min
@@ -975,6 +978,8 @@ steps:
   - tests/v1/shutdown
   - tests/v1/worker/test_worker_memory_snapshot.py
   commands:
+  # https://github.com/NVIDIA/nccl/issues/1838
+  - export NCCL_CUMEM_HOST_ENABLE=0
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
   - DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py

.github/CODEOWNERS

@@ -5,8 +5,8 @@
 /vllm/attention @LucasWilkinson
 /vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @22quinn
-/vllm/model_executor/layers/fused_moe @mgoin
-/vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256
+/vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
+/vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
 /vllm/model_executor/layers/mamba @tdoublep
 /vllm/model_executor/model_loader @22quinn
 /vllm/multimodal @DarkLight1337 @ywang96 @NickLucche
@@ -25,7 +25,8 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # vLLM V1
 /vllm/v1/attention @LucasWilkinson
-/vllm/v1/attention/backends/flashinfer.py @mgoin
+/vllm/v1/attention/backends/mla @pavanimajety
+/vllm/v1/attention/backends/flashinfer.py @mgoin @pavanimajety
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
 /vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat @heheda12345 @ApostaC
 /vllm/v1/sample @22quinn @houseroad @njhill
@@ -44,7 +45,7 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/kernels @mgoin @tlrmchlsmth @WoosukKwon @yewentao256
 /tests/models @DarkLight1337 @ywang96
 /tests/multimodal @DarkLight1337 @ywang96 @NickLucche
-/tests/quantization @mgoin @robertgshaw2-redhat @yewentao256
+/tests/quantization @mgoin @robertgshaw2-redhat @yewentao256 @pavanimajety
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm

.pre-commit-config.yaml

@@ -38,7 +38,7 @@ repos:
   rev: 0.9.1
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128, --python-platform, x86_64-manylinux_2_28]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu129, --python-platform, x86_64-manylinux_2_28]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:

CMakeLists.txt

@@ -49,8 +49,8 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.8.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.8.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.9.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.9.0")
 
 #
 # Try to find python package with an executable that exactly matches

cmake/external_projects/flashmla.cmake

@@ -19,7 +19,7 @@ else()
   FetchContent_Declare(
         flashmla
         GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
-        GIT_TAG 5f65b85703c7ed75fda01e06495077caad207c3f
+        GIT_TAG 28417e516fcbf6257a422ba117ef5b6f44da5682
         GIT_PROGRESS TRUE
         CONFIGURE_COMMAND ""
         BUILD_COMMAND ""
@@ -66,6 +66,7 @@ if(FLASH_MLA_ARCHS)
         ${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/extension/sm90/dense_fp8/flash_fwd_mla_metadata.cu
     )
 
     set(FlashMLA_INCLUDES

docker/Dockerfile

@@ -5,7 +5,7 @@
 # docs/contributing/dockerfile/dockerfile.md and
 # docs/assets/contributing/dockerfile-stages-dependency.png
 
-ARG CUDA_VERSION=12.8.1
+ARG CUDA_VERSION=12.9.1
 ARG PYTHON_VERSION=3.12
 
 # By parameterizing the base images, we allow third-party to use their own
@@ -132,7 +132,9 @@ 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 \
+    # TODO: remove apache-tvm-ffi once FlashInfer is fixed https://github.com/flashinfer-ai/flashinfer/issues/1962
+    uv pip install --python /opt/venv/bin/python3 --pre apache-tvm-ffi==0.1.0b15 \
+    && uv pip install --python /opt/venv/bin/python3 -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
@@ -273,6 +275,7 @@ WORKDIR /vllm-workspace
 ENV DEBIAN_FRONTEND=noninteractive
 ARG TARGETPLATFORM
 
+# TODO (huydhn): There is no prebuilt gdrcopy package on 12.9 at the moment
 ARG GDRCOPY_CUDA_VERSION=12.8
 # Keep in line with FINAL_BASE_IMAGE
 ARG GDRCOPY_OS_VERSION=Ubuntu22_04
@@ -353,9 +356,18 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Install vllm wheel first, so that torch etc will be installed.
 RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
     --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system dist/*.whl --verbose \
+    # TODO: remove apache-tvm-ffi once FlashInfer is fixed https://github.com/flashinfer-ai/flashinfer/issues/1962
+    uv pip install --system --pre apache-tvm-ffi==0.1.0b15 \
+    && uv pip install --system dist/*.whl --verbose \
         --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
 
+# TODO (huydhn): Remove this once xformers is released for 2.9.0
+RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
+    . /etc/environment
+    export TORCH_CUDA_ARCH_LIST='7.5 8.0+PTX 9.0a'
+    uv pip install --system --no-build-isolation "git+https://github.com/facebookresearch/[email protected]"
+BASH
+
 # Install FlashInfer pre-compiled kernel cache and binaries
 # https://docs.flashinfer.ai/installation.html
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -422,6 +434,7 @@ ARG PYTHON_VERSION
 
 ARG PIP_INDEX_URL UV_INDEX_URL
 ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
+ARG PYTORCH_CUDA_INDEX_BASE_URL
 
 # 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
@@ -434,7 +447,8 @@ ENV UV_LINK_MODE=copy
 RUN --mount=type=cache,target=/root/.cache/uv \
     CUDA_MAJOR="${CUDA_VERSION%%.*}"; \
     if [ "$CUDA_MAJOR" -ge 12 ]; then \
-        uv pip install --system -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 '.'); \
     fi
 
 # install development dependencies (for testing)

docker/Dockerfile.cpu

@@ -199,9 +199,13 @@ FROM base AS vllm-test-deps
 
 WORKDIR /workspace/vllm
 
+# TODO: Update to 2.9.0 when there is a new build for intel_extension_for_pytorch for that version
 RUN --mount=type=bind,src=requirements/test.in,target=requirements/test.in \
     cp requirements/test.in requirements/cpu-test.in && \
     sed -i '/mamba_ssm/d' requirements/cpu-test.in && \
+    sed -i 's/^torch==.*/torch==2.8.0/g' requirements/cpu-test.in && \
+    sed -i 's/torchaudio.*/torchaudio/g' requirements/cpu-test.in && \
+    sed -i 's/torchvision.*/torchvision/g' requirements/cpu-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 \

docker/Dockerfile.rocm_base

@@ -1,13 +1,13 @@
 ARG BASE_IMAGE=rocm/dev-ubuntu-22.04:7.0-complete
-ARG TRITON_BRANCH="f9e5bf54"
+ARG TRITON_BRANCH="57c693b6"
 ARG TRITON_REPO="https://github.com/ROCm/triton.git"
-ARG PYTORCH_BRANCH="b2fb6885"
+ARG PYTORCH_BRANCH="1c57644d"
 ARG PYTORCH_VISION_BRANCH="v0.23.0"
 ARG PYTORCH_REPO="https://github.com/ROCm/pytorch.git"
 ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
 ARG FA_BRANCH="0e60e394"
 ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
-ARG AITER_BRANCH="2ab9f4cd"
+ARG AITER_BRANCH="eef23c7f"
 ARG AITER_REPO="https://github.com/ROCm/aiter.git"
 
 FROM ${BASE_IMAGE} AS base

docs/assets/contributing/dockerfile-stages-dependency.png

@@ -87,7 +87,7 @@ is ineffective.
 
 While ongoing efforts like <https://github.com/vllm-project/vllm/issues/17419>
 address the long build time at its source, the current workaround is to set `VLLM_CI_BRANCH`
-to a custom branch provided by @khluu (`VLLM_CI_BRANCH=khluu/use_postmerge_q`)
+to a custom branch provided by @khluu (`VLLM_CI_BRANCH=khluu/long_build`)
 when manually triggering a build on Buildkite. This branch accomplishes two things:
 
 1. Increase the timeout limit to 10 hours so that the build doesn't time out.
@@ -100,35 +100,17 @@ to warm it up so that future builds are faster.
 
 ## Update dependencies
 
-Several vLLM dependencies, such as FlashInfer, also depend on PyTorch and need
+Several vLLM dependencies like xFormers depend on PyTorch and need
 to be updated accordingly. Rather than waiting for all of them to publish new
 releases (which would take too much time), they can be built from
 source to unblock the update process.
 
-### FlashInfer
-
-Here is how to build and install it from source with `torch2.7.0+cu128` in vLLM [Dockerfile](https://github.com/vllm-project/vllm/blob/27bebcd89792d5c4b08af7a65095759526f2f9e1/docker/Dockerfile#L259-L271):
-
-```bash
-export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.9 9.0 10.0+PTX'
-export FLASHINFER_ENABLE_SM90=1
-uv pip install --system \
-    --no-build-isolation "git+https://github.com/flashinfer-ai/[email protected]"
-```
-
-One caveat is that building FlashInfer from source adds approximately 30
-minutes to the vLLM build time. Therefore, it's preferable to cache the wheel in a
-public location for immediate installation, such as [this FlashInfer wheel link](https://download.pytorch.org/whl/cu128/flashinfer/flashinfer_python-0.2.6.post1%2Bcu128torch2.7-cp39-abi3-linux_x86_64.whl). For future releases, contact the PyTorch release
-team if you want to get the package published there.
-
 ### xFormers
 
-Similar to FlashInfer, here is how to build and install xFormers from source:
-
 ```bash
-export TORCH_CUDA_ARCH_LIST='7.0 7.5 8.0 8.9 9.0 10.0+PTX'
+export TORCH_CUDA_ARCH_LIST='7.5 8.0+PTX 9.0a'
 MAX_JOBS=16 uv pip install --system \
-    --no-build-isolation "git+https://github.com/facebookresearch/[email protected].30"
+    --no-build-isolation "git+https://github.com/facebookresearch/[email protected].32.post2"
 ```
 
 ## Update all the different vLLM platforms