[DO NOT MERGE] 2.9, Inductor partition, standalone compile, monkeypatch fix(es), spawn

.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
@@ -527,8 +529,7 @@ steps:
   # since torchao nightly is only compatible with torch nightly currently
   # 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
+  - pip install --pre torchao==0.15.0.dev20251014 --index-url https://download.pytorch.org/whl/nightly/cu128
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/
 
 - label: LM Eval Small Models # 53min
@@ -944,6 +945,8 @@ 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

.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, --extra-index-url, https://download.pytorch.org/whl/test/cu128, --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

docker/Dockerfile

@@ -55,7 +55,7 @@ ARG UV_INDEX_URL=${PIP_INDEX_URL}
 ARG UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
 
 # PyTorch provides its own indexes for standard and nightly builds
-ARG PYTORCH_CUDA_INDEX_BASE_URL=https://download.pytorch.org/whl
+ARG PYTORCH_CUDA_INDEX_BASE_URL=https://download.pytorch.org/whl/test
 ARG PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL=https://download.pytorch.org/whl/nightly
 
 # PIP supports multiple authentication schemes, including keyring
@@ -356,6 +356,13 @@ RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist
     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 +429,7 @@ 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 +442,8 @@ 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

@@ -111,9 +111,13 @@ 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 \

docs/contributing/ci/update_pytorch_version.md

@@ -87,7 +87,7 @@ is ineffective.
 
 While ongoing efforts like [#17419](gh-issue: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.
@@ -107,28 +107,24 @@ 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):
+After #25782, the pre-compiled FlashInfer wheel can be built using tools/flashinfer-build.sh
+script. The new wheel can then be uploaded to [PyTorch test index](https://download.pytorch.org/whl/test/cu128/flashinfer_python-0.3.1-cp39-abi3-linux_x86_64.whl) and used during the update.
+
+During PyTorch 2.9 update, using the old FlashInfer wheel built for
+2.8 led to a crash with the following error:
 
 ```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]"
+terminate called after throwing an instance of 'std::bad_array_new_length'
 ```
 
-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

pyproject.toml

@@ -6,7 +6,7 @@ requires = [
     "packaging>=24.2",
     "setuptools>=77.0.3,<80.0.0",
     "setuptools-scm>=8.0",
-    "torch == 2.8.0",
+    "torch == 2.9.0",
     "wheel",
     "jinja2",
 ]

requirements/build.txt

@@ -4,7 +4,7 @@ ninja
 packaging>=24.2
 setuptools>=77.0.3,<80.0.0
 setuptools-scm>=8
-torch==2.8.0
+torch==2.9.0
 wheel
 jinja2>=3.1.6
 regex

requirements/cuda.txt

@@ -5,11 +5,11 @@ numba == 0.61.2 # Required for N-gram speculative decoding
 
 # Dependencies for NVIDIA GPUs
 ray[cgraph]>=2.48.0 # Ray Compiled Graph, required for pipeline parallelism in V1.
-torch==2.8.0
-torchaudio==2.8.0
+torch==2.9.0
+torchaudio==2.9.0
 # These must be updated alongside torch
-torchvision==0.23.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
+torchvision==0.24.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
 # https://github.com/facebookresearch/xformers/releases/tag/v0.0.32.post1
-xformers==0.0.32.post1; platform_system == 'Linux' and platform_machine == 'x86_64'  # Requires PyTorch >= 2.8
+# xformers==0.0.32.post1; platform_system == 'Linux' and platform_machine == 'x86_64'  # Requires PyTorch >= 2.8
 # FlashInfer should be updated together with the Dockerfile
-flashinfer-python==0.4.0
+flashinfer-python==0.4.0

requirements/rocm-build.txt

@@ -1,10 +1,10 @@
 # Common dependencies
 -r common.txt
 
---extra-index-url https://download.pytorch.org/whl/rocm6.3
-torch==2.8.0
-torchvision==0.23.0
-torchaudio==2.8.0
+--extra-index-url https://download.pytorch.org/whl/test/rocm6.3
+torch==2.9.0
+torchvision==0.24.0
+torchaudio==2.9.0
 
 triton==3.3.0
 cmake>=3.26.1,<4

requirements/test.in

@@ -24,9 +24,9 @@ soundfile # required for audio tests
 jiwer # required for audio tests
 tblib # for pickling test exceptions
 timm >=1.0.17 # required for internvl and gemma3n-mm test
-torch==2.8.0
-torchaudio==2.8.0
-torchvision==0.23.0
+torch==2.9.0
+torchaudio==2.9.0
+torchvision==0.24.0
 transformers_stream_generator # required for qwen-vl test
 matplotlib # required for qwen-vl test
 mistral_common[image,audio] >= 1.8.5 # required for voxtral test