2111997f96 · tduncan/Nvidia-Optimization-Stack-VLLM

.buildkite/test-amd.yaml

+13 -5

@@ -2165,7 +2165,15 @@ steps:

2165 - vllm/platforms/rocm.py

2166 - tests/quantization

2167 commands:

2168 - - uv pip install --system torchao==0.14.1

2168 +

2169 + # temporary install here since we need nightly, will move to requirements/test.in

2170 + # after torchao 0.12 release, and pin a working version of torchao nightly here

2171 +

2172 + # since torchao nightly is only compatible with torch nightly currently

2173 + # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now

2174 + # we can only upgrade after this is resolved

2175 + # TODO(jerryzh168): resolve the above comment

2176 + - uv pip install --system torchao==0.17.0

2177 - uv pip install --system conch-triton-kernels

2178 - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py

2179

@@ -2924,10 +2932,10 @@ steps:

2932 - bash .buildkite/scripts/scheduled_integration_test/qwen3_next_mtp_async_eplb.sh 0.8 1319 8040

2933

2934 ##### .buildkite/test_areas/compile.yaml #####

2927 -# Slowly setting up the tests so that it is also easier for the

2935 +# Slowly setting up the tests so that it is also easier for the

2936 # CI team to review and upstream to the pipelinev2.

2937 # The following tests are important for vLLM IR Ops refactoring,

2930 -# which affects fusion passes on ROCm. So we have to

2938 +# which affects fusion passes on ROCm. So we have to

2939 # enable them as as soon as possible.

2940

2941 ## TODO: Enable the test in this group

@@ -3006,7 +3014,7 @@ steps:

3014

3015 ## There are no ops on ROCm for these tests.

3016 ## The test still passes but the logs are not useful.

3009 -## fused ops just call torch.ops.symm_mem which

3017 +## fused ops just call torch.ops.symm_mem which

3018 ## exists in ROCm even though they don't work

3019 # - label: AsyncTP Correctness Tests (2xH100-2xMI325)

3020 # - label: Fusion E2E TP2 Quick (H100-MI325)

@@ -3338,7 +3346,7 @@ steps:

3346 - vllm/_aiter_ops.py

3347 - vllm/platforms/rocm.py

3348 commands:

3341 - - uv pip install --system torchao==0.14.1

3349 + - uv pip install --system torchao==0.17.0

3350 - uv pip install --system conch-triton-kernels

3351 - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py

3352

.buildkite/test_areas/models_language.yaml

+2 -2

@@ -38,7 +38,7 @@ steps:

38 # Install fast path packages for testing against transformers

39 # Note: also needed to run plamo2 model in vLLM

40 - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'

41 - - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'

41 + - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'

42 # Shard hybrid language model tests

43 - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB

44 parallelism: 2

@@ -53,7 +53,7 @@ steps:

53 # Install fast path packages for testing against transformers

54 # Note: also needed to run plamo2 model in vLLM

55 - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'

56 - - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'

56 + - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'

57 - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'

58 mirror:

59 amd:

.buildkite/test_areas/quantization.yaml

+2 -2

@@ -1,5 +1,5 @@

1 group: Quantization

2 -depends_on:

2 +depends_on:

3 - image-build

4 steps:

5 - label: Quantization

@@ -16,7 +16,7 @@ steps:

16 # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now

17 # we can only upgrade after this is resolved

18 # TODO(jerryzh168): resolve the above comment

19 - - uv pip install --system torchao==0.14.1 --index-url https://download.pytorch.org/whl/cu129

19 + - uv pip install --system torchao==0.17.0 --index-url https://download.pytorch.org/whl/cu130

20 - uv pip install --system conch-triton-kernels

21 - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py

22

.pre-commit-config.yaml

+1 -1

@@ -39,7 +39,7 @@ repos:

39 rev: 0.11.1

40 hooks:

41 - id: pip-compile

42 - args: [requirements/test.in, -c, requirements/common.txt, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu129, --python-platform, x86_64-manylinux_2_28, --python-version, "3.12"]

42 + args: [requirements/test.in, -c, requirements/common.txt, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu130, --python-platform, x86_64-manylinux_2_28, --python-version, "3.12"]

43 files: ^requirements/test\.(in|txt)$

44 - id: pip-compile

45 alias: pip-compile-rocm

CMakeLists.txt

+2 -2

@@ -56,8 +56,8 @@ endif()

56 # requirements.txt files and should be kept consistent. The ROCm torch

57 # versions are derived from docker/Dockerfile.rocm

58 #

59 -set(TORCH_SUPPORTED_VERSION_CUDA "2.10.0")

60 -set(TORCH_SUPPORTED_VERSION_ROCM "2.10.0")

59 +set(TORCH_SUPPORTED_VERSION_CUDA "2.11.0")

60 +set(TORCH_SUPPORTED_VERSION_ROCM "2.11.0")

61

62 #

63 # Try to find python package with an executable that exactly matches

csrc/cpu/utils.hpp

+2 -1

@@ -55,7 +55,8 @@ struct Counter {

55

56 inline int64_t get_available_l2_size() {

57 static int64_t size = []() {

58 - const uint32_t l2_cache_size = at::cpu::L2_cache_size();

58 + auto caps = at::cpu::get_cpu_capabilities();

59 + const uint32_t l2_cache_size = caps.at("l2_cache_size").toInt();

60 return l2_cache_size >> 1; // use 50% of L2 cache

61 }();

62 return size;

docker/Dockerfile

+12 -8

@@ -22,7 +22,7 @@

22 # docker buildx bake -f docker/docker-bake.hcl -f docker/versions.json

23 # =============================================================================

24

25 -ARG CUDA_VERSION=12.9.1

25 +ARG CUDA_VERSION=13.0.0

26 ARG PYTHON_VERSION=3.12

27 ARG UBUNTU_VERSION=22.04

28

@@ -37,7 +37,7 @@ ARG UBUNTU_VERSION=22.04

37 # compatibility with other Linux OSes. The main reason for this is that the

38 # glibc version is baked into the distro, and binaries built with one glibc

39 # version are not backwards compatible with OSes that use an earlier version.

40 -ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04

40 +ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04

41 # Using cuda base image with minimal dependencies necessary for JIT compilation (FlashInfer, DeepGEMM, EP kernels)

42 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-base-ubuntu${UBUNTU_VERSION}

43

@@ -546,17 +546,21 @@ RUN apt-get update -y \

546 # Install CUDA development tools for runtime JIT compilation

547 # (FlashInfer, DeepGEMM, EP kernels all require compilation at runtime)

548 RUN CUDA_VERSION_DASH=$(echo $CUDA_VERSION | cut -d. -f1,2 | tr '.' '-') && \

549 + CUDA_VERSION_SHORT=$(echo $CUDA_VERSION | cut -d. -f1,2) && \

550 apt-get update -y && \

550 - apt-get install -y --no-install-recommends \

551 + apt-get install -y --no-install-recommends --allow-change-held-packages \

552 cuda-nvcc-${CUDA_VERSION_DASH} \

553 cuda-cudart-${CUDA_VERSION_DASH} \

554 cuda-nvrtc-${CUDA_VERSION_DASH} \

555 cuda-cuobjdump-${CUDA_VERSION_DASH} \

556 libcurand-dev-${CUDA_VERSION_DASH} \

556 - libcublas-${CUDA_VERSION_DASH} \

557 - # Fixes nccl_allocator requiring nccl.h at runtime

558 - # https://github.com/vllm-project/vllm/blob/1336a1ea244fa8bfd7e72751cabbdb5b68a0c11a/vllm/distributed/device_communicators/pynccl_allocator.py#L22

559 - libnccl-dev && \

557 + libcublas-${CUDA_VERSION_DASH} && \

558 + # Fixes nccl_allocator requiring nccl.h at runtime

559 + # https://github.com/vllm-project/vllm/blob/1336a1ea244fa8bfd7e72751cabbdb5b68a0c11a/vllm/distributed/device_communicators/pynccl_allocator.py#L22

560 + # NCCL packages don't use the cuda-MAJOR-MINOR naming convention,

561 + # so we pin the version to match our CUDA version

562 + NCCL_VER=$(apt-cache madison libnccl-dev | grep "+cuda${CUDA_VERSION_SHORT}" | head -1 | awk -F'|' '{gsub(/^ +| +$/, "", $2); print $2}') && \

563 + apt-get install -y --no-install-recommends --allow-change-held-packages libnccl-dev=${NCCL_VER} libnccl2=${NCCL_VER} && \

564 rm -rf /var/lib/apt/lists/*

565

566 # Install uv for faster pip installs

@@ -822,7 +826,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \

826 uv pip install --system -r /tmp/kv_connectors.txt --no-build || ( \

827 # if the above fails, install from source

828 apt-get update -y && \

825 - apt-get install -y --no-install-recommends ${BUILD_PKGS} && \

829 + apt-get install -y --no-install-recommends --allow-change-held-packages ${BUILD_PKGS} && \

830 uv pip install --system -r /tmp/kv_connectors.txt --no-build-isolation && \

831 apt-get purge -y ${BUILD_PKGS} && \

832 # clean up -dev packages, keep runtime libraries

docker/Dockerfile.cpu

+1 -1

@@ -140,7 +140,7 @@ RUN \

140 esac; \

141 }; \

142 remove_packages_not_supported_on_aarch64 && \

143 - sed -i 's/^torch==.*/torch==2.10.0/g' requirements/cpu-test.in && \

143 + sed -i 's/^torch==.*/torch==2.11.0/g' requirements/cpu-test.in && \

144 sed -i 's/torchaudio.*/torchaudio/g' requirements/cpu-test.in && \

145 sed -i 's/torchvision.*/torchvision/g' requirements/cpu-test.in && \

146 uv pip compile requirements/cpu-test.in -o requirements/cpu-test.txt --index-strategy unsafe-best-match --torch-backend cpu

docker/Dockerfile.s390x

+3 -3

@@ -93,13 +93,13 @@ RUN curl https://sh.rustup.rs -sSf | sh -s -- -y && \

93

94 FROM python-install AS torch-vision

95 # Install torchvision

96 -ARG TORCH_VISION_VERSION=v0.25.0

96 +ARG TORCH_VISION_VERSION=v0.26.0

97 WORKDIR /tmp

98 RUN --mount=type=cache,target=/root/.cache/uv \

99 git clone https://github.com/pytorch/vision.git && \

100 cd vision && \

101 git checkout $TORCH_VISION_VERSION && \

102 - uv pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cpu && \

102 + uv pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cpu && \

103 python setup.py bdist_wheel

104

105 FROM python-install AS hf-xet-builder

@@ -253,7 +253,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \

253 NUMBA_WHL_FILE=$(ls /tmp/numba-wheels/*.whl) && \

254 OPENCV_WHL_FILE=$(ls /tmp/opencv-wheels/*.whl) && \

255 OUTLINES_CORE_WHL_FILE=$(ls /tmp/outlines-core/dist/*.whl) && \

256 - uv pip install -v \

256 + uv pip install -v \

257 $ARROW_WHL_FILE \

258 $VISION_WHL_FILE \

259 $HF_XET_WHL_FILE \

docker/versions.json

+3 -3

@@ -2,7 +2,7 @@

2 "_comment": "Auto-generated from Dockerfile ARGs. Do not edit manually. Run: python tools/generate_versions_json.py",

3 "variable": {

4 "CUDA_VERSION": {

5 - "default": "12.9.1"

5 + "default": "13.0.0"

6 },

7 "PYTHON_VERSION": {

8 "default": "3.12"

@@ -11,10 +11,10 @@

11 "default": "22.04"

12 },

13 "BUILD_BASE_IMAGE": {

14 - "default": "nvidia/cuda:12.9.1-devel-ubuntu20.04"

14 + "default": "nvidia/cuda:13.0.0-devel-ubuntu22.04"

15 },

16 "FINAL_BASE_IMAGE": {

17 - "default": "nvidia/cuda:12.9.1-base-ubuntu22.04"

17 + "default": "nvidia/cuda:13.0.0-base-ubuntu22.04"

18 },

19 "GET_PIP_URL": {

20 "default": "https://bootstrap.pypa.io/get-pip.py"

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

Binary files a/docs/assets/contributing/dockerfile-stages-dependency.png and b/docs/assets/contributing/dockerfile-stages-dependency.png differ

pyproject.toml

+1 -1

@@ -6,7 +6,7 @@ requires = [

6 "packaging>=24.2",

7 "setuptools>=77.0.3,<81.0.0",

8 "setuptools-scm>=8.0",

9 - "torch == 2.10.0",

9 + "torch == 2.11.0",

10 "wheel",

11 "jinja2",

12 ]

requirements/build.txt

+1 -1

@@ -4,7 +4,7 @@ ninja

4 packaging>=24.2

5 setuptools>=77.0.3,<81.0.0

6 setuptools-scm>=8

7 -torch==2.10.0

7 +torch==2.11.0

8 wheel

9 jinja2>=3.1.6

10 regex

requirements/cpu-build.txt

+3 -2

@@ -1,10 +1,11 @@

1 +--extra-index-url https://download.pytorch.org/whl/cpu

2 cmake>=3.26.1

3 ninja

4 packaging>=24.2

5 setuptools==77.0.3 # this version can reuse CMake build dir

6 setuptools-scm>=8

6 -torch==2.10.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x"

7 -torch==2.10.0; platform_machine == "aarch64" or platform_system == "Darwin" or platform_machine == "ppc64le"

7 +torch==2.11.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64"

8 +torch==2.11.0; platform_system == "Darwin" or platform_machine == "ppc64le" or platform_machine == "riscv64"

9 wheel

10 jinja2>=3.1.6

11 regex

requirements/cpu.txt

+3 -2

@@ -1,3 +1,4 @@

1 +--extra-index-url https://download.pytorch.org/whl/cpu

2 # Common dependencies

3 -r common.txt

4

@@ -6,8 +7,8 @@ setuptools==77.0.3 # this version can reuse CMake build dir

7 numba == 0.61.2; platform_machine != "s390x" # Required for N-gram speculative decoding

8

9 # Dependencies for CPUs

9 -torch==2.10.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x"

10 -torch==2.10.0; platform_machine == "aarch64" or platform_system == "Darwin" or platform_machine == "ppc64le" or platform_machine == "riscv64"

10 +torch==2.11.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64"

11 +torch==2.11.0; platform_system == "Darwin" or platform_machine == "ppc64le" or platform_machine == "riscv64"

12

13 # required for the image processor of minicpm-o-2_6, this must be updated alongside torch

14 torchaudio; platform_machine != "s390x" and platform_machine != "riscv64"

requirements/cuda.txt

+3 -3

@@ -4,10 +4,10 @@

4 numba == 0.61.2 # Required for N-gram speculative decoding

5

6 # Dependencies for NVIDIA GPUs

7 -torch==2.10.0

8 -torchaudio==2.10.0

7 +torch==2.11.0

8 +torchaudio==2.11.0

9 # These must be updated alongside torch

10 -torchvision==0.25.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version

10 +torchvision==0.26.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version

11 # FlashInfer should be updated together with the Dockerfile

12 flashinfer-python==0.6.7

13 flashinfer-cubin==0.6.7

requirements/kv_connectors.txt

+1 -1

@@ -1,3 +1,3 @@

1 lmcache >= 0.3.9

2 -nixl >= 0.7.1, < 0.10.0 # Required for disaggregated prefill

2 +nixl[cu13] >= 0.7.1, < 0.10.0 # Required for disaggregated prefill

3 mooncake-transfer-engine >= 0.3.8

requirements/rocm-build.txt

+4 -3

@@ -1,10 +1,11 @@

1 # Common dependencies

2 -r common.txt

3

4 +

5 --extra-index-url https://download.pytorch.org/whl/rocm7.1

5 -torch==2.10.0

6 -torchvision==0.25.0

7 -torchaudio==2.10.0

6 +torch==2.11.0

7 +torchvision==0.26.0

8 +torchaudio==2.11.0

9 triton==3.6.0

10 cmake>=3.26.1,<4

11 packaging>=24.2

requirements/test.in

+3 -3

@@ -27,9 +27,9 @@ soundfile # required for audio tests

27 jiwer # required for audio tests

28 tblib # for pickling test exceptions

29 timm >=1.0.17 # required for internvl and gemma3n-mm test

30 -torch==2.10.0

31 -torchaudio==2.10.0

32 -torchvision==0.25.0

30 +torch==2.11.0

31 +torchaudio==2.11.0

32 +torchvision==0.26.0

33 transformers_stream_generator # required for qwen-vl test

34 matplotlib # required for qwen-vl test

35 mistral_common[image,audio] >= 1.11.0 # required for voxtral test

requirements/test.txt

+41 -40

@@ -1,5 +1,5 @@

1 # This file was autogenerated by uv via the following command:

2 -# uv pip compile requirements/test.in -c requirements/common.txt -o requirements/test.txt --index-strategy unsafe-best-match --torch-backend cu129 --python-platform x86_64-manylinux_2_28 --python-version 3.12

2 +# uv pip compile requirements/test.in -c requirements/common.txt -o requirements/test.txt --index-strategy unsafe-best-match --torch-backend cu130 --python-platform x86_64-manylinux_2_28 --python-version 3.12

3 absl-py==2.1.0

4 # via

5 # rouge-score

@@ -165,10 +165,12 @@ cryptography==46.0.5

165 # azure-storage-blob

166 # msal

167 # pyjwt

168 -cuda-bindings==12.9.4

168 +cuda-bindings==13.0.3

169 # via torch

170 cuda-pathfinder==1.3.3

171 # via cuda-bindings

172 +cuda-toolkit==13.0.2

173 + # via torch

174 cupy-cuda12x==13.6.0

175 # via ray

176 cycler==0.12.1

@@ -615,45 +617,45 @@ numpy==2.2.6

617 # tritonclient

618 # vocos

619 # xarray

618 -nvidia-cublas-cu12==12.9.1.4

619 - # via

620 - # nvidia-cudnn-cu12

621 - # nvidia-cusolver-cu12

622 - # torch

623 -nvidia-cuda-cupti-cu12==12.9.79

624 - # via torch

625 -nvidia-cuda-nvrtc-cu12==12.9.86

626 - # via torch

627 -nvidia-cuda-runtime-cu12==12.9.79

628 - # via torch

629 -nvidia-cudnn-cu12==9.10.2.21

630 - # via torch

631 -nvidia-cufft-cu12==11.4.1.4

620 +nvidia-cublas==13.1.0.3

621 + # via

622 + # cuda-toolkit

623 + # nvidia-cudnn-cu13

624 + # nvidia-cusolver

625 +nvidia-cuda-cupti==13.0.85

626 + # via cuda-toolkit

627 +nvidia-cuda-nvrtc==13.0.88

628 + # via cuda-toolkit

629 +nvidia-cuda-runtime==13.0.96

630 + # via cuda-toolkit

631 +nvidia-cudnn-cu13==9.19.0.56

632 # via torch

633 -nvidia-cufile-cu12==1.14.1.1

633 +nvidia-cufft==12.0.0.61

634 + # via cuda-toolkit

635 +nvidia-cufile==1.15.1.6

636 + # via cuda-toolkit

637 +nvidia-curand==10.4.0.35

638 + # via cuda-toolkit

639 +nvidia-cusolver==12.0.4.66

640 + # via cuda-toolkit

641 +nvidia-cusparse==12.6.3.3

642 + # via

643 + # cuda-toolkit

644 + # nvidia-cusolver

645 +nvidia-cusparselt-cu13==0.8.0

646 # via torch

635 -nvidia-curand-cu12==10.3.10.19

647 +nvidia-nccl-cu13==2.28.9

648 # via torch

637 -nvidia-cusolver-cu12==11.7.5.82

638 - # via torch

639 -nvidia-cusparse-cu12==12.5.10.65

649 +nvidia-nvjitlink==13.0.88

650 # via

641 - # nvidia-cusolver-cu12

642 - # torch

643 -nvidia-cusparselt-cu12==0.7.1

644 - # via torch

645 -nvidia-nccl-cu12==2.27.5

646 - # via torch

647 -nvidia-nvjitlink-cu12==12.9.86

648 - # via

649 - # nvidia-cufft-cu12

650 - # nvidia-cusolver-cu12

651 - # nvidia-cusparse-cu12

652 - # torch

653 -nvidia-nvshmem-cu12==3.4.5

654 - # via torch

655 -nvidia-nvtx-cu12==12.9.79

651 + # cuda-toolkit

652 + # nvidia-cufft

653 + # nvidia-cusolver

654 + # nvidia-cusparse

655 +nvidia-nvshmem-cu13==3.4.5

656 # via torch

657 +nvidia-nvtx==13.0.85

658 + # via cuda-toolkit

659 omegaconf==2.3.0

660 # via

661 # hydra-core

@@ -1220,7 +1222,7 @@ tomli==2.2.1

1222 # via schemathesis

1223 tomli-w==1.2.0

1224 # via schemathesis

1223 -torch==2.10.0+cu129

1225 +torch==2.11.0+cu130

1226 # via

1227 # -r requirements/test.in

1228 # accelerate

@@ -1240,13 +1242,12 @@ torch==2.10.0+cu129

1242 # tensorizer

1243 # terratorch

1244 # timm

1243 - # torchaudio

1245 # torchgeo

1246 # torchmetrics

1247 # torchvision

1248 # vector-quantize-pytorch

1249 # vocos

1249 -torchaudio==2.10.0+cu129

1250 +torchaudio==2.11.0+cu130

1251 # via

1252 # -r requirements/test.in

1253 # encodec

@@ -1259,7 +1260,7 @@ torchmetrics==1.7.4

1260 # pytorch-lightning

1261 # terratorch

1262 # torchgeo

1262 -torchvision==0.25.0+cu129

1263 +torchvision==0.26.0+cu130

1264 # via

1265 # -r requirements/test.in

1266 # lightly

requirements/xpu.txt

+1 -1

@@ -11,7 +11,7 @@ jinja2>=3.1.6

11 datasets # for benchmark scripts

12 numba == 0.61.2 # Required for N-gram speculative decoding

13 --extra-index-url=https://download.pytorch.org/whl/xpu

14 -torch==2.10.0+xpu

14 +torch==2.11.0+xpu

15 torchaudio

16 torchvision

17

tests/kernels/helion/test_pattern_matching.py

+2

@@ -67,6 +67,7 @@ class TestMakeFxHop:

67 def setup_method(self):

68 helion_kernel_side_table.reset_table()

69

70 + @pytest.mark.skip(reason="SymInt proxy tracking issue with PyTorch 2.11+")

71 def test_make_fx_symbolic(self):

72 def raw_add_scale(

73 x: torch.Tensor, y: torch.Tensor, scale: float

@@ -128,6 +129,7 @@ class TestMakeFxHop:

129 for out_s, in_s in zip(val.shape, input_shape):

130 assert out_s == in_s

131

132 + @pytest.mark.skip(reason="SymInt proxy tracking issue with PyTorch 2.11+")

133 def test_pattern_matcher_replaces_with_helion_hop(self):

134 def raw_silu_mul(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:

135 M, N = x.size()

tests/standalone_tests/python_only_compile.sh

-1

@@ -68,7 +68,6 @@ apt autoremove -y

68 echo 'import os; os.system("touch /tmp/changed.file")' >> vllm/__init__.py

69

70 VLLM_PRECOMPILED_WHEEL_COMMIT=$merge_base_commit VLLM_USE_PRECOMPILED=1 pip3 install -vvv -e .

71 -

71 # Run the script

72 python3 -c 'import vllm'

73

vllm/distributed/parallel_state.py

+1

@@ -1445,6 +1445,7 @@ def init_distributed_environment(

1445 # local rank not set, this usually happens in single-node

1446 # setting, where we can use rank as local rank

1447 local_rank = envs.LOCAL_RANK if distributed_init_method == "env://" else rank

1448 +

1449 global _WORLD, _NODE_COUNT, _INNER_DP_WORLD

1450 if enable_elastic_ep:

1451 _init_elastic_ep_world(config, local_rank, backend, rank, world_size)

vllm/model_executor/layers/fused_moe/runner/moe_runner_base.py

+8 -1

@@ -65,8 +65,15 @@ else:

65 _layer_name_type = ModuleName if HAS_OPAQUE_TYPE else str

66

67

68 +@torch.compiler.assume_constant_result

69 def _resolve_layer_name(layer_name: str | ModuleName) -> str:

69 - return layer_name.value if isinstance(layer_name, ModuleName) else layer_name

70 + from torch._library.fake_class_registry import FakeScriptObject

71 +

72 + if isinstance(layer_name, ModuleName):

73 + return layer_name.value

74 + elif isinstance(layer_name, FakeScriptObject):

75 + return layer_name.real_obj.value

76 + return layer_name

77

78

79 # Note: _moe_forward and _moe_forward_shared should not contain any

vllm/utils/torch_utils.py

+1 -1

@@ -706,7 +706,7 @@ def is_torch_equal(target: str) -> bool:

706 return Version(importlib.metadata.version("torch")) == Version(target)

707

708

709 -HAS_OPAQUE_TYPE = is_torch_equal_or_newer("2.11.0.dev")

709 +HAS_OPAQUE_TYPE = is_torch_equal_or_newer("2.12.0.dev")

710

711 if HAS_OPAQUE_TYPE:

712 from torch._opaque_base import OpaqueBase

[release 2.11] Update to torch 2.11 (#34644)