From 30e79c3a766c628fa005146d713ae72530be2b76 Mon Sep 17 00:00:00 2001 From: Andrey Talman Date: Wed, 14 Oct 2026 03:17:59 -0400 Subject: [PATCH] [Core] Update PyTorch to 2.14.0, torchvision to 0.29.0, triton to 3.8.0 (release channel) (#52183) Signed-off-by: Andrey Talman Signed-off-by: Andrey Talman Co-authored-by: Andrey Talman Co-authored-by: Claude Opus 5.5 --- .buildkite/test-amd.yaml | 2 +- .buildkite/test_areas/quantization.yaml | 2 +- CMakeLists.txt | 4 ++-- docker/Dockerfile.s390x | 6 +++--- docker/Dockerfile.zen | 2 +- pyproject.toml | 2 +- requirements/build/cpu.txt | 4 ++-- requirements/build/cuda.txt | 2 +- requirements/cpu.txt | 4 ++-- requirements/cuda.txt | 4 ++-- requirements/test/cpu.txt | 6 +++--- requirements/test/cuda.in | 4 ++-- requirements/test/cuda.txt | 10 +++++----- setup.py | 2 +- tests/models/language/generation/test_hybrid.py | 3 +++ tests/v1/e2e/general/test_kv_sharing_fast_prefill.py | 6 ++++++ 16 files changed, 36 insertions(+), 27 deletions(-) diff --git a/.buildkite/test-amd.yaml b/.buildkite/test-amd.yaml index 7f0a351524..22856d837f 100644 --- a/.buildkite/test-amd.yaml +++ b/.buildkite/test-amd.yaml @@ -5271,7 +5271,7 @@ steps: - tests/rocm/test_moe_weight_replay.py optional: true commands: - - uv pip install --system torchao==0.17.0 + - uv pip install --system torchao==0.18.0 - uv pip install --system conch-triton-kernels - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py --ignore quantization/test_rocm_moe.py --ignore quantization/test_gfx950_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT - if [ "$$BUILDKITE_PARALLEL_JOB" = "0" ]; then pytest -v -s rocm/test_moe_weight_replay.py; fi diff --git a/.buildkite/test_areas/quantization.yaml b/.buildkite/test_areas/quantization.yaml index 4b4f7858c5..cdf8e7bbe7 100644 --- a/.buildkite/test_areas/quantization.yaml +++ b/.buildkite/test_areas/quantization.yaml @@ -14,7 +14,7 @@ steps: - tests/quantization commands: # Pin torchao to the version compatible with the CI PyTorch and CUDA stack. - - uv pip install --system torchao==0.17.0 --index-url https://download.pytorch.org/whl/cu130 + - uv pip install --system torchao==0.18.0 --index-url https://download.pytorch.org/whl/cu130 - uv pip install --system conch-triton-kernels # The SM90-only checkpoint currently contains a removed weight_chan_scale # parameter. It was not exercised by the previous L4 job. diff --git a/CMakeLists.txt b/CMakeLists.txt index e27a031f26..d59f56c00a 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -68,8 +68,8 @@ endif() # requirements.txt files and should be kept consistent. The ROCm torch # versions are derived from docker/Dockerfile.rocm # -set(TORCH_SUPPORTED_VERSION_CUDA "2.13.0") -set(TORCH_SUPPORTED_VERSION_ROCM "2.13.0") +set(TORCH_SUPPORTED_VERSION_CUDA "2.14.0") +set(TORCH_SUPPORTED_VERSION_ROCM "2.14.0") # TORCH_NIGHTLY=1 builds run against unpinned nightly wheels, so the supported- # version check would always warn. Only treat it as a nightly build when the # value is exactly "1" (the bootstrap exports TORCH_NIGHTLY=0 by default, which diff --git a/docker/Dockerfile.s390x b/docker/Dockerfile.s390x index e04ff55ce6..024f3582a9 100644 --- a/docker/Dockerfile.s390x +++ b/docker/Dockerfile.s390x @@ -70,13 +70,13 @@ RUN curl -LO https://github.com/gperftools/gperftools/releases/download/gperftoo FROM python-install AS torch-vision # Install torchvision -ARG TORCH_VISION_VERSION=v0.28.0 +ARG TORCH_VISION_VERSION=v0.29.0 WORKDIR /tmp RUN --mount=type=cache,target=/root/.cache/uv \ git clone https://github.com/pytorch/vision.git && \ cd vision && \ git checkout $TORCH_VISION_VERSION && \ - uv pip install torch==2.13.0 --index-url https://download.pytorch.org/whl/cpu && \ + uv pip install torch==2.14.0 --index-url https://download.pytorch.org/whl/cpu && \ python setup.py bdist_wheel FROM python-install AS hf-xet-builder @@ -167,7 +167,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \ git clone --depth 1 --branch ${TORCHCODEC_VERSION} \ https://github.com/pytorch/torchcodec.git && \ cd torchcodec && \ - uv pip install torch==2.13.0 --index-url https://download.pytorch.org/whl/cpu && \ + uv pip install torch==2.14.0 --index-url https://download.pytorch.org/whl/cpu && \ uv pip install pybind11 setuptools wheel && \ pybind11_DIR="$(python -c 'import pybind11; print(pybind11.get_cmake_dir())')" && \ export pybind11_DIR CMAKE_PREFIX_PATH="${pybind11_DIR}:${CMAKE_PREFIX_PATH}" && \ diff --git a/docker/Dockerfile.zen b/docker/Dockerfile.zen index 8c9e49c5f8..d0ede7149a 100644 --- a/docker/Dockerfile.zen +++ b/docker/Dockerfile.zen @@ -62,7 +62,7 @@ COPY requirements/common.txt requirements/common.txt COPY requirements/test/cuda.in requirements/test/zen.in RUN sed -i '/mamba_ssm/d' requirements/test/zen.in && \ - sed -i 's/^torch==.*/torch==2.13.0/g' requirements/test/zen.in && \ + sed -i 's/^torch==.*/torch==2.14.0/g' requirements/test/zen.in && \ sed -i 's/torchaudio.*/torchaudio/g' requirements/test/zen.in && \ sed -i 's/torchvision.*/torchvision/g' requirements/test/zen.in && \ # zentorch parity: keep sentence-transformers pinned where vllm-test-deps did. diff --git a/pyproject.toml b/pyproject.toml index db5a2a3b9c..409207700e 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -7,7 +7,7 @@ requires = [ "setuptools>=77.0.3,<81.0.0", "setuptools-scm>=8.0", "setuptools-rust>=1.9.0", - "torch == 2.13.0", + "torch == 2.14.0", "wheel", "jinja2", ] diff --git a/requirements/build/cpu.txt b/requirements/build/cpu.txt index 7f2f4c7914..e9e03425f8 100644 --- a/requirements/build/cpu.txt +++ b/requirements/build/cpu.txt @@ -4,8 +4,8 @@ packaging>=24.2 setuptools==78.1.1 # this version can reuse CMake build dir; >=78.1.1 for GHSA-5rjg-fvgr-3xxf setuptools-scm>=8 setuptools-rust>=1.9.0 -torch==2.13.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64" -torch==2.13.0; platform_system == "Darwin" or platform_machine == "ppc64le" or platform_machine == "riscv64" +torch==2.14.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64" +torch==2.14.0; platform_system == "Darwin" or platform_machine == "ppc64le" or platform_machine == "riscv64" wheel jinja2>=3.1.6 regex diff --git a/requirements/build/cuda.txt b/requirements/build/cuda.txt index 3cc83287bc..6e6fa61892 100644 --- a/requirements/build/cuda.txt +++ b/requirements/build/cuda.txt @@ -5,7 +5,7 @@ packaging>=24.2 setuptools>=78.1.1,<81.0.0 setuptools-scm>=8 setuptools-rust>=1.9.0 -torch==2.13.0 +torch==2.14.0 wheel jinja2>=3.1.6 regex diff --git a/requirements/cpu.txt b/requirements/cpu.txt index 0570517e74..2ae87a40ee 100644 --- a/requirements/cpu.txt +++ b/requirements/cpu.txt @@ -6,8 +6,8 @@ setuptools==78.1.1 # this version can reuse CMake build dir; >=78.1.1 for GHSA-5 numba == 0.65.0 # Required for N-gram speculative decoding # Dependencies for CPUs -torch==2.13.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64" -torch==2.13.0; platform_system == "Darwin" or platform_machine == "ppc64le" or platform_machine == "riscv64" +torch==2.14.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x" or platform_machine == "aarch64" +torch==2.14.0; platform_system == "Darwin" or platform_machine == "ppc64le" or platform_machine == "riscv64" # required for the image processor of minicpm-o-2_6, this must be updated alongside torch torchaudio; platform_machine != "s390x" and platform_machine != "riscv64" diff --git a/requirements/cuda.txt b/requirements/cuda.txt index 5e193bac5a..48b59f560a 100644 --- a/requirements/cuda.txt +++ b/requirements/cuda.txt @@ -4,10 +4,10 @@ numba == 0.65.0 # Required for N-gram speculative decoding # Dependencies for NVIDIA GPUs -torch==2.13.0 +torch==2.14.0 torchaudio==2.11.0 # These must be updated alongside torch -torchvision==0.28.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version +torchvision==0.29.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version torchcodec >= 0.14 PyNvVideoCodec==2.0.4 flashinfer-python==0.7.0.post1 diff --git a/requirements/test/cpu.txt b/requirements/test/cpu.txt index 7d1f13970f..86bd822110 100644 --- a/requirements/test/cpu.txt +++ b/requirements/test/cpu.txt @@ -1170,7 +1170,7 @@ tokenizers==0.23.2 # -r requirements/test/../common.txt # cohere # transformers -torch==2.13.0+cpu +torch==2.14.0+cpu # via # -r requirements/test/cuda.in # accelerate @@ -1198,7 +1198,7 @@ torchaudio==2.11.0+cpu # vocos torchcodec==0.14.0+cpu # via -r requirements/test/cuda.in -torchvision==0.28.0+cpu +torchvision==0.29.0+cpu # via # -r requirements/test/cuda.in # open-clip-torch @@ -1240,7 +1240,7 @@ transformers==5.19.0 # xgrammar transformers-stream-generator==0.0.5 # via -r requirements/test/cuda.in -triton==3.6.0 +triton==3.8.0 # via xgrammar tritonclient==2.64.0 # via -r requirements/test/cuda.in diff --git a/requirements/test/cuda.in b/requirements/test/cuda.in index 4a32471652..53a9b70da5 100644 --- a/requirements/test/cuda.in +++ b/requirements/test/cuda.in @@ -28,9 +28,9 @@ 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==2.13.0 +torch==2.14.0 torchaudio==2.11.0 -torchvision==0.28.0 +torchvision==0.29.0 transformers_stream_generator # required for qwen-vl test matplotlib # required for qwen-vl test mistral_common[image,audio] >= 1.11.6 # required for voxtral test diff --git a/requirements/test/cuda.txt b/requirements/test/cuda.txt index 29553d78c7..e06e43a7f9 100644 --- a/requirements/test/cuda.txt +++ b/requirements/test/cuda.txt @@ -657,7 +657,7 @@ nvidia-cuda-nvrtc==13.0.88 # nvidia-cublas nvidia-cuda-runtime==13.0.96 # via cuda-toolkit -nvidia-cudnn-cu13==9.20.0.48 +nvidia-cudnn-cu13==9.24.0.43 # via torch nvidia-cufft==12.0.0.61 # via cuda-toolkit @@ -673,7 +673,7 @@ nvidia-cusparse==12.6.3.3 # nvidia-cusolver nvidia-cusparselt-cu13==0.8.1 # via torch -nvidia-nccl-cu13==2.29.7 +nvidia-nccl-cu13==2.30.7 # via torch nvidia-nvjitlink==13.0.88 # via @@ -1272,7 +1272,7 @@ tokenizers==0.23.2 # -r requirements/test/../common.txt # cohere # transformers -torch==2.13.0+cu130 +torch==2.14.0+cu130 # via # -c requirements/cuda.txt # -r requirements/test/cuda.in @@ -1304,7 +1304,7 @@ torchcodec==0.14.0+cu130 # via # -c requirements/cuda.txt # -r requirements/test/cuda.in -torchvision==0.28.0+cu130 +torchvision==0.29.0+cu130 # via # -c requirements/cuda.txt # -r requirements/test/cuda.in @@ -1348,7 +1348,7 @@ transformers==5.19.0 # xgrammar transformers-stream-generator==0.0.5 # via -r requirements/test/cuda.in -triton==3.7.1 +triton==3.8.0 # via # torch # xgrammar diff --git a/setup.py b/setup.py index 2a5b54f410..8958ab6dc7 100644 --- a/setup.py +++ b/setup.py @@ -1688,7 +1688,7 @@ setup( install_requires=get_requirements(), extras_require={ # AMD Zen CPU optimizations via zentorch - "zen": ["zentorch==2.13.0.1"], + "zen": ["zentorch==2.14.0.0"], "bench": ["pandas", "matplotlib", "seaborn", "datasets", "scipy", "plotly"], "tensorizer": ["tensorizer==2.10.1"], "fastsafetensors": ["fastsafetensors >= 0.3.3"], diff --git a/tests/models/language/generation/test_hybrid.py b/tests/models/language/generation/test_hybrid.py index 1b90429feb..865aa57f9f 100644 --- a/tests/models/language/generation/test_hybrid.py +++ b/tests/models/language/generation/test_hybrid.py @@ -342,10 +342,13 @@ def test_full_cuda_graph( example_prompts, max_tokens, num_logprobs ) + # Leave headroom for the HF model that ran earlier in this process on a + # 32.5 GiB MIG. with vllm_runner( model, max_num_seqs=MAX_NUM_SEQS, attention_backend=ATTN_BACKEND, + gpu_memory_utilization=0.9, enable_chunked_prefill=True, ) as vllm_model: vllm_outputs = vllm_model.generate_greedy_logprobs( diff --git a/tests/v1/e2e/general/test_kv_sharing_fast_prefill.py b/tests/v1/e2e/general/test_kv_sharing_fast_prefill.py index b03acc21af..00d0ff939e 100644 --- a/tests/v1/e2e/general/test_kv_sharing_fast_prefill.py +++ b/tests/v1/e2e/general/test_kv_sharing_fast_prefill.py @@ -74,6 +74,12 @@ def test_kv_sharing_fast_prefill( "ROCm: torch.compile produces incorrect output for gemma-3n's GELU " "with tanh approximation. Use enforce_eager=True instead." ) + if not enforce_eager and not use_v2_model_runner and current_platform.is_cuda(): + pytest.skip( + "cublasCreate fails with CUBLAS_STATUS_ALLOC_FAILED during memory " + "profiling with torch 2.14: " + "https://github.com/pytorch/pytorch/issues/199944" + ) sampling_params = SamplingParams(temperature=0.0, max_tokens=100) compilation_config = CompilationConfig(