2.9.1 PyTorch release update (#28495)

.buildkite/test-amd.yaml

@@ -740,7 +740,7 @@ 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==0.14.1
   - uv pip install --system conch-triton-kernels
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 

.buildkite/test-pipeline.yaml

@@ -658,7 +658,7 @@ 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 --index-url https://download.pytorch.org/whl/cu129
+  - uv pip install --system torchao==0.14.1 --index-url https://download.pytorch.org/whl/cu129
   - uv pip install --system conch-triton-kernels
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 

CMakeLists.txt

@@ -56,8 +56,8 @@ endif()
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.9.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.9.1")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.9.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.9.1")
 
 #
 # Try to find python package with an executable that exactly matches

pyproject.toml

@@ -6,7 +6,7 @@ requires = [
     "packaging>=24.2",
     "setuptools>=77.0.3,<81.0.0",
     "setuptools-scm>=8.0",
-    "torch == 2.9.0",
+    "torch == 2.9.1",
     "wheel",
     "jinja2",
 ]

requirements/build.txt

@@ -4,7 +4,7 @@ ninja
 packaging>=24.2
 setuptools>=77.0.3,<81.0.0
 setuptools-scm>=8
-torch==2.9.0
+torch==2.9.1
 wheel
 jinja2>=3.1.6
 regex

requirements/cuda.txt

@@ -5,9 +5,9 @@ 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==2.9.0
+torch==2.9.1
-torchaudio==2.9.0
+torchaudio==2.9.1
 # These must be updated alongside torch
-torchvision==0.24.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
+torchvision==0.24.1 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
 # FlashInfer should be updated together with the Dockerfile
 flashinfer-python==0.5.3

requirements/rocm-build.txt

@@ -2,11 +2,11 @@
 -r common.txt
 
 --extra-index-url https://download.pytorch.org/whl/rocm6.4
-torch==2.9.0
+torch==2.9.1
-torchvision==0.24.0
+torchvision==0.24.1
-torchaudio==2.9.0
+torchaudio==2.9.1
 
-triton==3.5.0
+triton==3.5.1
 cmake>=3.26.1,<4
 packaging>=24.2
 setuptools>=77.0.3,<80.0.0

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==2.9.0
+torch==2.9.1
-torchaudio==2.9.0
+torchaudio==2.9.1
-torchvision==0.24.0
+torchvision==0.24.1
 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

requirements/test.txt

@@ -1123,7 +1123,7 @@ tomli==2.2.1
     # via schemathesis
 tomli-w==1.2.0
     # via schemathesis
-torch==2.9.0+cu129
+torch==2.9.1+cu129
     # via
     #   -r requirements/test.in
     #   accelerate
@@ -1152,7 +1152,7 @@ torch==2.9.0+cu129
     #   torchvision
     #   vector-quantize-pytorch
     #   vocos
-torchaudio==2.9.0+cu129
+torchaudio==2.9.1+cu129
     # via
     #   -r requirements/test.in
     #   encodec
@@ -1165,7 +1165,7 @@ torchmetrics==1.7.4
     #   pytorch-lightning
     #   terratorch
     #   torchgeo
-torchvision==0.24.0+cu129
+torchvision==0.24.1+cu129
     # via
     #   -r requirements/test.in
     #   lightly
@@ -1206,7 +1206,7 @@ transformers==4.57.3
     #   transformers-stream-generator
 transformers-stream-generator==0.0.5
     # via -r requirements/test.in
-triton==3.5.0
+triton==3.5.1
     # via torch
 tritonclient==2.51.0
     # via

vllm/model_executor/layers/conv.py

@@ -251,6 +251,6 @@ class Conv3dLayer(ConvLayerBase):
         # See: https://github.com/vllm-project/vllm/issues/27406
         # and https://github.com/pytorch/pytorch/issues/166122
         # By default, we use CUDNN's convolution ops with optimization.
-        if self.enable_linear and is_torch_equal("2.9.0"):
+        if self.enable_linear and (is_torch_equal("2.9.0") or is_torch_equal("2.9.1")):
             return self._forward_mulmat(x)
         return self._forward_conv(x)