A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 11, 07:30 UTC
2.9.1 PyTorch release update (#28495)
Andrey Talman committed
Dec 17, 2025 at 15:20 UTC
e06d0bf0aa2af11220b5c3aa5ccc8f999d0e3161
10 files changed
+21
-21
.buildkite/test-amd.yaml
+1
-1
@@ -740,7 +740,7 @@ steps:
740
# https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now
741
# we can only upgrade after this is resolved
742
# TODO(jerryzh168): resolve the above comment
743
- - uv pip install --system torchao==0.13.0
743
+ - uv pip install --system torchao==0.14.1
744
- uv pip install --system conch-triton-kernels
745
- VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
746
.buildkite/test-pipeline.yaml
+1
-1
@@ -658,7 +658,7 @@ steps:
658
# https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now
659
# we can only upgrade after this is resolved
660
# TODO(jerryzh168): resolve the above comment
661
- - uv pip install --system torchao==0.13.0 --index-url https://download.pytorch.org/whl/cu129
661
+ - uv pip install --system torchao==0.14.1 --index-url https://download.pytorch.org/whl/cu129
662
- uv pip install --system conch-triton-kernels
663
- VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
664
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 "2.9.0")
60
-set(TORCH_SUPPORTED_VERSION_ROCM "2.9.0")
59
+set(TORCH_SUPPORTED_VERSION_CUDA "2.9.1")
60
+set(TORCH_SUPPORTED_VERSION_ROCM "2.9.1")
61
62
#
63
# Try to find python package with an executable that exactly matches
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 == 2.9.0",
9
+ "torch == 2.9.1",
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==2.9.0
7
+torch==2.9.1
8
wheel
9
jinja2>=3.1.6
10
regex
requirements/cuda.txt
+3
-3
@@ -5,9 +5,9 @@ numba == 0.61.2 # Required for N-gram speculative decoding
5
6
# Dependencies for NVIDIA GPUs
7
ray[cgraph]>=2.48.0 # Ray Compiled Graph, required for pipeline parallelism in V1.
8
-torch==2.9.0
9
-torchaudio==2.9.0
8
+torch==2.9.1
9
+torchaudio==2.9.1
10
# These must be updated alongside torch
11
-torchvision==0.24.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
11
+torchvision==0.24.1 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
12
# FlashInfer should be updated together with the Dockerfile
13
flashinfer-python==0.5.3
requirements/rocm-build.txt
+4
-4
@@ -2,11 +2,11 @@
2
-r common.txt
3
4
--extra-index-url https://download.pytorch.org/whl/rocm6.4
5
-torch==2.9.0
6
-torchvision==0.24.0
7
-torchaudio==2.9.0
5
+torch==2.9.1
6
+torchvision==0.24.1
7
+torchaudio==2.9.1
8
9
-triton==3.5.0
9
+triton==3.5.1
10
cmake>=3.26.1,<4
11
packaging>=24.2
12
setuptools>=77.0.3,<80.0.0
requirements/test.in
+3
-3
@@ -24,9 +24,9 @@ soundfile # required for audio tests
24
jiwer # required for audio tests
25
tblib # for pickling test exceptions
26
timm >=1.0.17 # required for internvl and gemma3n-mm test
27
-torch==2.9.0
28
-torchaudio==2.9.0
29
-torchvision==0.24.0
27
+torch==2.9.1
28
+torchaudio==2.9.1
29
+torchvision==0.24.1
30
transformers_stream_generator # required for qwen-vl test
31
matplotlib # required for qwen-vl test
32
mistral_common[image,audio] >= 1.8.5 # required for voxtral test
requirements/test.txt
+4
-4
@@ -1123,7 +1123,7 @@ tomli==2.2.1
1123
# via schemathesis
1124
tomli-w==1.2.0
1125
# via schemathesis
1126
-torch==2.9.0+cu129
1126
+torch==2.9.1+cu129
1127
# via
1128
# -r requirements/test.in
1129
# accelerate
@@ -1152,7 +1152,7 @@ torch==2.9.0+cu129
1152
# torchvision
1153
# vector-quantize-pytorch
1154
# vocos
1155
-torchaudio==2.9.0+cu129
1155
+torchaudio==2.9.1+cu129
1156
# via
1157
# -r requirements/test.in
1158
# encodec
@@ -1165,7 +1165,7 @@ torchmetrics==1.7.4
1165
# pytorch-lightning
1166
# terratorch
1167
# torchgeo
1168
-torchvision==0.24.0+cu129
1168
+torchvision==0.24.1+cu129
1169
# via
1170
# -r requirements/test.in
1171
# lightly
@@ -1206,7 +1206,7 @@ transformers==4.57.3
1206
# transformers-stream-generator
1207
transformers-stream-generator==0.0.5
1208
# via -r requirements/test.in
1209
-triton==3.5.0
1209
+triton==3.5.1
1210
# via torch
1211
tritonclient==2.51.0
1212
# via
vllm/model_executor/layers/conv.py
+1
-1
@@ -251,6 +251,6 @@ class Conv3dLayer(ConvLayerBase):
251
# See: https://github.com/vllm-project/vllm/issues/27406
252
# and https://github.com/pytorch/pytorch/issues/166122
253
# By default, we use CUDNN's convolution ops with optimization.
254
- if self.enable_linear and is_torch_equal("2.9.0"):
254
+ if self.enable_linear and (is_torch_equal("2.9.0") or is_torch_equal("2.9.1")):
255
return self._forward_mulmat(x)
256
return self._forward_conv(x)