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 12, 09:45 UTC
Revert "[CPU]Update CPU PyTorch to 2.9.0 (#29589)" (#29647)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Nov 28, 2025 at 14:43 UTC
b34e8775a31c1a077a1a24f22ffbf048b2a979f6
4 files changed
+12
-8
docker/Dockerfile.cpu
+4
@@ -119,6 +119,7 @@ FROM base AS vllm-test-deps
119
120
WORKDIR /workspace/vllm
121
122
+# TODO: Update to 2.9.0 when there is a new build for intel_extension_for_pytorch for that version
123
RUN --mount=type=bind,src=requirements/test.in,target=requirements/test.in \
124
cp requirements/test.in requirements/cpu-test.in && \
125
sed -i '/mamba_ssm/d' requirements/cpu-test.in && \
@@ -131,6 +132,9 @@ RUN --mount=type=bind,src=requirements/test.in,target=requirements/test.in \
132
esac; \
133
}; \
134
remove_packages_not_supported_on_aarch64 && \
135
+ sed -i 's/^torch==.*/torch==2.8.0/g' requirements/cpu-test.in && \
136
+ sed -i 's/torchaudio.*/torchaudio/g' requirements/cpu-test.in && \
137
+ sed -i 's/torchvision.*/torchvision/g' requirements/cpu-test.in && \
138
uv pip compile requirements/cpu-test.in -o requirements/cpu-test.txt --index-strategy unsafe-best-match --torch-backend cpu
139
140
RUN --mount=type=cache,target=/root/.cache/uv \
requirements/cpu-build.txt
+2
-2
@@ -4,9 +4,9 @@ packaging>=24.2
4
setuptools>=77.0.3,<81.0.0
5
setuptools-scm>=8
6
--extra-index-url https://download.pytorch.org/whl/cpu
7
-torch==2.9.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x"
7
+torch==2.8.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x"
8
torch==2.9.0; platform_system == "Darwin"
9
-torch==2.9.0; platform_machine == "ppc64le" or platform_machine == "aarch64"
9
+torch==2.8.0; platform_machine == "ppc64le" or platform_machine == "aarch64"
10
scons; platform_machine == "aarch64" # needed to build Arm Compute Library (ACL)
11
wheel
12
jinja2>=3.1.6
requirements/cpu.txt
+4
-4
@@ -7,17 +7,17 @@ numba == 0.61.2; platform_machine != "s390x" # Required for N-gram speculative d
7
packaging>=24.2
8
setuptools>=77.0.3,<81.0.0
9
--extra-index-url https://download.pytorch.org/whl/cpu
10
-torch==2.9.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x"
10
+torch==2.8.0+cpu; platform_machine == "x86_64" or platform_machine == "s390x"
11
torch==2.9.0; platform_system == "Darwin"
12
-torch==2.9.0; platform_machine == "ppc64le" or platform_machine == "aarch64"
12
+torch==2.8.0; platform_machine == "ppc64le" or platform_machine == "aarch64"
13
14
# required for the image processor of minicpm-o-2_6, this must be updated alongside torch
15
torchaudio; platform_machine != "ppc64le" and platform_machine != "s390x"
16
-torchaudio==2.9.0; platform_machine == "ppc64le"
16
+torchaudio==2.8.0; platform_machine == "ppc64le"
17
18
# required for the image processor of phi3v, this must be updated alongside torch
19
torchvision; platform_machine != "ppc64le" and platform_machine != "s390x"
20
-torchvision==0.24.0; platform_machine == "ppc64le"
20
+torchvision==0.23.0; platform_machine == "ppc64le"
21
datasets # for benchmark scripts
22
23
# Intel Extension for PyTorch, only for x86_64 CPUs
vllm/model_executor/models/qwen3_vl.py
+2
-2
@@ -1123,14 +1123,14 @@ class Qwen3LLMModel(Qwen3Model):
1123
assert intermediate_tensors is not None
1124
hidden_states = intermediate_tensors["hidden_states"]
1125
residual = intermediate_tensors["residual"]
1126
-
1126
+
1127
aux_hidden_states = []
1128
for layer_idx, layer in islice(
1129
enumerate(self.layers), self.start_layer, self.end_layer
1130
):
1131
if layer_idx in self.aux_hidden_state_layers:
1132
aux_hidden_states.append(hidden_states + residual)
1133
-
1133
+
1134
hidden_states, residual = layer(
1135
positions,
1136
hidden_states,