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 4, 17:30 UTC
[XPU] Upgrade torch 2.11 for xpu (#37947)
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Kunshang Ji committed
Apr 23, 2026 at 23:07 UTC
53ecc807c0e323aea2f2a48dfdae71be838c4f5c
4 files changed
+29
-29
docker/Dockerfile.xpu
+4
-4
@@ -50,9 +50,9 @@ RUN curl -LsSf https://astral.sh/uv/install.sh | sh
50
RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
51
ENV PATH="$VIRTUAL_ENV/bin:$PATH"
52
53
-# This oneccl contains the BMG support which is not the case for default version of oneapi 2025.2.
54
-ARG ONECCL_INSTALLER="intel-oneccl-2021.15.7.8_offline.sh"
55
-RUN wget "https://github.com/uxlfoundation/oneCCL/releases/download/2021.15.7/${ONECCL_INSTALLER}" && \
53
+# This oneccl contains the BMG support which is not the case for default version of oneapi 2025.3.
54
+ARG ONECCL_INSTALLER="intel-oneccl-2021.15.9.14_offline.sh"
55
+RUN wget "https://github.com/uxlfoundation/oneCCL/releases/download/2021.15.9/${ONECCL_INSTALLER}" && \
56
bash "${ONECCL_INSTALLER}" -a --silent --eula accept && \
57
rm "${ONECCL_INSTALLER}" && \
58
echo "source /opt/intel/oneapi/setvars.sh --force" >> /root/.bashrc && \
@@ -164,7 +164,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
164
# FIX triton
165
RUN --mount=type=cache,target=/root/.cache/uv \
166
uv pip uninstall triton triton-xpu && \
167
- uv pip install triton-xpu==3.6.0
167
+ uv pip install triton-xpu==3.7.0
168
169
# remove torch bundled oneccl to avoid conflicts
170
RUN --mount=type=cache,target=/root/.cache/uv \
docs/getting_started/installation/gpu.xpu.inc.md
+1
-1
@@ -46,7 +46,7 @@ pip install -v -r requirements/xpu.txt
46
47
!!! note
48
- `triton` (without suffix) is for NVIDIA GPUs only. On XPU, using it instead of `triton-xpu` can cause correctness or runtime issues.
49
- - For torch 2.10 (the version used in `requirements/xpu.txt`), the matching package is `triton-xpu==3.6.0`. If you use a different version of torch, check the corresponding `triton-xpu` version in [docker/Dockerfile.xpu](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.xpu).
49
+ - For torch 2.11 (the version used in `requirements/xpu.txt`), the matching package is `triton-xpu==3.7.0`. If you use a different version of torch, check the corresponding `triton-xpu` version in [docker/Dockerfile.xpu](https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.xpu).
50
51
- Finally, build and install vLLM XPU backend:
52
requirements/test/xpu.txt
+23
-23
@@ -93,7 +93,7 @@ docker==7.1.0
93
# via gpt-oss
94
docopt==0.6.2
95
# via num2words
96
-dpcpp-cpp-rt==2025.3.1
96
+dpcpp-cpp-rt==2025.3.2
97
# via
98
# onemkl-sycl-blas
99
# onemkl-sycl-dft
@@ -172,27 +172,27 @@ idna==3.11
172
# yarl
173
imageio==2.37.3
174
# via scikit-image
175
-impi-rt==2021.17.0
175
+impi-rt==2021.17.2
176
# via
177
# oneccl
178
# torch
179
iniconfig==2.3.0
180
# via pytest
181
-intel-cmplr-lib-rt==2025.3.1
181
+intel-cmplr-lib-rt==2025.3.2
182
# via
183
# intel-sycl-rt
184
# torch
185
-intel-cmplr-lib-ur==2025.3.1
185
+intel-cmplr-lib-ur==2025.3.2
186
# via
187
# intel-openmp
188
# intel-sycl-rt
189
# torch
190
-intel-cmplr-lic-rt==2025.3.1
190
+intel-cmplr-lic-rt==2025.3.2
191
# via
192
# intel-opencl-rt
193
# intel-sycl-rt
194
# torch
195
-intel-opencl-rt==2025.3.1
195
+intel-opencl-rt==2025.3.2
196
# via
197
# dpcpp-cpp-rt
198
# onemkl-sycl-blas
@@ -201,14 +201,14 @@ intel-opencl-rt==2025.3.1
201
# onemkl-sycl-rng
202
# onemkl-sycl-sparse
203
# torch
204
-intel-openmp==2025.3.1
204
+intel-openmp==2025.3.2
205
# via
206
# dpcpp-cpp-rt
207
# mkl
208
# torch
209
-intel-pti==0.15.0
209
+intel-pti==0.16.0
210
# via torch
211
-intel-sycl-rt==2025.3.1
211
+intel-sycl-rt==2025.3.2
212
# via
213
# dpcpp-cpp-rt
214
# oneccl
@@ -270,7 +270,7 @@ mistral-common==1.11.0
270
# via
271
# -c requirements/common.txt
272
# -r requirements/test/xpu.in
273
-mkl==2025.3.0
273
+mkl==2025.3.1
274
# via
275
# onemkl-sycl-blas
276
# onemkl-sycl-dft
@@ -335,28 +335,28 @@ numpy==2.2.6
335
# tifffile
336
# torchvision
337
# transformers
338
-oneccl==2021.17.1
338
+oneccl==2021.17.2
339
# via
340
# oneccl-devel
341
# torch
342
-oneccl-devel==2021.17.1
342
+oneccl-devel==2021.17.2
343
# via torch
344
-onemkl-license==2025.3.0
344
+onemkl-license==2025.3.1
345
# via
346
# mkl
347
# torch
348
-onemkl-sycl-blas==2025.3.0
348
+onemkl-sycl-blas==2025.3.1
349
# via
350
# onemkl-sycl-lapack
351
# onemkl-sycl-sparse
352
# torch
353
-onemkl-sycl-dft==2025.3.0
353
+onemkl-sycl-dft==2025.3.1
354
# via torch
355
-onemkl-sycl-lapack==2025.3.0
355
+onemkl-sycl-lapack==2025.3.1
356
# via torch
357
-onemkl-sycl-rng==2025.3.0
357
+onemkl-sycl-rng==2025.3.1
358
# via torch
359
-onemkl-sycl-sparse==2025.3.0
359
+onemkl-sycl-sparse==2025.3.1
360
# via torch
361
openai-harmony==0.0.8
362
# via
@@ -608,7 +608,7 @@ tabledata==1.3.4
608
# via pytablewriter
609
tabulate==0.10.0
610
# via sacrebleu
611
-tbb==2022.3.0
611
+tbb==2022.3.1
612
# via
613
# intel-opencl-rt
614
# mkl
@@ -645,7 +645,7 @@ tokenizers==0.22.2
645
# via
646
# -c requirements/common.txt
647
# transformers
648
-torch==2.10.0+xpu
648
+torch==2.11.0+xpu
649
# via
650
# -c requirements/xpu.txt
651
# accelerate
@@ -653,7 +653,7 @@ torch==2.10.0+xpu
653
# sentence-transformers
654
# timm
655
# torchvision
656
-torchvision==0.25.0+xpu
656
+torchvision==0.26.0+xpu
657
# via timm
658
tqdm==4.67.3
659
# via
@@ -671,7 +671,7 @@ transformers==5.5.3
671
# via
672
# -c requirements/common.txt
673
# sentence-transformers
674
-triton-xpu==3.6.0
674
+triton-xpu==3.7.0
675
# via torch
676
typepy==1.3.4
677
# via
@@ -710,7 +710,7 @@ typing-inspection==0.4.2
710
# via
711
# fastapi
712
# pydantic
713
-umf==1.0.2
713
+umf==1.0.3
714
# via
715
# intel-cmplr-lib-ur
716
# torch
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