@tduncan / Nvidia-Optimization-Stack-VLLM / commits / 5b29c958c7

[XPU] upgrade to torch 2.13 (#48677)

Signed-off-by: Yan Ma <yan.ma@intel.com> Signed-off-by: Kunshang Ji <kunshang.ji@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

Yan Ma committed Jul 29, 2026 at 16:24 UTC 5b29c958c713d367c7caf9313529d4fb2dd6b762
7 files changed +35 -92
.buildkite/scripts/hardware_ci/run-intel-test.sh
+1 -1
@@ -369,7 +369,7 @@ export HF_TOKEN ZE_AFFINITY_MASK
369 -e CMDS \
370 --name "${container_name}" \
371 "${IMAGE}" \
372 - bash -c 'set -e; source /opt/intel/oneapi/setvars.sh --force; source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force; echo "ZE_AFFINITY_MASK is ${ZE_AFFINITY_MASK:-}"; eval "$CMDS"' \
372 + bash -c 'set -e; echo "ZE_AFFINITY_MASK is ${ZE_AFFINITY_MASK:-}"; eval "$CMDS"' \
373 >/dev/null
374 } 9>/tmp/docker-pull.lock
375
docker/Dockerfile.xpu
+2 -35
@@ -63,21 +63,6 @@ RUN apt-get update -y && \
63 python3-pip && \
64 rm -rf /var/lib/apt/lists/*
65
66 -# Add oneAPI repo, pin oneAPI to 2025.3, then install pinned packages in one layer.
67 -RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null && \
68 - echo "deb [signed-by=/usr/share/keyrings/oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main" | tee /etc/apt/sources.list.d/oneAPI.list && \
69 - printf '%s\n' \
70 - 'Package: intel-oneapi-* intel-deep-learning-essentials* intel-pti*' \
71 - 'Pin: version 2025.3*' \
72 - 'Pin-Priority: 1001' \
73 - > /etc/apt/preferences.d/oneapi-2025.3.pref && \
74 - apt-get update -y && \
75 - apt-get install -y --no-install-recommends \
76 - intel-oneapi-compiler-dpcpp-cpp-2025.3 \
77 - intel-oneapi-mkl-devel-2025.3 \
78 - intel-oneapi-dnnl-devel-2025.3 && \
79 - rm -rf /var/lib/apt/lists/*
80 -
66 # Install UMD
67 RUN mkdir neo && \
68 cd neo && \
@@ -100,22 +85,6 @@ RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
85 && uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
86 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
87
103 -# This oneccl contains the BMG support which is not the case for default version of oneapi 2025.3.
104 -ARG ONECCL_INSTALLER="intel-oneccl-2021.15.9.14_offline.sh"
105 -RUN wget "https://github.com/uxlfoundation/oneCCL/releases/download/2021.15.9/${ONECCL_INSTALLER}" && \
106 - bash "${ONECCL_INSTALLER}" -a --silent --eula accept && \
107 - rm "${ONECCL_INSTALLER}" && \
108 - echo "source /opt/intel/oneapi/setvars.sh --force" >> /root/.bashrc && \
109 - echo "source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force" >> /root/.bashrc && \
110 - rm -f /opt/intel/oneapi/ccl/latest && \
111 - ln -s /opt/intel/oneapi/ccl/2021.15 /opt/intel/oneapi/ccl/latest && \
112 - printf '%s\n' \
113 - '/opt/intel/oneapi/ccl/2021.15/lib' \
114 - '/opt/intel/oneapi/mpi/2021.15/lib' \
115 - '/opt/intel/oneapi/compiler/2025.3/lib' \
116 - > /etc/ld.so.conf.d/oneapi-ccl.conf && \
117 - ldconfig
118 -
88 SHELL ["bash", "-c"]
89 CMD ["bash", "-c", "source /root/.bashrc && exec bash"]
90
@@ -135,8 +104,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
104 uv pip install --upgrade pip
105
106
138 -
139 -ENV LD_LIBRARY_PATH=/opt/intel/oneapi/ccl/2021.15/lib:/opt/intel/oneapi/mpi/2021.15/lib:/opt/intel/oneapi/compiler/2025.3/lib:/usr/local/lib
107 +ENV LD_LIBRARY_PATH=/opt/venv/lib:/usr/local/lib
108 CMD ["/bin/bash"]
109
110 ######################### UCX + NIXL BUILD STAGE #########################
@@ -216,8 +184,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
184 uv pip install -r /workspace/vllm/requirements/xpu.txt && \
185 uv pip install --no-build-isolation -r /workspace/vllm/requirements/test/xpu.txt && \
186 uv pip uninstall triton triton-xpu && \
219 - uv pip install triton-xpu==3.7.1 && \
220 - uv pip uninstall oneccl oneccl-devel
187 + uv pip install triton-xpu==3.7.2
188
189 # Keep source-dependent layers near the end so frequent code-only changes
190 # don't invalidate heavy dependency and UCX/NIXL layers.
docs/getting_started/installation/gpu.xpu.inc.md
+2 -2
@@ -42,12 +42,12 @@ pip install -v -r requirements/xpu.txt
42
43 ```bash
44 pip uninstall -y triton triton-xpu
45 - pip install triton-xpu==3.7.1 --extra-index-url https://download.pytorch.org/whl/xpu
45 + pip install triton-xpu==3.7.2 --extra-index-url https://download.pytorch.org/whl/xpu
46 ```
47
48 !!! note
49 - `triton` (without suffix) is for NVIDIA GPUs only. On XPU, using it instead of `triton-xpu` can cause correctness or runtime issues.
50 - - For torch 2.12 (the version used in `requirements/xpu.txt`), the matching package is `triton-xpu==3.7.1`. 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 + - For torch 2.13 (the version used in `requirements/xpu.txt`), the matching package is `triton-xpu==3.7.2`. 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).
51
52 - Finally, build and install vLLM XPU backend:
53
requirements/test/xpu.txt
+26 -24
@@ -140,7 +140,7 @@ docopt==0.6.2
140 # via num2words
141 docstring-parser==0.18.0
142 # via anthropic
143 -dpcpp-cpp-rt==2025.3.2
143 +dpcpp-cpp-rt==2026.0.0
144 # via
145 # onemkl-sycl-blas
146 # onemkl-sycl-dft
@@ -253,27 +253,27 @@ ijson==3.5.0
253 # via -r requirements/test/../common.txt
254 imageio==2.37.3
255 # via scikit-image
256 -impi-rt==2021.17.2
256 +impi-rt==2021.18.0
257 # via
258 # oneccl
259 # torch
260 iniconfig==2.3.0
261 # via pytest
262 -intel-cmplr-lib-rt==2025.3.2
262 +intel-cmplr-lib-rt==2026.0.0
263 # via
264 # intel-sycl-rt
265 # torch
266 -intel-cmplr-lib-ur==2025.3.2
266 +intel-cmplr-lib-ur==2026.0.0
267 # via
268 # intel-openmp
269 # intel-sycl-rt
270 # torch
271 -intel-cmplr-lic-rt==2025.3.2
271 +intel-cmplr-lic-rt==2026.0.0
272 # via
273 # intel-opencl-rt
274 # intel-sycl-rt
275 # torch
276 -intel-opencl-rt==2025.3.2
276 +intel-opencl-rt==2026.0.0
277 # via
278 # dpcpp-cpp-rt
279 # onemkl-sycl-blas
@@ -282,14 +282,14 @@ intel-opencl-rt==2025.3.2
282 # onemkl-sycl-rng
283 # onemkl-sycl-sparse
284 # torch
285 -intel-openmp==2025.3.2
285 +intel-openmp==2026.0.0
286 # via
287 # dpcpp-cpp-rt
288 # mkl
289 # torch
290 -intel-pti==0.16.0
290 +intel-pti==0.17.0
291 # via torch
292 -intel-sycl-rt==2025.3.2
292 +intel-sycl-rt==2026.0.0
293 # via
294 # dpcpp-cpp-rt
295 # oneccl
@@ -378,7 +378,7 @@ mistral-common==1.11.5
378 # -c requirements/common.txt
379 # -r requirements/test/../common.txt
380 # -r requirements/test/xpu.in
381 -mkl==2025.3.1
381 +mkl==2026.0.0
382 # via
383 # onemkl-sycl-blas
384 # onemkl-sycl-dft
@@ -453,28 +453,28 @@ numpy==2.2.6
453 # torchvision
454 # transformers
455 # xgrammar
456 -oneccl==2021.17.2
456 +oneccl==2022.0.0
457 # via
458 # oneccl-devel
459 # torch
460 -oneccl-devel==2021.17.2
460 +oneccl-devel==2022.0.0
461 # via torch
462 -onemkl-license==2025.3.1
462 +onemkl-license==2026.0.0
463 # via
464 # mkl
465 # torch
466 -onemkl-sycl-blas==2025.3.1
466 +onemkl-sycl-blas==2026.0.0
467 # via
468 # onemkl-sycl-lapack
469 # onemkl-sycl-sparse
470 # torch
471 -onemkl-sycl-dft==2025.3.1
471 +onemkl-sycl-dft==2026.0.0
472 # via torch
473 -onemkl-sycl-lapack==2025.3.1
473 +onemkl-sycl-lapack==2026.0.0
474 # via torch
475 -onemkl-sycl-rng==2025.3.1
475 +onemkl-sycl-rng==2026.0.0
476 # via torch
477 -onemkl-sycl-sparse==2025.3.1
477 +onemkl-sycl-sparse==2026.0.0
478 # via torch
479 openai==2.44.0
480 # via
@@ -719,6 +719,8 @@ pyyaml==6.0.3
719 # timm
720 # transformers
721 # uvicorn
722 +pyzes==0.1.1
723 + # via torch
724 pyzmq==27.1.0
725 # via
726 # -c requirements/common.txt
@@ -871,14 +873,14 @@ tabledata==1.3.4
873 # via pytablewriter
874 tabulate==0.10.0
875 # via sacrebleu
874 -tbb==2022.3.1
876 +tbb==2023.0.0
877 # via
878 # intel-opencl-rt
879 # mkl
880 # torch
881 tblib==3.1.0
882 # via -r requirements/test/xpu.in
881 -tcmlib==1.4.1
883 +tcmlib==1.5.0
884 # via
885 # tbb
886 # torch
@@ -910,7 +912,7 @@ tokenizers==0.22.2
912 # -c requirements/common.txt
913 # -r requirements/test/../common.txt
914 # transformers
913 -torch==2.12.0+xpu
915 +torch==2.13.0+xpu
916 # via
917 # -c requirements/xpu.txt
918 # accelerate
@@ -920,7 +922,7 @@ torch==2.12.0+xpu
922 # timm
923 # torchvision
924 # xgrammar
923 -torchvision==0.27.0+xpu
925 +torchvision==0.28.0+xpu
926 # via timm
927 tqdm==4.67.3
928 # via
@@ -946,7 +948,7 @@ transformers==5.14.1
948 # xgrammar
949 triton==3.7.1
950 # via xgrammar
949 -triton-xpu==3.7.1
951 +triton-xpu==3.7.2
952 # via torch
953 typepy==1.3.4
954 # via
@@ -1001,7 +1003,7 @@ typing-inspection==0.4.2
1003 # mcp
1004 # pydantic
1005 # pydantic-settings
1004 -umf==1.0.3
1006 +umf==1.1.0
1007 # via
1008 # intel-cmplr-lib-ur
1009 # torch
requirements/xpu.txt
+3 -3
@@ -12,10 +12,10 @@ jinja2>=3.1.6
12 datasets # for benchmark scripts
13 numba == 0.65.0 # Required for N-gram speculative decoding
14 --extra-index-url=https://download.pytorch.org/whl/xpu
15 -torch==2.12.0
15 +torch==2.13.0
16 torchaudio
17 torchvision
18 torchcodec >= 0.14 # Required for the torchcodec video decoding backend
19
20 -auto_round_lib==0.14.1
21 -vllm_xpu_kernels @ https://github.com/vllm-project/vllm-xpu-kernels/releases/download/v0.1.11.1/vllm_xpu_kernels-0.1.11.1-cp38-abi3-manylinux_2_28_x86_64.whl
20 +auto_round_lib==0.14.2
21 +vllm_xpu_kernels @ https://github.com/vllm-project/vllm-xpu-kernels/releases/download/v0.1.12/vllm_xpu_kernels-0.1.12-cp38-abi3-manylinux_2_28_x86_64.whl
setup.py
+1 -1
@@ -1122,7 +1122,7 @@ if _is_cuda() or _is_hip():
1122 # copying the relevant .py files from the source repository.
1123 ext_modules.append(CMakeExtension(name="vllm.triton_kernels", optional=True))
1124
1125 -if sys.version_info >= (3, 11):
1125 +if not _is_xpu() and sys.version_info >= (3, 11):
1126 ext_modules.append(CMakeExtension(name="vllm.spinloop"))
1127 ext_modules.append(CMakeExtension(name="vllm.fs_io_C"))
1128
vllm/_custom_ops.py
-26
@@ -2391,17 +2391,6 @@ def topk_softmax(
2391 e_score_correction_bias: torch.Tensor | None = None,
2392 is_padding: torch.Tensor | None = None,
2393 ) -> None:
2394 - if current_platform.is_xpu():
2395 - # TODO: Remove after vllm-xpu-kernels supports is_padding.
2396 - torch.ops._moe_C.topk_softmax(
2397 - topk_weights,
2398 - topk_ids,
2399 - token_expert_indices,
2400 - gating_output,
2401 - renormalize,
2402 - e_score_correction_bias,
2403 - )
2404 - return
2394 torch.ops._moe_C.topk_softmax(
2395 topk_weights,
2396 topk_ids,
@@ -2447,21 +2436,6 @@ def topk_hash_softplus_sqrt(
2436 hash_indices_table: torch.Tensor | None = None,
2437 is_padding: torch.Tensor | None = None,
2438 ) -> None:
2450 - if current_platform.is_xpu():
2451 - # TODO: Remove after vllm-xpu-kernels supports is_padding.
2452 - torch.ops._moe_C.topk_softplus_sqrt(
2453 - topk_weights,
2454 - topk_indices,
2455 - token_expert_indices,
2456 - gating_output,
2457 - renormalize,
2458 - routed_scaling_factor,
2459 - e_score_correction_bias,
2460 - input_tokens,
2461 - hash_indices_table,
2462 - )
2463 - return
2464 -
2439 torch.ops._moe_C.topk_softplus_sqrt(
2440 topk_weights,
2441 topk_indices,