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 14, 05:30 UTC
Revert "Install pre-built xformers-0.0.32.post2 built with pt-2.9.0" (#27714)
Simon Mo committed
Oct 28, 2025 at 20:58 UTC
9007bf57e6a25e46c0c2408e2510d6b897400895
2 files changed
+8
-1
docker/Dockerfile
+7
@@ -361,6 +361,13 @@ RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist
361
&& uv pip install --system dist/*.whl --verbose \
362
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
363
364
+# TODO (huydhn): Remove this once xformers is released for 2.9.0
365
+RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
366
+ . /etc/environment
367
+ export TORCH_CUDA_ARCH_LIST='7.5 8.0+PTX 9.0a'
368
+ uv pip install --system --no-build-isolation "git+https://github.com/facebookresearch/xformers@v0.0.32.post2"
369
+BASH
370
+
371
# Install FlashInfer pre-compiled kernel cache and binaries
372
# https://docs.flashinfer.ai/installation.html
373
RUN --mount=type=cache,target=/root/.cache/uv \
requirements/cuda.txt
+1
-1
@@ -10,7 +10,7 @@ torchaudio==2.9.0
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
12
# https://github.com/facebookresearch/xformers/releases/tag/v0.0.32.post1
13
-xformers==0.0.33+5d4b92a5.d20251026; platform_system == 'Linux' and platform_machine == 'x86_64' # Requires PyTorch >= 2.9
13
+# xformers==0.0.32.post1; platform_system == 'Linux' and platform_machine == 'x86_64' # Requires PyTorch >= 2.8
14
# FlashInfer should be updated together with the Dockerfile
15
flashinfer-python==0.4.1
16
# Triton Kernels are needed for mxfp4 fused moe. (Should be updated alongside torch)