[Build][NVIDIA] Update public Rubin dependencies and MSA compatibility (#56545)
Signed-off-by: Shang Wang <samshang.wang@mail.utoronto.ca> Co-authored-by: Codex <noreply@openai.com>
Shang Wang committed
Sep 15, 2026 at 16:29 UTC
031f5810c1957b6ec2a7666f65a3fe68c1f7e534
3 files changed
+18
-8
cmake/external_projects/fmha_sm100.cmake
+1
-1
@@ -17,7 +17,7 @@ else()
17
FetchContent_Declare(
18
fmha_sm100
19
GIT_REPOSITORY https://github.com/vllm-project/MSA.git
20
- GIT_TAG 087c161814d4d9c735b46c21212a09e5f8eb92fa
20
+ GIT_TAG f355c37eb4e1413f21ee2ad8bbad25079e6bef9d
21
GIT_PROGRESS TRUE
22
CONFIGURE_COMMAND ""
23
BUILD_COMMAND ""
docker/Dockerfile
+1
-1
@@ -641,7 +641,7 @@ RUN --mount=type=cache,target=/opt/uv/cache \
641
python3 use_existing_torch.py --prefix; \
642
fi \
643
&& if [ "${INSTALL_RUBIN_PRERELEASE}" = "true" ]; then \
644
- sed -Ei '/^[[:space:]]*(flashinfer-python|nvidia-cutlass-dsl\[cu13\])[[:space:]]*==/Id' \
644
+ sed -Ei '/^[[:space:]]*(flashinfer-python|nvidia-cutlass-dsl\[cu13\]|quack-kernels)[[:space:]]*==/Id' \
645
requirements/cuda.txt; \
646
fi
647
requirements/rubin-prerelease.txt
+16
-6
@@ -2,15 +2,25 @@
2
# Installed when INSTALL_RUBIN_PRERELEASE=true (see docker/Dockerfile).
3
# TODO(rubin-bringup): delete once Rubin APIs ship in regular PyTorch wheels
4
# (vllm-project/vllm#49735).
5
---pre
5
--extra-index-url https://download.pytorch.org/whl/nightly/cu134
6
+--extra-index-url https://flashinfer.ai/whl/cu134
7
+--extra-index-url https://flashinfer.ai/whl
8
8
-torch==2.15.0.dev20260903+cu134
9
-torchvision==0.30.0.dev20260903+cu134
10
-torchaudio==2.11.0.dev20260903+cu134
9
+# Match each architecture's torch pin to the latest torchvision wheel.
10
+torch==2.15.0.dev20260909+cu134; platform_machine == "x86_64"
11
+torch==2.15.0.dev20260908+cu134; platform_machine == "aarch64"
12
+torchvision==0.30.0.dev20260909+cu134
13
+torchaudio==2.11.0.dev20260911+cu134
14
12
-# docker/Dockerfile strips these two from requirements/cuda.txt whenever
15
+# docker/Dockerfile strips these packages from requirements/cuda.txt whenever
16
# INSTALL_RUBIN_PRERELEASE=true, so they must be pinned here instead to
17
# still get installed.
18
nvidia-cutlass-dsl[cu13]==4.8.0.dev0
16
-flashinfer-python==0.6.18
19
+flashinfer-python==0.6.18.post1
20
+quack-kernels==0.6.5
21
+
22
+cuda-python==13.4.1
23
+cuda-bindings==13.4.1
24
+flashinfer-jit-cache==0.6.18.post1+cu134
25
+flashinfer-cubin==0.6.18.post1
26
+nixl==1.4.1