From 031f5810c1957b6ec2a7666f65a3fe68c1f7e534 Mon Sep 17 00:00:00 2001 From: Shang Wang Date: Tue, 15 Sep 2026 16:29:13 -0700 Subject: [PATCH] [Build][NVIDIA] Update public Rubin dependencies and MSA compatibility (#56545) Signed-off-by: Shang Wang Co-authored-by: Codex --- cmake/external_projects/fmha_sm100.cmake | 2 +- docker/Dockerfile | 2 +- requirements/rubin-prerelease.txt | 22 ++++++++++++++++------ 3 files changed, 18 insertions(+), 8 deletions(-) diff --git a/cmake/external_projects/fmha_sm100.cmake b/cmake/external_projects/fmha_sm100.cmake index 442097dec7..ae5bfec7b4 100644 --- a/cmake/external_projects/fmha_sm100.cmake +++ b/cmake/external_projects/fmha_sm100.cmake @@ -17,7 +17,7 @@ else() FetchContent_Declare( fmha_sm100 GIT_REPOSITORY https://github.com/vllm-project/MSA.git - GIT_TAG 087c161814d4d9c735b46c21212a09e5f8eb92fa + GIT_TAG f355c37eb4e1413f21ee2ad8bbad25079e6bef9d GIT_PROGRESS TRUE CONFIGURE_COMMAND "" BUILD_COMMAND "" diff --git a/docker/Dockerfile b/docker/Dockerfile index b3e32777e7..d1a7afcbfb 100644 --- a/docker/Dockerfile +++ b/docker/Dockerfile @@ -641,7 +641,7 @@ RUN --mount=type=cache,target=/opt/uv/cache \ python3 use_existing_torch.py --prefix; \ fi \ && if [ "${INSTALL_RUBIN_PRERELEASE}" = "true" ]; then \ - sed -Ei '/^[[:space:]]*(flashinfer-python|nvidia-cutlass-dsl\[cu13\])[[:space:]]*==/Id' \ + sed -Ei '/^[[:space:]]*(flashinfer-python|nvidia-cutlass-dsl\[cu13\]|quack-kernels)[[:space:]]*==/Id' \ requirements/cuda.txt; \ fi diff --git a/requirements/rubin-prerelease.txt b/requirements/rubin-prerelease.txt index 18a644f6b1..8b9f7d3155 100644 --- a/requirements/rubin-prerelease.txt +++ b/requirements/rubin-prerelease.txt @@ -2,15 +2,25 @@ # Installed when INSTALL_RUBIN_PRERELEASE=true (see docker/Dockerfile). # TODO(rubin-bringup): delete once Rubin APIs ship in regular PyTorch wheels # (vllm-project/vllm#49735). ---pre --extra-index-url https://download.pytorch.org/whl/nightly/cu134 +--extra-index-url https://flashinfer.ai/whl/cu134 +--extra-index-url https://flashinfer.ai/whl -torch==2.15.0.dev20260903+cu134 -torchvision==0.30.0.dev20260903+cu134 -torchaudio==2.11.0.dev20260903+cu134 +# Match each architecture's torch pin to the latest torchvision wheel. +torch==2.15.0.dev20260909+cu134; platform_machine == "x86_64" +torch==2.15.0.dev20260908+cu134; platform_machine == "aarch64" +torchvision==0.30.0.dev20260909+cu134 +torchaudio==2.11.0.dev20260911+cu134 -# docker/Dockerfile strips these two from requirements/cuda.txt whenever +# docker/Dockerfile strips these packages from requirements/cuda.txt whenever # INSTALL_RUBIN_PRERELEASE=true, so they must be pinned here instead to # still get installed. nvidia-cutlass-dsl[cu13]==4.8.0.dev0 -flashinfer-python==0.6.18 +flashinfer-python==0.6.18.post1 +quack-kernels==0.6.5 + +cuda-python==13.4.1 +cuda-bindings==13.4.1 +flashinfer-jit-cache==0.6.18.post1+cu134 +flashinfer-cubin==0.6.18.post1 +nixl==1.4.1