PR #174065: Avoid wrong transform when auto-detecting cuda arch in cmake for SM121

• Repository: pytorch/pytorch
• Author: lygztq
• State: open | merged: False
• Link: https://github.com/pytorch/pytorch/pull/174065

Description (problem / solution / changelog)

current implementation of CUDA_DETECT_INSTALLED_GPUS will transform all cuda arch code 12.1 to 12.1(2.0), which will eventually feed a wrong argument -gencode;arch=compute_20,code=sm_121 to nvcc.

Changed files

• cmake/Modules_CUDA_fix/upstream/FindCUDA/select_compute_arch.cmake (modified, +1/-1)

PR #173754: [NVIDIA] DGX Spark: Fix wrong computing capabilities replacement in cmake

• Repository: pytorch/pytorch
• Author: Rtoax
• State: open | merged: False
• Link: https://github.com/pytorch/pytorch/pull/173754

Description (problem / solution / changelog)

Nvidia DGX Spark [0] has computing-cap = 12.1, however, due to an incorrect replacement (2.1 -> 2.1(2.0)), 12.1 has been replaced with 12.1(2.0), which will incorrectly set the nvcc parameter:

$ nvcc ... -gencode arch=compute_20,code=sm_121
nvcc fatal   : Unsupported gpu architecture 'compute_20'

[0] https://docs.nvidia.com/dgx/dgx-spark/

Changed files

• cmake/Modules_CUDA_fix/upstream/FindCUDA/select_compute_arch.cmake (modified, +2/-1)

PR #38484: [Build] Add SM121 (DGX Spark / GB10) to published build targets

• Repository: vllm-project/vllm
• Author: JCorners68
• State: open | merged: False
• Link: https://github.com/vllm-project/vllm/pull/38484

Description (problem / solution / changelog)

Summary

• Add compute capability 12.1 (SM121) to TORCH_CUDA_ARCH_LIST in all build configurations: release wheel CI, Dockerfile, docker-bake.hcl, versions.json, and FlashInfer AOT build
• CMakeLists.txt already lists 12.1 in CUDA_SUPPORTED_ARCHS for CUDA 12.8+ and 13.0+, and all kernel-specific arch lists (CUTLASS, Marlin FP8, MoE, NVFP4) already include 12.1 — but the build scripts never passed 12.1 in TORCH_CUDA_ARCH_LIST, so the intersection filtering excluded it from all published artifacts
• This is the minimal change needed so that published wheels and Docker images include SM121 kernels for DGX Spark / GB10 users out of the box

Context

The NVIDIA GB10 GPU in DGX Spark reports compute capability 12.1 (SM121). Without native kernels in published builds, users must either build from source or rely on PTX JIT compilation via the 12.0+PTX forward-compat path (slower first-run, no guarantee of correctness for all kernels).

Upstream PyTorch fix: The CMake auto-detection bug that corrupts 12.1 → 12.1(2.0) when building PyTorch from source on SM121 hardware has fixes in review: pytorch/pytorch#174065 and pytorch/pytorch#173754. I independently reproduced and verified both fixes on DGX Spark hardware (see comments on those PRs). PyTorch cu130 aarch64 nightly wheels already ship with 12.0+PTX which covers SM121 via JIT.

Relationship to existing PRs: #31740 includes these build config changes among 27 files of broader SM121 platform support, but has been CONFLICTING since January. This PR extracts just the build target additions — 5 files, 6 lines changed — to unblock SM121 in published artifacts independently.

Addresses #36821.

Test evidence

Tested on NVIDIA DGX Spark:

Confirmed that CMakeLists.txt already supports 12.1 in CUDA_SUPPORTED_ARCHS and all kernel-specific arch lists (SCALED_MM_ARCHS, FP4_ARCHS, MLA_ARCHS, CUTLASS_MOE_DATA_ARCHS, MARLIN_FP8_ARCHS, MARLIN_MOE_FP8_ARCHS). The only gap was the TORCH_CUDA_ARCH_LIST in build scripts.

Files changed

Test plan

• Verify docker buildx bake --print shows 12.1 in resolved torch_cuda_arch_list
• Verify wheel build with CUDA 13.0 includes SM121 kernels (cuobjdump check)
• Run vLLM inference on DGX Spark with published image (no --enforce-eager needed)

Changed files

• .github/workflows/scripts/build.sh (modified, +1/-1)
• docker/Dockerfile (modified, +2/-2)
• docker/docker-bake.hcl (modified, +1/-1)
• docker/versions.json (modified, +1/-1)
• tools/flashinfer-build.sh (modified, +1/-1)