From c9d8142a783b887d8b4db9900f7398331ee82ece Mon Sep 17 00:00:00 2001 From: Matt Van Horn Date: Thu, 16 Apr 2026 04:51:18 -0400 Subject: [PATCH] feat: add Blackwell GPU (sm_120) CUDA support (#401) Set TORCH_CUDA_ARCH_LIST in the CUDA build step to include 12.0+PTX for forward compatibility with Blackwell GPUs (RTX 5070 Ti, 5080, etc). Pre-built PyTorch cu128 wheels only ship native kernels for sm_80/86/89/90. Without this, Blackwell GPU users get "no kernel image is available for execution on the device" at runtime. Fixes #386 Related: #395, #396, #399, #400 Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com> --- .github/workflows/release.yml | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml index a3c224df..c3edd02c 100644 --- a/.github/workflows/release.yml +++ b/.github/workflows/release.yml @@ -203,6 +203,12 @@ jobs: - name: Build CUDA server binary (onedir) shell: bash working-directory: backend + env: + # Include Blackwell (sm_120) via PTX forward compatibility. + # Pre-built PyTorch cu128 wheels ship native kernels for sm_80/86/89/90 + # but not sm_120. Setting this env var causes torch.utils.cpp_extension + # (and any JIT-compiled kernels) to target Blackwell GPUs as well. + TORCH_CUDA_ARCH_LIST: "8.0;8.6;8.9;9.0;12.0+PTX" run: python build_binary.py --cuda - name: Package into server core + CUDA libs archives