From 27280ff648aaa0612436ee7bea130fa86ab7ba54 Mon Sep 17 00:00:00 2001 From: Bryce Adelstein Lelbach Date: Sat, 27 Jun 2026 16:18:02 +0000 Subject: [PATCH] Add cuTile (cuda-tile[tileiras]) to the Modal CUDA image Install NVIDIA's cuTile tile programming model into the shared `cuda_image` so submissions can `import cuda.tile`. The `[tileiras]` extra pulls the Tile IR assembler (cuda-toolkit[nvcc,nvvm,tileiras] 13.2-13.3) so kernels compile in-process. Pinned to 1.4.0, the current release, with a cp313 manylinux x86_64 wheel matching the image's Python 3.13. A probe submission to the `eigh` leaderboard (B200) confirmed cuTile is not currently importable on the runner: `import cuda.tile` raises ModuleNotFoundError and the `cuda` namespace exposes only bindings/core/ pathfinder. tileiras 13.2+ supports Blackwell (B200) and Ampere/Ada and needs driver r580+. Co-Authored-By: Claude Opus 4.8 (1M context) --- src/runners/modal_runner.py | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/src/runners/modal_runner.py b/src/runners/modal_runner.py index 488abf62e..82559cd27 100644 --- a/src/runners/modal_runner.py +++ b/src/runners/modal_runner.py @@ -62,6 +62,11 @@ "nvidia-cutlass-dsl==4.5.2", "cuda-core[cu13]", "cuda-python[all]==13.0", + # cuTile: the CUDA Tile programming model in Python (`import cuda.tile`). + # The [tileiras] extra adds the Tile IR assembler so kernels can be + # compiled in-process. tileiras 13.2+ targets Blackwell and Ampere/Ada + # and needs driver r580+, both satisfied by the B200 runners. + "cuda-tile[tileiras]==1.4.0", # "nvmath-python[cu13]~=0.4", # "numba-cuda[cu13]~=0.15", )