Show HN: modal-cuda – CLI to run CUDA .cu programs on Modal GPUs (github.com)

modal-cuda (command: mcc) is a tiny CLI that compiles and runs local CUDA C (.cu) programs on Modal cloud GPUs. Installable from PyPI, mcc ships your source to a disposable container on Modal, invokes nvcc inside the chosen GPU runtime, streams compiler and program stdout/stderr back to your terminal, and tears down artifacts when finished. It wraps the full workflow (upload → compile → execute → teardown) in a single command, surfacing compiler or runtime failures as non‑zero shell exits so you can iterate quickly without managing local GPU drivers or CUDA toolchains. For developers and researchers this is significant: it makes on-demand GPU compilation and testing easy across multiple hardware tiers (default T4, also L4, A10, A100/A100-40GB/80GB, L40S, H100, H200, B200), and lets you pick or supply container images (default nvidia/cuda:12.4.1-devel-ubuntu22.04). mcc accepts repeated --nvcc-arg flags for fine‑grained builds (e.g., -arch=sm_90 or -Xptxas -v), supports a 600s default timeout, and integrates with Modal accounts/CLI (modal token new) and Python 3.12+. It’s handy for CI, rapid kernel debugging, or trying kernels on hardware you don’t have locally; caveats include image pull times (use default or prebuilt custom images) and the need for Modal network access. Licensed MIT by ExpressGradient.