gpu_func_cli

Standalone CLI for running course exercises and custom CUDA kernels on a remote GPU through the GFAAS REST API. The local machine needs no CUDA, nvcc, Nsight Compute, or GPU. The CLI sends a self-contained job to a GFAAS worker, and the worker does the CUDA work.

Full documentation: GUIDE.md. It covers install, configuration, running a starter.zip exercise, custom kernels (with and without a harness), reports and feedback, command reference, and troubleshooting.

Quick start

uv tool install --editable /path/to/gpu_func_cli   # or: pip install .
export GFAAS_API_BASE="https://<hub-host>/api"
export GFAAS_API_KEY="<your-api-key>"
gpu_func_cli workers

Run a course exercise (starter.zip)

Unzip the starter.zip, edit the starter .cu, then run an action from inside the folder — the exercise is auto-detected from the cwd:

unzip 01-haxpy.zip -d 01-haxpy && cd 01-haxpy
# edit haxpy.cu (your solution), then:
gpu_func_cli test          # correctness tests
gpu_func_cli benchmark     # timing + GiB/s + % of peak
gpu_func_cli grade         # full suite: test + sanitizer + benchmark
# from elsewhere, point at the unzipped dir: --exercise-dir /path/to/01-haxpy

Run a custom kernel

# any self-contained .cu (has its own main()) — nothing else to bring:
gpu_func_cli custom run /path/to/your_kernel.cu --gpu B200

# kernel-only source? add a --harness that supplies main():
gpu_func_cli custom run kernel.cu --harness harness.cu --gpu B200

# profile on the GPU, then read the report locally:
gpu_func_cli custom profile your_kernel.cu --gpu B200 --artifact-dir ./out
gpu_func_cli report summary ./out/your_kernel.ncu-rep --per-kernel

New to the tool? Section 3 of GUIDE.md walks the starter.zip flow end to end, and the custom walkthrough creates its own test files so you don't need to bring a CUDA program.