pygpubench-modal-run

Files (2)

• SKILL.md
• run_pygpubench_modal.sh · scripts/

SKILL.md (first lines)

# PyGPUBench Modal Run

Run PyGPUBench on Modal B200 using:
- `tools/pygpubench/pygpubench_modal_app.py` (deployed app/class)
- `tools/pygpubench/run_pygpubench_modal.py` (deployed caller)

This flow is self-contained:
- Modal image installs `pygpubench` from GitHub (`ngc92/pygpubench`) at build time.
- No local `pygpubench` clone is required.

## Required steps

1. Source environment first:
```bash
set -a; source .env; set +a
```

2. Deploy once (or after app changes):
```bash
bash skills/codex/pygpubench-modal-run/scripts/run_pygpubench_modal.sh --deploy
```

3. Run the deployed wrapper:
```bash
bash skills/codex/pygpubench-modal-run/scripts/run_pygpubench_modal.sh \
  --harness /abs/path/to/harness.py \
  --submission /abs/path/to/submission.py \
  --profile candidate \
  --print-log
```

## Wrapper defaults

- GPU is fixed to `B200`.
- Profiles (all B200):
  - `smoke`: short timeout + low repeats + early-stop threshold
  - `candidate`: medium timeout/repeats
  - `final`: longer timeout/full repeats
- Output JSON defaults to:
  - `artifacts/pygpubench_modal_last_run.json`

## Harness contract

Your harness must:
- define `kernel_generator()` returning a callable from `submission.py`
- define `generate_test_case(args, seed)` returning:
  - `(kernel_input_tuple, (expected_tensor, rtol, atol))`
- call `pygpubench.do_bench_isolated(...)` in `if __name__ == "__main__":`

## Cost-focused knobs

- `--profile smoke|candidate|final`
- `--timeout-seconds N`
- `--repeats N`
- `--stage-repeats 8,16,32`
- `--early-stop-us X`
- `--env KEY=VALUE` (repeatable)

## Failure handling

- If a run fails, re-run once with the same command.
- If second run fails, treat as hard failure and preserve JSON/log output.

From cybernetic-physics/ttt_discover, MIT.

More from cybernetic-physics/ttt_discover