gpu-model-smoke

name	gpu-model-smoke
description	Run GPU smoke checks for lra and pdebench models in this repo (compile, instantiate, strict state_dict reload, and CUDA forward pass). Use this whenever validating that model wiring/loading/execution still works after refactors or syncs.

GPU Model Smoke

Use this skill when the user asks to validate that models in lra and pdebench compile, load, and run on GPU.

Preconditions

• Run from repo root.
• .venv exists with project dependencies.
• CUDA is available (torch.cuda.is_available()).

Command

source .venv/bin/activate && python tests/gpu/smoke_models_gpu.py

Expected Outcome

• compile_errors=0
• unexpected_failures=0
• pdebench.UPT may appear under expected_failures (known incomplete implementation in this repo).

If It Fails

• Read the per-case JSON lines to identify the exact model/backend failing.
• If failure is a test config incompatibility (not a real wiring bug), adjust only that case config in tests/gpu/smoke_models_gpu.py.
• If failure is a real model bug, patch model code and rerun the command until unexpected_failures=0.