AgentSkillsCN

gpu-model-smoke

在本仓库中，针对lra与pdebench模型运行GPU烟雾测试（编译、实例化、严格重载state_dict，并执行CUDA前向传播）。每当在重构或同步后，需要验证模型的连线、加载与执行是否依然正常时，均可使用此方法。

SKILL.md
--- frontmatter
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

bash
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.