AI-assisted development is always welcome, and if you’re not using AI these days, you may need to disclose that 😄.
If a PR changes runtime behavior, model implementation, backend logic, performance paths, or shared framework code, it should include actual runtime validation. Code inspection, static reasoning, generated tests, or “this should work” are not enough. AI hallucination can produce plausible-looking but incorrect changes.
One simple example: Why run ops X on the CPU instead of putting it in the CUDA graph? AI will almost always try to convince you that CUDA must be faster. But once you actually measure it, the gain might be only 0.1% while consuming GPU resources that are better used elsewhere.
At minimum, please run the affected path end to end and include the relevant command, environment/backend, and observed result. For optimization PRs, include before/after measurements. For behavior changes, include regression or output validation where appropriate.
Bug/optimization PRs auto-generated by AI agents with NO runtime validation will be given the lowest review priority, or may be closed if I can’t reach a real human behind the submission.
Please keep no more than 3 PRs open at the same time, including draft PRs. Once one is merged or closed, feel free to open another.
Thank you for understanding!
AI-assisted development is always welcome, and if you’re not using AI these days, you may need to disclose that 😄.
If a PR changes runtime behavior, model implementation, backend logic, performance paths, or shared framework code, it should include actual runtime validation. Code inspection, static reasoning, generated tests, or “this should work” are not enough. AI hallucination can produce plausible-looking but incorrect changes.
One simple example: Why run ops X on the CPU instead of putting it in the CUDA graph? AI will almost always try to convince you that CUDA must be faster. But once you actually measure it, the gain might be only 0.1% while consuming GPU resources that are better used elsewhere.
At minimum, please run the affected path end to end and include the relevant command, environment/backend, and observed result. For optimization PRs, include before/after measurements. For behavior changes, include regression or output validation where appropriate.
Bug/optimization PRs auto-generated by AI agents with NO runtime validation will be given the lowest review priority, or may be closed if I can’t reach a real human behind the submission.
Please keep no more than 3 PRs open at the same time, including draft PRs. Once one is merged or closed, feel free to open another.
Thank you for understanding!