Rewind a training run to the instant its loss spiked, localize the poisoned optimizer state, repair just that slice, and prove it via bit-identical replay forks. Instrument built + CUDA-verified on Kaggle T4; restarting on AMD MI300X. PyTorch research.
What is the jaztulsi/optimizer-autopsy GitHub project? Description: "Rewind a training run to the instant its loss spiked, localize the poisoned optimizer state, repair just that slice, and prove it via bit-identical replay forks. Instrument built + CUDA-verified on Kaggle T4; restarting on AMD MI300X. PyTorch research.". Written in Python. Explain what it does, its main use cases, key features, and who would benefit from using it.
Question is copied to clipboard — paste it after the AI opens.
Clone via HTTPS
Clone via SSH
Download ZIP
Download main.zipReport bugs or request features on the optimizer-autopsy issue tracker:
Open GitHub Issues