Fix corrupted ONNX checkpoints from save_model_onnx caching#368
Merged
Conversation
save_model_onnx cached the ONNX protobuf on the first call and only patched initializers on subsequent saves. With do_constant_folding=True, BatchNorm parameters get folded into preceding Conv weights, producing 21 transformed initializers from 55 state_dict tensors. The subsequent-save path then overwrote those folded initializers with the raw state_dict tensors, silently corrupting every checkpoint after model_0. Rust self-play consumes these ONNX files; the corruption produced garbage NN evaluations and collapsed training (raw_win_perc 18% vs Python's 94%). Always do a full torch.onnx.export instead; the cost is negligible against the seconds-per-training-step budget. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
alejandromarcu
approved these changes
May 15, 2026
alejandromarcu
left a comment
alejandromarcu
left a comment
Collaborator
There was a problem hiding this comment.
yey, great catch!
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
After this change, using
train_v2.pywith rust for self-play appears to be doing the right thing (good win_perc graph).save_model_onnx cached the ONNX protobuf on the first call and only patched initializers on subsequent saves. With do_constant_folding=True, BatchNorm parameters get folded into preceding Conv weights, producing 21 transformed initializers from 55 state_dict tensors. The subsequent-save path then overwrote those folded initializers with the raw state_dict tensors, silently corrupting every checkpoint after model_0.
Rust self-play consumes these ONNX files; the corruption produced garbage NN evaluations and collapsed training (raw_win_perc 18% vs Python's 94%). Always do a full torch.onnx.export instead; the cost is negligible against the seconds-per-training-step budget.