fix: prevent batch_size=1 crashes, secure torch.load, fix device/contiguity issues#901
Merged
jhnwu3 merged 1 commit intosunlabuiuc:masterfrom Mar 22, 2026
Conversation
…x device/contiguity issues 1. Fix bare .squeeze() calls that silently remove the batch dimension when batch_size=1, causing wrong results during single-sample inference: - concare.py: .squeeze() → .squeeze(dim=-1) and .squeeze(dim=1) - agent.py: .squeeze() → .squeeze(dim=-1) or removed (already 1-D after .sum/.mean) 2. Add weights_only=True to all torch.load() calls for PyTorch 2.6+ compatibility and security (prevents arbitrary code execution via pickle deserialization): - trainer.py, biot.py, tfm_tokenizer.py (2 calls), kg_base.py 3. Add .contiguous() before pack_padded_sequence in RNNLayer to prevent cuDNN errors with non-contiguous input tensors (fixes sunlabuiuc#800) 4. Fix StageNet device mismatch — tensors were created on CPU instead of the input tensor's device, causing crashes during GPU training: - torch.zeros/ones(...) → torch.zeros/ones(..., device=device) - time == None → time is None (PEP8)
jhnwu3
approved these changes
Mar 22, 2026
Collaborator
jhnwu3
left a comment
jhnwu3
left a comment
There was a problem hiding this comment.
Ah, nice catch! Thanks for the bug fixes. If you're on the discord, happy to chat more.
Contributor
Author
No problem at all! I'll join the Discord, happy to chat with you more there. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Four categories of bug fixes across 8 files, addressing silent correctness errors, security warnings, and device compatibility crashes.
1. Bare
.squeeze()→ explicit dim (prevents batch_size=1 crashes)When
batch_size=1, bare.squeeze()silently removes the batch dimension, causing all downstream operations to compute on wrong axes without any error.concare.py.squeeze().squeeze(dim=-1)concare.py.squeeze().squeeze(dim=1)agent.py.squeeze().squeeze(dim=-1)agent.py.sum(-1).squeeze().sum(dim=-1)(already 1-D)agent.py.mean(dim=-1).squeeze().mean(dim=-1)(already 1-D)2.
torch.load()+weights_only=True(5 call sites)Since PyTorch 2.6,
torch.load()defaults toweights_only=True. Without explicit opt-in, loading checkpoints produces deprecation warnings and will break in future versions. Also prevents arbitrary code execution via pickle deserialization (security).Files:
trainer.py,biot.py,tfm_tokenizer.py(×2),kg_base.py3. RNN
.contiguous()beforepack_padded_sequence(fixes #800)Non-contiguous tensors (from slicing/transposing) passed to
pack_padded_sequencecauseCUDNN_STATUS_NOT_SUPPORTEDerrors. Added.contiguous()call.4. StageNet device mismatch fix
torch.zeros()/torch.ones()were created on CPU regardless of input device, causing device mismatch crashes during GPU training. Now created onx.device. Also fixedtime == None→time is None(PEP8).Test plan