Fix ambiguous tensor truth-value check in train_rl.loss_func (use .it… #2085
+3
−2
Conversation
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
…em())
Fixes a runtime error in train_rl.py where a 0-d PyTorch tensor (total_tokens) was compared directly to 0 in an if-statement. Comparing a tensor truth value in Python is ambiguous and can raise an exception ("The truth value of a tensor is ambiguous"). The change uses total_tokens.item() for a safe scalar comparison and creates the fallback tensor with a matching dtype/device.
Motivation
In some RL training configurations (ex: empty bins / no non-padded tokens), loss_mask.sum() produces a 0-d tensor. The original code used if total_tokens == 0: which tries to evaluate the truth value of a tensor and can crash. This fix makes the check explicit and robust, avoiding failures during forward passes.
What I changed
In loss_func in train_rl.py:
Replaced: if total_tokens == 0:
With: if total_tokens.item() == 0: and create the fallback total_tokens with the same device and dtype as loss_mask_flat: total_tokens = torch.tensor(1.0, device=loss_mask_flat.device, dtype=loss_mask_flat.dtype)
No functional changes to the loss computation beyond avoiding the exception and ensuring dtype/device consistency for the fallback scalar.
Files modified
train_rl.py
One-line logic change in loss_func to safely compare the 0-d tensor and to create the fallback tensor with matching dtype/device.
Behavioral impact
Prevents a runtime exception when total_tokens is zero (empty or fully padded microbatches).
Keeps the original behavior of using 1 as the fallback token count to avoid division-by-zero.
Backwards compatible with existing training setups.
Notes
Cosmetic typos in comments were observed (ex: "Without out" -> "Without", "determinisic" -> "deterministic"). Those do not affect runtime; I left them untouched in this change but can include a follow-up cleanup commit if preferred.
tdene
added
the
Expert Review
tdene
requested changes
Nov 2, 2025
Contributor
tdene
left a comment
tdene
left a comment
There was a problem hiding this comment.
Thank you for the contribution!
| # Avoid division by zero for empty bins | ||
| if total_tokens == 0: | ||
| total_tokens = torch.tensor(1.0, device=loss_mask_flat.device) | ||
| # total_tokens is a 0-d tensor; use .item() to compare its scalar value safely in Python |
Contributor
There was a problem hiding this comment.
I think this comment is unnecessary, could you remove?
| total_tokens = torch.tensor(1.0, device=loss_mask_flat.device) | ||
| # total_tokens is a 0-d tensor; use .item() to compare its scalar value safely in Python | ||
| if total_tokens.item() == 0: | ||
| total_tokens = torch.tensor(1.0, device=loss_mask_flat.device, dtype=loss_mask_flat.dtype) |
Contributor
There was a problem hiding this comment.
Could you reformat this line so that it is less than 100 characters in length?
I know the RL files are not good at always following this linting convention, but we should at least try to make any new code follow it. We'll do a comprehensive cleanup pass later, and then add RL files to the auto-linter.
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.
Fixes a runtime error in train_rl.py where a 0-d PyTorch tensor (total_tokens) was compared directly to 0 in an if-statement. Comparing a tensor truth value in Python is ambiguous and can raise an exception ("The truth value of a tensor is ambiguous"). The change uses total_tokens.item() for a safe scalar comparison and creates the fallback tensor with a matching dtype/device. Motivation
In some RL training configurations (ex: empty bins / no non-padded tokens), loss_mask.sum() produces a 0-d tensor. The original code used if total_tokens == 0: which tries to evaluate the truth value of a tensor and can crash. This fix makes the check explicit and robust, avoiding failures during forward passes. What I changed
In loss_func in train_rl.py:
Replaced: if total_tokens == 0:
With: if total_tokens.item() == 0: and create the fallback total_tokens with the same device and dtype as loss_mask_flat: total_tokens = torch.tensor(1.0, device=loss_mask_flat.device, dtype=loss_mask_flat.dtype) No functional changes to the loss computation beyond avoiding the exception and ensuring dtype/device consistency for the fallback scalar.
Files modified
train_rl.py
One-line logic change in loss_func to safely compare the 0-d tensor and to create the fallback tensor with matching dtype/device. Behavioral impact
Prevents a runtime exception when total_tokens is zero (empty or fully padded microbatches). Keeps the original behavior of using 1 as the fallback token count to avoid division-by-zero. Backwards compatible with existing training setups.
Notes
Cosmetic typos in comments were observed (ex: "Without out" -> "Without", "determinisic" -> "deterministic"). Those do not affect runtime; I left them untouched in this change but can include a follow-up cleanup commit if preferred.