Fix Florence2 training-loss double-shift (same pattern as Moonshine #…#46898
Merged
zucchini-nlp merged 11 commits intoJun 29, 2026
Merged
Conversation
…uggingface#46784) Florence2ForConditionalGeneration shifts labels right via shift_tokens_right() to create decoder_input_ids, then calls self.loss_function() which maps to ForCausalLMLoss — that function shifts labels again internally. This causes the model to train against labels[..., 1:] instead of labels. Replace self.loss_function() with plain CrossEntropyLoss, matching the fix applied to Moonshine in huggingface#46784 and CohereASR.
Comment on lines
+905
to
+906
| loss = self.loss_function( | ||
| logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **kwargs | ||
| ) | ||
| loss_fct = CrossEntropyLoss() |
Member
There was a problem hiding this comment.
we can pass shift_labels=labels into loss-fn to not sift it twice
…as called (per reviewer)
| >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] | ||
| "A green car parked in front of a yellow building." | ||
| ```""" | ||
| do_shift_labels = False |
Member
There was a problem hiding this comment.
do_shift_labels is always True imo since users prepare labels by copying input ids, and shift decoder-input-ids when created by a collator. CMIIW, I can't remember a train use-case with non-shifted inputs
Contributor
Author
There was a problem hiding this comment.
Simplified - removed the do_shift_labels flag entirely, now always passing shift_labels=labels when labels are present. Thanks for the feedback!
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
Contributor
|
[For maintainers] Suggested jobs to run (before merge) run-slow: florence2 |
zucchini-nlp
enabled auto-merge
June 29, 2026 10:36
5 tasks
4 tasks
eliaghazal
added a commit
to eliaghazal/transformers
that referenced
this pull request
Jul 17, 2026
Moonshine's huggingface#46784 fix used a plain CrossEntropyLoss, dropping num_items_in_batch handling; PPFormulaNet overrides Florence2's forward and missed the huggingface#46898 fix, double-shifting labels. Both now pass labels as shift_labels through self.loss_function, keeping gradient accumulation reduction. Regression tests use a global token count that differs from the per-step count so the summed path is distinguishable.
ShamSaleem
added a commit
to ShamSaleem/transformers
that referenced
this pull request
Jul 23, 2026
PPFormulaNetForConditionalGeneration.forward() computes its loss via self.loss_function() (ForCausalLMLoss), which left-shifts labels by one (labels[..., 1:]). PPFormulaNet is an encoder-decoder model whose logits are already aligned with labels, so this extra shift drops the last target token and trains the model against misaligned labels. Pass shift_labels=labels so ForCausalLMLoss skips its internal shift and keeps labels aligned with logits. This is the same fix already merged for Florence2 in huggingface#46898; PPFormulaNet inherits from Florence2 but overrides forward(), so the fix did not propagate. Fixes huggingface#47317
5 tasks
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
3 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.
What does this PR do?
Fixes #46897
Fix double-shift training-loss bug in
Florence2ForConditionalGeneration- the exact same pattern fixed in Moonshine (#46784) and CohereASR (#46894).Florence2ForConditionalGeneration.forward()shifts labels right viashift_tokens_right()to createdecoder_input_ids, then callsself.loss_function()which maps toForCausalLMLoss. That function shifts labels again internally (labels[..., 1:]), so the model trains againstlabels[..., 1:]instead oflabels.The fix replaces
self.loss_function()with plainCrossEntropyLoss(), matching the approach used in Whisper, Bart, and the now-fixed Moonshine.Both
modular_florence2.py(source of truth) andmodeling_florence2.py(auto-generated) are updated.Who can review?
@zucchini-nlp @yonigozlan (multimodal/vision-language models)
CI recap
Dashboard: View test results in Grafana
Latest run: 28365382183:2
Result:
success| Jobs: 2 | Tests: 19 | Failures: 0 | Duration: 2m 2s