…t labels[..., 1:])

forward right-shifts the target into decoder_input_ids, then ForCausalLMLoss
shifts the labels again, so the model trains against labels[..., 1:]. Use a
plain CrossEntropyLoss instead, matching Whisper/Bart and the Moonshine fix
(huggingface#46784). Adds a regression test, with and without -100.

huggingface#46898)

* Fix Florence2 training-loss double-shift (same pattern as Moonshine huggingface#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.

* Fix import sort order for ruff I001

* Use shift_labels parameter instead of CrossEntropyLoss (per reviewer feedback)

* Fix ruff formatting (one param per line)

* Use do_shift_labels flag: only skip shift when shift_tokens_right() was called (per reviewer)

* Simplify: always pass shift_labels=labels (per reviewer feedback)

…t labels[..., 1:])

forward right-shifts the target into decoder_input_ids, then ForCausalLMLoss
shifts the labels again, so the model trains against labels[..., 1:]. Use a
plain CrossEntropyLoss instead, matching Whisper/Bart and the Moonshine fix
(huggingface#46784). Adds a regression test, with and without -100.

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.

…ngface#46784) (huggingface#46895)

* Fix CohereASR training-loss double-shift (same as Moonshine fix huggingface#46784)

CohereAsrForConditionalGeneration 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.

* Use shift_labels parameter instead of CrossEntropyLoss (per reviewer feedback)

* Use do_shift_labels flag: only skip shift when shift_tokens_right() was called (per reviewer)

* Add test for training loss no double shift

* Retrigger CI

* Fix test: use ParakeetEncoderModelOutput with attention_mask

* Simplify: always pass shift_labels=labels (per reviewer feedback)

Encoder-decoder LM heads right-shift their targets into decoder_input_ids, so
their logits are already position-aligned with `labels`. Two related bugs
mishandle that alignment:

- PPFormulaNet (and, pre-huggingface#46784, Moonshine) routed `labels` through
  ForCausalLMLoss without `shift_labels`, so the loss shifted a second time and
  the model trained against labels[..., 1:].
- Trainer._loss_shifts_labels counted num_items_in_batch over labels[..., 1:]
  for every ForCausalLMLoss head. Encoder-decoder heads have a real target at
  every non-`-100` position, so this under-counts targets (one position per
  sequence) and over-scales the gradient-accumulation loss.

Changes:
- trainer.py: _loss_shifts_labels additionally requires config.is_encoder_decoder
  to be false, so encoder-decoder heads count every valid label. Decoder-only
  heads are unchanged; Csm (decoder-only) is unaffected.
- pp_formulanet: compute the loss via self.loss_function with
  shift_labels=kwargs.pop("shift_labels", labels) (caller-safe, keeps
  num_items_in_batch normalization).
- moonshine / moonshine_streaming: route through self.loss_function with
  shift_labels=labels instead of a plain CrossEntropyLoss that ignored
  num_items_in_batch.

Tests: generic guard test_encoder_decoder_loss_no_double_shift in
tests/test_modeling_common.py, plus Trainer num_items, PPFormulaNet compute_loss
and Moonshine coverage.

Fixes huggingface#46901

* upstream/main: (39 commits)
  Remove deprecated training args and `is_fast` property (huggingface#46917)
  Consistent output shape from `get_image_features` (huggingface#46405)
  Fix multi-device mxfp4 dequantization race in `_convert_moe_packed_tensors` (huggingface#47423)
  fix failed test cases for qwen3_omni_moe model (huggingface#47449)
  Fix Hunyuan-VL PIL image resize parity with reference preprocessing (huggingface#47233)
  Move `value` padding into the attention interfaces that need it (huggingface#47451)
  Simplify function dispatch for linear attention (huggingface#47450)
  [cache] Allow sliding window layers to be roll-backed for speculative decoding (huggingface#47447)
  Fix double-shifted training loss in GitForCausalLM (huggingface#47395)
  Fix CohereASR training-loss double-shift (same as Moonshine fix huggingface#46784) (huggingface#46895)
  Warn when `group_by_length` is silently ignored for iterable datasets (huggingface#47379)
  Update bug report list (huggingface#46607)
  Fix shape mismatch in KyutaiSpeechToText `generate()` last window (huggingface#46952)
  Optimize flash attention max seqlen computation in vision attention (huggingface#47170)
  fix: remove unreachable return in special token builder (huggingface#47420)
  Add Harry to slow CI (huggingface#47454)
  BLT: vectorize patch length processing (huggingface#47385)
  Fix `TrackioCallback` fails to log evaluation metrics after training ends (huggingface#46935)
  [Kimi] add integration tests (huggingface#47383)
  Fix typo in `MusicgenForCausalLM.generate()` (huggingface#46974)
  ...

Encoder-decoder LM heads right-shift their targets into decoder_input_ids, so
their logits are already position-aligned with `labels`. Two related bugs
mishandle that alignment:

- PPFormulaNet (and, pre-huggingface#46784, Moonshine) routed `labels` through
  ForCausalLMLoss without `shift_labels`, so the loss shifted a second time and
  the model trained against labels[..., 1:].
- Trainer._loss_shifts_labels counted num_items_in_batch over labels[..., 1:]
  for every ForCausalLMLoss head. Encoder-decoder heads have a real target at
  every non-`-100` position, so this under-counts targets (one position per
  sequence) and over-scales the gradient-accumulation loss.

Changes:
- trainer.py: _loss_shifts_labels additionally requires config.is_encoder_decoder
  to be false, so encoder-decoder heads count every valid label. Decoder-only
  heads are unchanged; Csm (decoder-only) is unaffected.
- pp_formulanet: compute the loss via self.loss_function with labels=None and the
  aligned targets passed as shift_labels. shift_labels is popped from kwargs
  before the inner model call, so a caller-supplied value is honored, does not
  collide with the derived one, and does not leak into the inner model.
- moonshine / moonshine_streaming: same form, replacing a plain CrossEntropyLoss
  that ignored num_items_in_batch.

Tests:
- tests/test_modeling_common.py: generic guard
  test_encoder_decoder_loss_no_double_shift. It supplies the right-shifted
  decoder_input_ids itself and compares gradients of the reported loss w.r.t.
  output.logits against the aligned and the shifted cross-entropy, which stays
  decisive on testers whose initialization makes the logits near-uniform. No
  per-model branches.
- Dia and ProphetNet skip it via @unittest.skip in their own testers.
- tests/trainer/test_trainer.py: Bart encoder-decoder num_items_in_batch count.
- The Moonshine loss test also covers num_items_in_batch scaling and a
  caller-supplied shift_labels.

Fixes huggingface#46901

…els (huggingface#46903)

Fix encoder-decoder loss alignment and Trainer token counting

Encoder-decoder LM heads right-shift their targets into decoder_input_ids, so
their logits are already position-aligned with `labels`. Two related bugs
mishandle that alignment:

- PPFormulaNet (and, pre-huggingface#46784, Moonshine) routed `labels` through
  ForCausalLMLoss without `shift_labels`, so the loss shifted a second time and
  the model trained against labels[..., 1:].
- Trainer._loss_shifts_labels counted num_items_in_batch over labels[..., 1:]
  for every ForCausalLMLoss head. Encoder-decoder heads have a real target at
  every non-`-100` position, so this under-counts targets (one position per
  sequence) and over-scales the gradient-accumulation loss.

Changes:
- trainer.py: _loss_shifts_labels additionally requires config.is_encoder_decoder
  to be false, so encoder-decoder heads count every valid label. Decoder-only
  heads are unchanged; Csm (decoder-only) is unaffected.
- pp_formulanet: compute the loss via self.loss_function with labels=None and the
  aligned targets passed as shift_labels. shift_labels is popped from kwargs
  before the inner model call, so a caller-supplied value is honored, does not
  collide with the derived one, and does not leak into the inner model.
- moonshine / moonshine_streaming: same form, replacing a plain CrossEntropyLoss
  that ignored num_items_in_batch.

Tests:
- tests/test_modeling_common.py: generic guard
  test_encoder_decoder_loss_no_double_shift. It supplies the right-shifted
  decoder_input_ids itself and compares gradients of the reported loss w.r.t.
  output.logits against the aligned and the shifted cross-entropy, which stays
  decisive on testers whose initialization makes the logits near-uniform. No
  per-model branches.
- Dia and ProphetNet skip it via @unittest.skip in their own testers.
- tests/trainer/test_trainer.py: Bart encoder-decoder num_items_in_batch count.
- The Moonshine loss test also covers num_items_in_batch scaling and a
  caller-supplied shift_labels.

Fixes huggingface#46901