logits_to_keep bool mask for all ForXXX classes (#48825)
* feat(gemma4): bool mask logits_to_keep, expose last_hidden_state
logits_to_keep only supported an int (keep the last N) or a flat
index tensor for a single packed sequence. SFT training with
sequence packing needs to exclude non-contiguous spans per example
in a batch, which neither form can express.
Accept a 2D bool mask shaped like input_ids (batch_size,
sequence_length); logits are computed only for the True positions,
flattened in input_ids order. Existing int/tensor behavior is
unchanged.
Also add last_hidden_state to Gemma4CausalLMOutputWithPast,
independent of output_hidden_states, since packed-sequence loss
computation needs the full hidden states regardless of that flag.
Closes #48784
* feat(gemma4): extend logits_to_keep mask support to all ForXXX classes, fix negative layer indices
Extend the bool-mask logits_to_keep support added for
Gemma4ForConditionalGeneration to the rest of the family
(Gemma4ForCausalLM, Gemma4UnifiedForCausalLM,
Gemma4UnifiedForConditionalGeneration), settling on a plain if/else
with a single lm_head call per branch to match the convention already
used elsewhere in the codebase (Gemma4ForCausalLM itself, biogpt, csm,
llava_onevision, etc.).
Also fix output_hidden_states selective layer capture
(src/transformers/utils/output_capturing.py, #48081) to resolve
negative indices (e.g. -1 for the last layer). The hook fires
progressively during the forward pass and can't know the total layer
count in advance, so count how many hidden_states hooks get installed
per model during the one-time hook installation walk, store it on the
model, and resolve negative indices against that count before the
forward pass runs. No change to the per-layer hook logic itself.
Refs #48784
* feat: extend logits_to_keep bool-mask support to all ForCausalLM/ForConditionalGeneration classes
* remove last_hidden_state field, rely on output_hidden_states indices instead
* revert CSM to original logits_to_keep handling, drop stray gemma4-only docstring
* revert output_capturing.py
* Add: slice_logits_to_keep to modeling_utils.py
* apply to `bit`, `falcon`, `hyperclovax`, `hyperclovax_vision_v2`, `nemotron_h`, `opt`
* Add: 2d idx process
* Address review feedback: https://github.com/huggingface/transformers/pull/48825#pullrequestreview-5573728421
* Address feedback: https://github.com/huggingface/transformers/pull/48825#issuecomment-6272236980
* Address feedback: https://github.com/huggingface/transformers/pull/48825#issuecomment-6272236980
* Update src/transformers/modeling_utils.py
Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com>
* Update src/transformers/modeling_utils.py
Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com>
* Apply slice_logits_to_keep to the models that were missed
* dont like the newline that was added
* more newline
* more
* remove more newline and nonsense assignment
* unrelated
* more based on latest main
---------
Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com>
Co-authored-by: Cyril Vallez <cyril.vallez@huggingface.co>