Skip to content

[generate] Fix vocab_size access for multimodal models #37937

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
gante merged 2 commits into from
May 5, 2025
Merged

[generate] Fix vocab_size access for multimodal models #37937

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6b8939a
Implements last migrations for generation from `config.vocab_size` to…
May 3, 2025
40b5c0c
Merge branch 'main' into mm-generation-vocab-size
May 5, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 6 additions & 4 deletions src/transformers/generation/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -968,7 +968,7 @@ def _get_candidate_generator(
atm_translator = AssistantVocabTranslatorCache.get_translator(
target_tokenizer,
assistant_tokenizer,
self.config.vocab_size,
self.config.get_text_config().vocab_size,
assistant_model=assistant_model,
assistant_prune_lm_head=True, # prune LM head of assistant model
)
Expand Down Expand Up @@ -1234,7 +1234,9 @@ def _get_logits_processor(
# Watermarking should be after all logits processing is finished (see #34630)
if generation_config.watermarking_config is not None:
processors.append(
generation_config.watermarking_config.construct_processor(self.config.vocab_size, device)
generation_config.watermarking_config.construct_processor(
self.config.get_text_config().vocab_size, device
)
)

# `LogitNormalization` should always be the last logit processor, when present
Expand Down Expand Up @@ -1412,7 +1414,7 @@ def compute_transition_scores(

# 3. Optionally normalize the logits (across the vocab dimension)
if normalize_logits:
scores = scores.reshape(-1, self.config.vocab_size, scores.shape[-1])
scores = scores.reshape(-1, self.config.get_text_config().vocab_size, scores.shape[-1])
scores = torch.nn.functional.log_softmax(scores, dim=1)
scores = scores.reshape(-1, scores.shape[-1])

Expand All @@ -1426,7 +1428,7 @@ def compute_transition_scores(
beam_indices[beam_indices_mask] = 0

# 6. multiply beam_indices with vocab size to gather correctly from scores
beam_sequence_indices = beam_indices * self.config.vocab_size
beam_sequence_indices = beam_indices * self.config.get_text_config().vocab_size

# 7. Define which indices contributed to scores
cut_idx = sequences.shape[-1] - max_beam_length
Expand Down