Skip to content

mask losses for sequence_first #207

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
merged 1 commit into from
Mar 27, 2025
Merged

mask losses for sequence_first #207

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
25 changes: 16 additions & 9 deletions fast_llm/models/gpt/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -249,21 +249,28 @@ def preprocess(
TransformerKwargs.presents: presents,
}
if phase != PhaseType.inference:
sequence_offset = sequence_k - sequence_q + 1
if sequence_first:
labels = batch.token_ids[sequence_k - sequence_q + 1 : sequence_k + 1]
labels = batch.token_ids[sequence_offset : sequence_k + 1]
else:
# TODO: Avoid multiple contiguous calls?
labels = batch.token_ids[:, sequence_k - sequence_q + 1 : sequence_k + 1].contiguous()
# We set label indices to -100 for masked spans, inline with ignore_index in torch.nn.CrossEntropyLoss
# TODO: take ignore_index from config
if batch.loss_masking_spans is not None:
for i, spans_i in enumerate(batch.loss_masking_spans):
mask_indices = (
torch.cat([torch.arange(s - 1, e) for s, e in spans_i])
if len(spans_i)
else torch.tensor([], dtype=torch.int64)
)
labels[i, mask_indices] = -100
if batch.loss_masking_spans is not None:
for i, spans in enumerate(batch.loss_masking_spans):
if not spans.numel():
continue
valid_spans = spans[(spans[:, 0] <= sequence_k) & (spans[:, 1] >= sequence_offset)]
if valid_spans.numel():
valid_spans[:, 0].clamp_(min=sequence_offset)
valid_spans[:, 1].clamp_(max=sequence_k)
valid_spans -= sequence_offset
for start, end in valid_spans:
if sequence_first:
labels[start : end + 1, i] = -100
else:
labels[i, start : end + 1] = -100
kwargs[LanguageModelKwargs.labels] = labels
if self._config.use_absolute_position_embeddings:
self._position_embedding_preprocessor.preprocess(kwargs)
Expand Down