Substantially reduce memory usage in _update_causal_mask for large batches by using .expand instead of .repeat #29413
try to fix gemma mem use
51c8efde
fix: handle attention mask dim==2 case
ac01cea0
remove logits=logits.float()
5cd54d0d
clean up + add llama
92d46824
nqgl
changed the title Substantially reduce memory usage in _update_causal_mask for large batches by using .expand instead of .repeat [WIP - need tests+sanity check] Substantially reduce memory usage in _update_causal_mask for large batches by using .expand instead of .repeat [needs tests+sanity check] 2 years ago
apply formatting
23935d95
readability edit: swap order of items being multiplied
7ed6b7f1
Merge branch 'huggingface:main' into causal_mask_fix
3420f9ab
revert change unrelated to PR
473f6cb4
revert black autoformat
7dc28df6
Merge pull request #2 from nqgl/remove_repeat_to
b2a1c99c
switch to one .to
654f6a5a
Merge pull request #3 from nqgl/remove_repeat_to
16753281
Accept style edits
0a907afe
nqgl
changed the title Substantially reduce memory usage in _update_causal_mask for large batches by using .expand instead of .repeat [needs tests+sanity check] Substantially reduce memory usage in _update_causal_mask for large batches by using .expand instead of .repeat 2 years ago
Assignees
No one assigned
Login to write a write a comment.
Login via GitHub