transformers
perf: avoid recomputing rotary_emb for each layer in some Google and ModernBERT models
#45555
Merged
Merged
perf: avoid recomputing rotary_emb for each layer in some Google and ModernBERT models #45555
vasqu
merged
2 commits into
huggingface:main
from
casinca:gemma3-optim
perf: avoid redundant rotary_emb calls for `position_embeddings`
67620c54
vasqu
approved these changes
on 2026-04-21
perf: avoid redundant rotary_emb calls for `position_embeddings` dict
31bb4ddc
casinca
changed the title
zucchini-nlp
approved these changes
on 2026-04-22
Login to write a write a comment.
Login via GitHub