Add overlap scheduling for embeddings code path (#14032)

Co-authored-by: Satyam Kumar <satyamk@linkedin.com>
This commit is contained in:
satyamk7054
2025-12-24 18:24:18 -08:00
committed by GitHub
parent ecd2d09a6c
commit 38dd4fbb66
6 changed files with 63 additions and 19 deletions

View File

@@ -1996,7 +1996,10 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
self.orig_seq_lens = self.orig_seq_lens[keep_indices_device]
self.out_cache_loc = None
self.seq_lens_sum = self.seq_lens.sum().item()
self.output_ids = self.output_ids[keep_indices_device]
if self.output_ids is not None:
self.output_ids = self.output_ids[keep_indices_device]
self.mamba_track_indices = None
self.mamba_track_mask = None
self.mamba_track_seqlens = None