[Performance] e2e overheads reduction: Small followup diff (#7364)
This commit is contained in:
@@ -336,9 +336,9 @@ class BlockSpaceManagerV1(BlockSpaceManager):
|
||||
|
||||
# Assign the self-attention block tables for each sequence.
|
||||
if len(wait_seqs) == 1:
|
||||
self.block_tables[wait_seqs[0].seq_id] = block_table
|
||||
self.block_tables[seq.seq_id] = block_table
|
||||
else:
|
||||
for seq in seq_group.get_seqs(status=SequenceStatus.WAITING):
|
||||
for seq in wait_seqs:
|
||||
self.block_tables[seq.seq_id] = block_table.copy()
|
||||
|
||||
# Allocate encoder sequence
|
||||
|
||||
Reference in New Issue
Block a user