[AMD] Work around HIP TPOT regression from Event.wait() in MTP seq lens resolution (#26672)
This commit is contained in:
@@ -204,6 +204,10 @@ class FutureMap:
|
||||
if fi is None:
|
||||
return
|
||||
if self.publish_ready is not None:
|
||||
if _is_hip:
|
||||
# Temporary workaround: Event.wait() regresses TPOT on AMD MI355.
|
||||
self.publish_ready.synchronize()
|
||||
else:
|
||||
self.publish_ready.wait()
|
||||
batch.seq_lens = self.new_seq_lens_buf[fi]
|
||||
|
||||
|
||||
Reference in New Issue
Block a user