Fix core (MI300X) with --enable-overlap (#2048)

This commit is contained in:
HAI
2024-11-15 21:24:42 -08:00
committed by GitHub
parent 023d0a73df
commit e5c6715003

View File

@@ -588,14 +588,14 @@ class ScheduleBatch:
) )
if not decoder_out_cache_loc: if not decoder_out_cache_loc:
self.out_cache_loc = torch.empty(0, dtype=torch.int32).to( self.out_cache_loc = torch.zeros(0, dtype=torch.int32).to(
self.device, non_blocking=True self.device, non_blocking=True
) )
else: else:
self.out_cache_loc = torch.cat(decoder_out_cache_loc) self.out_cache_loc = torch.cat(decoder_out_cache_loc)
if not encoder_out_cache_loc: if not encoder_out_cache_loc:
self.encoder_out_cache_loc = torch.empty(0, dtype=torch.int32).to( self.encoder_out_cache_loc = torch.zeros(0, dtype=torch.int32).to(
self.device, non_blocking=True self.device, non_blocking=True
) )
else: else: