Enable evict swa with piecewise cuda graph (#21754)
This commit is contained in:
@@ -2383,13 +2383,6 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
|
||||
sliding_window_size = self.tree_cache.sliding_window_size
|
||||
server_args = get_global_server_args()
|
||||
|
||||
if (
|
||||
self.forward_mode.is_decode()
|
||||
and not server_args.disable_piecewise_cuda_graph
|
||||
and not self.tree_cache.is_chunk_cache()
|
||||
):
|
||||
return
|
||||
|
||||
for idx, req in enumerate(self.reqs):
|
||||
if self.forward_mode.is_decode():
|
||||
# We set evict_swa condition here with two reasons:
|
||||
|
||||
Reference in New Issue
Block a user