Remove unnecessary kernels of num_token_non_padded (#6965)
This commit is contained in:
@@ -131,9 +131,6 @@ class TboCudaGraphRunnerPlugin:
|
||||
def replay_prepare(
|
||||
self, forward_mode: ForwardMode, bs: int, num_token_non_padded: int
|
||||
):
|
||||
if not global_server_args_dict["enable_two_batch_overlap"]:
|
||||
return
|
||||
|
||||
tbo_split_seq_index, tbo_split_token_index = (
|
||||
compute_split_indices_for_cuda_graph_replay(
|
||||
forward_mode=forward_mode,
|
||||
|
||||
Reference in New Issue
Block a user