fix high qps crash when enable mtp (#3592)
Co-authored-by: ispobock <ispobaoke@hotmail.com>
This commit is contained in:
@@ -263,7 +263,10 @@ class ForwardBatch:
|
||||
ret.extend_prefix_lens = torch.tensor(
|
||||
batch.extend_prefix_lens, dtype=torch.int32
|
||||
).to(device, non_blocking=True)
|
||||
if model_runner.server_args.attention_backend != "torch_native":
|
||||
if (
|
||||
model_runner.server_args.attention_backend != "torch_native"
|
||||
and model_runner.server_args.speculative_algorithm != "NEXTN"
|
||||
):
|
||||
ret.extend_num_tokens = batch.extend_num_tokens
|
||||
positions, ret.extend_start_loc = compute_position_triton(
|
||||
ret.extend_prefix_lens, ret.extend_seq_lens, ret.extend_num_tokens
|
||||
|
||||
Reference in New Issue
Block a user