[perf] tiny optimize select_index op for draft extend (#29078)
This commit is contained in:
@@ -821,8 +821,12 @@ class EagleDraftWorker(EagleDraftWorkerBase):
|
||||
num_tokens_for_logprob_per_req=self.speculative_num_draft_tokens,
|
||||
)
|
||||
select_index = (
|
||||
torch.arange(len(batch.seq_lens), device=self.device)
|
||||
* self.speculative_num_draft_tokens
|
||||
torch.arange(
|
||||
0,
|
||||
len(batch.seq_lens) * self.speculative_num_draft_tokens,
|
||||
self.speculative_num_draft_tokens,
|
||||
device=self.device,
|
||||
)
|
||||
+ batch_result.accept_lens
|
||||
- 1
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user