[Perf] DFLASH: build the draft num_token_non_padded scalar with torch.full (#44119)
This commit is contained in:
@@ -2599,7 +2599,7 @@ class DFlashWorkerV2(BaseSpecWorker):
|
||||
spec_info=self._draft_block_spec_info,
|
||||
capture_hidden_mode=CaptureHiddenMode.NULL,
|
||||
num_token_non_padded=(
|
||||
torch.tensor(bs * block_size, dtype=torch.int32, device=device)
|
||||
torch.full((), bs * block_size, dtype=torch.int32, device=device)
|
||||
if enable_num_token_non_padded()
|
||||
else None
|
||||
),
|
||||
|
||||
Reference in New Issue
Block a user