[Feature] PD-Multiplexing Context and Scheduler. (#11592)

This commit is contained in:
ykcombat
2025-10-28 00:54:43 +08:00
committed by GitHub
parent 0cae873fcd
commit 05ad28f25e
9 changed files with 464 additions and 24 deletions

View File

@@ -135,10 +135,7 @@ class LogitsMetadata:
@classmethod
def from_forward_batch(cls, forward_batch: ForwardBatch):
if (
(
forward_batch.forward_mode.is_extend()
or forward_batch.forward_mode.is_split_prefill()
)
forward_batch.forward_mode.is_extend()
and forward_batch.return_logprob
and not forward_batch.forward_mode.is_target_verify()
):
@@ -389,8 +386,8 @@ class LogitsProcessor(nn.Module):
input_logprob_indices = None
elif (
logits_metadata.forward_mode.is_extend()
or logits_metadata.forward_mode.is_split_prefill()
) and not logits_metadata.extend_return_logprob:
and not logits_metadata.extend_return_logprob
):
# Prefill without input logprobs.
if logits_metadata.padded_static_len < 0:
last_index = torch.cumsum(logits_metadata.extend_seq_lens, dim=0) - 1