Fix potential eos bug on decode instance when PD is enabled (#12206)

Signed-off-by: Shangming Cai <csmthu@gmail.com>
This commit is contained in:
Shangming Cai
2025-10-28 16:29:02 +08:00
committed by GitHub
parent a1f2dc90e4
commit 41a113356a

View File

@@ -773,8 +773,12 @@ class DecodeTransferQueue:
indices_to_remove.add(i)
decode_req.req.time_stats.wait_queue_entry_time = time.perf_counter()
# special handling for sampling_params.max_new_tokens == 1
if decode_req.req.sampling_params.max_new_tokens == 1:
# special handling for corner cases
should_finish = (
decode_req.req.sampling_params.max_new_tokens == 1
or output_id in decode_req.req.eos_token_ids
)
if should_finish:
# finish immediately
decode_req.req.time_stats.forward_entry_time = (
decode_req.req.time_stats.completion_time