minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

Fix spec v2 + logprob when max_num_token is set (#20799)

Browse Source
This commit is contained in:
Qiaolin Yu
2026-04-02 01:55:16 -07:00
committed by GitHub
parent e55a35fbcd
commit b684b0b72f
1 changed files with 10 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/managers/scheduler_output_processor_mixin.py
+10 -9
View File
@@ -1034,6 +1034,8 @@ class SchedulerOutputProcessorMixin:
and not req.input_logprob_sent and not req.input_logprob_sent
# Decode server does not send input logprobs # Decode server does not send input logprobs
and self.disaggregation_mode != DisaggregationMode.DECODE and self.disaggregation_mode != DisaggregationMode.DECODE
# Only send when input logprobs have been computed (after prefill)
and req.input_token_logprobs_val is not None
): ):
input_token_logprobs_val.append(req.input_token_logprobs_val) input_token_logprobs_val.append(req.input_token_logprobs_val)
input_token_logprobs_idx.append(req.input_token_logprobs_idx) input_token_logprobs_idx.append(req.input_token_logprobs_idx)
@@ -1055,39 +1057,38 @@ class SchedulerOutputProcessorMixin:
input_token_ids_logprobs_idx.append([]) input_token_ids_logprobs_idx.append([])
if req.return_logprob: if req.return_logprob:
logprob_end = max(len(output_ids_), 1)
output_token_logprobs_val.append( output_token_logprobs_val.append(
req.output_token_logprobs_val[ req.output_token_logprobs_val[
send_output_token_logprobs_offset: send_output_token_logprobs_offset:logprob_end
] ]
) )
output_token_logprobs_idx.append( output_token_logprobs_idx.append(
req.output_token_logprobs_idx[ req.output_token_logprobs_idx[
send_output_token_logprobs_offset: send_output_token_logprobs_offset:logprob_end
] ]
) )
output_top_logprobs_val.append( output_top_logprobs_val.append(
req.output_top_logprobs_val[ req.output_top_logprobs_val[
send_output_token_logprobs_offset: send_output_token_logprobs_offset:logprob_end
] ]
) )
output_top_logprobs_idx.append( output_top_logprobs_idx.append(
req.output_top_logprobs_idx[ req.output_top_logprobs_idx[
send_output_token_logprobs_offset: send_output_token_logprobs_offset:logprob_end
] ]
) )
output_token_ids_logprobs_val.append( output_token_ids_logprobs_val.append(
req.output_token_ids_logprobs_val[ req.output_token_ids_logprobs_val[
send_output_token_logprobs_offset: send_output_token_logprobs_offset:logprob_end
] ]
) )
output_token_ids_logprobs_idx.append( output_token_ids_logprobs_idx.append(
req.output_token_ids_logprobs_idx[ req.output_token_ids_logprobs_idx[
send_output_token_logprobs_offset: send_output_token_logprobs_offset:logprob_end
] ]
) )
req.send_output_token_logprobs_offset = len( req.send_output_token_logprobs_offset = logprob_end
req.output_token_logprobs_val
)
else: else:
output_token_logprobs_val.append([]) output_token_logprobs_val.append([])
output_token_logprobs_idx.append([]) output_token_logprobs_idx.append([])