chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

fix(grpc): handle embedding requests in _handle_batch_output (#19221)

Browse Source 
Signed-off-by: Simo Lin <linsimo.mark@gmail.com>
This commit is contained in:
Simo Lin
2026-02-23 22:32:16 -08:00
committed by GitHub
parent 2e053d6eb6
commit 88a7e48108
1 changed files with 5 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/grpc/grpc_request_manager.py
View File

@@ -659,12 +659,14 @@ class GrpcRequestManager:
}
# Accumulate logprobs (following tokenizer_manager pattern)
if state.obj.return_logprob:
# Use getattr for safe access - not all request types have return_logprob
# (e.g., TokenizedEmbeddingReqInput)
if getattr(state.obj, "return_logprob", False):
self._convert_logprob_style(state, batch_out, i)
# Send input logprobs based if available
if (
state.obj.return_logprob
getattr(state.obj, "return_logprob", False)
and state.obj.logprob_start_len >= 0
and state.input_token_logprobs_val
):
@@ -688,7 +690,7 @@ class GrpcRequestManager:
# Send output logprobs if available
if (
state.obj.return_logprob
getattr(state.obj, "return_logprob", False)
and batch_out.output_token_logprobs_val
and i < len(batch_out.output_token_logprobs_val)
):