scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-12 23:01:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 87

fix: strip no-content chunks before stream_chunk_builder, keep usage-bearing ones

Browse source 
Remove chunks where choices==[] AND usage is absent — these cause stream_chunk_builder
to IndexError on chunks[-1]["choices"][0] and silently drop the partial billing record.
Usage-bearing empty-choices chunks (OpenAI stream_options include_usage final chunk)
are explicitly preserved so real provider token counts are not replaced by estimates.
This commit is contained in:
Sameer Kankute 2026-06-29 15:14:45 +05:30
parent 924a245edd
commit e4f66c01a0
No known key found for this signature in database
1 changed files with 16 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

16
litellm/proxy/common_request_processing.py
View file

@ -2465,6 +2465,22 @@ class ProxyBaseLLMRequestProcessing:
return
elif not isinstance(first_chunk, str) and not hasattr(first_chunk, "choices"):
return
# Remove pure no-content chunks (choices=[] and no usage payload) that can
# cause stream_chunk_builder to IndexError on chunks[-1]["choices"][0].
# Usage-bearing empty-choices chunks (OpenAI stream_options include_usage)
# are intentionally kept so real token counts are not replaced by estimates.
def _is_discardable_chunk(c: Any) -> bool:
if isinstance(c, dict):
return c.get("choices") == [] and not c.get("usage")
return (
hasattr(c, "choices")
and getattr(c, "choices", None) == []
and not getattr(c, "usage", None)
)
chunks = [c for c in chunks if not _is_discardable_chunk(c)]
if not chunks:
return
# Optimization, not a correctness guard: dispatch_success_handlers is the
# authoritative de-dup via has_dispatched_final_stream_success. This just
# skips the stream_chunk_builder assembly when completion already logged.