scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-10-07 02:59:05 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 103

fix(deepseek): aggregate reasoning_content placeholder warning to one per request (#37629)

Browse source
This commit is contained in:
OSS Agent 2026-08-20 18:55:05 +02:00 • committed by pacocartones
parent 08b94986fc
commit 4c3f0306f2
1 changed files with 21 additions and 11 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

32
litellm/llms/deepseek/chat/transformation.py
View file

@ -72,7 +72,8 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
(LiteLLM stores provider-specific response fields there).
2. Otherwise inject a single space — the minimum value the API accepts.
"""
result: Final[list[AllMessageValues]] = []
result: list[AllMessageValues] = []
placeholder_injections = 0
for msg in messages:
if msg.get("role") == "assistant" and not msg.get("reasoning_content"):
patched = dict(cast(dict, msg))
@ -84,20 +85,29 @@ class DeepSeekChatConfig(OpenAIGPTConfig):
cleaned.pop("reasoning_content", None)
patched["provider_specific_fields"] = cleaned
else:
litellm.verbose_logger.warning(
"DeepSeek thinking mode: assistant message is missing "
"`reasoning_content` and none was saved in "
"`provider_specific_fields`. A single-space placeholder "
"is being injected to satisfy API validation, but the "
"model will receive a blank reasoning chain for this turn, "
"which may silently degrade multi-turn response quality. "
"Preserve `reasoning_content` from the original assistant "
"response when building multi-turn conversation history."
)
placeholder_injections += 1
patched["reasoning_content"] = " "
result.append(cast(AllMessageValues, patched))
else:
result.append(msg)
# Emit a single aggregated warning per request instead of one per
# affected message. The old per-message warning produced an identical
# line for every historical assistant turn (issue #37629): a 6-message
# conversation logged 6 copies on every request, drowning out real
# errors. Aggregating keeps the diagnostic while reporting the count.
if placeholder_injections:
litellm.verbose_logger.warning(
"DeepSeek thinking mode: %d assistant message(s) were missing "
"`reasoning_content` and none was saved in "
"`provider_specific_fields`. A single-space placeholder was "
"injected for each to satisfy API validation, but the model "
"will receive a blank reasoning chain for those turns, which "
"may silently degrade multi-turn response quality. Preserve "
"`reasoning_content` from the original assistant response when "
"building multi-turn conversation history.",
placeholder_injections,
)
return result
@overload