From aea9678f6198002fc6f914733320d921a5aff558 Mon Sep 17 00:00:00 2001 From: kerry Date: Wed, 16 Sep 2026 23:28:28 +0000 Subject: [PATCH] refactor(streaming): compute the reasoning token cap once Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- .../litellm_core_utils/streaming_chunk_builder_utils.py | 9 +++------ 1 file changed, 3 insertions(+), 6 deletions(-) diff --git a/litellm/litellm_core_utils/streaming_chunk_builder_utils.py b/litellm/litellm_core_utils/streaming_chunk_builder_utils.py index 9bb2d489f83..0ca93fe08b3 100644 --- a/litellm/litellm_core_utils/streaming_chunk_builder_utils.py +++ b/litellm/litellm_core_utils/streaming_chunk_builder_utils.py @@ -1095,8 +1095,8 @@ class ChunkProcessor: returned_usage.completion_tokens_details = completion_tokens_details if reasoning_tokens is not None: + capped_reasoning_tokens: Final = min(max(0, reasoning_tokens), returned_usage.completion_tokens) if returned_usage.completion_tokens_details is None: - capped_reasoning_tokens: Final = min(max(0, reasoning_tokens), returned_usage.completion_tokens) returned_usage.completion_tokens_details = CompletionTokensDetailsWrapper( reasoning_tokens=capped_reasoning_tokens, text_tokens=returned_usage.completion_tokens - capped_reasoning_tokens, @@ -1105,13 +1105,10 @@ class ChunkProcessor: returned_usage.completion_tokens_details is not None and returned_usage.completion_tokens_details.reasoning_tokens is None ): - existing_capped_reasoning_tokens: Final = min( - max(0, reasoning_tokens), returned_usage.completion_tokens - ) - returned_usage.completion_tokens_details.reasoning_tokens = existing_capped_reasoning_tokens + returned_usage.completion_tokens_details.reasoning_tokens = capped_reasoning_tokens if returned_usage.completion_tokens_details.text_tokens is None: returned_usage.completion_tokens_details.text_tokens = ( - returned_usage.completion_tokens - existing_capped_reasoning_tokens + returned_usage.completion_tokens - capped_reasoning_tokens ) if prompt_tokens_details is not None: returned_usage.prompt_tokens_details = prompt_tokens_details