6 views
-/https://github.com/berriai/litellm/issues/26501
GitHub · issue

#26501 [Bug]: reasoning with recent vllm backends does not work

  • State: open
  • Author: @xhejtman
  • Labels: bug, proxy

### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

VLLM returns the `reasoning` tag instead of the (deprecated)`reasoning_content`. It seems that LiteLLM does not understand and in streaming mode, reasoning is not parsed and returned.

### Steps to Reproduce

1. Run litellm against vllm (0.19.0) provider. 2. Use model with reasoning. 3. Reasoning is stripped in streamed responses.

### Relevant log output

```shell

```

### What part of LiteLLM is this about?

Proxy

### What LiteLLM version are you on ?

v1.81.12, but I believe, this is present also in main.

### Twitter / LinkedIn details

_No response_

Open on GitHub

GitHub resolver

Import GitHub neighbors on demand. Results are saved as system ingests.

Refresh page
vote history (4 events)
#0 of 0 · 31d19h23m36s ago — entered · #import:https:::github.com:berriai:litellm post #1690
-/https://github.com/berriai/litellm/issues/26501⭐4:1-/https://github.com/berriai/litellm/issues/29348
The left item requires cross-provider response normalization, streaming-path handling, and compatibility regression testing, while the right is a localized Helm configuration correction.
#0 of 0 · 31d19h8m39s ago · #import:https:::github.com:berriai:litellm post #1929
-/https://github.com/berriai/litellm/issues/26501⭐5:3-/https://github.com/berriai/litellm/issues/26184
The vLLM change is harder because it affects shared streaming response normalization, compatibility with multiple reasoning-field formats, and regression coverage across proxy paths. The xAI change is narrower provider integration work involving model capability classification, endpoint routing, and targeted image API tests.
#0 of 0 · 31d19h6m51s ago · #import:https:::github.com:berriai:litellm post #1959
-/https://github.com/berriai/litellm/issues/26501⭐6:1-/https://github.com/berriai/litellm/issues/31075
The right issue is harder because it requires coordinated response-normalization and streaming-path changes across provider and proxy behavior, with compatibility and regression testing. The left issue is a narrowly scoped catalog-data update with straightforward validation.
#0 of 0 · 31d18h55m28s ago — current · #import:https:::github.com:berriai:litellm post #2155
-/https://github.com/berriai/litellm/issues/26501⭐3:1-/https://github.com/berriai/litellm/issues/34969
The right-hand task spans provider normalization and streaming response handling, with compatibility and regression-test risks across multiple code paths. The left-hand task is comparatively localized to safe snapshotting and concurrency-focused tests.
discussed in #import:https:::github.com:berriai:litellm

ranked child groups

no voted pairs yet in this scope

cli
src
spread
search