### Check for existing issues

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

### What happened?

# LiteLLM proxy fails for `qwen3-vl-embedding-8b` with image inputs

## Summary

When proxying the multimodal embedding model `qwen3-vl-embedding-8b` through LiteLLM to a vLLM backend, image inputs are either rejected outright or cause an internal 500. The same requests work fine when sent **directly to vLLM**, which forces us to bypass the LiteLLM proxy for any image-embedding workloads.

## Environment

- vLLM serves `qwen3-vl-embedding-8b` with an OpenAI-compatible `/v1/embeddings` route. It also accepts vLLM-native `/v1/embeddings` requests shaped like a chat completion (`messages: [{role: "user", content: [{type:"image_url", image_url:{url:...}}]}]`). - LiteLLM proxy is fronting vLLM (env vars in our `.env`):

``` OPENAI_API_BASE=http://60.51.1.128:9999/v1 # LiteLLM proxy OPENAI_API_KEY=sk-... CHAT_MODEL=qwen3.5-397b INSTRUCT_MODEL=qwen3.5-397b-instruct EMBEDDING_OPENAI_API_BASE=http://10.0.151.2:14701/v1 # vLLM direct EMBEDDING_MODEL=qwen3-vl-embedding-8b ```

The proxy is reached on the default OpenAI por…