How to make litellm proxy return the headers from llm provider?

[lyajedi]

according to https://docs.litellm.ai/docs/proxy/response_headers#response-headers-from-llm-providers

LiteLLM also returns the original response headers from the LLM provider. These headers are prefixed with llm_provider- to distinguish them from LiteLLM's headers.

However, I never see any llm_provider- prefix headers in response, the provider returns lots of additional headers like with x- in it, but they are all removed in the proxy response.

[Ultroncn]

An upstream header starting with x- should not, by itself, prevent the llm_provider- prefix. The current header helper prefixes the headers passed into it, including arbitrary x-... names. The important question is whether your provider adapter/route passes those headers through to that point.

I would narrow it down in this order:

1. Make a non-streaming request directly to the LiteLLM proxy and inspect raw HTTP headers (curl -D - ...). Use the same configured model as the failing request and avoid a response-cache hit while comparing.
2. Compare the actual upstream response with the proxy response. If the upstream has x-example but the proxy has no llm_provider-x-example, record the LiteLLM image tag, provider, route, and streaming flag. That isolates adapter/version handling from the generic prefix helper.
3. If curl sees the header but browser JavaScript does not, inspect CORS Access-Control-Expose-Headers. A header can be on the wire and visible in DevTools while still being unavailable through browser fetch().headers.
4. If an SDK is hiding the HTTP layer, inspect its raw-response interface rather than the parsed completion object; HTTP headers are not completion JSON fields.

The documentation describes the intended forwarding behavior, but that does not prove every provider/streaming route in an older release implemented it identically. I have not reproduced your provider/image, so I would not attribute the missing headers to one configuration flag without that comparison.