Summary

When LiteLLM proxies an OpenAI-compatible embedding endpoint serving a model whose name does not contain text-embedding-3 (e.g., a vLLM server hosting Qwen/Qwen3-Embedding-0.6B), passing dimensions in the request always raises UnsupportedParamsError regardless of drop_params. The raised error message instructs the user to set litellm.drop_params = True — but setting it has no effect on this code path. The only working escape hatch is additional_drop_params: ["dimensions"] per-model, which is undocumented for this scenario.

Version

litellm == 1.83.7 (also affects main at the time of filing). Running as a Docker proxy via ghcr.io/berriai/litellm.

Where the bug lives

litellm/utils.py :: get_optional_params_embeddings, in the elif custom_llm_provider == "openai": branch — roughly lines 3300–3310 in 1.83.7:

elif custom_llm_provider == "openai":
    # 'dimensions` is only supported in `text-embedding-3` and later models
    if (
        model is not None
        and "text-embedding-3" not in model
        and "dimensions" in non_default_params.keys()
        and "dimensions" not in (allowed_openai_params or [])
    ):
        raise UnsupportedParamsError(
            status_code=500,
            message="Setting dimensions is not supported for OpenAI `text-embedding-3` and later models. To drop it from the call, set `litellm.drop_params = True`.",
        )

This branch hard-raises without checking drop_params (per-call) or litellm.drop_params (global), even though both the inline comment and the error message direct the user to set drop_params. By contrast, _check_valid_arg defined just above in the same function does honor drop_params:

if litellm.drop_params is True or (drop_params is not None and drop_params is True):
    pass
else:
    raise UnsupportedParamsError(...)

Reproduction

Minimal config.yaml:

model_list:
  - model_name: qwen3-embedding-0.6b
    litellm_params:
      model: openai/Qwen/Qwen3-Embedding-0.6B
      api_base: http://<vllm-host>:8002/v1
      api_key: "EMPTY"
      drop_params: true        # has no effect
    model_info:
      mode: embedding

litellm_settings:
  drop_params: true            # also has no effect

Request:

curl -X POST http://127.0.0.1:8000/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3-embedding-0.6b","input":"hello","dimensions":1024}'

Response: HTTP 400

litellm.UnsupportedParamsError: Setting dimensions is not supported for OpenAI
`text-embedding-3` and later models. To drop it from the call, set
`litellm.drop_params = True`.

…even though litellm.drop_params is set (both globally in litellm_settings and per-model in litellm_params).

Workaround (for anyone hitting this via search)

Use additional_drop_params per-model — that one is consumed by embedding_pre_process_non_default_params before the dimensions check fires, so the param gets stripped cleanly:

litellm_params:
  model: openai/Qwen/Qwen3-Embedding-0.6B
  api_base: http://<vllm-host>:8002/v1
  api_key: "EMPTY"
  additional_drop_params: ["dimensions"]

Note: allowed_openai_params: ["dimensions"] also gets past the LiteLLM check, but then the request is forwarded to the backend, which (in vLLM's case for non-matryoshka models) returns its own 400 — so whitelisting is not equivalent to dropping.

Suggested fix

In the OpenAI-provider branch of get_optional_params_embeddings, when dimensions would otherwise be rejected, consult drop_params (per-call) and litellm.drop_params (global) the same way _check_valid_arg does in the same function — silently strip dimensions from non_default_params and continue, instead of raising. This makes the error message's advice actually work.

A test that exercises dimensions + a non-text-embedding-3 OpenAI-provider model + drop_params=True (expecting silent drop) would lock the behavior in.

Why it matters

This commonly hits anyone using LiteLLM as a translation layer between an OpenAI-shaped client and a self-hosted embedding server (vLLM, TEI, Ollama-compat, etc.) — exactly the use case the proxy is designed for. The error message also misleads operators into spending time on a config setting that doesn't fix it.