PR #28308: fix(embedding): normalize list encoding_format from model config to first element

• Repository: BerriAI/litellm
• Author: VANDRANKI
• State: open | merged: False
• Link: https://github.com/BerriAI/litellm/pull/28308

Description (problem / solution / changelog)

Root Cause

When encoding_format is set as a list in a model's litellm_params (e.g. to declare which formats the model supports):

model_list:
  - model_name: my-embeddings
    litellm_params:
      model: hosted_vllm/bge-m3
      api_base: http://localhost:8000
      encoding_format:
        - float
        - base64
        - ubyte
        - int8

the list is forwarded verbatim to the upstream embedding API. vLLM rejects it with validation errors because it expects a single string, not an array.

The proxy "Test Connection" flow and any request that does not explicitly set encoding_format both hit this path.

Fix

Normalize encoding_format at the top of embedding() in litellm/main.py, before get_optional_params_embeddings builds optional_params:

if isinstance(encoding_format, list):
    encoding_format = encoding_format[0] if encoding_format else None

• A non-empty list → use the first element as the concrete format (natural default).
• An empty list → None (no format specified, let the backend choose).
• A plain string or None → unchanged.

This matches the encoding_format: Optional[str] contract on the function signature and applies uniformly to all providers (hosted_vllm, openai, azure, etc.).

Test Plan

• litellm.embedding(model="hosted_vllm/...", input=["hello"], encoding_format=["float","base64"]) no longer raises
• Model configured with encoding_format: [float] in litellm_params: proxy request without explicit encoding_format succeeds
• Model configured with encoding_format: [float, base64]: explicit request encoding_format=base64 still uses base64
• encoding_format="float" (string) continues to work unchanged

Fixes #28239

Changed files

• litellm/main.py (modified, +2/-0)