Description

When using litellm.aembedding() with a hosted_vllm provider (e.g., hosted_vllm/qwen/qwen3-embedding-8b via OpenRouter), the dimensions parameter is always passed through to the API, even when drop_params=True. This causes 422 errors from providers that don't support dimensions.

Root Cause

In litellm/utils.py, get_optional_params_embeddings() has an elif custom_llm_provider in openai_compatible_providers branch that passes all parameters through without validation or drop_params handling.

The OpenAI-specific branch correctly limits dimensions to text-embedding-3-* models, but this logic is not applied to openai_compatible_providers (which includes hosted_vllm).

Call chain:

1. get_optional_params_embeddings() — hosted_vllm hits the openai_compatible_providers branch, all params pass through unconditionally
2. litellm/llms/openai_like/embedding/handler.py L108 — data = {"model": model, "input": input, **optional_params} includes unsupported dimensions
3. Provider returns 422

Steps to Reproduce

import litellm

response = await litellm.aembedding(
    model="hosted_vllm/qwen/qwen3-embedding-8b",
    input=["Hello world"],
    dimensions=4096,
    api_base="https://openrouter.ai/api/v1",
    api_key="...",
    drop_params=True,  # has no effect
)
# Expected: dimensions is dropped, embedding succeeds
# Actual: 422 Unprocessable Entity

Environment

• litellm version: 1.80+
• Python: 3.12
• Provider: OpenRouter (but affects any hosted_vllm endpoint that doesn't support dimensions)

Suggested Fix

Add dimensions validation to the openai_compatible_providers branch in get_optional_params_embeddings():

elif custom_llm_provider in openai_compatible_providers:
    optional_params = non_default_params.copy()
    if "dimensions" in optional_params:
        model_name = model or ""
        if "text-embedding-3" not in model_name.lower():
            if litellm.drop_params or drop_params:
                optional_params.pop("dimensions")
            else:
                raise UnsupportedParamsError(...)

Relevant files:

• litellm/utils.py — get_optional_params_embeddings() (~L2937)
• litellm/constants.py — openai_compatible_providers list
• litellm/llms/openai_like/embedding/handler.py — request body construction