Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

The Feature

Add a public introspection surface that tells callers whether a given custom_llm_provider honors the api_base kwarg, or templates its URL internally from vertex_project + vertex_location / aws_region / etc.

Smallest viable shape:

litellm.provider_honors_api_base("vertex_ai")  # -> False
litellm.provider_honors_api_base("bedrock")    # -> False
litellm.provider_honors_api_base("ollama")     # -> True
litellm.provider_honors_api_base("openai")     # -> True

Or as an attribute on the LlmProviders enum.

Bigger but more useful alternatives:

1. Make vertex_ai's URL builder honor api_base as the host portion only, and still template the path ({api_base}/v1/projects/.../locations/.../models/...:generateContent). Then api_base passthrough would be safe across all providers.
2. At minimum, raise a typed error when api_base is passed to a provider that ignores it, instead of silently 404ing — e.g. LiteLLMUnsupportedKwarg("api_base", provider="vertex_ai").

Happy to send a PR for option 1 (the introspection function) if maintainers think it's the right shape.

Motivation, pitch

I'm building a wrapper around litellm.completion(...) that exposes a single "endpoint URL" config field to its users. The wrapper has no public way to know whether forwarding that URL as api_base is safe for the configured provider, and the failure mode for the wrong choice is silent and hard to diagnose.

Repro on litellm 1.83.12, Python 3.11:

from dotenv import load_dotenv; load_dotenv()
import litellm

# Works — SDK templates the URL from vertex_project + vertex_location.
ok = litellm.completion(
    model="vertex_ai/gemini-2.5-flash",
    messages=[{"role": "user", "content": "pong"}],
    vertex_project="my-project",
    vertex_location="global",
)
print(ok.choices[0].message.content)  # -> "pong"

# Fails — same args plus a valid-looking api_base. 404 with empty body.
litellm.completion(
    model="vertex_ai/gemini-2.5-flash",
    messages=[{"role": "user", "content": "pong"}],
    vertex_project="my-project",
    vertex_location="global",
    api_base="https://aiplatform.googleapis.com/",
)
# litellm.NotFoundError: Vertex_aiException -
# (status_code=404, body is empty)

The exception class is NotFoundError but the message and body are blank. A successful call's _hidden_params["api_base"] shows the canonical templated URL: global-aiplatform.googleapis.com/v1/projects/.../locations/global/publishers/google/models/gemini-2.5-flash:generateContent, which differs from the bare aiplatform.googleapis.com host the user would naturally declare.

What I tried for runtime detection (all dead ends):

1. litellm.get_llm_provider(model, api_base=sentinel)[3] — echoes the sentinel back unchanged for every provider; can't distinguish templated from override-friendly.
2. litellm.openai_compatible_providers — third-party OpenAI-compat list only; doesn't cover vertex_ai/bedrock/ollama/openai/anthropic.
3. ProviderConfigManager.get_provider_chat_config(model, provider).get_complete_url(...) — for vertex_ai this just echoes the input api_base back, because the real URL builder lives in litellm.llms.vertex_ai.vertex_llm_base._get_token_and_url, which is private.
4. Method-presence heuristics (get_api_base etc.) — ollama's config class lacks get_api_base but Ollama does honor api_base, so the signal is unreliable.

This is a real gap for anyone building configuration layers on top of litellm; right now we have to either hardcode a per-provider list or expose the routing detail to our own users.

What part of LiteLLM is this about?

SDK (litellm Python package)

LiteLLM is hiring a founding backend engineer, are you interested in joining us and shipping to all our users?

No

Twitter / LinkedIn details

No response