Gemini provider: model names not prefixed for LiteLLM Gemini API routing

[Jos-Jerus]

Description

When using the remote::gemini provider with a Gemini API key (GEMINI_API_KEY), LiteLLM auto-detects model names like gemini-2.5-flash as Vertex AI models and attempts to use Application Default Credentials instead of the API key.

Expected Behavior

When GEMINI_API_KEY is set, the Gemini provider should route requests to the Gemini API (Google AI Studio), not Vertex AI.

Actual Behavior

LiteLLM receives gemini-2.5-flash and auto-detects it as a Vertex AI model, failing with:

litellm.APIConnectionError: Your default credentials were not found. 
To set up Application Default Credentials, see https://cloud.google.com/docs/authentication/external/set-up-adc

Root Cause

LiteLLM uses model name prefixes to determine routing:

• gemini/gemini-2.5-flash → Gemini API (uses GEMINI_API_KEY)
• gemini-2.5-flash (no prefix) → auto-detected as Vertex AI (uses ADC)

The Gemini provider passes model names without the gemini/ prefix to LiteLLM.

Suggested Fix

When the Gemini provider is configured with an API key (not Vertex AI), it should prefix model names with gemini/ before passing to LiteLLM:

# In the Gemini provider's inference call
if self.api_key:
    litellm_model = f"gemini/{model_id}"
else:
    litellm_model = model_id  # Vertex AI path

Workaround

Currently working around this by injecting a sitecustomize.py that sets litellm.model_alias_map:

import litellm
litellm.model_alias_map = {
    'gemini-2.5-flash': 'gemini/gemini-2.5-flash',
    'gemini-2.5-pro': 'gemini/gemini-2.5-pro',
    # ... other models
}

Environment

• llama-stack version: 0.3.x (via lightspeed-stack 0.3.1)
• LiteLLM version: latest bundled
• Python: 3.12

[mattf]

cc @major

[major]

@mattf I can have a look if you want, but this seems to be a gemini issue whereas I was working with the vertexai provider.

[major]

I think this might have fixed by @mattf with the OpenAIMixin work back in October: d23ed2623

[mattf]

@major it's also a red hat internal version: 0.3.x (via lightspeed-stack 0.3.1)