Check for existing issues

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

What happened?

LiteLLM's packaged model metadata for the Vertex Qwen model

vertex_ai/qwen/qwen3-235b-a22b-instruct-2507-maas

does not include us-south1 in supported_regions.

My LiteLLM proxy config uses this model in us-south1:

- model_name: qwen3-paygo
  litellm_params:
    model: vertex_ai/qwen/qwen3-235b-a22b-instruct-2507-maas
    vertex_project: os.environ/VERTEXAI_PROJECT
    vertex_location: us-south1
    vertex_credentials: os.environ/GOOGLE_APPLICATION_CREDENTIALS
    extra_headers:
      X-Vertex-AI-LLM-Request-Type: shared

- model_name: qwen3-pt
  litellm_params:
    model: vertex_ai/qwen/qwen3-235b-a22b-instruct-2507-maas
    vertex_project: os.environ/VERTEXAI_PROJECT
    vertex_location: us-south1
    vertex_credentials: os.environ/GOOGLE_APPLICATION_CREDENTIALS
    extra_headers:
      X-Vertex-AI-LLM-Request-Type: dedicated

To work around this, I had to patch LiteLLM's local model cost map by adding us-south1 to:

vertex_ai/qwen/qwen3-235b-a22b-instruct-2507-maas.supported_regions

and then mount that patched JSON over LiteLLM's packaged backup cost map.

Expected behavior: LiteLLM should ship the correct supported_regions metadata for this Vertex Qwen model so us-south1 works without a local cost-map override.

Steps to Reproduce

1. Start LiteLLM Proxy without any local cost-map override, with a config that routes this model to us-south1:

- model_name: qwen3-paygo
  litellm_params:
    model: vertex_ai/qwen/qwen3-235b-a22b-instruct-2507-maas
    vertex_project: os.environ/VERTEXAI_PROJECT
    vertex_location: us-south1
    vertex_credentials: os.environ/GOOGLE_APPLICATION_CREDENTIALS
    extra_headers:
      X-Vertex-AI-LLM-Request-Type: shared

- model_name: qwen3-pt
  litellm_params:
    model: vertex_ai/qwen/qwen3-235b-a22b-instruct-2507-maas
    vertex_project: os.environ/VERTEXAI_PROJECT
    vertex_location: us-south1
    vertex_credentials: os.environ/GOOGLE_APPLICATION_CREDENTIALS
    extra_headers:
      X-Vertex-AI-LLM-Request-Type: dedicated

1. Send a request through the proxy to one of those aliases.
2. Observe that LiteLLM warns that the model does not support us-south1 and routes to global instead.
3. Inspect the packaged LiteLLM cost map inside the image:

docker run --rm --entrypoint python ghcr.io/berriai/litellm:v1.81.9-stable -c 'import json; p="/usr/lib/python3.13/site-packages/litellm/model_prices_and_context_window_backup.json"; d=json.load(open(p)); print(d["ver tex_ai/qwen/qwen3-235b-a22b-instruct-2507-maas"].get("supported_regions"))'

This prints:

['global']

1. Apply a workaround by patching the cost map:

."vertex_ai/qwen/qwen3-235b-a22b-instruct-2507-maas".supported_regions |= ((. // []) + ["us-south1"] | unique)

1. Mount the patched JSON over LiteLLM's packaged backup cost map and enable LITELLM_LOCAL_MODEL_COST_MAP=True.

After that, the warning goes away and the model metadata shows ["global", "us-south1"].

Relevant log output

{"message": "Vertex AI model 'qwen/qwen3-235b-a22b-instruct-2507-maas' does not support region 'us-south1' (supported: ['global']). Routing to 'global'.", "level": "WARNING", "timestamp": "<redacted-timestamp>"}

Packaged LiteLLM metadata inside v1.81.9-stable:

['global']

Packaged LiteLLM metadata inside main-stable:

['global']

{ "vertex_ai/qwen/qwen3-235b-a22b-instruct-2507-maas": { "supported_regions": [ "global", "us-south1" ] } }

That makes the root cause much clearer:

• LiteLLM’s built-in metadata says only global
• LiteLLM logs a warning and reroutes to global
• adding us-south1 to the cost map fixes it

Relevant log output

litellm-proxy  | {"message": "Vertex AI model 'qwen/qwen3-235b-a22b-instruct-2507-maas' does not support region 'us-south1' (supported: ['global']). Routing to 'global'.", "level": "WARNING", "timestamp": "2026-04-08T20:03:23.481696"}

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

v1.82.3

Twitter / LinkedIn details

No response