Fix get_model_from_request() to extract model ID from Vertex AI passthrough URLs

[krisxia0506]

Title

Fix get_model_from_request() to extract model ID from Vertex AI passthrough URLs

Relevant issues

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/litellm/ directory, Adding at least 1 test is a hard requirement - see details
• I have added a screenshot of my new test passing locally
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem

Type

🐛 Bug Fix

Changes

Summary

This PR enhances the model extraction logic in the proxy authentication module to properly handle Vertex AI passthrough routes. Previously, the get_model_from_request() function could not extract model IDs from Vertex AI passthrough URLs, which follows a distinct pattern compared to OpenAI deployments.

Problem

Vertex AI passthrough endpoints have a URL structure like:

/vertex_ai/v1/projects/{project}/locations/{location}/publishers/google/models/{model_id}:{method}

The existing model extraction logic only handled OpenAI deployments (/openai/deployments/{model}/), leaving Vertex AI models unidentified.

Solution

Added regex-based model extraction for Vertex AI routes:

• Pattern: /vertex_ai/.../models/{model_id}:*
• Extracts model ID using the regex pattern: /models/([^/:]+)
• Handles both versioned endpoints (v1, v1beta1) and different Google Cloud regions
• Maintains precedence: request body model > extracted from URL
• Does not interfere with existing OpenAI deployment extraction

Testing

Added comprehensive test coverage with 7 parametrized test cases in test_get_model_from_request_vertex_ai_passthrough():

1. ✅ Vertex AI v1 endpoint with generateContent method
2. ✅ Vertex AI v1beta1 endpoint with streamGenerateContent method
3. ✅ Different region locations (us-central1, asia-southeast1)
4. ✅ Model extraction without method suffix (no colon)
5. ✅ Request body model takes precedence over URL pattern
6. ✅ Non-Vertex routes are not affected
7. ✅ Azure deployment pattern still works as expected

Files Changed

• litellm/proxy/auth/auth_utils.py: Added Vertex AI model extraction logic
• tests/local_testing/test_auth_utils.py: Added parametrized test cases
• tests/test_litellm/proxy/pass_through_endpoints/test_vertex_passthrough_auth.py: New test file with additional Vertex AI passthrough tests

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Review	Updated (UTC)
litellm	Ready	Preview, Comment	Dec 15, 2025 5:17am

[Sameerlite]

@krisxia0506 Reverted this PR. Can you make sure your tests pass. All o them are failing

[krisxia0506]

@krisxia0506 Reverted this PR. Can you make sure your tests pass. All o them are failing

I have run both test files in this PR successfully and attached screenshots in the PR description. Could you please let me know which tests did not pass and specify them clearly?

[Sameerlite]

[krisxia0506]

Thank you sir, I accidentally integrated another method from a PR #17526 that hasn't been merged yet into my unit test method get_vertex_model_id_from_url, I will remove it.

[krisxia0506]

@Sameerlite Hello, sir, I have resubmitted my changes in a new PR #18097. Please help me review them.