[litellm-agent] Staging → litellm_internal_staging (5/17/2026)

[oss-pr-review-agent-shin]

Automated staging PR created by litellm-agent.

This branch collects PRs approved by the agent on 5/17/2026.

⚠️ Human review required before CI. Convert from draft to ready when you've reviewed the diff.

[oss-pr-review-agent-shin]

@greptile please review

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 3 committers have signed the CLA.

✅ yuneng-berri
❌ voidborne-d
❌ shin-berri
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

This PR fixes issue #28084 by removing a spurious import vertexai gate from VertexAIPartnerModels.count_tokens. Partner model token counting (Claude, Mistral, Llama on Vertex AI) routes through VertexAIPartnerModelsTokenCounter, which talks to the :rawPredict endpoint over plain httpx and has no dependency on the Gemini SDK — the gate was an accidental hard dependency that broke /v1/messages/count_tokens for any LiteLLM installation without google-cloud-aiplatform>=1.38.

• main.py: Deletes ~12 lines of vertexai import/version-check boilerplate and replaces them with a clear explanatory comment; the rest of count_tokens is unchanged.
• New test: Two fully-mocked regression tests pin the absence of the gate — one simulates vertexai being unimportable and verifies normal execution continues, the other asserts the handler module itself never pulls the Gemini SDK into sys.modules.

Confidence Score: 5/5

Safe to merge — the change removes dead code that was incorrectly gating a network-only path, and the new tests mock all I/O correctly with no real network calls.

The diff is small and surgical: it deletes an import check that was never needed by the actual token-counting implementation and adds two well-structured regression tests that fully stub the auth and HTTP layers. The handler itself is unchanged, the rest of count_tokens is unchanged, and all tests respect the no-real-network-calls constraint for this test folder.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/vertex_ai/vertex_ai_partner_models/main.py	Removes the now-unnecessary vertexai (Gemini SDK) import gate from count_tokens; replaces it with an explanatory comment. The actual token-counting path uses VertexAIPartnerModelsTokenCounter over plain httpx and never needed that import.
tests/test_litellm/llms/vertex_ai/vertex_ai_partner_models/count_tokens/test_count_tokens_no_vertexai_sdk.py	New regression test file with two fully-mocked tests: one verifies count_tokens proceeds past the old import gate when vertexai is unimportable; the other asserts the handler module itself never loads the Gemini SDK. All network calls are stubbed.

_{Reviews (1): Last reviewed commit: "fix(vertex_ai/partner_models): drop unus..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!