Skip to content

Try again to fix GCP Vertex batch tests#7521

Open
Aaron1011 wants to merge 1 commit into
mainfrom
aaron/more-batch-fixes
Open

Try again to fix GCP Vertex batch tests#7521
Aaron1011 wants to merge 1 commit into
mainfrom
aaron/more-batch-fixes

Conversation

@Aaron1011
Copy link
Copy Markdown
Member

@Aaron1011 Aaron1011 commented May 28, 2026

I've bumped the max_tokens again (we're still seeing intermittent 'max_length' finish reasons), as well as disabled the 'tool_choice=none' test that interacts badly with Gemini


Note

Low Risk
Changes are limited to e2e fixtures and test skips; no production inference or provider runtime logic is modified.

Overview
Stabilizes GCP Vertex Gemini batch e2e by raising max_tokens on the gcp-vertex-gemini-flash variant from 200 → 400 in both basic_test fixture configs (file-based and config-in-db), aiming to cut intermittent max_length finish reasons.

In batch.rs, the tool_choice="none" tool-use checks are skipped when model_provider_name == "gcp_vertex_gemini" (in two parallel test paths). Vertex can return UNEXPECTED_TOOL_CALL with empty content and no candidatesTokenCount, which fails the existing output_tokens > 0 assertion until finish-reason mapping and assertions are relaxed.

Reviewed by Cursor Bugbot for commit cc59500. Configure here.

I've bumped the max_tokens again (we're still seeing intermittent
'max_length' finish reasons), as well as disabled the 'tool_choice=none'
test that interacts badly with Gemini
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants