Repository navigation
Replace Llama with Gemma 4 maas - #171
Merged
Merged
Conversation
Gemma 4 on Vertex MaaS is served only from the global location, whose endpoint host has no region prefix, so build the base URL accordingly. Only send the Llama Guard safety settings to Llama models; Gemma rejects the field with a 400 INVALID_ARGUMENT. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Gemma 4 answered READY on the first turn without ever invoking search, so every analysis ran with zero sources. With this instruction it searches on every claim (English and French); Llama 3.3 behaviour is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
deploy.sh had drifted from how production is actually deployed: - push to the regional Artifact Registry the cluster pulls from, not gcr.io - pass the commit-tagged image to Terraform via TF_VAR_image_uri - refuse to deploy uncommitted or unpushed commits (tag must match code) - show the Terraform plan and ask before applying, instead of -auto-approve - only log in when not already authenticated; fail fast on missing vars - wait for rollout with a timeout and check /v1/health afterwards Terraform rewrote two resources on every apply: - Cloud SQL: declare edition = ENTERPRISE so it stops sending an update - Ingress: ignore the pre-shared-cert annotation GKE manages itself Keep env, .env*, service-account.json and saved plans out of the image. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
NardinNakhla
approved these changes
Oct 19, 2026
…check - DEPLOY_TAG=<sha> redeploys an existing image with the current Terraform config instead of checking out (and running) an old copy of this script - skip the build when the commit's image already exists - refuse to build from a checkout whose shell scripts have CRLF endings - warn before deploying a commit that is not on origin/main - fix the plan-file cleanup path after leaving the Terraform directory Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Replace the Llama model with Gemma 4 26B A4B IT to ensure compatibility after Oct 21st. Changed variables, fixed Llama guard incompatibility, and added a statement to the prompt to ensure web searching.
Environment variable changes:
LLAMA_MODEL_NAME=google/gemma-4-26b-a4b-it-maasVERTEX_AI_LOCATION=global