[AMD][CI] Add Gemma 4 nightly accuracy tests for MI30x and MI35x

[michaelzhang-ai]

Motivation

Add nightly accuracy test for Gemma 4 on AMD MI30x (MI325/MI300X), matching AMD Day 0 support announcement.

Changes

New test file

• test/registered/amd/accuracy/mi30x/test_gemma4_eval_amd.py

Model tested

Model	Architecture	TP	mgsm_en Threshold	Observed Score
google/gemma-4-31B-it	Dense 31B	1	0.90	0.984

Server configuration

• --attention-backend triton (required for bidirectional image-token attention on AMD)
• --watchdog-timeout 1200 (for first-time weight download)
• Step installs transformers from the commit required by [New Model] Gemma 4 #21952

Workflow integration

Gemma 4 test added as a step (not a standalone job) inside the existing nightly-accuracy-2-gpu job in both workflows:

• nightly-test-amd.yml — step: "Accuracy Test (Gemma 4)"
• nightly-test-amd-rocm720.yml — step: "Accuracy Test ROCm 7.2 (Gemma 4)"

Step uses if: always() so it runs even if the prior GSM8K step fails.

CI Results

Run	Platform	Result
Run 7	MI30x (default ROCm)	✅ pass
Run 7	MI35x (default ROCm)	✅ pass
Run 8	MI30x (ROCm 7.2)	✅ pass
Run 8	MI35x (ROCm 7.2)	✅ pass

Test plan

• YAML validation passes locally
• Single commit, 3 files changed (2 workflow edits + 1 test file)
• Gemma 4 accuracy verified: 0.984 on MI30x (threshold 0.90)
• if: always() confirmed working — Gemma 4 step passed despite prior step failure
• Passed on all 4 platform/ROCm combinations

[gemini-code-assist]

Code Review

This pull request introduces accuracy evaluation tests for Gemma 4 models on AMD MI300X/MI325 and MI35x platforms using the mgsm_en benchmark. The review feedback highlights several improvement opportunities, including refactoring the tests to eliminate code duplication between the two new files, extracting common configuration arguments into constants to improve maintainability, and adopting a more idiomatic approach for passing evaluation arguments by using types.SimpleNamespace instead of the type() constructor.