fix: refine scanned layers unroll logic for Gemma 3

[mesakhcienet]

Description

The unroll_gemma_scanned_weights workaround—specifically designed to unstack Gemma 3 scanned blocks due to Tunix's limitations—was previously formatting the unrolled keys incorrectly (using raw integers for the layer index). This PR fixes :

1. the key generation logic to correctly format the unrolled layers as strings (e.g., changing the tuple injection from (0,) to ("layers_0",)), which aligns with the expected dictionary structure for downstream weight mapping.
2. it refines the scan_length detection to safely check for the mlp.wi_0 architectural pattern.

Tests

We run train_rl.py for both scanned and unscanned layers. Both should return normal sentences

Scanned layers

Config: scan_layers=true and use_standalone_converter=True

• script
• log

Unscanned layers

Config : scan_layers=false

• script
• log

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[codecov]

Codecov Report

❌ Patch coverage is 7.69231% with 12 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...c/maxtext/integration/vllm/maxtext_vllm_rollout.py	7.69%	12 Missing ⚠️

📢 Thoughts on this report? Let us know!

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.