Add Qwen3-VL-30B-A3B

[lydhr]

Description

This PR adds support for the Qwen/Qwen3-VL-30B-A3B-Instruct model in MaxText, integrating this advanced Multimodal Vision-Language Model (VLM) which features a Mixture-of-Experts (MoE) text decoder.

Key Changes & Context

• New Architecture Enablement: Brought up qwen3-vl-30b-a3b using standard Qwen3-style decoder pathways paired with custom Vision Layer subclasses (Qwen3VLVisionEncoder, Qwen3VLVisionProjector).
• Feature Additions:
  • Added modern architecture configurations in Standard yaml directories.
  • Hooked required configurations into parameter converters, processing pipes, and tensor reshape/merge methods.
  • Resolved conflicts regarding positional Embedding/attention parameters unique to MOE scaling.
• Why this pipeline: Enables parity with the state-of-the-art vision-language pipelines in typical Maxtext environments smoothly bypassing legacy backend limits during Deepstack injection steps.

Implementation Specifics

• New configuration injected: src/maxtext/configs/models/qwen3-vl-30b-a3b.yml.
• Support is wired as headless decoder layers, with downstream projector transformations feeding standard decoding passes.
• Weight-transformation rules in param_mapping.py are enriched for correct layer replication vs scanning.

Tests

The changes have been thoroughly tested on a remote TPU VM environment using custom assertions and standard quick-tracing mechanisms.

1. Quick E2E Tracing: Verified standalone processing pipelines function without memory leaks/runtime locks by bootstrapping reduced configurations with mock
   pipelines.
2. Layer-wise equivalence (Numerical correctness): Validated JAX vision encoders and processing projectors yield output and decoder-state representations
   mathematically Equivalent to Torch baseline (HuggingFace implementations).

Checklist

• 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.

[codecov]

Codecov Report

❌ Patch coverage is 15.38462% with 11 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/maxtext/multimodal/processor.py	0.00%	8 Missing ⚠️
...xtext/checkpoint_conversion/utils/param_mapping.py	0.00%	2 Missing ⚠️
src/maxtext/layers/decoders.py	0.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!