Support vllm_decode for Qwen3-VL

[hengtaoguo]

Description

Start with a short description of what the PR does and how this is a change from
the past.

Tests

python -m maxtext.inference.vllm_decode src/maxtext/configs/base.yml run_name=vllm_decode_qwen3_vl model_name=qwen3-vl-2b load_parameters_path=/home/hengtaoguo_google_com/projects/checkpoints/qwen3-vl-2b/unscanned/2026-07-01-11-55/0/items tokenizer_path=Qwen/Qwen3-VL-2B-Instruct prompt=\"Describe\ this\ image:\ \" image_path=/home/hengtaoguo_google_com/projects/maxtext/tests/assets/test_image.jpg max_prefill_predict_length=1024 max_target_length=2048 per_device_batch_size=1 scan_layers=false weight_dtype=bfloat16 max_num_batched_tokens=16384 use_chat_template=true vllm_hf_overrides=\{\"architectures\":\[\"MaxTextForCausalLM\"\]\}

Rendering prompts: 100%|████████████████████████| 1/1 [00:09<00:00,  9.43s/it]
Processed prompts: 100%|█| 1/1 [00:21<00:00, 21.93s/it, est. speed input: 16.1
I0727 04:42:04.155640 140668460825728 vllm_decode.py:229] Prompt: <|im_start|>user
<|vision_start|><|image_pad|><|vision_end|>Describe this image: <|im_end|>
<|im_start|>assistant
, Generated text: This is a panoramic photograph of the Seattle city skyline on a clear, sunny day.

The most prominent feature is the **Space Needle**, a distinctive observation deck that stands out in the center of the image. It is surrounded by a dense collection of modern skyscrapers and high-rise buildings that form the city's downtown core and business district. The buildings have a variety of architectural styles and glass facades, giving the skyline a dynamic and vibrant appearance.

In the foreground, the city is populated with a mix of residential buildings, including apartment complexes and smaller housing units, nestled among lush green trees and parks. These natural elements provide a contrast to the concrete and steel of the city's architecture.

In the background, a range of mountains is visible, with snow-capped peaks, suggesting the scenic beauty of the Pacific Northwest. The sky is a brilliant, clear blue, with a few soft, white clouds drifting across it.

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 15.88785% with 90 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...t/integration/vllm/maxtext_vllm_adapter/adapter.py	13.92%	67 Missing and 1 partial ⚠️
.../integration/vllm/maxtext_vllm_adapter/__init__.py	17.64%	11 Missing and 3 partials ⚠️
src/maxtext/layers/attentions.py	25.00%	6 Missing ⚠️
src/maxtext/utils/model_creation_utils.py	33.33%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!