[Docs] Add Wan2.1-T2V as supported video generation models

[SamitHuang]

Purpose

Add Wan-AI/Wan2.1-T2V-1.3B-Diffusers and Wan-AI/Wan2.1-T2V-14B-Diffusers as a supported video generation model in vLLM-Omni documentation.

Wan2.1-T2V-1.3B shares the same WanPipeline architecture (_class_name: "WanPipeline" in model_index.json) and WanTransformer3DModel as Wan2.2, so it works out-of-the-box with the existing pipeline — no code changes required, only different CLI args (--flow-shift 3.0 --boundary-ratio 0.0).

This PR was created using the add-diffusion-model skill from vllm-omni-skills, which guided the analysis: checking model_index.json → identifying the model reuses the existing WanPipeline registry entry → testing → updating docs.

Test Plan

Ran offline inference with the existing text_to_video.py script:

CUDA_VISIBLE_DEVICES=0 python examples/offline_inference/text_to_video/text_to_video.py \
    --model Wan-AI/Wan2.1-T2V-1.3B-Diffusers \
    --prompt "A cute cat playing with a ball of yarn" \
    --num-inference-steps 20 \
    --height 480 --width 832 \
    --num-frames 33 \
    --flow-shift 3.0 \
    --boundary-ratio 0.0 \
    --guidance-scale 5.0 \
    --enforce-eager \
    --fps 16

Test Result

• Model loaded successfully (13.7 GiB, 22.2s load time)
• 20 denoising steps at ~1.97 it/s on single H800
• Total generation: 13.1 seconds
• Output video saved successfully (221 KB, 33 frames at 480x832)
• No code changes needed — existing Wan2.2 pipeline handles Wan2.1 out-of-the-box

[Stage-0] INFO: Model loading took 13.7214 GiB and 22.188893 seconds
[Stage-0] INFO: Generation completed successfully.
[Stage-0] INFO: DiffusionEngine.step breakdown: preprocess=0.01 ms, add_req_and_wait=12096.52 ms, postprocess=105.56 ms
Total generation time: 13.0964 seconds (13096.43 ms)
Saved generated video to /tmp/wan21_test_output.mp4

Wan-AI/Wan2.1-T2V-1.3B-Diffusers :

wan21_1d3b.mp4

Wan-AI/Wan2.1-T2V-14B-Diffusers

t2v_out_wan21_14b.mp4

[lishunyang12]

Hi, nice addition! It would be great if you could include a visual comparison between the 1.3B and 14B model outputs (e.g. side-by-side frames or short gifs with the same prompt) so readers can see the quality tradeoff.

[SamitHuang]

Hi, nice addition! It would be great if you could include a visual comparison between the 1.3B and 14B model outputs (e.g. side-by-side frames or short gifs with the same prompt) so readers can see the quality tradeoff.

sure, added @lishunyang12

[wtomin]

LGTM.

[wtomin]

Please also update:

• docs/user_guide/diffusion_acceleration.md
• docs/models/supported_models.md

[SamitHuang]

Please also update:

• docs/user_guide/diffusion_acceleration.md
• docs/models/supported_models.md

done