[Doc] Add diffusion attention backend docs

[david6666666]

Summary

This PR adds user-facing documentation for diffusion attention backend selection in vLLM-Omni.

What Changed

• add docs/user_guide/diffusion/attention_backends.md
• document DIFFUSION_ATTENTION_BACKEND selection and backend options
• document SageAttention source installation and usage examples
• add a link from the text-to-video offline inference guide to the new backend guide
• add the new guide to the MkDocs navigation

Why

Diffusion users need one place to understand how to switch attention backends, how to install SageAttention, and what to validate when comparing SAGE_ATTN against the default FlashAttention path.

Validation

• pre-commit run --all-files
• pytest -q tests/diffusion/attention/test_flash_attn.py
• local offline validation on H20 GPUs for:
  • HunyuanVideo-1.5
  • Wan2.2-TI2V-5B-Diffusers

[david6666666]

Validation

I ran local offline validation on H20 GPUs after rebuilding SageAttention from upstream main.

HunyuanVideo-1.5

Config:

• model: hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v
• 33 frames, 480x832, 8 steps, TP=1, same prompt / seed

Results:

• FLASH_ATTN: forward ~= 28.36s, total ~= 28.84s
• SAGE_ATTN: forward ~= 24.86s, total ~= 25.33s
• output diff vs FLASH_ATTN: PSNR ~= 16.96 dB, MAE ~= 22.96

FLASH_ATTN

hv15_retest_fa3_steps8.mp4

SAGE_ATTN

hv15_retest_sage_steps8.mp4

Wan2.2 TI2V 5B

Config:

• model: Wan-AI/Wan2.2-TI2V-5B-Diffusers
• 49 frames, 704x1280, 30 steps, TP=2, same prompt / seed

Results:

• FLASH_ATTN: diffuse ~= 36.20s, forward ~= 44.30s, total ~= 45.06s
• SAGE_ATTN: diffuse ~= 32.89s, forward ~= 41.09s, total ~= 41.83s
• output diff vs FLASH_ATTN: PSNR ~= 27.96 dB, MAE ~= 3.51

FLASH_ATTN

wan22_fa3.mp4

SAGE_ATTN

wan22_sage.mp4

Notes

• pre-commit run --all-files passed locally before publishing the docs changes.
• For Wan2.2, SageAttention provided a real speedup with a relatively small output difference.
• For HunyuanVideo, SageAttention was faster but still showed a noticeably larger output drift relative to the FlashAttention baseline.

[david6666666]

@ZJY0516 @lishunyang12 ptal thx

[hsliuustc0106]

Good documentation. Covers all diffusion attention backends with clear installation and usage examples.