feat: Add FastVideo evaluation support for Wan models

[yixinhuang48]

Summary

Add support for evaluating FastVideo models (FastWan and base Wan) in the VideoScience benchmark framework.

Changes

1. backend/api_providers.py: Fix attention backend selection

   • FastWan DMD models → VIDEO_SPARSE_ATTN
   • Base Wan models → FLASH_ATTN

2. scripts/prepare_fastvideo_experiments.py: New script to prepare evaluation experiments

   • Supports FastWan-1.3B (3-step DMD distilled, ~17x faster)
   • Supports Wan-1.3B (50-step base model)
   • Skip-existing generated video logic to resume interrupted runs
   • Multi-run support for variance analysis

Usage

# Prepare experiments
python scripts/prepare_fastvideo_experiments.py \
    --data-file data/database/data_filtered.jsonl \
    --output-dir out/fastvideo_eval \
    --models both  # or fastwan-1.3b, wan-1.3b

# Run generation
python frontend.py --json out/fastvideo_eval/experiments.json --workers 1

Supported Models

Model	Inference Steps	Attention Backend
FastVideo/FastWan2.1-T2V-1.3B-Diffusers	3	VIDEO_SPARSE_ATTN
Wan-AI/Wan2.1-T2V-1.3B-Diffusers	50	FLASH_ATTN

Test plan

• Verified attention backend selection for both model types
• Tested video generation with FastWan-1.3B
• Tested video generation with Wan-1.3B base
• Verified skip-existing logic works correctly