README.md

---
title: Wan2.1 T2V 1.3B - 8x A100
emoji: 🎬
colorFrom: indigo
colorTo: purple
sdk: docker
app_file: Dockerfile
short_description: Wan2.1 T2V-1.3B on 8x A100, FSDP + xDiT USP
startup_duration_timeout: 1h
---

# Wan2.1 (T2V-1.3B) — Multi-GPU Gradio demo

Gradio demo for testing [Wan-AI/Wan2.1-T2V-1.3B](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B)
running on **8x A100 (640 GB)** with the official multi-GPU recipe from the model card:

```
torchrun --nproc_per_node=8 generate.py --task t2v-1.3B --size 832*480 \
    --ckpt_dir ./Wan2.1-T2V-1.3B --dit_fsdp --t5_fsdp --ring_size 8 \
    --sample_shift 8 --sample_guide_scale 6
```

- **FSDP** shards the DiT and the UMT5-XXL text encoder across all 8 GPUs
- **xDiT USP** (ring sequence parallelism, `ring_size=8`) splits the sequence across 8 GPUs for faster generation — Ulysses is not used here because the 1.3B model has 12 attention heads (must be divisible by `ulysses_size`; `12 % 8 != 0`), which is exactly the config from the official efficiency table for the 1.3B model on 8 GPUs
- ~17.5 GB of checkpoints are baked into the Docker image

Default settings follow the official demo: 480P (832*480), 50 steps, guide scale 6, shift 8, 16 fps, 81 frames (5 s).

## Related

- Model: https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B
- Code: https://github.com/Wan-Video/Wan2.1
- License: Apache 2.0