Text-to-Video
Diffusers
Safetensors
WanPipeline
wlsaidhi commited on
Commit
38277dd
·
verified ·
1 Parent(s): d53f8ab

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +71 -1
README.md CHANGED
@@ -1,7 +1,77 @@
1
  ---
2
  license: apache-2.0
 
 
 
3
  ---
4
 
5
  # FastWan-QAD-1.3B
 
 
 
 
 
 
 
 
 
 
6
 
7
- NVFP4 linear layers with SageAttention3 FP4 backend
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: apache-2.0
3
+ pipeline_tag: text-to-video
4
+ library_name: diffusers
5
+ arxiv: 2603.00040
6
  ---
7
 
8
  # FastWan-QAD-1.3B
9
+ <p align="center">
10
+ <img src="https://raw.githubusercontent.com/hao-ai-lab/FastVideo/main/assets/logos/logo.svg" width="200"/>
11
+ </p>
12
+ <div>
13
+ <div align="center">
14
+ <a href="https://github.com/hao-ai-lab/FastVideo">Github</a> |
15
+ <a href="https://haoailab.com/blogs/fastwan-qad/">Blog</a> |
16
+ <a href="https://hao-ai-lab.github.io/FastVideo">Documentation</a>
17
+ </div>
18
+ </div>
19
 
20
+ ## Introduction
21
+
22
+ FastWan-QAD-1.3B is the fastest variant of the FastWan-QAD series, targeting RTX 5090 users. It uses **NVFP4 quantized linear layers** paired with the **SageAttention3 FP4 attention backend**, achieving end-to-end generation of a 5-second 480p video in **1.78 seconds** — over 3.4× faster than prior distilled models on the same hardware.
23
+
24
+ The model is built on [Wan-AI/Wan2.1-T2V-1.3B-Diffusers](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B-Diffusers) and trained with **quantization-aware distillation (QAD)**, jointly optimizing for low-bit precision and 3-step inference quality.
25
+
26
+ > **Hardware requirement:** RTX 5090 (sm100+). NVFP4 is a Blackwell-native format and is not supported on older GPUs. See [FastWan-QAD-1.3B-SA2](https://huggingface.co/FastVideo/FastWan-QAD-1.3B-SA2) for an alternative using SageAttention2++ or [FastWan-QAD-FP8-1.3B](https://huggingface.co/FastVideo/FastWan-QAD-FP8-1.3B) for RTX 4090 support.
27
+
28
+ ---
29
+
30
+ ## Model Overview
31
+
32
+ - **3-step inference** via quantization-aware distillation
33
+ - **NVFP4 linear layers** for maximum throughput on Blackwell GPUs
34
+ - **SageAttention3 FP4 backend** for attention computation
35
+ - Trained at **480p (832×480)** resolution, 81 frames (5 seconds at 16 fps)
36
+ - No classifier-free guidance at inference time
37
+ - Fast decoding via [TAEHV](https://github.com/madebyollin/taehv) tiny autoencoder
38
+
39
+ ## Performance
40
+
41
+ | Model | Hardware | Generation Time (5s 480p) |
42
+ |---|---|---|
43
+ | FastWan-QAD-1.3B | RTX 5090 | **1.78s** |
44
+ | FastWan-QAD-1.3B-SA2 | RTX 5090 | ~2.0s |
45
+ | FastWan-QAD-FP8-1.3B | RTX 4090 | ~3.4s |
46
+ | TurboDiffusion | RTX 5090 | 6.10s |
47
+ | LightX2V | RTX 5090 | 6.91s |
48
+
49
+ ## Inference
50
+
51
+ ```bash
52
+ # Install Tiny Autoencoder
53
+ git clone https://github.com/madebyollin/taehv.git
54
+ uv pip install -e taehv/
55
+
56
+ git clone https://github.com/hao-ai-lab/FastVideo.git
57
+ cd FastVideo
58
+ uv pip install -e .
59
+ cd examples/inference/optimizations
60
+ python nvfp4_qat_wan2_1_1_3b.py --taehv-checkpoint /path/to/taehv/taew2_1.pth
61
+ ```
62
+
63
+ ## Training
64
+
65
+ More details coming soon.
66
+
67
+ ---
68
+
69
+ It would be greatly appreciated if you cite our paper:
70
+ ```
71
+ @article{Zhang2026AttnQAT,
72
+ title={Attn-QAT: 4-Bit Attention With Quantization-Aware Training},
73
+ author={Zhang, Peiyuan and Noto, Matthew and Tan, Wenxuan and Jiang, Chengquan and Lin, Will and Zhou, Wei and Zhang, Hao},
74
+ journal={arXiv preprint arXiv:2603.00040},
75
+ year={2026}
76
+ }
77
+ ```