Quick Start#
This page uses Wan2.1 as an example to show how to run text-to-video inference with MindIE SD. For more model-specific inference details, see Modelers - MindIE/Wan2.1.
Prerequisites#
Before running inference, complete the environment preparation and install MindIE SD by following the Installation Guide.
Run inference#
1. Get inference scripts#
Clone the Wan2.1 inference script repository from Modelers and install its dependencies:
git clone https://modelers.cn/MindIE/Wan2.1.git && cd Wan2.1
pip install -r requirements.txt
2. Get model weights#
The repository above contains inference scripts and does not include model weights. Weights must be downloaded separately. For Wan2.1, the following models are available:
Model |
Description |
Weights |
|---|---|---|
Wan2.1-T2V-14B |
Text-to-video |
|
Wan2.1-I2V-14B-480P |
Image-to-video (480P) |
|
Wan2.1-I2V-14B-720P |
Image-to-video (720P) |
After downloading, the weight directory should look like this (using Wan2.1-T2V-14B as an example):
Wan2.1-T2V-14B/
├── config.json
├── model_index.json
├── models/
│ ├── dit/
│ ├── vae/
│ └── text_encoder/
└── ...
Note
Ascend-adapted models are also available on Modelers - MindIE.
For other models (FLUX.1-dev, HunyuanVideo, etc.), see the links in the supported matrix.
3. Run inference#
Set the weight path via the model_base parameter and run the inference script. Parameter details are documented in parameter_config.md.
# 8-card inference for Wan2.1-T2V-14B
cp MindIE-SD/examples/wan/infer_t2v.sh ./
export model_base="/path/to/Wan2.1-T2V-14B"
bash infer_t2v.sh
Acceleration results#
The following Wan2.1 example shows the effect of different acceleration features on an Atlas 800I A2 inference server (1*64G), including both single-card and multi-card runs.
Where:
Cache refers to the AttentionCache feature.
TP refers to the Tensor Parallel feature.
FA sparse refers to the RainFusion optimization under FA sparsity.
CFG refers to the CFG Parallel feature.
Ulysses refers to the Ulysses Sequence Parallel feature. The generated video resolution is 832*480 and
sample_stepsis 50.
Single-card acceleration#
Cache acceleration
Baseline |
+ Cache ratio 1.6 |
+ Cache ratio 2.0 |
+ Cache ratio 2.4 |
|---|---|---|---|
860.2s |
631.7s 1.36x |
541.8s 1.59x |
516.9s *1.66x |
|
|
|
|
Parallel strategy results#
Two-card single-strategy results
Model |
Cards |
Parallel strategy |
Output resolution |
Operator optimization |
Cache optimization |
FA sparse |
50-step E2E time (s) |
Speedup |
|---|---|---|---|---|---|---|---|---|
Wan2.1 |
2 |
VAE |
832*480 |
√ |
√ |
√ |
548.8 |
1.02x |
Wan2.1 |
2 |
TP |
832*480 |
√ |
√ |
√ |
502.8 |
1.12x |
Wan2.1 |
2 |
CFG |
832*480 |
√ |
√ |
√ |
332.6 |
1.69x |
Wan2.1 |
2 |
Ulysses |
832*480 |
√ |
√ |
√ |
327.6 |
*1.71x |
Note: * marks the best acceleration result.
Multi-card combined-strategy results
Model |
Cards |
Parallel strategy |
Output resolution |
Operator optimization |
Cache optimization |
FA sparse |
50-step E2E time (s) |
Speedup |
|---|---|---|---|---|---|---|---|---|
Wan2.1 |
4 |
TP=4, VAE |
832*480 |
√ |
√ |
√ |
204.0 |
2.754x |
Wan2.1 |
4 |
CFG=2, TP=2, VAE |
832*480 |
√ |
√ |
√ |
175.8 |
3.19x |
Wan2.1 |
4 |
Ulysses=4, VAE |
832*480 |
√ |
√ |
√ |
151.1 |
3.71x |
Wan2.1 |
4 |
CFG=2, Ulysses=2, VAE |
832*480 |
√ |
√ |
√ |
147.9 |
*3.79x |
Wan2.1 |
8 |
TP=8, VAE |
832*480 |
√ |
√ |
√ |
141.5 |
3.96x |
Wan2.1 |
8 |
CFG=2, TP=4, VAE |
832*480 |
√ |
√ |
√ |
102.9 |
5.45x |
Wan2.1 |
8 |
Ulysses=8, VAE |
832*480 |
√ |
√ |
√ |
78.1 |
7.18x |
Wan2.1 |
8 |
CFG=2, Ulysses=4, VAE |
832*480 |
√ |
√ |
√ |
76.4 |
*7.34x |
Note: * marks the best acceleration result.