TeaCache+xDiT

April 18, 2025 ยท View on GitHub

This repository implements multi-GPU parallelization of TeaCache within the xDiT framework. The supported models include:

Please run the appropriate script (run_[your_model].sh) to get started.

Note: All following experiments are conducted on the A800

Performance on Wan 2.1

ModelMethod8x
Wan 2.1-T2V-14BUSP473
TeaCache(0.2)241

Performance on CogVideoX

ModelMethod1x2x4x8x
CogvideoXUlysses133.577.759.4136.68
TeaCache(slow)103.360.946.829.6
TeaCache(fast)7846.736.523.7

Performance on HunyuanVideo

ModelMethod1x2x4x8x
HunyuanVideoUlysses3086.51624.4856.6480.4
TeaCache(slow)1884.51001540.8312.9
TeaCache(fast)1395.2753.9418.4254.1

Performance on Flux

ModelMethod1x2x4x8x
FluxUlysses12.97.914.43.61
TeaCache(slow)12.14.612.682.24
TeaCache(fast)7.92.451.531.3

Performance on ConsistID

ModelMethod1x2x3x6x
ConsistIDUlysses218.64123.8488.654.09
TeaCache(slow)-0.1142.1881.3960.7139.77
TeaCache(fast)-0.290.3453.8241.3427.18

Acknowledgement

This respotory is built based on TeaCache and xDiT, thanks for their contributors!