TeaCache+xDiT

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

• CogVideoX
• ConsisID
• HunyuanVideo
• Flux.1 [dev]
• Wan 2.1

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

Model	Method	8x
Wan 2.1-T2V-14B	USP	473
	TeaCache(0.2)	241

Performance on CogVideoX

Model	Method	1x	2x	4x	8x
CogvideoX	Ulysses	133.5	77.7	59.41	36.68
	TeaCache(slow)	103.3	60.9	46.8	29.6
	TeaCache(fast)	78	46.7	36.5	23.7

Performance on HunyuanVideo

Model	Method	1x	2x	4x	8x
HunyuanVideo	Ulysses	3086.5	1624.4	856.6	480.4
	TeaCache(slow)	1884.5	1001	540.8	312.9
	TeaCache(fast)	1395.2	753.9	418.4	254.1

Performance on Flux

Model	Method	1x	2x	4x	8x
Flux	Ulysses	12.9	7.91	4.4	3.61
	TeaCache(slow)	12.1	4.61	2.68	2.24
	TeaCache(fast)	7.9	2.45	1.53	1.3

Performance on ConsistID

Model	Method	1x	2x	3x	6x
ConsistID	Ulysses	218.64	123.84	88.6	54.09
	TeaCache(slow)-0.1	142.18	81.39	60.71	39.77
	TeaCache(fast)-0.2	90.34	53.82	41.34	27.18

Acknowledgement

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