[New Feasture]: Support SP + DP parallal on Wan training#1223

Open

mahaocong90 wants to merge 4 commits into

modelscope:mainfrom

mahaocong90:dev-support-sequence-parallal-for-wan-training

Commits on Jan 21, 2026

Implement SP (spatial parallelism) support for WAN training, enabling DP + SP grouping configuration. The computational logic follows: DP = world_size / SP.

System Administrator
authored and
System Administrator
committed
Merge branch 'main' into dev-support-sequence-parallal-for-wan-training
mahaocong90
authored
[Bugfix] Fix the code issues reviewed by gemini-code-assist about sequence parallal on wan training.
mahaocong90
committed

Commits on Feb 14, 2026

Add a DeepSpeed configuration example for wan2.2-TI2V-5B, enabling SP and DeepSpeed zero2, where train_micro_batch_size_per_gpu must be specified in the json.
mahaocong90
committed