Diff: origin/master...HEAD, staged and unstaged changes
Total: 13 lines
Missing: 1 line
Coverage: 92%
Source File
Diff Coverage (%)
Missing Lines
hyper_parallel/core/utils/clip_grad.py
92.3%
865
hyper_parallel/core/utils/clip_grad.py
861862863864865866867868869
buf=torch.cat(chunks)forpdim,reduce_op,needs_avginpartial_info:group=mesh.get_group(pdim)buf=_all_reduce_offloaded(buf,reduce_op,group)ifneeds_avg:buf/=dist.get_world_size(group=group)# Extract views for params with actual gradients.