• Xin Yao's avatar
    [Refactor] Replace third_party/nccl with PyTorch's NCCL backend (#4989) · 8d5d8962
    Xin Yao authored
    * expose GeneratePermutation
    
    * add sparse_all_to_all_push
    
    * add sparse_all_to_all_pull
    
    * add unit test
    
    * handle world_size=1
    
    * remove python nccl wrapper
    
    * remove the nccl dependency
    
    * use pinned memory to speedup D2H copy
    
    * fix lint
    
    * resolve comments
    
    * fix lint
    
    * fix ut
    
    * resolve comments
    8d5d8962
sparse_optim.py 34.7 KB