Rename nccl_p2p extension to nccl_p2p_cuda

apex/contrib/bottleneck/bottleneck.py

@@ -2,7 +2,7 @@ import torch
 import torch.distributed as dist
 from torch import nn
 import fast_bottleneck
-import nccl_p2p as inc
+import nccl_p2p_cuda as inc
 
 def kaiming_uniform_(tensor, a=0, mode='fan_in', nonlinearity='leaky_relu'):
     weight_tensor_nchw = tensor

apex/contrib/bottleneck/halo_exchangers.py

 import torch
 import torch.distributed as dist
 from torch import nn
-import nccl_p2p as inc
+import nccl_p2p_cuda as inc
 import peer_memory_cuda as pm
 
 # Communication free halo exchanger.

setup.py

@@ -646,7 +646,7 @@ if "--nccl_p2p" in sys.argv:
     raise_if_cuda_home_none("--nccl_p2p")
     ext_modules.append(
         CUDAExtension(
-            name="nccl_p2p",
+            name="nccl_p2p_cuda",
             sources=[
                 "apex/contrib/csrc/nccl_p2p/nccl_p2p_cuda.cu",
                 "apex/contrib/csrc/nccl_p2p/nccl_p2p.cpp",