Merge pull request #1159 from NVIDIA/more_bug_fixes

Bug fixes

Merge pull request #1159 from NVIDIA/more_bug_fixes
Bug fixes
0506fe36 · Thor Johnsen · GitHub · 0cb1cb3b · 8c4a0075 · 0506fe36
Unverified Commit 0506fe36 authored Sep 02, 2021 by Thor Johnsen Committed by GitHub Sep 02, 2021
Show whitespace changes
Inline Side-by-side

Showing with 1 addition and 1 deletion

apex/contrib/bottleneck/bottleneck.py apex/contrib/bottleneck/bottleneck.py +1 -1

No files found.
--- a/apex/contrib/bottleneck/bottleneck.py
+++ b/apex/contrib/bottleneck/bottleneck.py
@@ -354,7 +354,7 @@ class SpatialBottleneckFunction(torch.autograd.Function):
                    btm_halo = all_halos[ctx.local_rank+1][:,:1,:,:]
                    fat_halo[:,:2,:,:].copy_(grad_out2[:,Hs-2:,:,:])
                    fat_halo[:,2:,:,:].copy_(btm_halo)
-                    relu_halo[:,:2,:,:].copy_(relu1[:,Hs-2,:,:])
+                    relu_halo[:,:2,:,:].copy_(relu1[:,Hs-2:,:,:])
                    relu_halo[:,2:,:,:].zero_()
                    btm_grad_out1_halo = fast_bottleneck.backward_grad_out1_halo(ctx.nhwc, ctx.stride_1x1, t_list, grads, fat_halo, relu_halo)
                    btm_grad_out1_halo = btm_grad_out1_halo[:,1:2,:,:]