Merge pull request #1233 from jjsjann123/nhwc_cond_conv2d

fixing channels_last on cond_conv2d; update nvfuser debug env variable

timm/models/layers/cond_conv2d.py

@@ -91,7 +91,8 @@ class CondConv2d(nn.Module):
             bias = torch.matmul(routing_weights, self.bias)
             bias = bias.view(B * self.out_channels)
         # move batch elements with channels so each batch element can be efficiently convolved with separate kernel
-        x = x.view(1, B * C, H, W)
+        # reshape instead of view to work with channels_last input
+        x = x.reshape(1, B * C, H, W)
         if self.dynamic_padding:
             out = conv2d_same(
                 x, weight, bias, stride=self.stride, padding=self.padding,

timm/utils/jit.py

@@ -34,9 +34,9 @@ def set_jit_fuser(fuser):
         torch._C._jit_override_can_fuse_on_gpu(True)
         torch._C._jit_set_texpr_fuser_enabled(False)
     elif fuser == "nvfuser" or fuser == "nvf":
-        os.environ['PYTORCH_CUDA_FUSER_DISABLE_FALLBACK'] = '1'
-        os.environ['PYTORCH_CUDA_FUSER_DISABLE_FMA'] = '1'
-        os.environ['PYTORCH_CUDA_FUSER_JIT_OPT_LEVEL'] = '0'
+        os.environ['PYTORCH_NVFUSER_DISABLE_FALLBACK'] = '1'
+        os.environ['PYTORCH_NVFUSER_DISABLE_FMA'] = '1'
+        os.environ['PYTORCH_NVFUSER_JIT_OPT_LEVEL'] = '0'
         torch._C._jit_set_texpr_fuser_enabled(False)
         torch._C._jit_set_profiling_executor(True)
         torch._C._jit_set_profiling_mode(True)