Merge pull request #4918 from brkirch/pytorch-fixes

Fixes for PyTorch 1.12.1 when using MPS

modules/devices.py

@@ -2,9 +2,10 @@ import sys, os, shlex
 import contextlib
 import torch
 from modules import errors
+from packaging import version
 
 
-# has_mps is only available in nightly pytorch (for now) and MasOS 12.3+.
+# has_mps is only available in nightly pytorch (for now) and macOS 12.3+.
 # check `getattr` and try it for compatibility
 def has_mps() -> bool:
     if not getattr(torch, 'has_mps', False):
@@ -99,9 +100,25 @@ def autocast(disable=False):
 
 
 # MPS workaround for https://github.com/pytorch/pytorch/issues/79383
-def mps_contiguous(input_tensor, device):
-    return input_tensor.contiguous() if device.type == 'mps' else input_tensor
-
-
-def mps_contiguous_to(input_tensor, device):
-    return mps_contiguous(input_tensor, device).to(device)
+orig_tensor_to = torch.Tensor.to
+def tensor_to_fix(self, *args, **kwargs):
+    if self.device.type != 'mps' and \
+       ((len(args) > 0 and isinstance(args[0], torch.device) and args[0].type == 'mps') or \
+       (isinstance(kwargs.get('device'), torch.device) and kwargs['device'].type == 'mps')):
+        self = self.contiguous()
+    return orig_tensor_to(self, *args, **kwargs)
+
+
+# MPS workaround for https://github.com/pytorch/pytorch/issues/80800 
+orig_layer_norm = torch.nn.functional.layer_norm
+def layer_norm_fix(*args, **kwargs):
+    if len(args) > 0 and isinstance(args[0], torch.Tensor) and args[0].device.type == 'mps':
+        args = list(args)
+        args[0] = args[0].contiguous()
+    return orig_layer_norm(*args, **kwargs)
+
+
+# PyTorch 1.13 doesn't need these fixes but unfortunately is slower and has regressions that prevent training from working
+if has_mps() and version.parse(torch.__version__) < version.parse("1.13"):
+    torch.Tensor.to = tensor_to_fix
+    torch.nn.functional.layer_norm = layer_norm_fix

modules/esrgan_model.py

@@ -199,7 +199,7 @@ def upscale_without_tiling(model, img):
     img = img[:, :, ::-1]
     img = np.ascontiguousarray(np.transpose(img, (2, 0, 1))) / 255
     img = torch.from_numpy(img).float()
-    img = devices.mps_contiguous_to(img.unsqueeze(0), devices.device_esrgan)
+    img = img.unsqueeze(0).to(devices.device_esrgan)
     with torch.no_grad():
         output = model(img)
     output = output.squeeze().float().cpu().clamp_(0, 1).numpy()

modules/scunet_model.py

@@ -54,7 +54,7 @@ class UpscalerScuNET(modules.upscaler.Upscaler):
         img = img[:, :, ::-1]
         img = np.moveaxis(img, 2, 0) / 255
         img = torch.from_numpy(img).float()
-        img = devices.mps_contiguous_to(img.unsqueeze(0), device)
+        img = img.unsqueeze(0).to(device)
 
         with torch.no_grad():
             output = model(img)

modules/swinir_model.py

@@ -111,7 +111,7 @@ def upscale(
     img = img[:, :, ::-1]
     img = np.moveaxis(img, 2, 0) / 255
     img = torch.from_numpy(img).float()
-    img = devices.mps_contiguous_to(img.unsqueeze(0), devices.device_swinir)
+    img = img.unsqueeze(0).to(devices.device_swinir)
     with torch.no_grad(), precision_scope("cuda"):
         _, _, h_old, w_old = img.size()
         h_pad = (h_old // window_size + 1) * window_size - h_old