Merge pull request #5810 from brkirch/fix-training-mps

Training fixes for MPS

modules/devices.py

@@ -125,7 +125,16 @@ def layer_norm_fix(*args, **kwargs):
     return orig_layer_norm(*args, **kwargs)
 
 
+# MPS workaround for https://github.com/pytorch/pytorch/issues/90532
+orig_tensor_numpy = torch.Tensor.numpy
+def numpy_fix(self, *args, **kwargs):
+    if self.requires_grad:
+        self = self.detach()
+    return orig_tensor_numpy(self, *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
+    torch.Tensor.numpy = numpy_fix

modules/safe.py

@@ -37,16 +37,16 @@ def find_class(self, module, name):
 
         if module == 'collections' and name == 'OrderedDict':
             return getattr(collections, name)
-        if module == 'torch._utils' and name in ['_rebuild_tensor_v2', '_rebuild_parameter']:
+        if module == 'torch._utils' and name in ['_rebuild_tensor_v2', '_rebuild_parameter', '_rebuild_device_tensor_from_numpy']:
             return getattr(torch._utils, name)
-        if module == 'torch' and name in ['FloatStorage', 'HalfStorage', 'IntStorage', 'LongStorage', 'DoubleStorage', 'ByteStorage']:
+        if module == 'torch' and name in ['FloatStorage', 'HalfStorage', 'IntStorage', 'LongStorage', 'DoubleStorage', 'ByteStorage', 'float32']:
             return getattr(torch, name)
         if module == 'torch.nn.modules.container' and name in ['ParameterDict']:
             return getattr(torch.nn.modules.container, name)
-        if module == 'numpy.core.multiarray' and name == 'scalar':
-            return numpy.core.multiarray.scalar
-        if module == 'numpy' and name == 'dtype':
-            return numpy.dtype
+        if module == 'numpy.core.multiarray' and name in ['scalar', '_reconstruct']:
+            return getattr(numpy.core.multiarray, name)
+        if module == 'numpy' and name in ['dtype', 'ndarray']:
+            return getattr(numpy, name)
         if module == '_codecs' and name == 'encode':
             return encode
         if module == "pytorch_lightning.callbacks" and name == 'model_checkpoint':