Merge pull request #14046 from hidenorly/AddFP32FallbackSupportOnSdVa…

…eApprox

Add FP32 fallback support on sd_vae_approx

modules/mac_specific.py

@@ -1,6 +1,7 @@
 import logging
 
 import torch
+from torch import Tensor
 import platform
 from modules.sd_hijack_utils import CondFunc
 from packaging import version
@@ -51,6 +52,17 @@ def cumsum_fix(input, cumsum_func, *args, **kwargs):
     return cumsum_func(input, *args, **kwargs)
 
 
+# MPS workaround for https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14046
+def interpolate_with_fp32_fallback(orig_func, *args, **kwargs) -> Tensor:
+    try:
+        return orig_func(*args, **kwargs)
+    except RuntimeError as e:
+        if "not implemented for" in str(e) and "Half" in str(e):
+            input_tensor = args[0]
+            return orig_func(input_tensor.to(torch.float32), *args[1:], **kwargs).to(input_tensor.dtype)
+        else:
+            print(f"An unexpected RuntimeError occurred: {str(e)}")
+
 if has_mps:
     if platform.mac_ver()[0].startswith("13.2."):
         # MPS workaround for https://github.com/pytorch/pytorch/issues/95188, thanks to danieldk (https://github.com/explosion/curated-transformers/pull/124)
@@ -77,6 +89,9 @@ def cumsum_fix(input, cumsum_func, *args, **kwargs):
         # MPS workaround for https://github.com/pytorch/pytorch/issues/96113
         CondFunc('torch.nn.functional.layer_norm', lambda orig_func, x, normalized_shape, weight, bias, eps, **kwargs: orig_func(x.float(), normalized_shape, weight.float() if weight is not None else None, bias.float() if bias is not None else bias, eps).to(x.dtype), lambda _, input, *args, **kwargs: len(args) == 4 and input.device.type == 'mps')
 
+        # MPS workaround for https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14046
+        CondFunc('torch.nn.functional.interpolate', interpolate_with_fp32_fallback, None)
+
         # MPS workaround for https://github.com/pytorch/pytorch/issues/92311
         if platform.processor() == 'i386':
             for funcName in ['torch.argmax', 'torch.Tensor.argmax']: