Implemented workaround to allow the use of seeds with the mps/metal backend. Fixed img2img's use of unsupported precision float64 with mps backend.

1 files changed, 31 insertions, 7 deletions

diff --git a/modules/processing.py b/modules/processing.py
index cf2e13d3..80bf7cc0 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -1,3 +1,6 @@
+# Metal backend fixes written and placed

+# into the public domain by Elias Oenal <sd@eliasoenal.com>

+

 import contextlib

 import json

 import math

@@ -105,18 +108,32 @@ def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, see
     for i, seed in enumerate(seeds):

         noise_shape = shape if seed_resize_from_h <= 0 or seed_resize_from_w <= 0 else (shape[0], seed_resize_from_h//8, seed_resize_from_w//8)

 

+        # Pytorch currently doesn't handle seeting randomness correctly when the metal backend is used.

+        if shared.device.type == 'mps':

+            g = torch.Generator(device='cpu')

+

         subnoise = None

         if subseeds is not None:

             subseed = 0 if i >= len(subseeds) else subseeds[i]

-            torch.manual_seed(subseed)

-            subnoise = torch.randn(noise_shape, device=shared.device)

+            if shared.device.type == 'mps':

+                g.manual_seed(subseed)

+                subnoise = torch.randn(noise_shape, generator=g, device='cpu').to('mps')

+            else: # cpu or cuda

+                torch.manual_seed(subseed)

+                subnoise = torch.randn(noise_shape, device=shared.device)

 

         # randn results depend on device; gpu and cpu get different results for same seed;

         # the way I see it, it's better to do this on CPU, so that everyone gets same result;

         # but the original script had it like this, so I do not dare change it for now because

         # it will break everyone's seeds.

-        torch.manual_seed(seed)

-        noise = torch.randn(noise_shape, device=shared.device)

+        # When using the mps backend falling back to the cpu device is needed, since mps currently

+        # does not implement seeding properly.

+        if shared.device.type == 'mps':

+            g.manual_seed(seed)

+            noise = torch.randn(noise_shape, generator=g, device='cpu').to('mps')

+        else: # cpu or cuda

+            torch.manual_seed(seed)

+            x = torch.randn(shape, device=shared.device)

 

         if subnoise is not None:

             #noise = subnoise * subseed_strength + noise * (1 - subseed_strength)

@@ -127,8 +144,12 @@ def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, see
             # noise_shape = (64, 80)

             # shape = (64, 72)

 

-            torch.manual_seed(seed)

-            x = torch.randn(shape, device=shared.device)

+            if shared.device.type == 'mps':

+                g.manual_seed(seed)

+                x = torch.randn(shape, generator=g, device='cpu').to('mps')

+            else:

+                torch.manual_seed(seed)

+                x = torch.randn(shape, device=shared.device)

             dx = (shape[2] - noise_shape[2]) // 2 # -4

             dy = (shape[1] - noise_shape[1]) // 2

             w = noise_shape[2] if dx >= 0 else noise_shape[2] + 2 * dx

@@ -463,7 +484,10 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing):
         if self.image_mask is not None:

             init_mask = latent_mask

             latmask = init_mask.convert('RGB').resize((self.init_latent.shape[3], self.init_latent.shape[2]))

-            latmask = np.moveaxis(np.array(latmask, dtype=np.float64), 2, 0) / 255

+            if shared.device.type == 'mps': # mps backend does not support float64

+                latmask = np.moveaxis(np.array(latmask, dtype=np.float32), 2, 0) / 255

+            else:

+                latmask = np.moveaxis(np.array(latmask, dtype=np.float64), 2, 0) / 255

             latmask = latmask[0]

             latmask = np.around(latmask)

             latmask = np.tile(latmask[None], (4, 1, 1))