Merge pull request #294 from EliasOenal/master

Fixes for mps/Metal: use of seeds, img2img, CodeFormer

modules/codeformer_model.py

@@ -47,6 +47,8 @@ def setup_codeformer():
             def __init__(self):
                 self.net = None
                 self.face_helper = None
+                if shared.device.type == 'mps': # CodeFormer currently does not support mps backend
+                    shared.device_codeformer = torch.device('cpu')
 
             def create_models(self):
 
@@ -54,13 +56,13 @@ def setup_codeformer():
                     self.net.to(shared.device)
                     return self.net, self.face_helper
 
-                net = net_class(dim_embd=512, codebook_size=1024, n_head=8, n_layers=9, connect_list=['32', '64', '128', '256']).to(shared.device)
+                net = net_class(dim_embd=512, codebook_size=1024, n_head=8, n_layers=9, connect_list=['32', '64', '128', '256']).to(shared.device_codeformer)
                 ckpt_path = load_file_from_url(url=pretrain_model_url, model_dir=os.path.join(path, 'weights/CodeFormer'), progress=True)
                 checkpoint = torch.load(ckpt_path)['params_ema']
                 net.load_state_dict(checkpoint)
                 net.eval()
 
-                face_helper = FaceRestoreHelper(1, face_size=512, crop_ratio=(1, 1), det_model='retinaface_resnet50', save_ext='png', use_parse=True, device=shared.device)
+                face_helper = FaceRestoreHelper(1, face_size=512, crop_ratio=(1, 1), det_model='retinaface_resnet50', save_ext='png', use_parse=True, device=shared.device_codeformer)
 
                 self.net = net
                 self.face_helper = face_helper
@@ -82,7 +84,7 @@ def setup_codeformer():
                 for idx, cropped_face in enumerate(self.face_helper.cropped_faces):
                     cropped_face_t = img2tensor(cropped_face / 255., bgr2rgb=True, float32=True)
                     normalize(cropped_face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
-                    cropped_face_t = cropped_face_t.unsqueeze(0).to(shared.device)
+                    cropped_face_t = cropped_face_t.unsqueeze(0).to(shared.device_codeformer)
 
                     try:
                         with torch.no_grad():

modules/processing.py

@@ -103,18 +103,33 @@ 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.
+        generator = torch
+        if shared.device.type == 'mps':
+            shared.device_seed_type = 'cpu'
+            generator = torch.Generator(device=shared.device_seed_type)
+
         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)
+            generator.manual_seed(subseed)
+
+            if shared.device.type != shared.device_seed_type:
+                subnoise = torch.randn(noise_shape, generator=generator, device=shared.device_seed_type).to(shared.device)
+            else:
+                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.
+        generator.manual_seed(seed)
+        if shared.device.type != shared.device_seed_type:
+            noise = torch.randn(noise_shape, generator=generator, device=shared.device_seed_type).to(shared.device)
+        else:
+            noise = torch.randn(noise_shape, device=shared.device)
 
         if subnoise is not None:
             #noise = subnoise * subseed_strength + noise * (1 - subseed_strength)
@@ -124,9 +139,11 @@ def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, see
             #noise = torch.nn.functional.interpolate(noise.unsqueeze(1), size=shape[1:], mode="bilinear").squeeze()
             # noise_shape = (64, 80)
             # shape = (64, 72)
-
-            torch.manual_seed(seed)
-            x = torch.randn(shape, device=shared.device)
+            generator.manual_seed(seed)
+            if shared.device.type != shared.device_seed_type:
+                x = torch.randn(shape, generator=generator, device=shared.device_seed_type).to(shared.device)
+            else:
+                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
@@ -465,7 +482,10 @@ 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
+            precision = np.float64
+            if shared.device.type == 'mps': # mps backend does not support float64
+                precision = np.float32
+            latmask = np.moveaxis(np.array(latmask, dtype=precision), 2, 0) / 255
             latmask = latmask[0]
             latmask = np.around(latmask)
             latmask = np.tile(latmask[None], (4, 1, 1))

modules/shared.py

@@ -49,6 +49,8 @@ parser.add_argument("--opt-channelslast", action='store_true', help="change memo
 cmd_opts = parser.parse_args()
 
 device = get_optimal_device()
+device_codeformer = device
+device_seed_type = device
 
 batch_cond_uncond = cmd_opts.always_batch_cond_uncond or not (cmd_opts.lowvram or cmd_opts.medvram)
 parallel_processing_allowed = not cmd_opts.lowvram and not cmd_opts.medvram