Merge pull request #5194 from brkirch/autocast-and-mps-randn-fixes

Use devices.autocast() and fix MPS randn issues

modules/devices.py

@@ -66,24 +66,15 @@ dtype_vae = torch.float16
 
 
 def randn(seed, shape):
-    # Pytorch currently doesn't handle setting randomness correctly when the metal backend is used.
-    if device.type == 'mps':
-        generator = torch.Generator(device=cpu)
-        generator.manual_seed(seed)
-        noise = torch.randn(shape, generator=generator, device=cpu).to(device)
-        return noise
-
     torch.manual_seed(seed)
+    if device.type == 'mps':
+        return torch.randn(shape, device=cpu).to(device)
     return torch.randn(shape, device=device)
 
 
 def randn_without_seed(shape):
-    # Pytorch currently doesn't handle setting randomness correctly when the metal backend is used.
     if device.type == 'mps':
-        generator = torch.Generator(device=cpu)
-        noise = torch.randn(shape, generator=generator, device=cpu).to(device)
-        return noise
-
+        return torch.randn(shape, device=cpu).to(device)
     return torch.randn(shape, device=device)
 
 

modules/hypernetworks/hypernetwork.py

@@ -495,7 +495,7 @@ def train_hypernetwork(hypernetwork_name, learn_rate, batch_size, gradient_step,
                 if shared.state.interrupted:
                     break
 
-                with torch.autocast("cuda"):
+                with devices.autocast():
                     x = batch.latent_sample.to(devices.device, non_blocking=pin_memory)
                     if tag_drop_out != 0 or shuffle_tags:
                         shared.sd_model.cond_stage_model.to(devices.device)

modules/interrogate.py

@@ -148,8 +148,7 @@ class InterrogateModels:
 
             clip_image = self.clip_preprocess(pil_image).unsqueeze(0).type(self.dtype).to(devices.device_interrogate)
 
-            precision_scope = torch.autocast if shared.cmd_opts.precision == "autocast" else contextlib.nullcontext
-            with torch.no_grad(), precision_scope("cuda"):
+            with torch.no_grad(), devices.autocast():
                 image_features = self.clip_model.encode_image(clip_image).type(self.dtype)
 
                 image_features /= image_features.norm(dim=-1, keepdim=True)

modules/sd_hijack.py

@@ -183,11 +183,7 @@ def register_buffer(self, name, attr):
 
     if type(attr) == torch.Tensor:
         if attr.device != devices.device:
-
-            if devices.has_mps():
-                attr = attr.to(device="mps", dtype=torch.float32)
-            else:
-                attr = attr.to(devices.device)
+            attr = attr.to(device=devices.device, dtype=(torch.float32 if devices.device.type == 'mps' else None))
 
     setattr(self, name, attr)
 

modules/sd_samplers.py

@@ -6,6 +6,7 @@ import tqdm
 from PIL import Image
 import inspect
 import k_diffusion.sampling
+import torchsde._brownian.brownian_interval
 import ldm.models.diffusion.ddim
 import ldm.models.diffusion.plms
 from modules import prompt_parser, devices, processing, images
@@ -364,7 +365,23 @@ class TorchHijack:
             if noise.shape == x.shape:
                 return noise
 
-        return torch.randn_like(x)
+        if x.device.type == 'mps':
+            return torch.randn_like(x, device=devices.cpu).to(x.device)
+        else:
+            return torch.randn_like(x)
+
+
+# MPS fix for randn in torchsde
+def torchsde_randn(size, dtype, device, seed):
+    if device.type == 'mps':
+        generator = torch.Generator(devices.cpu).manual_seed(int(seed))
+        return torch.randn(size, dtype=dtype, device=devices.cpu, generator=generator).to(device)
+    else:
+        generator = torch.Generator(device).manual_seed(int(seed))
+        return torch.randn(size, dtype=dtype, device=device, generator=generator)
+
+
+torchsde._brownian.brownian_interval._randn = torchsde_randn
 
 
 class KDiffusionSampler:
@@ -415,8 +432,7 @@ class KDiffusionSampler:
         self.model_wrap.step = 0
         self.eta = p.eta or opts.eta_ancestral
 
-        if self.sampler_noises is not None:
-            k_diffusion.sampling.torch = TorchHijack(self.sampler_noises)
+        k_diffusion.sampling.torch = TorchHijack(self.sampler_noises if self.sampler_noises is not None else [])
 
         extra_params_kwargs = {}
         for param_name in self.extra_params:

modules/swinir_model.py

@@ -13,10 +13,6 @@ from modules.swinir_model_arch import SwinIR as net
 from modules.swinir_model_arch_v2 import Swin2SR as net2
 from modules.upscaler import Upscaler, UpscalerData
 
-precision_scope = (
-    torch.autocast if cmd_opts.precision == "autocast" else contextlib.nullcontext
-)
-
 
 class UpscalerSwinIR(Upscaler):
     def __init__(self, dirname):
@@ -112,7 +108,7 @@ def upscale(
     img = np.moveaxis(img, 2, 0) / 255
     img = torch.from_numpy(img).float()
     img = img.unsqueeze(0).to(devices.device_swinir)
-    with torch.no_grad(), precision_scope("cuda"):
+    with torch.no_grad(), devices.autocast():
         _, _, h_old, w_old = img.size()
         h_pad = (h_old // window_size + 1) * window_size - h_old
         w_pad = (w_old // window_size + 1) * window_size - w_old

modules/textual_inversion/dataset.py

@@ -82,7 +82,7 @@ class PersonalizedBase(Dataset):
             torchdata = torch.from_numpy(npimage).permute(2, 0, 1).to(device=device, dtype=torch.float32)
             latent_sample = None
 
-            with torch.autocast("cuda"):
+            with devices.autocast():
                 latent_dist = model.encode_first_stage(torchdata.unsqueeze(dim=0))
 
             if latent_sampling_method == "once" or (latent_sampling_method == "deterministic" and not isinstance(latent_dist, DiagonalGaussianDistribution)):
@@ -101,7 +101,7 @@ class PersonalizedBase(Dataset):
                 entry.cond_text = self.create_text(filename_text)
 
             if include_cond and not (self.tag_drop_out != 0 or self.shuffle_tags):
-                with torch.autocast("cuda"):
+                with devices.autocast():
                     entry.cond = cond_model([entry.cond_text]).to(devices.cpu).squeeze(0)
 
             self.dataset.append(entry)

modules/textual_inversion/textual_inversion.py

@@ -316,7 +316,7 @@ def train_embedding(embedding_name, learn_rate, batch_size, gradient_step, data_
                 if shared.state.interrupted:
                     break
 
-                with torch.autocast("cuda"):
+                with devices.autocast():
                     # c = stack_conds(batch.cond).to(devices.device)
                     # mask = torch.tensor(batch.emb_index).to(devices.device, non_blocking=pin_memory)
                     # print(mask)