(requested) support AMD 8GB GPUs via Windows DirectML

this update is requested by users

ldm_patched/ldm/models/autoencoder.py

@@ -8,6 +8,7 @@ from ldm_patched.ldm.modules.distributions.distributions import DiagonalGaussian
 
 from ldm_patched.ldm.util import instantiate_from_config
 from ldm_patched.ldm.modules.ema import LitEma
+import ldm_patched.modules.ops
 
 class DiagonalGaussianRegularizer(torch.nn.Module):
     def __init__(self, sample: bool = True):
@@ -161,12 +162,12 @@ class AutoencodingEngineLegacy(AutoencodingEngine):
             },
             **kwargs,
         )
-        self.quant_conv = torch.nn.Conv2d(
+        self.quant_conv = ldm_patched.modules.ops.disable_weight_init.Conv2d(
             (1 + ddconfig["double_z"]) * ddconfig["z_channels"],
             (1 + ddconfig["double_z"]) * embed_dim,
             1,
         )
-        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.post_quant_conv = ldm_patched.modules.ops.disable_weight_init.Conv2d(embed_dim, ddconfig["z_channels"], 1)
         self.embed_dim = embed_dim
 
     def get_autoencoder_params(self) -> list:

ldm_patched/ldm/modules/diffusionmodules/model.py

@@ -41,7 +41,7 @@ def nonlinearity(x):
 
 
 def Normalize(in_channels, num_groups=32):
-    return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+    return ops.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
 
 
 class Upsample(nn.Module):

ldm_patched/ldm/modules/diffusionmodules/upscaling.py

@@ -43,8 +43,8 @@ class AbstractLowScaleModel(nn.Module):
 
     def q_sample(self, x_start, t, noise=None):
         noise = default(noise, lambda: torch.randn_like(x_start))
-        return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
-                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+        return (extract_into_tensor(self.sqrt_alphas_cumprod.to(x_start.device), t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod.to(x_start.device), t, x_start.shape) * noise)
 
     def forward(self, x):
         return x, None

ldm_patched/ldm/modules/diffusionmodules/util.py

@@ -51,9 +51,9 @@ class AlphaBlender(nn.Module):
         if self.merge_strategy == "fixed":
             # make shape compatible
             # alpha = repeat(self.mix_factor, '1 -> b () t  () ()', t=t, b=bs)
-            alpha = self.mix_factor
+            alpha = self.mix_factor.to(image_only_indicator.device)
         elif self.merge_strategy == "learned":
-            alpha = torch.sigmoid(self.mix_factor)
+            alpha = torch.sigmoid(self.mix_factor.to(image_only_indicator.device))
             # make shape compatible
             # alpha = repeat(alpha, '1 -> s () ()', s = t * bs)
         elif self.merge_strategy == "learned_with_images":
@@ -61,7 +61,7 @@ class AlphaBlender(nn.Module):
             alpha = torch.where(
                 image_only_indicator.bool(),
                 torch.ones(1, 1, device=image_only_indicator.device),
-                rearrange(torch.sigmoid(self.mix_factor), "... -> ... 1"),
+                rearrange(torch.sigmoid(self.mix_factor.to(image_only_indicator.device)), "... -> ... 1"),
             )
             alpha = rearrange(alpha, self.rearrange_pattern)
             # make shape compatible

ldm_patched/ldm/modules/encoders/noise_aug_modules.py

@@ -15,12 +15,12 @@ class CLIPEmbeddingNoiseAugmentation(ImageConcatWithNoiseAugmentation):
 
     def scale(self, x):
         # re-normalize to centered mean and unit variance
-        x = (x - self.data_mean) * 1. / self.data_std
+        x = (x - self.data_mean.to(x.device)) * 1. / self.data_std.to(x.device)
         return x
 
     def unscale(self, x):
         # back to original data stats
-        x = (x * self.data_std) + self.data_mean
+        x = (x * self.data_std.to(x.device)) + self.data_mean.to(x.device)
         return x
 
     def forward(self, x, noise_level=None):

ldm_patched/ldm/modules/temporal_ae.py

@@ -82,14 +82,14 @@ class VideoResBlock(ResnetBlock):
 
             x = self.time_stack(x, temb)
 
-            alpha = self.get_alpha(bs=b // timesteps)
+            alpha = self.get_alpha(bs=b // timesteps).to(x.device)
             x = alpha * x + (1.0 - alpha) * x_mix
 
             x = rearrange(x, "b c t h w -> (b t) c h w")
         return x
 
 
-class AE3DConv(torch.nn.Conv2d):
+class AE3DConv(ops.Conv2d):
     def __init__(self, in_channels, out_channels, video_kernel_size=3, *args, **kwargs):
         super().__init__(in_channels, out_channels, *args, **kwargs)
         if isinstance(video_kernel_size, Iterable):
@@ -97,7 +97,7 @@ class AE3DConv(torch.nn.Conv2d):
         else:
             padding = int(video_kernel_size // 2)
 
-        self.time_mix_conv = torch.nn.Conv3d(
+        self.time_mix_conv = ops.Conv3d(
             in_channels=out_channels,
             out_channels=out_channels,
             kernel_size=video_kernel_size,
@@ -167,7 +167,7 @@ class AttnVideoBlock(AttnBlock):
         emb = emb[:, None, :]
         x_mix = x_mix + emb
 
-        alpha = self.get_alpha()
+        alpha = self.get_alpha().to(x.device)
         x_mix = self.time_mix_block(x_mix, timesteps=timesteps)
         x = alpha * x + (1.0 - alpha) * x_mix  # alpha merge