Patch timestep embedding to create tensor on-device

modules/sd_hijack_unet.py

 import torch
 from packaging import version
+from einops import repeat
+import math
 
 from modules import devices
 from modules.sd_hijack_utils import CondFunc
@@ -48,6 +50,30 @@ def apply_model(orig_func, self, x_noisy, t, cond, **kwargs):
         return orig_func(self, x_noisy.to(devices.dtype_unet), t.to(devices.dtype_unet), cond, **kwargs).float()
 
 
+# Monkey patch to create timestep embed tensor on device, avoiding a block.
+def timestep_embedding(_, timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=timesteps.device) / half
+        )
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    return embedding
+
+
 class GELUHijack(torch.nn.GELU, torch.nn.Module):
     def __init__(self, *args, **kwargs):
         torch.nn.GELU.__init__(self, *args, **kwargs)
@@ -69,6 +95,7 @@ def hijack_ddpm_edit():
 
 unet_needs_upcast = lambda *args, **kwargs: devices.unet_needs_upcast
 CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.apply_model', apply_model, unet_needs_upcast)
+CondFunc('ldm.modules.diffusionmodules.openaimodel.timestep_embedding', timestep_embedding)
 CondFunc('ldm.modules.diffusionmodules.openaimodel.timestep_embedding', lambda orig_func, timesteps, *args, **kwargs: orig_func(timesteps, *args, **kwargs).to(torch.float32 if timesteps.dtype == torch.int64 else devices.dtype_unet), unet_needs_upcast)
 if version.parse(torch.__version__) <= version.parse("1.13.2") or torch.cuda.is_available():
     CondFunc('ldm.modules.diffusionmodules.util.GroupNorm32.forward', lambda orig_func, self, *args, **kwargs: orig_func(self.float(), *args, **kwargs), unet_needs_upcast)