Bump ComfyUI version to v0.3.12

comfy/ldm/cosmos/model.py

@@ -293,7 +293,7 @@ class GeneralDIT(nn.Module):
         x_B_T_H_W_D = self.x_embedder(x_B_C_T_H_W)
 
         if self.extra_per_block_abs_pos_emb:
-            extra_pos_emb = self.extra_pos_embedder(x_B_T_H_W_D, fps=fps, device=x_B_C_T_H_W.device)
+            extra_pos_emb = self.extra_pos_embedder(x_B_T_H_W_D, fps=fps, device=x_B_C_T_H_W.device, dtype=x_B_C_T_H_W.dtype)
         else:
             extra_pos_emb = None
 

comfy/ldm/cosmos/position_embedding.py

@@ -41,12 +41,12 @@ def normalize(x: torch.Tensor, dim: Optional[List[int]] = None, eps: float = 0)
 
 
 class VideoPositionEmb(nn.Module):
-    def forward(self, x_B_T_H_W_C: torch.Tensor, fps=Optional[torch.Tensor], device=None) -> torch.Tensor:
+    def forward(self, x_B_T_H_W_C: torch.Tensor, fps=Optional[torch.Tensor], device=None, dtype=None) -> torch.Tensor:
         """
         It delegates the embedding generation to generate_embeddings function.
         """
         B_T_H_W_C = x_B_T_H_W_C.shape
-        embeddings = self.generate_embeddings(B_T_H_W_C, fps=fps, device=device)
+        embeddings = self.generate_embeddings(B_T_H_W_C, fps=fps, device=device, dtype=dtype)
 
         return embeddings
 
@@ -104,6 +104,7 @@ class VideoRopePosition3DEmb(VideoPositionEmb):
         w_ntk_factor: Optional[float] = None,
         t_ntk_factor: Optional[float] = None,
         device=None,
+        dtype=None,
     ):
         """
         Generate embeddings for the given input size.
@@ -189,13 +190,12 @@ class LearnablePosEmbAxis(VideoPositionEmb):
         self.pos_emb_w = nn.Parameter(torch.empty(len_w, model_channels, device=device, dtype=dtype))
         self.pos_emb_t = nn.Parameter(torch.empty(len_t, model_channels, device=device, dtype=dtype))
 
-
+    def generate_embeddings(self, B_T_H_W_C: torch.Size, fps=Optional[torch.Tensor], device=None, dtype=None) -> torch.Tensor:
-    def generate_embeddings(self, B_T_H_W_C: torch.Size, fps=Optional[torch.Tensor], device=None) -> torch.Tensor:
         B, T, H, W, _ = B_T_H_W_C
         if self.interpolation == "crop":
-            emb_h_H = self.pos_emb_h[:H].to(device=device)
+            emb_h_H = self.pos_emb_h[:H].to(device=device, dtype=dtype)
-            emb_w_W = self.pos_emb_w[:W].to(device=device)
+            emb_w_W = self.pos_emb_w[:W].to(device=device, dtype=dtype)
-            emb_t_T = self.pos_emb_t[:T].to(device=device)
+            emb_t_T = self.pos_emb_t[:T].to(device=device, dtype=dtype)
             emb = (
                 repeat(emb_t_T, "t d-> b t h w d", b=B, h=H, w=W)
                 + repeat(emb_h_H, "h d-> b t h w d", b=B, t=T, w=W)

comfy/ldm/cosmos/vae.py

@@ -18,6 +18,7 @@ import logging
 import torch
 from torch import nn
 from enum import Enum
+import math
 
 from .cosmos_tokenizer.layers3d import (
     EncoderFactorized,
@@ -105,17 +106,23 @@ class CausalContinuousVideoTokenizer(nn.Module):
         z, posteriors = self.distribution(moments)
         latent_ch = z.shape[1]
         latent_t = z.shape[2]
-        dtype = z.dtype
+        in_dtype = z.dtype
-        mean = self.latent_mean.view(latent_ch, -1)[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=dtype, device=z.device)
+        mean = self.latent_mean.view(latent_ch, -1)
-        std = self.latent_std.view(latent_ch, -1)[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=dtype, device=z.device)
+        std = self.latent_std.view(latent_ch, -1)
+
+        mean = mean.repeat(1, math.ceil(latent_t / mean.shape[-1]))[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=in_dtype, device=z.device)
+        std = std.repeat(1, math.ceil(latent_t / std.shape[-1]))[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=in_dtype, device=z.device)
         return ((z - mean) / std) * self.sigma_data
 
     def decode(self, z):
         in_dtype = z.dtype
         latent_ch = z.shape[1]
         latent_t = z.shape[2]
-        mean = self.latent_mean.view(latent_ch, -1)[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=in_dtype, device=z.device)
+        mean = self.latent_mean.view(latent_ch, -1)
-        std = self.latent_std.view(latent_ch, -1)[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=in_dtype, device=z.device)
+        std = self.latent_std.view(latent_ch, -1)
+
+        mean = mean.repeat(1, math.ceil(latent_t / mean.shape[-1]))[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=in_dtype, device=z.device)
+        std = std.repeat(1, math.ceil(latent_t / std.shape[-1]))[:, : latent_t].reshape([1, latent_ch, -1, 1, 1]).to(dtype=in_dtype, device=z.device)
 
         z = z / self.sigma_data
         z = z * std + mean

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.11"
+__version__ = "0.3.12"

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.11"
+version = "0.3.12"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"