Offload merge

README.md

@@ -212,6 +212,8 @@ Python 3.14 works but you may encounter issues with the torch compile node. The
 
 Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12
 
+torch 2.4 and above is supported but some features might only work on newer versions. We generally recommend using the latest major version of pytorch unless it is less than 2 weeks old.
+
 ### Instructions:
 
 Git clone this repo.

app/model_manager.py

@@ -44,7 +44,7 @@ async def get_model_folders(request):
         @routes.get("/experiment/models/{folder}")
         async def get_all_models(request):
             folder = request.match_info.get("folder", None)
-            if not folder in folder_paths.folder_names_and_paths:
+            if folder not in folder_paths.folder_names_and_paths:
                 return web.Response(status=404)
             files = self.get_model_file_list(folder)
             return web.json_response(files)
@@ -55,7 +55,7 @@ async def get_model_preview(request):
             path_index = int(request.match_info.get("path_index", None))
             filename = request.match_info.get("filename", None)
 
-            if not folder_name in folder_paths.folder_names_and_paths:
+            if folder_name not in folder_paths.folder_names_and_paths:
                 return web.Response(status=404)
 
             folders = folder_paths.folder_names_and_paths[folder_name]

comfy/clip_model.py

@@ -2,6 +2,25 @@
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.ops
 
+def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True):
+    image = image[:, :, :, :3] if image.shape[3] > 3 else image
+    mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
+    std = torch.tensor(std, device=image.device, dtype=image.dtype)
+    image = image.movedim(-1, 1)
+    if not (image.shape[2] == size and image.shape[3] == size):
+        if crop:
+            scale = (size / min(image.shape[2], image.shape[3]))
+            scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3]))
+        else:
+            scale_size = (size, size)
+
+        image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True)
+        h = (image.shape[2] - size)//2
+        w = (image.shape[3] - size)//2
+        image = image[:,:,h:h+size,w:w+size]
+    image = torch.clip((255. * image), 0, 255).round() / 255.0
+    return (image - mean.view([3,1,1])) / std.view([3,1,1])
+
 class CLIPAttention(torch.nn.Module):
     def __init__(self, embed_dim, heads, dtype, device, operations):
         super().__init__()

comfy/clip_vision.py

@@ -1,6 +1,5 @@
 from .utils import load_torch_file, transformers_convert, state_dict_prefix_replace
 import os
-import torch
 import json
 import logging
 
@@ -17,24 +16,7 @@ def __getitem__(self, key):
     def __setitem__(self, key, item):
         setattr(self, key, item)
 
-def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True):
-    image = image[:, :, :, :3] if image.shape[3] > 3 else image
-    mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
-    std = torch.tensor(std, device=image.device, dtype=image.dtype)
-    image = image.movedim(-1, 1)
-    if not (image.shape[2] == size and image.shape[3] == size):
-        if crop:
-            scale = (size / min(image.shape[2], image.shape[3]))
-            scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3]))
-        else:
-            scale_size = (size, size)
-
-        image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True)
-        h = (image.shape[2] - size)//2
-        w = (image.shape[3] - size)//2
-        image = image[:,:,h:h+size,w:w+size]
-    image = torch.clip((255. * image), 0, 255).round() / 255.0
-    return (image - mean.view([3,1,1])) / std.view([3,1,1])
+clip_preprocess = comfy.clip_model.clip_preprocess  # Prevent some stuff from breaking, TODO: remove eventually
 
 IMAGE_ENCODERS = {
     "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
@@ -73,7 +55,7 @@ def get_sd(self):
 
     def encode_image(self, image, crop=True):
         comfy.model_management.load_model_gpu(self.patcher)
-        pixel_values = clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
+        pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
         out = self.model(pixel_values=pixel_values, intermediate_output='all' if self.return_all_hidden_states else -2)
 
         outputs = Output()

comfy/context_windows.py

@@ -188,6 +188,12 @@ def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: Inde
                                 audio_cond = cond_value.cond
                                 if audio_cond.ndim > 1 and audio_cond.size(1) == x_in.size(self.dim):
                                     new_cond_item[cond_key] = cond_value._copy_with(window.get_tensor(audio_cond, device, dim=1))
+                            # Handle vace_context (temporal dim is 3)
+                            elif cond_key == "vace_context" and hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
+                                vace_cond = cond_value.cond
+                                if vace_cond.ndim >= 4 and vace_cond.size(3) == x_in.size(self.dim):
+                                    sliced_vace = window.get_tensor(vace_cond, device, dim=3, retain_index_list=self.cond_retain_index_list)
+                                    new_cond_item[cond_key] = cond_value._copy_with(sliced_vace)
                             # if has cond that is a Tensor, check if needs to be subset
                             elif hasattr(cond_value, "cond") and isinstance(cond_value.cond, torch.Tensor):
                                 if  (self.dim < cond_value.cond.ndim and cond_value.cond.size(self.dim) == x_in.size(self.dim)) or \

comfy/hooks.py

@@ -527,7 +527,8 @@ def prepare_current_keyframe(self, curr_t: float, transformer_options: dict[str,
                         if self._current_keyframe.get_effective_guarantee_steps(max_sigma) > 0:
                             break
                     # if eval_c is outside the percent range, stop looking further
-                    else: break
+                    else:
+                        break
         # update steps current context is used
         self._current_used_steps += 1
         # update current timestep this was performed on

comfy/k_diffusion/sampling.py

@@ -74,6 +74,9 @@ def get_ancestral_step(sigma_from, sigma_to, eta=1.):
 
 def default_noise_sampler(x, seed=None):
     if seed is not None:
+        if x.device == torch.device("cpu"):
+            seed += 1
+
         generator = torch.Generator(device=x.device)
         generator.manual_seed(seed)
     else:

comfy/latent_formats.py

@@ -407,6 +407,9 @@ def __init__(self):
 
         self.latent_rgb_factors_bias = [-0.0571, -0.1657, -0.2512]
 
+class LTXAV(LTXV):
+    pass
+
 class HunyuanVideo(LatentFormat):
     latent_channels = 16
     latent_dimensions = 3

comfy/ldm/chroma_radiance/model.py

@@ -270,7 +270,7 @@ def radiance_get_override_params(self, overrides: dict) -> ChromaRadianceParams:
         bad_keys = tuple(
             k
             for k, v in overrides.items()
-            if type(v) != type(getattr(params, k)) and (v is not None or k not in nullable_keys)
+            if not isinstance(v, type(getattr(params, k))) and (v is not None or k not in nullable_keys)
         )
         if bad_keys:
             e = f"Invalid value(s) in transformer_options chroma_radiance_options: {', '.join(bad_keys)}"

comfy/ldm/hunyuan_video/upsampler.py

@@ -3,7 +3,8 @@
 import torch.nn.functional as F
 from comfy.ldm.modules.diffusionmodules.model import ResnetBlock, VideoConv3d
 from comfy.ldm.hunyuan_video.vae_refiner import RMS_norm
-import model_management, model_patcher
+import model_management
+import model_patcher
 
 class SRResidualCausalBlock3D(nn.Module):
     def __init__(self, channels: int):