nodes/kijai-ComfyUI-WanVideoWrapper
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix uni3c reloading

Browse Source
This commit is contained in:
kijai
2025-12-14 16:48:55 +02:00
parent 0fa5383106
commit 3611341339
3 changed files with 6 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nodes_model_loading.py
+1 -1
View File
@@ -863,7 +863,7 @@ def load_weights(transformer, sd=None, weight_dtype=None, base_dtype=None,
except Exception: except Exception:
block_idx = None block_idx = None
if "loras" in name: if "loras" in name or "uni3c" in name:
continue continue
# GGUF: skip GGUFParameter params # GGUF: skip GGUFParameter params
nodes_sampler.py
+1 -1
View File
@@ -890,7 +890,7 @@ class WanVideoSampler:
#uni3c #uni3c
uni3c_data = uni3c_data_input = None uni3c_data = uni3c_data_input = None
if uni3c_embeds is not None: if uni3c_embeds is not None:
transformer.controlnet = uni3c_embeds["controlnet"] transformer.uni3c_controlnet = uni3c_embeds["controlnet"]
render_latent = uni3c_embeds["render_latent"].to(device) render_latent = uni3c_embeds["render_latent"].to(device)
if render_latent.shape != noise.shape: if render_latent.shape != noise.shape:
render_latent = torch.nn.functional.interpolate(render_latent, size=(noise.shape[1], noise.shape[2], noise.shape[3]), mode='trilinear', align_corners=False) render_latent = torch.nn.functional.interpolate(render_latent, size=(noise.shape[1], noise.shape[2], noise.shape[3]), mode='trilinear', align_corners=False)
wanvideo/modules/model.py
+4 -4
View File
@@ -2939,15 +2939,15 @@ class WanModel(torch.nn.Module):
if uni3c_data is not None: if uni3c_data is not None:
if (uni3c_data["start"] <= current_step_percentage <= uni3c_data["end"]) or \ if (uni3c_data["start"] <= current_step_percentage <= uni3c_data["end"]) or \
(uni3c_data["end"] > 0 and current_step == 0 and current_step_percentage >= uni3c_data["start"]): (uni3c_data["end"] > 0 and current_step == 0 and current_step_percentage >= uni3c_data["start"]):
self.controlnet.to(self.main_device) self.uni3c_controlnet.to(self.main_device)
with torch.autocast(device_type=mm.get_autocast_device(device), dtype=self.base_dtype, enabled=True): with torch.autocast(device_type=mm.get_autocast_device(device), dtype=self.base_dtype, enabled=True):
uni3c_controlnet_states = self.controlnet( uni3c_controlnet_states = self.uni3c_controlnet(
render_latent=render_latent.to(self.main_device, self.controlnet.dtype), render_latent=render_latent.to(self.main_device, self.uni3c_controlnet.dtype),
render_mask=uni3c_data["render_mask"], render_mask=uni3c_data["render_mask"],
camera_embedding=uni3c_data["camera_embedding"], camera_embedding=uni3c_data["camera_embedding"],
temb=e.to(self.main_device), temb=e.to(self.main_device),
device=self.offload_device) device=self.offload_device)
self.controlnet.to(self.offload_device) self.uni3c_controlnet.to(self.offload_device)
# Asynchronous block offloading with CUDA streams and events # Asynchronous block offloading with CUDA streams and events
if torch.cuda.is_available(): if torch.cuda.is_available():