Skip to content

ModelPatcherDynamic: force load non leaf weights #12433

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
comfyanonymous merged 1 commit into from
Feb 13, 2026
Merged

ModelPatcherDynamic: force load non leaf weights #12433

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 10 additions & 9 deletions comfy/model_patcher.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -679,18 +679,19 @@ def unpin_all_weights(self):
for key in list(self.pinned):
self.unpin_weight(key)

def _load_list(self, prio_comfy_cast_weights=False):
def _load_list(self, prio_comfy_cast_weights=False, default_device=None):
loading = []
for n, m in self.model.named_modules():
params = []
skip = False
for name, param in m.named_parameters(recurse=False):
params.append(name)
default = False
params = { name: param for name, param in m.named_parameters(recurse=False) }
for name, param in m.named_parameters(recurse=True):
if name not in params:
skip = True # skip random weights in non leaf modules
default = True # default random weights in non leaf modules
break
if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
if default and default_device is not None:
for param in params.values():
param.data = param.data.to(device=default_device)
if not default and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
module_mem = comfy.model_management.module_size(m)
module_offload_mem = module_mem
if hasattr(m, "comfy_cast_weights"):
Expand Down Expand Up @@ -1495,7 +1496,7 @@ def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False
#with pin and unpin syncrhonization which can be expensive for small weights
#with a high layer rate (e.g. autoregressive LLMs).
#prioritize the non-comfy weights (note the order reverse).
loading = self._load_list(prio_comfy_cast_weights=True)
loading = self._load_list(prio_comfy_cast_weights=True, default_device=device_to)
loading.sort(reverse=True)

for x in loading:
Expand Down Expand Up @@ -1579,7 +1580,7 @@ def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=Fals
return 0 if vbar is None else vbar.free_memory(memory_to_free)

def partially_unload_ram(self, ram_to_unload):
loading = self._load_list(prio_comfy_cast_weights=True)
loading = self._load_list(prio_comfy_cast_weights=True, default_device=self.offload_device)
for x in loading:
_, _, _, _, m, _ = x
ram_to_unload -= comfy.pinned_memory.unpin_memory(m)
Expand Down