Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

mm/mp: always unload re-used but modified models (#10724)

rattus18e7d6dba5f1
diff --git a/comfy/model_management.py b/comfy/model_management.py
index d8913082a..a21df54b3 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -503,10 +503,7 @@ class LoadedModel:
use_more_vram = lowvram_model_memory
if use_more_vram == 0:
use_more_vram = 1e32
- if use_more_vram > 0:
- self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights)
- else:
- self.model.partially_unload(self.model.offload_device, -use_more_vram, force_patch_weights=force_patch_weights)
+ self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights)
real_model = self.model.model
diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py
index 68b0a9192..cf1b0d441 100644
--- a/comfy/model_patcher.py
+++ b/comfy/model_patcher.py
@@ -928,6 +928,9 @@ class ModelPatcher:
extra_memory += (used - self.model.model_loaded_weight_memory)
self.patch_model(load_weights=False)
+ if extra_memory < 0 and not unpatch_weights:
+ self.partially_unload(self.offload_device, -extra_memory, force_patch_weights=force_patch_weights)
+ return 0
full_load = False
if self.model.model_lowvram == False and self.model.model_loaded_weight_memory > 0:
self.apply_hooks(self.forced_hooks, force_apply=True)