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.

More proper fix for the memory issue.

comfyanonymous2dda7c11a345
diff --git a/comfy/model_management.py b/comfy/model_management.py
index d5d77d7d7..6f667dfc5 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -314,6 +314,9 @@ class LoadedModel:
def model_memory(self):
return self.model.model_size()
+ def model_loaded_memory(self):
+ return self.model.loaded_size()
+
def model_offloaded_memory(self):
return self.model.model_size() - self.model.loaded_size()
@@ -504,8 +507,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
lowvram_model_memory = 0
if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM) and not force_full_load:
model_size = loaded_model.model_memory_required(torch_dev)
- current_free_mem = get_free_memory(torch_dev)
- lowvram_model_memory = max(1, (current_free_mem - minimum_memory_required), min(current_free_mem * 0.4, current_free_mem - minimum_inference_memory()))
+ loaded_memory = loaded_model.model_loaded_memory()
+ current_free_mem = get_free_memory(torch_dev) + loaded_memory
+ lowvram_model_memory = max(64 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * 0.4, current_free_mem - minimum_inference_memory()))
+ lowvram_model_memory = max(0.1, lowvram_model_memory - loaded_memory)
if model_size <= lowvram_model_memory: #only switch to lowvram if really necessary
lowvram_model_memory = 0