The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Move unet to device right after loading on highvram mode.
comfyanonymous62db11683b72
diff --git a/comfy/model_management.py b/comfy/model_management.pyindex 4e0e6a0ae..4f3f28571 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -216,6 +216,11 @@ current_gpu_controlnets = []model_accelerated = False+def unet_offload_device():+ if vram_state == VRAMState.HIGH_VRAM or vram_state == VRAMState.SHARED:+ return get_torch_device()+ else:+ return torch.device("cpu")def unload_model():global current_loaded_model@@ -228,10 +233,9 @@ def unload_model():accelerate.hooks.remove_hook_from_submodules(current_loaded_model.model)model_accelerated = False- #never unload models from GPU on high vram- if vram_state != VRAMState.HIGH_VRAM:- current_loaded_model.model.cpu()- current_loaded_model.model_patches_to("cpu")++ current_loaded_model.model.to(unet_offload_device())+ current_loaded_model.model_patches_to(unet_offload_device())current_loaded_model.unpatch_model()current_loaded_model = Nonediff --git a/comfy/sd.py b/comfy/sd.pyindex 52d016b10..542f704a6 100644--- a/comfy/sd.py+++ b/comfy/sd.py@@ -1142,6 +1142,7 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, oclipvision = clip_vision.load_clipvision_from_sd(sd, model_config.clip_vision_prefix, True)model = model_config.get_model(sd)+ model = model.to(model_management.unet_offload_device())model.load_model_weights(sd, "model.diffusion_model.")if output_vae: