The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Fix some issues with TE being in lowvram mode.
comfyanonymous74e124f4d784
diff --git a/comfy/model_management.py b/comfy/model_management.pyindex fdf3308ba..152a76f37 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -434,7 +434,7 @@ def free_memory(memory_required, device, keep_loaded=[]):soft_empty_cache()return unloaded_models-def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None):+def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None, force_full_load=False):global vram_stateinference_memory = minimum_inference_memory()@@ -513,7 +513,7 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimuelse:vram_set_state = vram_statelowvram_model_memory = 0- if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):+ 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(64 * (1024 * 1024), (current_free_mem - minimum_memory_required), min(current_free_mem * 0.4, current_free_mem - minimum_inference_memory()))diff --git a/comfy/model_patcher.py b/comfy/model_patcher.pyindex 39d00d82e..1edbf24ab 100644--- a/comfy/model_patcher.py+++ b/comfy/model_patcher.py@@ -411,7 +411,7 @@ class ModelPatcher:logging.debug("lowvram: loaded module regularly {} {}".format(n, m))if lowvram_counter > 0:- logging.info("loaded partially {} {}".format(lowvram_model_memory / (1024 * 1024), patch_counter))+ logging.info("loaded partially {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), patch_counter))self.model.model_lowvram = Trueelse:logging.info("loaded completely {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024)))diff --git a/comfy/sd.py b/comfy/sd.pyindex 13909d674..edd0b51d8 100644--- a/comfy/sd.py+++ b/comfy/sd.py@@ -86,7 +86,7 @@ class CLIP:self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)if params['device'] == load_device:- model_management.load_model_gpu(self.patcher)+ model_management.load_models_gpu([self.patcher], force_full_load=True)self.layer_idx = Nonelogging.debug("CLIP model load device: {}, offload device: {}, current: {}".format(load_device, offload_device, params['device']))@@ -585,7 +585,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_cmodel_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device())if inital_load_device != torch.device("cpu"):logging.info("loaded straight to GPU")- model_management.load_model_gpu(model_patcher)+ model_management.load_models_gpu([model_patcher], force_full_load=True)return (model_patcher, clip, vae, clipvision)