The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Make ComfyUI split batches a higher priority than weight offload.
comfyanonymousd965474aaae2
diff --git a/comfy/model_management.py b/comfy/model_management.pyindex b4f32d648..da0b989a8 100644--- a/comfy/model_management.py+++ b/comfy/model_management.py@@ -379,11 +379,15 @@ def free_memory(memory_required, device, keep_loaded=[]):if mem_free_torch > mem_free_total * 0.25:soft_empty_cache()-def load_models_gpu(models, memory_required=0, force_patch_weights=False):+def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimum_memory_required=None):global vram_stateinference_memory = minimum_inference_memory()extra_mem = max(inference_memory, memory_required)+ if minimum_memory_required is None:+ minimum_memory_required = extra_mem+ else:+ minimum_memory_required = max(inference_memory, minimum_memory_required)models = set(models)@@ -446,8 +450,8 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False):if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):model_size = loaded_model.model_memory_required(torch_dev)current_free_mem = get_free_memory(torch_dev)- lowvram_model_memory = int(max(64 * (1024 * 1024), (current_free_mem - extra_mem)))- if model_size <= (current_free_mem - inference_memory): #only switch to lowvram if really necessary+ lowvram_model_memory = int(max(64 * (1024 * 1024), (current_free_mem - minimum_memory_required)))+ if model_size <= lowvram_model_memory: #only switch to lowvram if really necessarylowvram_model_memory = 0if vram_set_state == VRAMState.NO_VRAM:diff --git a/comfy/sampler_helpers.py b/comfy/sampler_helpers.pyindex a18abd9e9..4a2ec123b 100644--- a/comfy/sampler_helpers.py+++ b/comfy/sampler_helpers.py@@ -61,7 +61,9 @@ def prepare_sampling(model, noise_shape, conds):device = model.load_devicereal_model = Nonemodels, inference_memory = get_additional_models(conds, model.model_dtype())- comfy.model_management.load_models_gpu([model] + models, model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:])) + inference_memory)+ memory_required = model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:])) + inference_memory+ minimum_memory_required = model.memory_required([noise_shape[0]] + list(noise_shape[1:])) + inference_memory+ comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required)real_model = model.modelreturn real_model, conds, models