From 167dbc6411a1caff76e0453f9f8ddaa8c7a454d3 Mon Sep 17 00:00:00 2001 From: lllyasviel Date: Thu, 22 Feb 2024 06:31:21 -0800 Subject: [PATCH] safe value for new memory peak --- ldm_patched/modules/model_management.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/ldm_patched/modules/model_management.py b/ldm_patched/modules/model_management.py index 30f53f07..c1447cc9 100644 --- a/ldm_patched/modules/model_management.py +++ b/ldm_patched/modules/model_management.py @@ -435,7 +435,7 @@ def load_models_gpu(models, memory_required=0): 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 - 1024 * (1024 * 1024)) / 1.3 )) + lowvram_model_memory = int(max(64 * (1024 * 1024), (current_free_mem - 1.5 * 1024 * (1024 * 1024)) / 1.3)) if model_size > (current_free_mem - inference_memory): #only switch to lowvram if really necessary vram_set_state = VRAMState.LOW_VRAM else: @@ -443,9 +443,6 @@ def load_models_gpu(models, memory_required=0): if vram_set_state == VRAMState.NO_VRAM: lowvram_model_memory = 64 * 1024 * 1024 - - # TODO: New offload system seems to have unpredicted risk to OOM on some 2GB devices - # maybe lowvram_model_memory *= 1.5 later if user reports that cur_loaded_model = loaded_model.model_load(lowvram_model_memory) current_loaded_models.insert(0, loaded_model)