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.

Cap lowvram to half of free memory.

comfyanonymous1e68002b87a3
diff --git a/comfy/model_management.py b/comfy/model_management.py
index c0fb15095..2008229f2 100644
--- a/comfy/model_management.py
+++ b/comfy/model_management.py
@@ -450,7 +450,8 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
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 - minimum_memory_required)))
+ lowvram_model_memory = max(64 * (1024 * 1024), (current_free_mem - minimum_memory_required))
+ lowvram_model_memory = int(min(current_free_mem * 0.5, lowvram_model_memory))
if model_size <= lowvram_model_memory: #only switch to lowvram if really necessary
lowvram_model_memory = 0