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.

Workaround for lora OOM on lowvram mode.

comfyanonymouscb7c4b4be3b3
diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py
index 430b59879..1ef49308f 100644
--- a/comfy/model_patcher.py
+++ b/comfy/model_patcher.py
@@ -348,8 +348,8 @@ class ModelPatcher:
m.comfy_cast_weights = True
else:
if hasattr(m, "weight"):
- self.patch_weight_to_device(weight_key, device_to)
- self.patch_weight_to_device(bias_key, device_to)
+ self.patch_weight_to_device(weight_key) #TODO: speed this up without causing OOM
+ self.patch_weight_to_device(bias_key)
m.to(device_to)
mem_counter += comfy.model_management.module_size(m)
logging.debug("lowvram: loaded module regularly {} {}".format(n, m))