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.

Try to fix slow load issue on low ram hardware with pinned mem. (#10536)

comfyanonymous25de7b1bfa22
diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py
index 119119e96..74b9e48bc 100644
--- a/comfy/model_patcher.py
+++ b/comfy/model_patcher.py
@@ -658,6 +658,7 @@ class ModelPatcher:
loading = self._load_list()
load_completely = []
+ offloaded = []
loading.sort(reverse=True)
for x in loading:
n = x[1]
@@ -699,8 +700,7 @@ class ModelPatcher:
patch_counter += 1
cast_weight = True
- for param in params:
- self.pin_weight_to_device("{}.{}".format(n, param))
+ offloaded.append((module_mem, n, m, params))
else:
if hasattr(m, "comfy_cast_weights"):
wipe_lowvram_weight(m)
@@ -741,6 +741,12 @@ class ModelPatcher:
for x in load_completely:
x[2].to(device_to)
+ for x in offloaded:
+ n = x[1]
+ params = x[3]
+ for param in params:
+ self.pin_weight_to_device("{}.{}".format(n, param))
+
if lowvram_counter > 0:
logging.info("loaded partially {} {} {}".format(lowvram_model_memory / (1024 * 1024), mem_counter / (1024 * 1024), patch_counter))
self.model.model_lowvram = True