Speed up lora compute and lower memory usage by doing it in fp16. (#11161)

Author: comfyanonymous

comfy/model_management.py

@@ -1492,6 +1492,20 @@ def extended_fp16_support():
 
     return True
 
+LORA_COMPUTE_DTYPES = {}
+def lora_compute_dtype(device):
+    dtype = LORA_COMPUTE_DTYPES.get(device, None)
+    if dtype is not None:
+        return dtype
+
+    if should_use_fp16(device):
+        dtype = torch.float16
+    else:
+        dtype = torch.float32
+
+    LORA_COMPUTE_DTYPES[device] = dtype
+    return dtype
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:

comfy/model_patcher.py

@@ -614,10 +614,11 @@ def patch_weight_to_device(self, key, device_to=None, inplace_update=False):
         if key not in self.backup:
             self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
 
+        temp_dtype = comfy.model_management.lora_compute_dtype(device_to)
         if device_to is not None:
-            temp_weight = comfy.model_management.cast_to_device(weight, device_to, torch.float32, copy=True)
+            temp_weight = comfy.model_management.cast_to_device(weight, device_to, temp_dtype, copy=True)
         else:
-            temp_weight = weight.to(torch.float32, copy=True)
+            temp_weight = weight.to(temp_dtype, copy=True)
         if convert_func is not None:
             temp_weight = convert_func(temp_weight, inplace=True)