To be really simple CheckpointLoaderSimple should pick the right type.

comfy/model_management.py

@@ -173,6 +173,30 @@ def maximum_batch_area():
     memory_free = get_free_memory() / (1024 * 1024)
     area = ((memory_free - 1024) * 0.9) / (0.6)
     return int(max(area, 0))
+
+def cpu_mode():
+    global vram_state
+    return vram_state == CPU
+
+def should_use_fp16():
+    if cpu_mode():
+        return False #TODO ?
+
+    if torch.cuda.is_bf16_supported():
+        return True
+
+    props = torch.cuda.get_device_properties()
+    if props.major < 7:
+        return False
+
+    #FP32 is faster on those cards?
+    nvidia_16_series = ["1660", "1650", "1630"]
+    for x in nvidia_16_series:
+        if x in props.name:
+            return False
+
+    return True
+
 #TODO: might be cleaner to put this somewhere else
 import threading
 

comfy/sd.py

@@ -656,12 +656,14 @@ def load_checkpoint(config_path, ckpt_path, output_vae=True, output_clip=True, e
     return (ModelPatcher(model), clip, vae)
 
 
-def load_checkpoint_guess_config(ckpt_path, fp16=False, output_vae=True, output_clip=True, embedding_directory=None):
+def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, embedding_directory=None):
     sd = load_torch_file(ckpt_path)
     sd_keys = sd.keys()
     clip = None
     vae = None
 
+    fp16 = model_management.should_use_fp16()
+
     class WeightsLoader(torch.nn.Module):
         pass