Only enable fp16 on z image models that actually support it. (Comfy-Org#12065)

Author: comfyanonymous

comfy/ldm/lumina/model.py

@@ -451,6 +451,7 @@ def __init__(
         device=None,
         dtype=None,
         operations=None,
+        **kwargs,
     ) -> None:
         super().__init__()
         self.dtype = dtype

comfy/model_detection.py

@@ -444,6 +444,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["ffn_dim_multiplier"] = (8.0 / 3.0)
             dit_config["z_image_modulation"] = True
             dit_config["time_scale"] = 1000.0
+            try:
+                dit_config["allow_fp16"] = torch.std(state_dict['{}layers.{}.ffn_norm1.weight'.format(key_prefix, dit_config["n_layers"] - 2)], unbiased=False).item() < 0.42
+            except Exception:
+                pass
             if '{}cap_pad_token'.format(key_prefix) in state_dict_keys:
                 dit_config["pad_tokens_multiple"] = 32
             sig_weight = state_dict.get('{}siglip_embedder.0.weight'.format(key_prefix), None)

comfy/supported_models.py

@@ -1093,7 +1093,7 @@ class ZImage(Lumina2):
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
-        if comfy.model_management.extended_fp16_support():
+        if comfy.model_management.extended_fp16_support() and unet_config.get("allow_fp16", False):
             self.supported_inference_dtypes = self.supported_inference_dtypes.copy()
             self.supported_inference_dtypes.insert(1, torch.float16)