Wan seems to work with fp16.

Author: comfyanonymous

comfy/supported_models.py

@@ -911,7 +911,7 @@ class WAN21_T2V(supported_models_base.BASE):
 
     memory_usage_factor = 1.0
 
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     vae_key_prefix = ["vae."]
     text_encoder_key_prefix = ["text_encoders."]