Always use fp16 for the text encoders.

comfy/model_management.py

@@ -546,10 +546,8 @@ def text_encoder_dtype(device=None):
     if is_device_cpu(device):
         return torch.float16
 
-    if should_use_fp16(device, prioritize_performance=False):
-        return torch.float16
-    else:
-        return torch.float32
+    return torch.float16
+
 
 def intermediate_device():
     if args.gpu_only: