diff --git a/modules/comfy_ops.py b/modules/comfy_ops.py
index d11d4d9..8fdf9e1 100644
--- a/modules/comfy_ops.py
+++ b/modules/comfy_ops.py
@@ -24,16 +24,20 @@ def dequantize_block_weight(weight, scale_weight):
         dequantized_weight = dequantized_weight.view_as(weight)
     return dequantized_weight
 
-def block_scaled_fp8_ops(override_dtype=None, block_size=16):
+def block_scaled_fp8_ops(override_dtype=None, block_size=64):
     logging.info(f"Using block scaled fp8: fp8 dtype: {override_dtype}, block size: {block_size}")
+
     class block_scaled_fp8_op(manual_cast):
         class Linear(manual_cast.Linear):
             def __init__(self, *args, **kwargs):
+                # override Linear dtype if specified
                 if override_dtype is not None:
                     kwargs['dtype'] = override_dtype
                 super().__init__(*args, **kwargs)
 
+            # override Linear's reset_parameters
             def reset_parameters(self):
+                # add scale_weight parameter to load scale factor for each block
                 if not hasattr(self, 'scale_weight'):
                     out_features, in_features = self.weight.shape
                     if in_features % block_size != 0:
@@ -46,7 +50,8 @@ def block_scaled_fp8_ops(override_dtype=None, block_size=16):
                     )
 
                 return None
-            
+
+            # override manual_cast.Linear's forward_comfy_cast_weights
             def forward_comfy_cast_weights(self, input):
                 weight, bias = cast_bias_weight(self, input)
                 device, dtype = weight.device, weight.dtype
@@ -54,14 +59,16 @@ def block_scaled_fp8_ops(override_dtype=None, block_size=16):
 
                 return torch.nn.functional.linear(input, dequantize_block_weight(weight, self.scale_weight), bias)
 
+            # comfyui load lora will call convert_weight
             def convert_weight(self, weight, inplace=False, **kwargs):
                 scale_weight = self.scale_weight.to(device=weight.device, dtype=weight.dtype)
                 if inplace:
                     weight = dequantize_block_weight(weight, scale_weight)
                     return weight
                 else:
-                    return dequantize_block_weight(weight, self.scale_weight)
+                    return dequantize_block_weight(weight, scale_weight)
 
+            # comfyui load lora will call set_weight
             def set_weight(self, weight, inplace_update=False, seed=None, **kwargs):
                 scale_weight = self.scale_weight.to(device=weight.device, dtype=weight.dtype)
                 weight = comfy.float.stochastic_rounding(quantize_block_weight(weight, scale_weight), self.weight.dtype, seed=seed)
diff --git a/modules/nodes.py b/modules/nodes.py
index 6cd89ba..caaf5e6 100644
--- a/modules/nodes.py
+++ b/modules/nodes.py
@@ -11,8 +11,6 @@ from .comfy_ops import block_scaled_fp8_ops
 
 
 class MusubiUNETLoader:
-    """
-    """
     @classmethod
     def INPUT_TYPES(cls):
         return {