Update rp.py

rp.py

@@ -1,7 +1,7 @@
 import torch
 from safetensors.torch import load_file, save_file
 
-def reduce_key_size(input_file, output_file, reduction_factor=0.30):
+def reduce_key_size(input_file, output_file, reduction_factor=0.50):
     # Load the model
     model_data = load_file(input_file)
 
@@ -15,12 +15,17 @@ def reduce_key_size(input_file, output_file, reduction_factor=0.30):
         # Resize the tensor (this could vary depending on your requirements)
         if new_size > 0:  # Ensure new size is positive
             reduced_tensor = original_tensor[:new_size]
-            model_data[key] = reduced_tensor
+            
+            # Convert to FP8 (assuming your environment supports FP8)
+            # Note: PyTorch does not have built-in FP8 support; you may need to use a custom implementation
+            # Here's an example of converting a tensor to float16, then quantizing it
+            fp8_tensor = torch.quantize_per_tensor(reduced_tensor.to(torch.float16), scale=1.0, zero_point=0, dtype=torch.qint8)
+            model_data[key] = fp8_tensor
     
     # Save the modified model
     save_file(model_data, output_file)
 
 # Usage example
-input_file = 'merged_model2.safetensors'  # Replace with your input model file
-output_file = 'merged_model_216.safetensors'  # Desired output file name
+input_file = 'merged_model_16.safetensors'  # Replace with your input model file
+output_file = 'merged_model_8.safetensors'  # Desired output file name
 reduce_key_size(input_file, output_file)