Update rp.py

rp.py

@@ -16,10 +16,11 @@ def reduce_key_size(input_file, output_file, reduction_factor=0.50):
         if new_size > 0:  # Ensure new size is positive
             reduced_tensor = original_tensor[:new_size]
             
-            # 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)
+            # Convert to FP32 before quantizing
+            fp32_tensor = reduced_tensor.to(torch.float32)
+            
+            # Quantize the tensor to 8-bit integer format (or FP8 if available)
+            fp8_tensor = torch.quantize_per_tensor(fp32_tensor, scale=1.0, zero_point=0, dtype=torch.qint8)
             model_data[key] = fp8_tensor
     
     # Save the modified model