From 6f45390be8307b1aa84093518c54896a95a8c2bb Mon Sep 17 00:00:00 2001
From: Christian Medina <37550954+cmedinasoriano@users.noreply.github.com>
Date: Fri, 3 Jul 2026 01:24:25 -0400
Subject: [PATCH] fix: remove quantization_config from config.json to prevent
 re-quantization

---
 train.py | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/train.py b/train.py
index 13f344e..472f836 100644
--- a/train.py
+++ b/train.py
@@ -64,6 +64,18 @@ def train(config_path):
                 json.dump(index, f)
             print(f"✓ Created index: {index_file} ({len(weight_map)} weights mapped)")
         
+        # Remove quantization_config from config.json to prevent re-quantization
+        import json as json_mod
+        config_json_path = Path(config["base_model"]) / "config.json"
+        if config_json_path.exists():
+            with open(config_json_path, 'r') as f:
+                config_data = json_mod.load(f)
+            if 'quantization_config' in config_data:
+                del config_data['quantization_config']
+                with open(config_json_path, 'w') as f:
+                    json_mod.dump(config_data, f)
+                print(f"✓ Removed quantization_config from config.json")
+        
         print(f"\n[INFO] Loading pre-quantized BnB 4-bit model...")
         model = AutoModelForCausalLM.from_pretrained(
             config["base_model"],