Upload train_lora.py with huggingface_hub

train_lora.py

@@ -95,7 +95,7 @@ def main():
         "attn_implementation": "flash_attention_2",
     }
 
-    if USE_4BIT:
+    if USE_4BIT and "deepseek" not in MODEL_NAME.lower():
         print("  Using 4-bit quantization (QLoRA)")
         bnb_config = BitsAndBytesConfig(
             load_in_4bit=True,
@@ -113,7 +113,7 @@ def main():
 
     model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, **model_kwargs)
 
-    if USE_4BIT:
+    if USE_4BIT and "deepseek" not in MODEL_NAME.lower():
         model = prepare_model_for_kbit_training(model)
 
     # ---- LoRA config ----