fix: remove quantization_config after loading to bypass SFTTrainer check

training/configs/ornith-35b-lora.yaml

@@ -36,7 +36,7 @@ train_params:
   learning_rate: 2e-4
   lr_scheduler_type: cosine
   weight_decay: 0.01
-  warmup_ratio: 0.03
+  warmup_ratio: 0.03  # Will be converted to warmup_steps by TrainingArguments
   max_seq_length: 1024
   logging_steps: 10
   save_steps: 100

training/scripts/train.py

@@ -40,6 +40,8 @@ def train(config_path):
         device_map="cpu",  # Load to CPU first
         trust_remote_code=True,
     )
+    # Remove quantization config to avoid SFTTrainer validation error
+    model.config.quantization_config = None
     print("Model loaded and converted to bf16.")
     
     # Add LoRA