fix: use Ornith-1.0-35B bf16 base model (PEFT compatible)

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

@@ -1,7 +1,7 @@
 # LoRA Training Configuration for Llama-2-7b
 # Dataset: cyron_summary_lora_dataset (20k examples)
 
-base_model: /data/models/Ornith-1.0-35B-NVFP4
+base_model: /data/models/Ornith-1.0-35B
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 

training/scripts/train.py

@@ -32,17 +32,15 @@ def train(config_path):
     
     print(f"Loading model: {config['base_model']}")
     
-    # Load model and convert to bf16 (remove NVFP4 quantization)
+    # Load bf16 model
     print(f"Loading model: {config['base_model']}")
     model = AutoModelForCausalLM.from_pretrained(
         config["base_model"],
         torch_dtype=torch.bfloat16,
         device_map="cpu",  # Load to CPU first
         trust_remote_code=True,
-        # Override any quantization config (NVFP4 -> bf16)
-        _fast_init=False,
     )
-    print("Model loaded and converted to bf16.")
+    print("Model loaded.")
     
     # Add LoRA
     lora_config = LoraConfig(