Upload folder using huggingface_hub

examples/prt14_qwen25vl/requirements.txt

@@ -2,10 +2,9 @@ transformers
 peft
 accelerate
 datasets
-torch
-torchvision
 pillow
 requests
+bitsandbytes
 tqdm
 qwen_vl_utils
 hydra-core

examples/prt14_qwen25vl/train_prt14.py

@@ -17,6 +17,7 @@ from transformers import (
     Trainer,
     TrainingArguments,
     get_scheduler,
+    BitsAndBytesConfig,
 )
 
 # Configure logging
@@ -59,6 +60,7 @@ class PRTTrainer(Trainer):
         if self.ref_model.device != device:
              self.ref_model.to(device)
 
+
         with torch.no_grad():
             ref_outputs = self.ref_model(**inputs)
             ref_logits = ref_outputs.logits
@@ -167,19 +169,27 @@ def main(cfg: DictConfig):
     # 2. Load Models (Heavy)
     logger.info(f"Loading Models: {cfg.model.model_id}")
     
-    # Reference Model (Frozen)
+    # Configure 4-bit quantization
+    bnb_config = BitsAndBytesConfig(
+        load_in_4bit=True,
+        bnb_4bit_quant_type="nf4",
+        bnb_4bit_compute_dtype=torch.bfloat16 if cfg.model.bf16 else torch.float16,
+        bnb_4bit_use_double_quant=True,
+    )
+
+    # Reference Model (Frozen, 4-bit)
     ref_model = Qwen2VLForConditionalGeneration.from_pretrained(
         cfg.model.model_id,
-        torch_dtype=torch.bfloat16 if cfg.model.bf16 else torch.float16,
-        device_map=None, 
+        quantization_config=bnb_config,
+        device_map={"": 0}, # Explicitly put on GPU 0
         trust_remote_code=True,
     )
     
-    # Reward Model (Trainable Base)
+    # Reward Model (Trainable Base, 4-bit)
     reward_model = Qwen2VLForConditionalGeneration.from_pretrained(
         cfg.model.model_id,
-        torch_dtype=torch.bfloat16 if cfg.model.bf16 else torch.float16,
-        device_map=None,
+        quantization_config=bnb_config,
+        device_map={"": 0}, # Explicitly put on GPU 0
         trust_remote_code=True,
     )
 
@@ -197,7 +207,9 @@ def main(cfg: DictConfig):
         reward_model.print_trainable_parameters()
     else:
         logger.info("Full Fine-Tuning Mode")
-        reward_model.gradient_checkpointing_enable()
+    
+    # Unconditionally enable gradient checkpointing for memory efficiency
+    reward_model.gradient_checkpointing_enable()
 
     # Define training arguments
     training_args = PRTTrainingArguments(