Upload folder using huggingface_hub

examples/prt14_qwen25vl/train_prt14.py

@@ -158,7 +158,13 @@ def main(cfg: DictConfig):
     
     # 0. Load processor (Lightweight, needed for dataset)
     logger.info(f"Loading Processor: {cfg.model.model_id}")
-    processor = AutoProcessor.from_pretrained(cfg.model.model_id, trust_remote_code=True)
+    # Strictly limit resolution for T4 (Avoids OOM in SDPA)
+    processor = AutoProcessor.from_pretrained(
+        cfg.model.model_id, 
+        trust_remote_code=True,
+        min_pixels=256*256,
+        max_pixels=640*640, # Restrict to ~640px to reduce token count drastically
+    )
     dataset_cls = PRTDataset
 
     # 1. Load dataset (Fail fast)