Upload sft_train.py with huggingface_hub

sft_train.py

@@ -378,6 +378,7 @@ def load_sft_config(config_path):
     cfg = {
         "auto_config":     raw.get("auto_config", True),
         "hf_model_repo":   raw.get("hf_model_repo", "ASTERIZER/LUNA-100M"),
+        "hf_model_file":   raw.get("hf_model_file", "latest.pt"),
         "hf_dataset_repo": raw.get("hf_dataset_repo", "ASTERIZER/Luna_Dataset"),
         "pretrained_ckpt": raw.get("pretrained_ckpt", "Base/out/pretrain/luna_100m/latest.pt"),
         "train_json":      raw.get("train_json", "Base/Datasets/sft_clean/train.json"),
@@ -467,15 +468,18 @@ def sft_train(cfg):
     if not ckpt_path.exists() and cfg.get("hf_model_repo"):
         # Auto-download from HuggingFace model repo
         print(f"\n  Pretrained checkpoint not found locally.")
-        print(f"  Downloading from HuggingFace: {cfg['hf_model_repo']}")
+        print(f"  Downloading from HuggingFace: {cfg['hf_model_repo']} ({cfg['hf_model_file']})")
         from huggingface_hub import hf_hub_download
         ckpt_path.parent.mkdir(parents=True, exist_ok=True)
-        hf_hub_download(
+        downloaded = hf_hub_download(
             repo_id=cfg["hf_model_repo"],
-            filename="latest.pt",
+            filename=cfg["hf_model_file"],
             local_dir=str(ckpt_path.parent),
             token=os.environ.get("HF_TOKEN"),
         )
+        downloaded_path = Path(downloaded)
+        if not ckpt_path.exists() and downloaded_path.exists():
+            ckpt_path = downloaded_path
         print(f"  Downloaded to: {ckpt_path}")
 
     if ckpt_path.exists():