Use login_and_download_hf_lm in finetuning path

Signed-off-by: Mamta Singh <[email protected]>

QEfficient/cloud/finetune.py

@@ -30,6 +30,7 @@
     get_preprocessed_dataset,
 )
 from QEfficient.finetune.utils.train_utils import get_longest_seq_length, print_model_size, train
+from QEfficient.utils._utils import login_and_download_hf_lm
 
 try:
     import torch_qaic  # noqa: F401
@@ -62,9 +63,9 @@ def main(**kwargs):
         # TODO: may have to init qccl backend, next try run with torchrun command
         torch_device = torch.device(device)
         assert torch_device.type != "cpu", "Host doesn't support single-node DDP"
-        assert torch_device.index is None, (
-            f"DDP requires specification of device type only, however provided device index as well: {torch_device}"
-        )
+        assert (
+            torch_device.index is None
+        ), f"DDP requires specification of device type only, however provided device index as well: {torch_device}"
         dist.init_process_group(backend=train_config.dist_backend)
         # from here onward "qaic/cuda" will automatically map to "qaic:i/cuda:i", where i = process rank
         getattr(torch, torch_device.type).set_device(dist.get_rank())
@@ -76,8 +77,9 @@ def main(**kwargs):
 
     # Load the pre-trained model and setup its configuration
     # config = AutoConfig.from_pretrained(train_config.model_name)
+    pretrained_model_path = login_and_download_hf_lm(train_config.model_name)
     model = AutoModelForCausalLM.from_pretrained(
-        train_config.model_name,
+        pretrained_model_path,
         use_cache=False,
         attn_implementation="sdpa",
         torch_dtype=torch.float16,