diff --git a/openllm-python/pyproject.toml b/openllm-python/pyproject.toml
index 4a4bb9d3..3c9f97c7 100644
--- a/openllm-python/pyproject.toml
+++ b/openllm-python/pyproject.toml
@@ -109,6 +109,7 @@ fine-tune = [
   "accelerate",
   "trl",
   "scipy",
+  "huggingface-hub",
 ]
 flan-t5 = ["transformers>=4.34.0"]
 full = [
diff --git a/openllm-python/src/openllm/_llm.py b/openllm-python/src/openllm/_llm.py
index 5aedae22..739dae82 100644
--- a/openllm-python/src/openllm/_llm.py
+++ b/openllm-python/src/openllm/_llm.py
@@ -80,12 +80,7 @@ def _resolve_peft_config_type(adapter_map: dict[str, str]) -> AdapterMap:
     raise RuntimeError(
       "LoRA adapter requires 'peft' to be installed. Make sure to do 'pip install \"openllm[fine-tune]\"'"
     )
-  try:
-    from huggingface_hub import hf_hub_download
-  except ImportError:
-    raise MissingDependencyError(
-      "Failed to import 'huggingface_hub'. Make sure to do 'pip install \"openllm[fine-tune]\"'"
-    ) from None
+  from huggingface_hub import hf_hub_download
 
   resolved: AdapterMap = {}
   for path_or_adapter_id, name in adapter_map.items():
diff --git a/tools/dependencies.py b/tools/dependencies.py
index ce55507c..096cb6a6 100755
--- a/tools/dependencies.py
+++ b/tools/dependencies.py
@@ -159,7 +159,7 @@ _BASE_DEPENDENCIES = [
   Dependencies(name='bitsandbytes', upper_constraint='0.42'),  # 0.41 works with CUDA 11.8
 ]
 
-FINE_TUNE_DEPS = ['peft>=0.6.0', 'bitsandbytes', 'datasets', 'accelerate', 'trl', 'scipy']
+FINE_TUNE_DEPS = ['peft>=0.6.0', 'bitsandbytes', 'datasets', 'accelerate', 'trl', 'scipy', 'huggingface-hub']
 FLAN_T5_DEPS = ['transformers>=4.34.0']
 OPT_DEPS = ['transformers>=4.34.0']
 GRPC_DEPS = ['openllm-client[grpc]']