Download the models without instantiating them.

app.py

@@ -1,4 +1,5 @@
 import gradio
+import huggingface_hub
 import spaces
 
 import psaiops.common.model
@@ -9,12 +10,13 @@ import psaiops.score.human.app as _app
 # META #########################################################################
 
 # additional args used when loading the model
-_CONFIG = {}
+_MODEL_CFG = {}
+_LOAD_CFG = {'repo_type': 'model', 'ignore_patterns': ['*.onnx', '*.tflite', '*.msgpack'],}
 _app.MODELS = ['qwen/qwen3.5-9b', 'qwen/qwen3.5-27b']
 
 # frontload the models on the CPU to avoid downloading them from the GPU slot
 for __m in _app.MODELS:
-    psaiops.common.model.get_model(name=__m, device='cpu', **_CONFIG)
+    huggingface_hub.snapshot_download(repo_id=__m, **_LOAD_CFG)
 
 # but do not instantiate unless necessary
 _MODEL = None
@@ -34,7 +36,7 @@ def fetch_model() -> object:
     global _MODEL
     # control when the model is downloaded to avoid moving it to the CPU
     if _MODEL is None:
-        _MODEL = psaiops.common.model.get_model(name=_app.MODELS[0], device='cuda', **_CONFIG)
+        _MODEL = psaiops.common.model.get_model(name=_app.MODELS[0], device='cuda', **_MODEL_CFG)
     # tuple of objects or (None, None)
     return _MODEL
 

requirements.txt

@@ -1,5 +1,6 @@
 accelerate>=1.11
 gradio>=6.0
+huggingface_hub>=1.5
 kernels>=0.11
 psaiops>=0.10.3
 spaces>=0.47