Improve model download performance on HF Spaces

- Enable HF_HUB_ENABLE_HF_TRANSFER for faster downloads
- Add informative logging about download size and expected time
- Helps prevent timeout perception on T4 instances with slower network
- Model download is ~1GB and may take several minutes on first run

app.py

@@ -21,6 +21,9 @@ from mosaic.gradio_app import download_and_process_models
 from mosaic.ui import launch_gradio
 
 if __name__ == "__main__":
+    # Enable HF transfer for faster downloads on Spaces
+    os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
+    
     # Install flash-attn for H100 acceleration (if on Spaces)
     if os.environ.get("SPACE_ID"):
         try:

src/mosaic/gradio_app.py

@@ -66,6 +66,7 @@ def download_and_process_models():
     # Pre-download Optimus model from bioptimus/H-optimus-0
     # This ensures it's cached at startup since it's needed for every slide
     logger.info("Pre-downloading Optimus model from bioptimus/H-optimus-0...")
+    logger.info("This may take several minutes on first run - downloading ~1GB model...")
     from mussel.models import ModelType, get_model_factory
 
     optimus_factory = get_model_factory(ModelType.OPTIMUS)
@@ -75,7 +76,7 @@ def download_and_process_models():
         use_gpu=False,  # Just download, don't load to GPU yet
         gpu_device_id=None,
     )
-    logger.info("✓ Optimus model cached")
+    logger.info("✓ Optimus model cached successfully")
 
     model_map = pd.read_csv(
         Path(cache_dir) / "paladin_model_map.csv",