Fix HF auth: use login() to authenticate globally from HF_TOKEN secret

app.py

@@ -9,16 +9,21 @@ from fastapi import FastAPI, File, UploadFile, Form
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import RedirectResponse
 from PIL import Image
-from huggingface_hub import snapshot_download
+from huggingface_hub import snapshot_download, login
 from transformers import AutoProcessor, AutoModelForImageTextToText
 
 HF_TOKEN = os.environ.get("HF_TOKEN")
+if HF_TOKEN:
+    login(token=HF_TOKEN)
+    print("Authenticated with HF token.", flush=True)
+else:
+    print("WARNING: HF_TOKEN not set — gated models will fail.", flush=True)
 
 # ---------------------------------------------------------------------------
 # MedImageInsight — CLIP-style encoder for zero-shot label scoring
 # ---------------------------------------------------------------------------
 print("Downloading MedImageInsights repo...", flush=True)
-repo_path = snapshot_download("lion-ai/MedImageInsights", token=HF_TOKEN)
+repo_path = snapshot_download("lion-ai/MedImageInsights")
 print(f"Downloaded to: {repo_path}", flush=True)
 
 sys.path.insert(0, repo_path)
@@ -41,14 +46,13 @@ print("MedImageInsight ready.", flush=True)
 MEDGEMMA_ID = "google/medgemma-1.5-4b-it"
 
 print("Loading MedGemma processor...", flush=True)
-gemma_processor = AutoProcessor.from_pretrained(MEDGEMMA_ID, token=HF_TOKEN)
+gemma_processor = AutoProcessor.from_pretrained(MEDGEMMA_ID)
 
 print("Loading MedGemma model (bfloat16)...", flush=True)
 gemma_model = AutoModelForImageTextToText.from_pretrained(
     MEDGEMMA_ID,
     torch_dtype=torch.bfloat16,
     device_map="auto",
-    token=HF_TOKEN,
 )
 gemma_model.eval()
 print("MedGemma ready.", flush=True)