Update app.py

app.py

@@ -1,4 +1,5 @@
 import os
+import traceback
 from fastapi import FastAPI, HTTPException
 from pydantic import BaseModel
 from huggingface_hub import hf_hub_download
@@ -24,19 +25,22 @@ def get_model():
         return _model
 
     print("📥 Downloading model...")
-    local_path = hf_hub_download(
+    local_file = hf_hub_download(
         repo_id=REPO_ID,
         filename=FILENAME,
         cache_dir=CACHE_DIR,
         local_dir_use_symlinks=False,
     )
-    print("✅ Model downloaded at", local_path)
+    local_dir = os.path.dirname(local_file)
+    print("✅ Model downloaded at", local_file)
 
     print("🔄 Loading model into memory...")
     _model = AutoModelForCausalLM.from_pretrained(
-        local_path,
+        local_dir,
+        model_file=FILENAME,
         model_type=MODEL_TYPE,
-        gpu_layers=0  # 0 = CPU only
+        gpu_layers=0,  # CPU only
+        threads=os.cpu_count() or 2
     )
     print("✅ Model loaded")
     return _model
@@ -57,6 +61,17 @@ def generate_text(req: PromptRequest):
             max_new_tokens=req.max_new_tokens,
             temperature=req.temperature
         )
-        return {"response": output}
+        return {"ok": True, "response": output}
     except Exception as e:
-        raise HTTPException(status_code=500, detail=str(e))
+        tb = traceback.format_exc()
+        print("❌ ERROR in /generate:\n", tb)
+        raise HTTPException(status_code=500, detail={"error": str(e), "trace": tb})
+
+# --- Health Check ---
+@app.get("/health")
+def health():
+    try:
+        _ = get_model()
+        return {"ok": True}
+    except Exception as e:
+        return {"ok": False, "error": str(e)}