Isolate GTE startup packing from completed 0.6B model

app.py

@@ -27,11 +27,10 @@ TASK = (
 
 
 tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, padding_side="left")
-model = AutoModel.from_pretrained(
-    MODEL_PATH,
-    torch_dtype=torch.bfloat16,
-    trust_remote_code=True,
-).to("cuda").eval()
+# The completed 0.6B phase is intentionally not resident while the larger GTE
+# startup path is validated.  Packing both models together exceeded the
+# current Space startup envelope at 6.16 GB.
+model = None
 
 # ZeroGPU optimizes CUDA placements made during module startup.  The previous
 # larger-family path loaded this model inside the decorated call and exhausted
@@ -90,6 +89,8 @@ def retrieval_metrics(queries: np.ndarray, documents: np.ndarray) -> tuple[dict,
 
 @spaces.GPU(duration=300)
 def run_bf16_control() -> dict:
+    if model is None:
+        raise RuntimeError("0.6B BF16 control is offline during the GTE phase")
     pairs = json.loads(INPUT_PATH.read_text())
     query_texts = [detailed_instruction(item["query"]) for item in pairs]
     document_texts = [item["document"] for item in pairs]