Disable unused KV cache for embedding controls

app.py

@@ -56,7 +56,7 @@ def encode_with(active_model, text: str, active_tokenizer=None) -> np.ndarray:
     )
     batch = {key: value.to("cuda") for key, value in batch.items()}
     with torch.inference_mode():
-        output = active_model(**batch)
+        output = active_model(**batch, use_cache=False)
         vector = F.normalize(output.last_hidden_state[:, -1, :].float(), p=2, dim=-1)
     return vector[0].cpu().numpy()