Disable caching in model generation

Author: rhymeswithlion

module-18-reinforcement-fine-tuning-on-foundation-models/exercises/starter/teach-an-llm-to-spell-with-grpo-starter.ipynb

@@ -204,7 +204,7 @@
         ") -> (str, str):\n",
         "    inputs = tokenizer(prompt, return_tensors=\"pt\").to(device)\n",
         "    gen = model.generate(\n",
-        "        **inputs, max_new_tokens=max_new_tokens\n",
+        "        **inputs, max_new_tokens=max_new_tokens, use_cache=False\n",
         "    )  # No parameters = greedy search\n",
         "    output = tokenizer.decode(gen[0], skip_special_tokens=True)\n",
         "\n",
@@ -700,4 +700,4 @@
   },
   "nbformat": 4,
   "nbformat_minor": 5
-}
+}