made some minor updates to the notebook

Author: Chillee

notebooks/aot_autograd_optimizations.ipynb

@@ -374,6 +374,18 @@
    "source": [
     "We observe that both forward and backward latency improve over the default partitioner (and a lot better than eager). Fewer outputs in the forward pass and fewer inputs in the backward pass, along with fusion, allows better memory bandwidth utilization leading to further speedups."
    ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Actual Usage\n",
+    "For actual usage on CUDA devices, we've wrapped AOTAutograd in a convenient wrapper - `memory_efficient_fusion`. Use this for fusion on GPU!\n",
+    "\n",
+    "```\n",
+    "from functorch.compile import memory_efficient_fusion\n",
+    "```\n"
+   ]
   }
  ],
  "metadata": {
@@ -395,7 +407,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.8.12"
+   "version": "3.8.13"
   },
   "orig_nbformat": 4
  },