From f8ee88664df0ba3a7460c0a103efcf6b38e70d0a Mon Sep 17 00:00:00 2001
From: Hyunsu Cho <chohyu01@cs.washington.edu>
Date: Fri, 21 May 2021 15:35:29 -0700
Subject: [PATCH 1/2] Document ability to export cuML RF to predict on other
 machines

---
 docs/source/pickling_cuml_models.ipynb | 66 +++++++++++++++++++++++++-
 1 file changed, 65 insertions(+), 1 deletion(-)

diff --git a/docs/source/pickling_cuml_models.ipynb b/docs/source/pickling_cuml_models.ipynb
index c7da3524b7..211cdb502d 100644
--- a/docs/source/pickling_cuml_models.ipynb
+++ b/docs/source/pickling_cuml_models.ipynb
@@ -183,6 +183,70 @@
    "source": [
     "single_gpu_model.cluster_centers_"
    ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Exporting cuML Random Forest models for inferencing on machines without GPUs"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Starting with cuML version 21.06, you can export cuML Random Forest models and run predictions with them on machines without an NVIDIA GPUs. The [Treelite](https://github.com/dmlc/treelite) package defines an efficient exchange format that lets you portably move the cuML Random Forest models to other machines. We will refer to the exchange format as \"checkpoints.\"\n",
+    "\n",
+    "Here are the steps to export the model:\n",
+    "\n",
+    "1. Call `to_treelite_checkpoint()` to obtain the checkpoint file from the cuML Random Forest model."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from cuml.ensemble import RandomForestClassifier as cumlRandomForestClassifier\n",
+    "from sklearn.datasets import load_iris\n",
+    "import numpy as np\n",
+    "\n",
+    "X, y = load_iris(return_X_y=True)\n",
+    "X, y = X.astype(np.float32), y.astype(np.int32)\n",
+    "clf = cumlRandomForestClassifier(max_depth=3, random_state=0, n_estimators=10)\n",
+    "clf.fit(X, y)\n",
+    "\n",
+    "checkpoint_path = './checkpoint.tl'\n",
+    "# Export cuML RF model as Treelite checkpoint\n",
+    "clf.convert_to_treelite_model().to_treelite_checkpoint(checkpoint_path)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "2. Copy the generated checkpoint file `checkpoint.tl` to another machine on which you'd like to run predictions.\n",
+    "\n",
+    "3. On the target machine, install Treelite by running `pip install treelite` or `conda install -c conda-forge treelite`. The machine does not need to have an NVIDIA GPUs and does not need to have cuML installed.\n",
+    "\n",
+    "4. You can now load the model from the checkpoint, by running the following on the target machine:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import treelite\n",
+    "\n",
+    "# The checkpoint file has been copied over\n",
+    "checkpoint_path = './checkpoint.tl'\n",
+    "tl_model = treelite.Model.deserialize(checkpoint_path)\n",
+    "out_prob = treelite.gtil.predict(tl_model, X, pred_margin=True)\n",
+    "print(out_prob)"
+   ]
   }
  ],
  "metadata": {
@@ -201,7 +265,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.8.6"
+   "version": "3.8.8"
   }
  },
  "nbformat": 4,

From 52d03017065eecc1237efc22306ac8c7943e5af5 Mon Sep 17 00:00:00 2001
From: Hyunsu Cho <chohyu01@cs.washington.edu>
Date: Wed, 26 May 2021 21:57:00 -0700
Subject: [PATCH 2/2] Add link to docstring

---
 python/cuml/ensemble/randomforestclassifier.pyx | 5 +++++
 python/cuml/ensemble/randomforestregressor.pyx  | 5 +++++
 2 files changed, 10 insertions(+)

diff --git a/python/cuml/ensemble/randomforestclassifier.pyx b/python/cuml/ensemble/randomforestclassifier.pyx
index 1772b9678e..36272270ab 100644
--- a/python/cuml/ensemble/randomforestclassifier.pyx
+++ b/python/cuml/ensemble/randomforestclassifier.pyx
@@ -134,6 +134,11 @@ class RandomForestClassifier(BaseRandomForestModel,
       histogram-based algorithm to determine splits, rather than an exact
       count. You can tune the size of the histograms with the n_bins parameter.
 
+    .. note:: You can export cuML Random Forest models and run predictions
+      with them on machines without an NVIDIA GPUs. See
+      https://docs.rapids.ai/api/cuml/nightly/pickling_cuml_models.html
+      for more details.
+
     **Known Limitations**: This is an early release of the cuML
     Random Forest code. It contains a few known limitations:
 
diff --git a/python/cuml/ensemble/randomforestregressor.pyx b/python/cuml/ensemble/randomforestregressor.pyx
index 48d066e7f1..6eca660ac1 100644
--- a/python/cuml/ensemble/randomforestregressor.pyx
+++ b/python/cuml/ensemble/randomforestregressor.pyx
@@ -117,6 +117,11 @@ class RandomForestRegressor(BaseRandomForestModel,
       histogram-based algorithm to determine splits, rather than an exact
       count. You can tune the size of the histograms with the n_bins parameter.
 
+    .. note:: You can export cuML Random Forest models and run predictions
+      with them on machines without an NVIDIA GPUs. See
+      https://docs.rapids.ai/api/cuml/nightly/pickling_cuml_models.html
+      for more details.
+
     **Known Limitations**: This is an early release of the cuML
     Random Forest code. It contains a few known limitations: