Chapter 9 cuGraph: Correct PageRank examples and backend setup.

Install nx-cugraph before NetworkX performs backend discovery, select CPU and GPU implementations explicitly, and make the benchmark reproducible. Correct the directed toy graph and Karate dataset example while preserving a meaningful benchmark size.

Co-authored-by: myli5 <cloudmyli5@gmail.com>
Signed-off-by: Bryce Adelstein Lelbach <brycelelbach@gmail.com>

Author: brycelelbach

Accelerated_Python_User_Guide/notebooks/Chapter_09_Intro_to_cuGraph.ipynb

@@ -66,7 +66,7 @@
     "```\n",
     "pip install \\\n",
     "    --extra-index-url=https://pypi.nvidia.com \\\n",
-    "    cudf-cu12==24.10.* cugraph-cu12==24.10.* \n",
+    "    cudf-cu12==26.8.* cugraph-cu12==26.8.* nx-cugraph-cu12==26.8.* \n",
     "```\n",
     "\n",
     "\n",
@@ -93,9 +93,11 @@
     "# Install a cudf / cugraph build that matches this environment's CUDA and Python.\n",
     "# The 24.10 pin in the original notebook is far too old (no wheels for recent\n",
     "# Python); use a current RAPIDS release instead.\n",
-    "!pip install \\\n",
+    "%pip install \\\n",
     "    --extra-index-url=https://pypi.nvidia.com \\\n",
-    "    \"cudf-cu12==26.8.*\" \"cugraph-cu12==26.8.*\""
+    "    \"cudf-cu12==26.8.*\" \\\n",
+    "    \"cugraph-cu12==26.8.*\" \\\n",
+    "    \"nx-cugraph-cu12==26.8.*\""
    ]
   },
   {
@@ -209,8 +211,9 @@
     "    'dst': [1, 2, 2, 0]\n",
     "})\n",
     "\n",
-    "# Create the graph\n",
-    "G = cugraph.Graph()\n",
+    "# Create the graph. The edge list is directed (note the 2 -> 0 \"connects back\"\n",
+    "# edge), so build a directed Graph.\n",
+    "G = cugraph.Graph(directed=True)\n",
     "G.from_cudf_edgelist(edge_list, source='src', destination='dst')"
    ]
   },
@@ -310,16 +313,21 @@
    },
    "outputs": [],
    "source": [
-    "# Perform PageRank on the weighted graph\n",
+    "# Perform PageRank on the graph we built above\n",
     "pagerank_result = cugraph.pagerank(G)\n",
     "\n",
     "# Display the PageRank values\n",
     "print(pagerank_result)\n",
     "\n",
-    "G = karate.get_graph(download=True)\n",
+    "# Now try it on a larger, real-world graph: Zachary's Karate Club, which ships\n",
+    "# with cuGraph as a sample dataset.\n",
+    "from cugraph.datasets import karate\n",
+    "\n",
+    "karate_G = karate.get_graph(download=True)\n",
     "\n",
     "# Call cugraph.pagerank to get the pagerank scores\n",
-    "gdf_page = cugraph.pagerank(G)\n"
+    "gdf_page = cugraph.pagerank(karate_G)\n",
+    "print(gdf_page.sort_values('pagerank', ascending=False).head())"
    ]
   },
   {
@@ -329,7 +337,7 @@
    "source": [
     "## NetworkX x cuGraph\n",
     "\n",
-    "Let's start by installing the zero-code change NetworkX cuGraph package:"
+    "The prerequisites cell installs the zero-code-change NetworkX cuGraph package before NetworkX is imported. Verify the installed backend version:"
    ]
   },
   {
@@ -343,7 +351,8 @@
    },
    "outputs": [],
    "source": [
-    "!pip install \"nx-cugraph-cu12==26.8.*\" --extra-index-url=https://pypi.nvidia.com"
+    "from importlib.metadata import version\n",
+    "print(f\"using nx-cugraph version {version('nx-cugraph-cu12')}\")"
    ]
   },
   {
@@ -365,12 +374,8 @@
    },
    "outputs": [],
    "source": [
-    "#%env NX_CUGRAPH_AUTOCONFIG=True\n",
-    "\n",
     "import networkx as nx\n",
-    "print(f\"using networkx version {nx.__version__}\")\n",
-    "\n",
-    "#nx.config.warnings_to_ignore.add(\"cache\")"
+    "print(f\"using networkx version {nx.__version__}\")"
    ]
   },
   {
@@ -384,21 +389,22 @@
    },
    "outputs": [],
    "source": [
-    "G = nx.gnm_random_graph(5000, 40000)\n",
+    "# A reproducible graph large enough for a meaningful backend comparison.\n",
+    "benchmark_graph = nx.gnm_random_graph(200_000, 2_000_000, seed=42)\n",
     "\n",
-    "import time \n",
+    "import time\n",
     "start_time = time.time()\n",
-    "pr = nx.pagerank(G)\n",
-    "elapsed_time = time.time() - start_time\n",
-    "print(elapsed_time)"
+    "cpu_pagerank = nx.pagerank(benchmark_graph, backend=\"networkx\")\n",
+    "cpu_elapsed = time.time() - start_time\n",
+    "print(f\"NetworkX CPU time: {cpu_elapsed:.3f} seconds\")"
    ]
   },
   {
    "cell_type": "markdown",
    "id": "d3040a65",
    "metadata": {},
    "source": [
-    "Then let's try to set the backend to cuGraph by default instead of using the non-accelerated backend: "
+    "Then run the same calculation explicitly with the cuGraph backend:"
    ]
   },
   {
@@ -412,27 +418,26 @@
    },
    "outputs": [],
    "source": [
-    "%env NX_CUGRAPH_AUTOCONFIG=True\n",
-    "\n",
-    "import networkx as nx\n",
-    "print(f\"using networkx version {nx.__version__}\")\n",
-    "\n",
-    "#nx.config.warnings_to_ignore.add(\"cache\")\n",
+    "import time\n",
+    "start_time = time.time()\n",
+    "gpu_pagerank = nx.pagerank(benchmark_graph, backend=\"cugraph\")\n",
+    "gpu_elapsed = time.time() - start_time\n",
     "\n",
+    "assert len(gpu_pagerank) == len(cpu_pagerank)\n",
+    "assert abs(sum(cpu_pagerank.values()) - 1.0) < 1e-4\n",
+    "assert abs(sum(gpu_pagerank.values()) - 1.0) < 1e-4\n",
+    "max_error = max(abs(gpu_pagerank[node] - score) for node, score in cpu_pagerank.items())\n",
     "\n",
-    "import time \n",
-    "start_time = time.time()\n",
-    "pr = nx.pagerank(G)\n",
-    "elapsed_time = time.time() - start_time\n",
-    "print(elapsed_time)"
+    "print(f\"cuGraph GPU time: {gpu_elapsed:.3f} seconds\")\n",
+    "print(f\"Maximum PageRank difference: {max_error:.3e}\")"
    ]
   },
   {
    "cell_type": "markdown",
    "id": "f3d1af59",
    "metadata": {},
    "source": [
-    "Now that we've configured our cuGraph setup with NetworkX, let's start experimenting with its functionalities using a real-world example!\n"
+    "Now that the cuGraph backend is available through NetworkX, let's start experimenting with a real-world example!\n"
    ]
   },
   {
@@ -593,7 +598,7 @@
    },
    "outputs": [],
    "source": [
-    "pagerank_scores = nx.pagerank(C)\n",
+    "pagerank_scores = nx.pagerank(C, backend=\"cugraph\")\n",
     "pagerank_df = cudf.DataFrame({'node_id': pagerank_scores.keys(), 'score': pagerank_scores.values()})\n",
     "pagerank_df.tail()"
    ]