Add inductor graph partition mode to vLLM benchmark (#166)

Generate benchmark variants with compilation config
{"cudagraph_mode": "PIECEWISE", "use_inductor_graph_partition": true}
via the new --include-inductor-graph-partition flag.

This will be used to populate another set of data on the dashboard.
The current speedup on the dashboard is for default vLLM compile over
Eager. And the new data can be thought as "advanced vLLM compile"
over Eager.

Author: desertfire

.github/scripts/setup_vllm_benchmark.py

@@ -32,6 +32,12 @@
 # dashboard baseline.
 EAGER_COMPILATION_CONFIG = {"backend": "eager"}
 
+# Piecewise cudagraph mode with inductor graph partition enabled
+INDUCTOR_GRAPH_PARTITION_COMPILATION_CONFIG = {
+    "cudagraph_mode": "PIECEWISE",
+    "use_inductor_graph_partition": True,
+}
+
 
 def apply_compilation_config(
     config: Dict[str, Any],
@@ -115,6 +121,12 @@ def parse_args() -> Any:
         default=False,
         help="also generate eager mode variants of all benchmarks",
     )
+    parser.add_argument(
+        "--include-inductor-graph-partition",
+        action="store_true",
+        default=False,
+        help="also generate inductor graph partition variants of all benchmarks",
+    )
     parser.add_argument(
         "--compilation-config",
         type=str,
@@ -132,6 +144,7 @@ def setup_benchmark_configs(
     device: str,
     compilation_config: Optional[Dict[str, Any]] = None,
     include_eager_mode: bool = False,
+    include_inductor_graph_partition: bool = False,
 ) -> None:
     """
     Setup the benchmark configs to run on this runner.
@@ -169,6 +182,11 @@ def setup_benchmark_configs(
                     config, EAGER_COMPILATION_CONFIG, "_eager"
                 )
                 benchmark_configs.append(eager_config)
+            if include_inductor_graph_partition:
+                inductor_graph_partition_config = apply_compilation_config(
+                    config, INDUCTOR_GRAPH_PARTITION_COMPILATION_CONFIG, "_inductor_graph_partition"
+                )
+                benchmark_configs.append(inductor_graph_partition_config)
 
         if benchmark_configs:
             with open(os.path.join(to_benchmark_configs_dir, filename), "w") as f:
@@ -187,6 +205,7 @@ def main() -> None:
         args.device,
         compilation_config,
         args.include_eager_mode,
+        args.include_inductor_graph_partition,
     )
 
 

.github/workflows/vllm-benchmark.yml

@@ -287,7 +287,8 @@ jobs:
             --to-benchmark-configs-dir vllm-benchmarks/vllm/.buildkite/performance-benchmarks/tests \
             --models "${MODELS}" \
             --device "${DEVICE_NAME}" \
-            --include-eager-mode
+            --include-eager-mode \
+            --include-inductor-graph-partition
 
           pushd vllm-benchmarks/vllm
           ls -lah .buildkite/performance-benchmarks/tests