From c410c510b3048d6e5232fd3844ffe52759a395f5 Mon Sep 17 00:00:00 2001 From: Max Bremer Date: Fri, 10 Oct 2025 21:15:18 -0700 Subject: [PATCH] Use active driver benchmarker for do_bench measurements (#537) Summary: Sometimes the triton backend may set a profiler to be different from the standard `do_bench` implmentation. For cuda this has no functional change, i.e. `triton.runtime.driver.active.get_benchmarker()` still points at `do_bench`. Reviewed By: xuzhao9 Differential Revision: D84215437 --- tritonbench/components/do_bench/run.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tritonbench/components/do_bench/run.py b/tritonbench/components/do_bench/run.py index e93f37d5..29acba7c 100644 --- a/tritonbench/components/do_bench/run.py +++ b/tritonbench/components/do_bench/run.py @@ -475,7 +475,7 @@ def do_bench_wrapper( if latency_measure_mode == "profiler" else _do_bench_inductor if latency_measure_mode == "inductor_benchmarker" - else triton.testing.do_bench + else triton.runtime.driver.active.get_benchmarker() ) return Latency(