Don't use implicitly 'elapsed_time' in autotuner

Signed-off-by: Anatoly Myachev <[email protected]>

Author: anmyachev

benchmarks/triton_kernels_benchmark/benchmark_testing.py

@@ -2,6 +2,7 @@
 import itertools
 import os
 from typing import Any, Dict, List
+import triton
 
 USE_IPEX_OPTION = os.getenv("USE_IPEX", "1") == "1"
 if USE_IPEX_OPTION:
@@ -236,6 +237,8 @@ def do_bench_upstream_pytorch_profiler(fn, n_warmup=25, n_repeat=100, grad_to_no
 else:
     raise NotImplementedError(f"BENCHMARKING_METHOD: {BENCHMARKING_METHOD} isn't implemented")
 
+triton.testing.do_bench = do_bench
+
 
 def assert_close(x, y, atol=None, rtol=None, err_msg=""):
     import numpy as np

python/triton/runtime/autotuner.py

@@ -117,8 +117,8 @@ def _post_hook(kwargs, exception):
             import triton.testing
             self.do_bench = lambda kernel_call, quantiles: triton.testing.do_bench(
                 kernel_call,
-                warmup=warmup if warmup is not None else 25,
-                rep=rep if rep is not None else 100,
+                warmup=warmup if warmup is not None else 10,
+                rep=rep if rep is not None else 10,
                 quantiles=quantiles,
             )
             return