[Benchmark] use flashinfer bench_gpu_time instead of triton do_bench (#20305)

benchmark/kernels/deepseek/benchmark_deepgemm_fp8_gemm.py

@@ -11,6 +11,7 @@ from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     w8a8_block_fp8_matmul as vllm_w8a8_block_fp8_matmul,
 )
 
+from sglang.benchmark.bench_utils import run_bench
 from sglang.srt.layers.quantization.fp8_kernel import (
     w8a8_block_fp8_matmul_deepgemm as w8a8_block_fp8_matmul,
 )
@@ -303,10 +304,10 @@ def get_benchmark(tp_size):
         y_fp8, y_scale = per_block_cast_to_fp8(y)
         x_scale_col_major = get_mn_major_tma_aligned_tensor(x_scale.clone())
 
-        quantiles = [0.5, 0.2, 0.8]
+        quantiles = (0.5, 0.2, 0.8)
 
         if provider == "deepgemm":
-            ms, min_ms, max_ms = triton.testing.do_bench(
+            ms, min_ms, max_ms = run_bench(
                 lambda: fp8_gemm_deepgemm(
                     x_fp8.clone(),
                     x_scale_col_major.clone(),
@@ -319,7 +320,7 @@ def get_benchmark(tp_size):
                 quantiles=quantiles,
             )
         elif provider == "sglang":
-            ms, min_ms, max_ms = triton.testing.do_bench(
+            ms, min_ms, max_ms = run_bench(
                 lambda: fp8_gemm_sglang(
                     x_fp8.clone(),
                     x_scale.clone(),
@@ -334,7 +335,7 @@ def get_benchmark(tp_size):
         else:  # tilelang
             tilelang_func = tl_gemm(m, n, k, "e4m3_float8", "bfloat16", "float32")
             tilelang_kernel = tilelang.compile(tilelang_func, out_idx=[-1])
-            ms, min_ms, max_ms = triton.testing.do_bench(
+            ms, min_ms, max_ms = run_bench(
                 lambda: tilelang_kernel(
                     x_fp8.clone(),
                     x_scale.clone(),