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

benchmark/kernels/quantization/bench_fp4_quant.py

@@ -9,6 +9,7 @@ from flashinfer import (
 )
 from sgl_kernel.elementwise import silu_and_mul
 
+from sglang.benchmark.bench_utils import run_bench
 from sglang.srt.layers import deep_gemm_wrapper
 from sglang.srt.layers.moe.ep_moe.kernels import silu_and_mul_masked_post_quant_fwd
 
@@ -75,9 +76,9 @@ def benchmark(M, K, provider):
         dtype=torch.float32,
     )
 
-    quantiles = [0.5, 0.2, 0.8]
+    quantiles = (0.5, 0.2, 0.8)
     if provider == "triton_fp8":
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+        ms, min_ms, max_ms = run_bench(
             lambda: silu_and_mul_masked_post_quant_fwd(
                 x,
                 fp8_out,
@@ -89,7 +90,7 @@ def benchmark(M, K, provider):
             quantiles=quantiles,
         )
     if provider == "cuda_unfused_fp4":
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+        ms, min_ms, max_ms = run_bench(
             lambda: scaled_fp4_grouped_quantize(
                 silu_and_mul(x),
                 masks,
@@ -98,7 +99,7 @@ def benchmark(M, K, provider):
             quantiles=quantiles,
         )
     if provider == "cuda_fused_fp4":
-        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+        ms, min_ms, max_ms = run_bench(
             lambda: silu_and_mul_scaled_nvfp4_experts_quantize(
                 x,
                 masks,

benchmark/kernels/quantization/bench_int8_quant.py

@@ -4,6 +4,7 @@ import torch
 import triton
 from vllm._custom_ops import scaled_int8_quant as vllm_scaled_int8_quant
 
+from sglang.benchmark.bench_utils import run_bench
 from sglang.srt.layers.quantization.int8_kernel import per_token_quant_int8
 
 
@@ -59,19 +60,19 @@ def benchmark(batch_size, provider):
     M, K = batch_size, 16384
     x = torch.randn(M, K, dtype=torch.float16, device="cuda") * 1000
 
-    quantiles = [0.5, 0.2, 0.8]
+    quantiles = (0.5, 0.2, 0.8)
     if provider == "vllm op":
-        ms, min_ms, max_ms = triton.testing.do_bench(
+        ms, min_ms, max_ms = run_bench(
             lambda: vllm_scaled_int8_quant(x, symmetric=True),
             quantiles=quantiles,
         )
     if provider == "triton":
-        ms, min_ms, max_ms = triton.testing.do_bench(
+        ms, min_ms, max_ms = run_bench(
             lambda: per_token_quant_int8(x),
             quantiles=quantiles,
         )
     if provider == "torch.compile":
-        ms, min_ms, max_ms = triton.testing.do_bench(
+        ms, min_ms, max_ms = run_bench(
             lambda: torch_int8_quant(x),
             quantiles=quantiles,
         )