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

benchmark/kernels/fused_moe_triton/benchmark_sglang_fused_moe_triton.py

@@ -5,6 +5,7 @@ import torch
 import triton
 from common_utils import get_model_config
 
+from sglang.benchmark.bench_utils import run_bench
 from sglang.srt.distributed.parallel_state import (
     destroy_distributed_environment,
     destroy_model_parallel,
@@ -181,8 +182,8 @@ def benchmark(
     else:
         bench_lambda = lambda: api_func(**api_kwargs)
 
-    quantiles = [0.5, 0.2, 0.8]
-    ms, min_ms, max_ms = triton.testing.do_bench(bench_lambda, quantiles=quantiles)
+    quantiles = (0.5, 0.2, 0.8)
+    ms, min_ms, max_ms = run_bench(bench_lambda, quantiles=quantiles)
     return ms, min_ms, max_ms
 
 

benchmark/kernels/fused_moe_triton/benchmark_torch_compile_fused_moe.py

@@ -6,6 +6,7 @@ import triton
 from torch.nn import functional as F
 from transformers import AutoConfig
 
+from sglang.benchmark.bench_utils import run_bench
 from sglang.srt.layers.moe.fused_moe_triton.fused_moe import (
     fused_moe as fused_moe_triton,
 )
@@ -258,8 +259,8 @@ def benchmark(batch_size, provider, model_config, use_fp8_w8a8=False):
         )
     torch.cuda.synchronize()
 
-    quantiles = [0.5, 0.2, 0.8]
-    ms, min_ms, max_ms = triton.testing.do_bench(
+    quantiles = (0.5, 0.2, 0.8)
+    ms, min_ms, max_ms = run_bench(
         lambda: api_func(
             x,
             w1,

benchmark/kernels/fused_moe_triton/benchmark_vllm_vs_sglang_fused_moe_triton.py

@@ -5,6 +5,7 @@ import torch
 import triton
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_moe as fused_moe_vllm
 
+from sglang.benchmark.bench_utils import run_bench
 from sglang.srt.distributed.parallel_state import (
     destroy_distributed_environment,
     destroy_model_parallel,
@@ -190,8 +191,8 @@ def benchmark(batch_size, provider, model_config, use_fp8_w8a8=False):
         )
     torch.cuda.synchronize()
 
-    quantiles = [0.5, 0.2, 0.8]
-    ms, min_ms, max_ms = triton.testing.do_bench(
+    quantiles = (0.5, 0.2, 0.8)
+    ms, min_ms, max_ms = run_bench(
         lambda: api_func(
             x,
             w1,