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

python/sglang/benchmark/bench_utils.py

@@ -0,0 +1,23 @@
+"""Triton do_bench/do_bench_cudagraph compatible wrapper using flashinfer.testing.bench_gpu_time."""
+
+import numpy as np
+from flashinfer.testing import bench_gpu_time
+
+
+def run_bench(
+    fn,
+    use_cuda_graph: bool = True,
+    quantiles=(0.5, 0.2, 0.8),
+    warmup_ms: int = 25,
+    rep_ms: int = 100,
+):
+    """Returns (ms, min_ms, max_ms) or (median,) when quantiles=None."""
+    times = bench_gpu_time(
+        fn=fn,
+        use_cuda_graph=use_cuda_graph,
+        dry_run_time_ms=warmup_ms,
+        repeat_time_ms=rep_ms,
+    )
+    if quantiles is None:
+        return (float(np.median(times)),)
+    return tuple(float(np.percentile(times, q * 100)) for q in quantiles)