Run 01a02032 · KernelIndex

Primary measurement

24.6msmean

Rank 5 in its comparison group · source-native comparison · observed 2025-10-16

Compare with #1 →

6.44× reference

Identity

implementationgpt-o3_cuda_1dc3e4

projectgpt-o3

revisionda915083d4c7c5e61aa3005e3d17ae488e0fc71c

workloadtotal_q = 287 · num_pages = 989669 · len_indptr = 2 · num_kv_indices = 288 · bf16

comparison keysha256:e54a1953753f63ca…

sourceFlashInfer-Bench

external idtrace/gpt-o3_cuda_1dc3e4/7e602557-ff89-41a1-8650-6c3e48fdf52a

sha256:134949db9994d3727ae2fa9…

Correctness

comparatorflashinfer_bench_eval

max absolute error0.0078125

max relative error3.934890031814575

resultpassed

Workload

total_q287

num_pages989669

len_indptr2

num_kv_indices288

q_pebf16 [287, 16, 64]

q_nopebf16 [287, 16, 512]

ckv_cachebf16 [989669, 1, 512]

kpe_cachebf16 [989669, 1, 64]

kv_indptrint32 [2]

qo_indptrint32 [2]

kv_indicesint32 [288]

definition comparatorsol_execbench_eval

Measurements

latency · mean24.6 ms

Protocol

harnessflashinfer-bench

timerharness_reported

primaryStatisticmean

correctnessReferencedefinition reference implementation

comparabilityFamilyflashinfer_bench_trace

Environment

gpuNVIDIA B200 (unknown)

cudaToolkit12.8

frameworkpytorch 2.8.0+cu128

Artifacts

logssha256:bd46cc25665ac7d3… · text/plain · unavailable

Replications and notes

No attestations yet.

Community attestations. They never change the evidence level; only a KernelIndex-controlled rerun does.

Add a reproduction or note

Canonical manifest

Show manifest

{
  "run": {
    "kind": "BenchmarkRun",
    "spec": {
      "status": "passed",
      "timing": {
        "latencyNs": {
          "mean": 24552997
        },
        "primaryStatistic": "mean"
      },
      "evidence": {
        "logs": {
          "uri": "kernelindex:artifact/sha256:bd46cc25665ac7d3f563d3728f8e435901139b28de89ac0a8ba0e0c13543adb8",
          "digest": "sha256:bd46cc25665ac7d3f563d3728f8e435901139b28de89ac0a8ba0e0c13543adb8"
        }
      },
      "observedAt": "2025-10-16T11:10:07.953Z",
      "correctness": {
        "comparator": "sol_execbench_eval",
        "maximumAbsoluteError": 0.0078125,
        "maximumRelativeError": 3.934890031814575
      },
      "sourceNative": {
        "source": "flashinfer-bench",
        "metrics": {
          "latency_ms": 24.55299727121989,
          "speedup_factor": 6.435119435101167,
          "reference_latency_ms": 158.00146993001303
        },
        "benchmark": "mla_paged_prefill_causal_h16_ckv512_kpe64_ps1"
      },
      "protocolDigest": "sha256:7d44d206ad086b4bece35ce31f6531ebb11f07e0d9dcfb74c75be86d3ec02095",
      "workloadDigest": "sha256:7d5979c6c810890ad6a7aa6a66b668f23d449a922e980757d86bba77d12797a2",
      "environmentDigest": "sha256:f4292a8be127b7230cd3995c1d605fb6c644e2c2f94566e1d28cc1be766ca5d5",
      "implementationDigest": "sha256:cfd188b3111ac75c36302be921a813023797585305dd7e81f334d0668f4d9fc0"
    },
    "metadata": {
      "name": "sol-trace-mla-paged-prefill-causal-h16-ckv512-kpe64-ps1-7e602557",
      "title": "gpt-o3_cuda_1dc3e4 · mla_paged_prefill_causal_h16_ckv512_kpe64_ps1"
    },
    "apiVersion": "kernelindex.dev/v1alpha1"
  },
  "protocol": {
    "kind": "BenchmarkProtocol",
    "spec": {
      "harness": {
        "name": "flashinfer-bench",
        "repository": "https://github.com/flashinfer-ai/flashinfer-bench"
      },
      "correctness": {
        "reference": "definition reference implementation",
        "comparator": "flashinfer_bench_eval"
      },
      "measurement": {
        "timer": "harness_reported",
        "primaryStatistic": "mean"
      },
      "comparability": {
        "family": "flashinfer_bench_trace"
      }
    },
    "metadata": {
      "name": "flashinfer-bench-trace",
      "title": "FlashInfer-Bench trace harness"
    },
    "apiVersion": "kernelindex.dev/v1alpha1"
  },
  "environment": {
    "kind": "ExecutionEnvironment",
    "spec": {
      "hardware": {
        "vendor": "nvidia",
        "product": "NVIDIA B200",
        "architecture": "unknown"
      },
      "software": {
        "framework": {
          "name": "pytorch",
          "version": "2.8.0+cu128"
        },
        "libraries": {
          "triton": "3.4.0"
        },
        "cudaToolkit": "12.8"
      }
    },
    "metadata": {
      "name": "sol-trace-nvidia-b200-2-8-0-cu128-3-4-0-12-8",
      "title": "NVIDIA B200 (trace-reported software)"
    },
    "apiVersion": "kernelindex.dev/v1alpha1"
  }
}

Cite this record (permalink, digest, access date)

Report an issue with this run

Published 2026-08-20 · FlashInfer-Bench · Apache-2.0All results for MLA paged prefill causal h16 ckv512 kpe64 ps1 →JSON

gpt-o3 / cuda1dc3e4