Run 01a0202b · KernelIndex

Primary measurement

2.58msmedian

Rank 2 in its comparison group · source-native comparison · observed 2025-05-27

Compare with #1 →

ms 20 2.5763840675354004 · ms 50 2.5815041065216064 · ms 80 2.5960447788238525

Identity

implementationPyTorch

projectPyTorch

revisionunknown

workloadseq = 512 · batch = 4 · heads = 12 · hidden = 768 · dilation = 1 · kernel_size = 7 · fp32

comparison keysha256:f7cb6e69c7ddcf90…

sourceLiger-Kernel benchmarks

external idfused_neighborhood_attention/backward/torch/nvidia-geforce-rtx-3090/batch4-dilation1-heads12-hidden768-kernel_size7-seq512-fp32/2025-05-27-15-10-05

sha256:561d72d045ecafcc5e3db19…

Correctness

Marked passed by the source; the correctness policy was not published.

Workload

seq512

batch4

heads12

hidden768

dilation1

kernel_size7

k_biasfp32 [768]

q_biasfp32 [768]

v_biasfp32 [768]

k_weightfp32 [768, 768]

out_biasfp32 [768]

q_weightfp32 [768, 768]

v_weightfp32 [768, 768]

out_weightfp32 [768, 768]

hidden_statesfp32 [4, 512, 768]

definition comparatornot_asserted

Measurements

latency · median2.58 ms

latency · p202.58 ms

latency · p802.60 ms

Protocol

harnessLiger-Kernel benchmark scripts

timercuda_events

primaryStatisticmedian

comparabilityFamilyliger_kernel_bench

Environment

gpuNVIDIA GeForce RTX 3090 (sm_86)

Artifacts

No artifacts published with this run.

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": {
          "median": 2581504
        },
        "primaryStatistic": "median"
      },
      "observedAt": "2025-05-27T15:10:05.000Z",
      "measurements": [
        {
          "unit": "ns",
          "value": 2576384,
          "metric": "latency",
          "statistic": "p20"
        },
        {
          "unit": "ns",
          "value": 2596045,
          "metric": "latency",
          "statistic": "p80"
        }
      ],
      "sourceNative": {
        "source": "liger-kernel-bench",
        "metrics": {
          "ms_20": 2.5763840675354004,
          "ms_50": 2.5815041065216064,
          "ms_80": 2.5960447788238525
        },
        "benchmark": "fused_neighborhood_attention/backward",
        "externalId": "fused_neighborhood_attention/backward/torch/nvidia-geforce-rtx-3090/batch4-dilation1-heads12-hidden768-kernel_size7-seq512-fp32/2025-05-27-15-10-05"
      },
      "protocolDigest": "sha256:080e6e66e4ec75a131aa0c9121c74b8e8eed64e7b6d621d567c6a28f79548be8",
      "workloadDigest": "sha256:2996636bf104b8fc14802e19943113137e8240cf8057a52becaea6e649be6084",
      "environmentDigest": "sha256:b12fdfab6a337672499b10212a1e986e49e4383777bd7cbb58321b09af4bf6a3",
      "implementationDigest": "sha256:0c3e4e2a4849358a1d5183daed7f408a3422f6cc411cc6bd7d4ebeead8e57e00"
    },
    "metadata": {
      "name": "liger-fused-neighborhood-attention-backward-torch-nvidia-geforce-rtx-3090-batch4-dilation1-heads12-h",
      "title": "Fused neighborhood attention · torch · backward",
      "labels": {
        "liger_version": "0.5.10"
      }
    },
    "apiVersion": "kernelindex.dev/v1alpha1"
  },
  "protocol": {
    "kind": "BenchmarkProtocol",
    "spec": {
      "harness": {
        "name": "Liger-Kernel benchmark scripts",
        "repository": "https://github.com/linkedin/Liger-Kernel"
      },
      "measurement": {
        "timer": "cuda_events",
        "primaryStatistic": "median"
      },
      "comparability": {
        "notes": "Timed: the backward pass of the module output given a random upstream gradient, via triton.testing.do_bench with quantiles 0.5/0.2/0.8 (median with p20/p80 spread). The upstream CSV records no CUDA, driver, or torch version; the Liger release rides each run's labels. Comparable only within one kernel, workload, timed pass, and GPU.",
        "family": "liger_kernel_bench"
      }
    },
    "metadata": {
      "name": "liger-bench-do-bench-backward",
      "title": "Liger benchmark harness · backward pass"
    },
    "apiVersion": "kernelindex.dev/v1alpha1"
  },
  "environment": {
    "kind": "ExecutionEnvironment",
    "spec": {
      "hardware": {
        "vendor": "nvidia",
        "product": "NVIDIA GeForce RTX 3090",
        "architecture": "sm_86"
      },
      "software": {}
    },
    "metadata": {
      "name": "liger-bench-nvidia-geforce-rtx-3090",
      "title": "Liger benchmark host · GeForce RTX 3090"
    },
    "apiVersion": "kernelindex.dev/v1alpha1"
  }
}

Cite this record (permalink, digest, access date)

Report an issue with this run

Published 2026-08-20 · Liger-Kernel benchmarks · BSD-2-ClauseAll results for Fused neighborhood attention →JSON