Liger
Llama 4 rotary position embedding · seq = 2048 · hidden = 512 · q_heads = 32 · kv_heads = 8 · bf16 · run 01a0202b-232b…
●passedReported evidence
Not re-observed recently.
Primary measurement
82.5µsmedian
Rank 2 in its comparison group · source-native comparison · observed 2025-08-07
ms 20 0.08102399855852127 · ms 50 0.08249600231647491 · ms 80 0.08432000130414963
Identity
implementationLiger
projectLiger-Kernel
revisionunknown
workloadseq = 2048 · hidden = 512 · q_heads = 32 · kv_heads = 8 · bf16
comparison keysha256:f2d4336649cdd957…
sourceLiger-Kernel benchmarks
external idllama4_rope/forward/liger/nvidia-h100-80gb-hbm3/hidden512-kv_heads8-q_heads32-seq2048-bf16/2025-08-07-21-42-01
sha256:be51fb2e16d617d988df83c…
Correctness
Marked passed by the source; the correctness policy was not published.
Workload
seq2048
hidden512
q_heads32
kv_heads8
kbf16 [1, 2048, 8, 16]
qbf16 [1, 2048, 32, 16]
definition comparatornot_asserted
Measurements
latency · median82.5 µs
latency · p2081.0 µs
latency · p8084.3 µs
Protocol
harnessLiger-Kernel benchmark scripts
timercuda_events
primaryStatisticmedian
comparabilityFamilyliger_kernel_bench
Environment
gpuNVIDIA H100 (sm_90)
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 manifestHide manifest
{
"run": {
"kind": "BenchmarkRun",
"spec": {
"status": "passed",
"timing": {
"latencyNs": {
"median": 82496
},
"primaryStatistic": "median"
},
"observedAt": "2025-08-07T21:42:01.000Z",
"measurements": [
{
"unit": "ns",
"value": 81024,
"metric": "latency",
"statistic": "p20"
},
{
"unit": "ns",
"value": 84320,
"metric": "latency",
"statistic": "p80"
}
],
"sourceNative": {
"source": "liger-kernel-bench",
"metrics": {
"ms_20": 0.08102399855852127,
"ms_50": 0.08249600231647491,
"ms_80": 0.08432000130414963
},
"benchmark": "llama4_rope/forward",
"externalId": "llama4_rope/forward/liger/nvidia-h100-80gb-hbm3/hidden512-kv_heads8-q_heads32-seq2048-bf16/2025-08-07-21-42-01"
},
"protocolDigest": "sha256:a79eaf4df66c97c6adc1fb2d09df78682155d1ea86652ae0b70199875bde7747",
"workloadDigest": "sha256:98c5e76b9fa318ddeea51b56d90bc6fb39d72e404e2d01dbb467e67f57ff63e8",
"environmentDigest": "sha256:2750ae381d6de582f1eb6b9008e0987463d19ede8af127d97d76c5ae09121393",
"implementationDigest": "sha256:7b16e7caa07229eb63274c52969e9742bd849258f08f13b57ebaa55928740b6d"
},
"metadata": {
"name": "liger-llama4-rope-forward-liger-nvidia-h100-80gb-hbm3-hidden512-kv-heads8-q-heads32-seq2048-bf16-202",
"title": "Llama 4 rotary position embedding · liger · forward",
"labels": {
"liger_version": "0.6.1"
}
},
"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 module forward pass, 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-forward",
"title": "Liger benchmark harness · forward pass"
},
"apiVersion": "kernelindex.dev/v1alpha1"
},
"environment": {
"kind": "ExecutionEnvironment",
"spec": {
"hardware": {
"vendor": "nvidia",
"product": "NVIDIA H100",
"formFactor": "80GB HBM3",
"architecture": "sm_90"
},
"software": {}
},
"metadata": {
"name": "liger-bench-nvidia-h100-80gb-hbm3",
"title": "Liger benchmark host · H100 80GB HBM3"
},
"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 Llama 4 rotary position embedding →JSON