Liger-Kernel · KernelIndex

Records held

565as of 2026-10-01 14:30 UTC

Kernels

Runs

801

Is this you? Claim this project ›

A claim grants attribution and metadata maintenance, never the right to edit evidence. Evidence claims are reviewed by a maintainer.

Current records

Operation · workload

History

Record

Implementation

Hardware

Since

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 2048]

98.4µs

8.1% faster · was 107.1 µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 4096]

140.8µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 8192]

265.0µs

0.7% faster · was 266.8 µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 16384]

500.6µs

0.4% faster · was 502.5 µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 32768]

971.0µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 65536]

1.91ms

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 2048]

120.1µs

1.7% faster · was 122.2 µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 4096]

140.3µs

0.1% faster · was 140.5 µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 8192]

266.3µs

0.2% faster · was 266.8 µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 16384]

502.1µs

0.1% faster · was 502.6 µs

Liger

H100

2026-07-26

Show all 40 records ›

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 32768]

971.0µs

0.0% faster · was 971.5 µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 65536]

1.91ms

0.0% faster · was 1.91 ms

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 2048]

23.1µs

0.8% faster · was 23.3 µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 4096]

39.6µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 8192]

72.5µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 16384]

137.0µs

0.1% faster · was 137.1 µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 32768]

266.8µs

0.1% faster · was 267.0 µs

Liger

H100

2026-07-26

SwiGLU activation, Megatron layoutbf16 · [2048, 4, 65536]

527.3µs

0.0% faster · was 527.5 µs

Liger

H100

2026-07-26

RMSNorm, Megatron layoutbf16 · [4096, 1, 8192]

302.0µs

Liger

H100

2026-06-10

RMSNorm, Megatron layoutbf16 · [4096, 1, 16384]

455.6µs

Liger

H100

2026-06-10

RMSNorm, Megatron layoutbf16 · [4096, 1, 8192]

321.0µs

Liger

H100

2026-06-10

RMSNorm, Megatron layoutbf16 · [4096, 1, 16384]

587.7µs

Liger

H100

2026-06-10

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 4096]

138.6µs

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 8192]

262.6µs

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 16384]

415.8µs

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 32768]

809.8µs

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 65536]

1.53ms

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 131072]

2.97ms

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 4096]

349.8µs

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 8192]

585.4µs

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 16384]

1.05ms

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 32768]

2.06ms

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 65536]

4.02ms

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 131072]

7.96ms

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 4096]

349.4µs

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 8192]

585.5µs

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 16384]

1.05ms

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 32768]

2.06ms

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 65536]

4.02ms

Liger

H100

2026-06-15

Vocab-parallel cross-entropy, Megatron layoutbf16 · [2048, 4, 131072]

7.96ms

Liger

H100

2026-06-15

525 more records · Open in the records ledger →

Measured kernels

Implementation

Operation

Runtime

Best latency

Evidence

Availability

LigerLiger-Kernel

Softmax

triton

6.34µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

ReLU squared

triton

7.33µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Rotary position embedding

triton

11.4µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

RMSNorm, Megatron layout

triton

12.2µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

RMSNorm

triton

13.6µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Multi-token attention

triton

17.4µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

GroupNorm

triton

17.5µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

LayerNorm

triton

18.8µs

Reported

BSD-2-Clause · source

Liger fusedLiger-Kernel

Fused add + RMSNorm

triton

20.5µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

SwiGLU activation, Megatron layout

triton

23.1µs

Reported

BSD-2-Clause · source

Show all 31 kernels ›

LigerLiger-Kernel

Embedding lookup

triton

42.6µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Sparsemax

triton

44.0µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Llama 4 rotary position embedding

triton

74.2µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Vocab-parallel cross-entropy, Megatron layout

triton

138.6µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Fused neighborhood attention

triton

155.2µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Total variation distance loss

triton

275.7µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

KL-divergence loss

triton

306.4µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Jensen-Shannon divergence loss

triton

465.2µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Cross-entropy loss

triton

532.4µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Fused MoE expert MLP

triton

588.8µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

SwiGLU MLP

triton

659.5µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

GeGLU MLP

triton

661.4µs

Reported

BSD-2-Clause · source

LigerLiger-Kernel

GRPO loss, token-level importance sampling

triton

1.81ms

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Fused linear + KTO loss

triton

4.00ms

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Chunked distillation JSD loss

triton

7.74ms

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Chunked distillation cosine loss

triton

13.8ms

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Fused linear + SimPO loss

triton

30.3ms

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Fused linear + ORPO loss

triton

31.0ms

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Fused linear + CPO loss

triton

31.5ms

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Fused linear + JSD

triton

110.0ms

Reported

BSD-2-Clause · source

LigerLiger-Kernel

Fused linear + cross-entropy

triton

119.5ms

Reported

BSD-2-Clause · source

Record activity

55Nov

Dec

Jan

18Feb

38Mar

56Apr

14May

27Jun

31Jul

Aug

Sep

Oct

Sources: Liger-Kernel benchmarks (2026-07-26) · BSD-2-Clauselast observed 2026-07-26JSON