Skip to content
KernelIndex
Search⌘K
Sign in

GQA paged decode h32 kv8 d128 ps64

18 eligible runs
gqa-paged-attention

Batched Grouped Query Attention decode with a paged KV cache (page_size=64). Captured from Llama-3.1-8B.

Source baseline · unbeaten
28.8µsmean
flashinfer / wrapperad4135FlashInfer-Bench baselines · Apache-2.0 · python

Reported evidence · last observed 2026-04-30. The source's designated baseline implementation. Reported by source; not independently reproduced.

No package · vendor the source →View source →Run detail →

Current records

Workloadnum_pages = 349 · batch_size = 4 · len_indptr = 5 · num_kv_indices = 212 · bf1618 cases
num_kv_indices
num_pages
batch_size
len_indptr
dtype
21234945bf1623238945bf1626839845bf1652883989bf1657690289bf1664091089bf16103316861617bf16114517351617bf16125717941617bf1613783151143233bf1614103151783233bf1614145152333132bf1626582291896465bf1627222294416465bf1627392295696364bf164641451792128129bf164743852033128129bf164812052481127128bf16
scrolls · 18 cases total

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload num_pages = 349 · batch_size = 4 · len_indptr = 5 · num_kv_indices = 212 · bf16 · CUDA 12.8 · Framework pytorch 2.9.1+cu128 · Protocol flashinfer-bench · mean · 1 results · last observed 2026-04-30Record history →
Estimated floor 11.4 µs · record 2.51× above itestimate, not evidence ›
DRAM 11.4 µs · bandwidth-bound on B200
every declared tensor crosses HBM exactly once (8,000 GB/s, B200 datasheet)
no arithmetic formula for this family: bandwidth floor only
headroom-v1: a lower bound from declared tensors and datasheet peaks. A kernel can sit well above it for good reasons.
#
Implementation
Latency
vs #1
Trust
Observed
1
flashinfer / wrapperad4135baselineFlashInfer-Bench baselines
28.8µs
1.00×
Reported · Apache-2.0 · source
2026-04-30

Measured exactly what you asked. The source's designated baseline implementation. Reported by source; not independently reproduced.

source mirroredApache-2.0no install recipeView source →Run detail →

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
flashinfer / wrapperad4135FlashInfer-Bench baselines
python
28.8µs
1.00×
Reported
Apache-2.0 · source

Semantics

Inputs and outputs
qbf16 [batch_size, num_qo_heads, head_dim]
k_cachebf16 [num_pages, page_size, num_kv_heads, head_dim]
v_cachebf16 [num_pages, page_size, num_kv_heads, head_dim]
kv_indptrint32 [len_indptr]
kv_indicesint32 [num_kv_indices]
kv_last_page_lenint32 [batch_size]
sm_scalefp32 scalar
outputbf16 [batch_size, num_qo_heads, head_dim]
lsefp32 [batch_size, num_qo_heads]
Axes and behavior
head_dimconstant = 128
num_pagesvariable
page_sizeconstant = 64
batch_sizevariable
len_indptrvariable
num_kv_headsconstant = 8
num_qo_headsconstant = 32
num_kv_indicesvariable
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliasgqa_paged_decode_h32_kv8_d128_ps64modelllama-3-1-8bsha256c19c9983aa5f…
Sources: FlashInfer-Bench (2026-04-30) · Apache-2.0last observed 2026-04-30How records are decidedJSON