Skip to content
KernelIndex
Search⌘K
Sign in

GQA paged decode h24 kv4 d128 ps64

20 eligible runs
gqa-paged-attention

Batched Grouped Query Attention decode with a paged KV cache (page_size=64). Captured from Mixtral 8x22B at TP=2. 24 q-heads, 4 kv-heads, head_dim=128.

Source baseline · unbeaten
359.8µsmean
flashinfer / wrapper1b7890FlashInfer-Bench baselines · Apache-2.0 · python

Reported evidence · last observed 2026-03-31. 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 = 425 · batch_size = 1 · len_indptr = 2 · num_kv_indices = 361 · bf1620 cases
num_kv_indices
num_pages
batch_size
len_indptr
dtype
130134823bf16247161045bf16275172945bf1636142512bf16471645889bf16527678589bf16882126612bf16993149131617bf161954206733233bf162178218893233bf163176888889bf163939229136465bf164387250896465bf164583618045bf166068149771617bf16786527201128129bf1679411427189bf16876131809128129bf1612162207373233bf1616593197131617bf16
scrolls · 20 cases total

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload num_pages = 425 · batch_size = 1 · len_indptr = 2 · num_kv_indices = 361 · bf16 · CUDA 12.8 · Framework pytorch 2.9.1+cu128 · Protocol flashinfer-bench · mean · 1 results · last observed 2026-03-31Record history →
Estimated floor 6.96 µs · record 51.67× above itestimate, not evidence ›
DRAM 6.96 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 / wrapper1b7890baselineFlashInfer-Bench baselines
359.8µs
1.00×
Reported · Apache-2.0 · source
2026-03-31stale

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 / wrapper1b7890FlashInfer-Bench baselines
python
359.8µs
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 = 4
num_qo_headsconstant = 24
num_kv_indicesvariable
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliasgqa_paged_decode_h24_kv4_d128_ps64modelmixtral-8x22bsha256074d68773e74…
Sources: FlashInfer-Bench (2026-03-31) · Apache-2.0last observed 2026-03-31How records are decidedJSON