Skip to content
KernelIndex
Search⌘K
Sign in

GQA paged prefill causal h16 kv1 d128 ps64

14 eligible runs
gqa-paged-attention

Batched Grouped Query Attention prefill with a paged KV cache (page_size=64). Causal mask applied. From Qwen3 235B A22B at TP=4. 16 q-heads, 1 kv-heads, head_dim=128.

Source baseline · unbeaten
16.7µsmean
flashinfer / wrappera3c91fFlashInfer-Bench baselines · Apache-2.0 · python

Reported evidence · last observed 2026-04-05. The source's designated baseline implementation. Reported by source; not independently reproduced.

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

Current records

Workloadtotal_q = 1 · num_pages = 16 · len_indptr = 2 · num_kv_indices = 1 · bf1620 cases
num_pages
num_kv_indices
total_q
len_indptr
dtype
7662bf1616112bf165134342bf1621334342bf1635817172bf1645717172bf1660162622bf1679162622bf16101293932bf16123393932bf1616312702702bf1620292702702bf16217316162bf16231716162bf16246217172bf16260717172bf162873552bf16319734342bf16416696996916bf16628470702bf16
scrolls · 20 cases total

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload total_q = 1 · num_pages = 16 · len_indptr = 2 · num_kv_indices = 1 · bf16 · CUDA 12.8 · Framework pytorch 2.9.1+cu128 · Protocol flashinfer-bench · mean · 1 results · last observed 2026-04-05Record history →
Estimated floor 66 ns · record 252.58× above itestimate, not evidence ›
DRAM 66 ns · 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 / wrappera3c91fbaselineFlashInfer-Bench baselines
16.7µs
1.00×
Reported · Apache-2.0 · source
2026-04-05stale

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 / wrappera3c91fFlashInfer-Bench baselines
python
16.7µs
1.00×
Reproduction-ready
Apache-2.0 · source

Semantics

Inputs and outputs
qbf16 [total_q, 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]
qo_indptrint32 [len_indptr]
kv_indptrint32 [len_indptr]
kv_indicesint32 [num_kv_indices]
kv_last_page_lenint32 [len_indptr]
sm_scalefp32 scalar
outputbf16 [total_q, num_qo_heads, head_dim]
lsefp32 [total_q, num_qo_heads]
Axes and behavior
total_qvariable
head_dimconstant = 128
num_pagesvariable
page_sizeconstant = 64
len_indptrvariable
num_kv_headsconstant = 1
num_qo_headsconstant = 16
num_kv_indicesvariable
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliasgqa_paged_prefill_causal_h16_kv1_d128_ps64modelqwen3-235bsha256a830f990e947…
Sources: FlashInfer-Bench (2026-04-05) · Apache-2.0last observed 2026-04-05How records are decidedJSON