GQA paged decode h20 kv4 d128 ps64
40 eligible runs
gqa-paged-attention
Batched Grouped Query Attention decode with a paged KV cache (page_size=64). Captured from Qwen3 14B at TP=2 (20 q-heads and 4 kv-heads per device). GQA ratio 5:1, head_dim=128.
Source baseline · unbeaten
95.2µsmean
flashinfer_wrapper_bcbabfFlashInfer-Bench baselines · Apache-2.0 · python
Reported evidence · last observed 2026-03-28. The source's designated baseline implementation. Reported by source; not independently reproduced.
Current records
Workloadnum_pages = 52 · batch_size = 1 · len_indptr = 2 · num_kv_indices = 51 · bf1620 cases
num_kv_indices
num_pages
batch_size
len_indptr
dtype
scrolls · 20 cases total
Not measured on H100 for this workload. Challenges →
Source-native comparison · GPU NVIDIA B200 · Workload num_pages = 52 · batch_size = 1 · len_indptr = 2 · num_kv_indices = 51 · bf16 · CUDA 12.8 · Framework pytorch 2.9.1+cu128 · Protocol flashinfer-bench · mean · 2 results · last observed 2026-03-28Record history →
Estimated floor 853 ns · record 111.67× above itestimate, not evidence ›
DRAM 853 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
195.2µs1.00×Reported · Apache-2.0 · source2026-03-28stale
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 →
2171.5µs1.80×Reported · Apache-2.0 · source2026-03-28stale
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
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 = 20
num_kv_indicesvariable
determinismunspecified
constraintsNo mutation or aliasing
Identity
Sources: FlashInfer-Bench (2026-03-28) · Apache-2.0last observed 2026-03-28How records are decidedJSON