GQA paged decode h32 kv8 d64 ps1
8 eligible runs
gqa-paged-attention
Batched Grouped Query Attention decode with a paged KV cache. Captured from Llama-3.2-1B.
Source baseline · unbeaten
8.15µsmean
flashinfer / wrapper3f9411FlashInfer-Bench baselines · Apache-2.0 · python
Reported evidence · last observed 2026-04-20. The source's designated baseline implementation. Reported by source; not independently reproduced.
Current records
Workloadnum_pages = 9 · batch_size = 1 · len_indptr = 2 · num_kv_indices = 8 · bf16num_pages = 73 · batch_size = 1 · len_indptr = 2 · num_kv_indices = 56 · bf16num_pages = 18 · batch_size = 1 · len_indptr = 2 · num_kv_indices = 2 · bf16num_pages = 16 · batch_size = 1 · len_indptr = 2 · num_kv_indices = 15 · bf16
Not measured on H100 for this workload. Challenges →
Source-native comparison · GPU NVIDIA B200 · Workload num_pages = 9 · batch_size = 1 · len_indptr = 2 · num_kv_indices = 8 · bf16 · CUDA 13.0 · Framework pytorch 2.11.0+cu130 · Protocol flashinfer-bench · mean · 2 results · last observed 2026-04-20Record history →
Estimated floor 3 ns · record 2890.82× above itestimate, not evidence ›
DRAM 3 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
18.15µs1.00×Reported · Apache-2.0 · source2026-04-20
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 →
28.16µs1.00×Reported · Apache-2.0 · source2026-04-20
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]
sm_scalefp32 scalar
outputbf16 [batch_size, num_qo_heads, head_dim]
lsefp32 [batch_size, num_qo_heads]
Axes and behavior
head_dimconstant = 64
num_pagesvariable
page_sizeconstant = 1
batch_sizevariable
len_indptrvariable
num_kv_headsconstant = 8
num_qo_headsconstant = 32
num_kv_indicesvariable
determinismunspecified
constraintsNo mutation or aliasing
Identity
Sources: FlashInfer-Bench (2026-04-20) · Apache-2.0last observed 2026-04-20How records are decidedJSON