GQA paged prefill causal h32 kv4 d128 ps64
5 eligible runs
gqa-paged-attention
Batched Grouped Query Attention prefill with a paged KV cache (page_size=64). Causal mask is applied. Captured from Qwen3-30B-A3B during incremental prefill.
Source baseline · unbeaten
64.5µsmean
flashinfer / wrapper50ca11FlashInfer-Bench baselines · Apache-2.0 · python
Reproduction-ready evidence · last observed 2026-04-09. The source's designated baseline implementation. Reported by source; not independently reproduced.
Current records
Workloadtotal_q = 992 · num_pages = 16 · batch_size = 2 · len_indptr = 2 · num_kv_indices = 16 · bf16total_q = 3968 · num_pages = 64 · batch_size = 5 · len_indptr = 5 · num_kv_indices = 64 · bf16total_q = 7936 · num_pages = 128 · batch_size = 9 · len_indptr = 9 · num_kv_indices = 128 · bf16total_q = 15872 · num_pages = 256 · batch_size = 17 · len_indptr = 17 · num_kv_indices = 256 · bf16total_q = 31744 · num_pages = 512 · batch_size = 33 · len_indptr = 33 · num_kv_indices = 512 · bf16
Not measured on H100 for this workload. Challenges →
Source-native comparison · GPU NVIDIA B200 · Workload total_q = 992 · num_pages = 16 · batch_size = 2 · len_indptr = 2 · num_kv_indices = 16 · bf16 · CUDA 13.0 · Framework pytorch 2.11.0+cu130 · Protocol flashinfer-bench · mean · 1 results · last observed 2026-04-09Record history →
Estimated floor 1.28 µs · record 50.48× above itestimate, not evidence ›
DRAM 1.28 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
164.5µs1.00×Reproduction-ready · Apache-2.0 · source2026-04-09
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 / wrapper50ca11FlashInfer-Bench baselines
python
64.5µs
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 [batch_size]
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
batch_sizevariable
len_indptrvariable
num_kv_headsconstant = 4
num_qo_headsconstant = 32
num_kv_indicesvariable
determinismunspecified
constraintsNo mutation or aliasing
Identity
Sources: FlashInfer-Bench (2026-04-09) · Apache-2.0last observed 2026-04-09How records are decidedJSON