GQA ragged prefill causal h32 kv16 d128
20 eligible runs
gqa-ragged-attention
Batched Grouped Query Attention prefill with ragged (variable-length) inputs. Causal mask is applied. Captured from Gemma 3 27B during total prefill. GQA ratio 2:1 (32 q, 16 kv, head_dim=128).
Source baseline · unbeaten
10.2µsmean
flashinfer / wrapper463e16FlashInfer-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 = 7 · total_kv = 7 · len_indptr = 2 · bf1616 cases
total_kv
total_q
len_indptr
dtype
scrolls · 16 cases total
Not measured on H100 for this workload. Challenges →
Source-native comparison · GPU NVIDIA B200 · Workload total_q = 7 · total_kv = 7 · len_indptr = 2 · bf16 · CUDA 13.0 · Framework pytorch 2.11.0+cu130 · Protocol flashinfer-bench · mean · 2 results · last observed 2026-04-09Record history →
Estimated floor 14 ns · record 714.19× above itestimate, not evidence ›
DRAM 14 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
110.2µs1.00×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 →
1=10.2µs1.00×Reported · Apache-2.0 · source2026-04-09
Measured exactly what you asked; tied ranks share a number. 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 / wrapper463e16FlashInfer-Bench baselines
python
10.2µs
1.00×
Reproduction-ready
Apache-2.0 · source
Semantics
Inputs and outputs
qbf16 [total_q, num_qo_heads, head_dim]
kbf16 [total_kv, num_kv_heads, head_dim]
vbf16 [total_kv, num_kv_heads, head_dim]
qo_indptrint32 [len_indptr]
kv_indptrint32 [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
total_kvvariable
len_indptrvariable
num_kv_headsconstant = 16
num_qo_headsconstant = 32
determinismunspecified
constraintsNo mutation or aliasing
Identity
Sources: FlashInfer-Bench (2026-04-09) · Apache-2.0last observed 2026-04-09How records are decidedJSON