Skip to content
KernelIndex
Search⌘K
Sign in

GQA ragged prefill causal h20 kv4 d128

10 eligible runs
gqa-ragged-attention

Batched Grouped Query Attention prefill with ragged (variable-length) inputs. Causal mask is applied. Captured from Qwen3 14B at TP=2 during total prefill. 20 q-heads, 4 kv-heads per device, head_dim=128.

Source baseline · unbeaten
84.2µsmean
flashinfer_wrapper_023122FlashInfer-Bench baselines · Apache-2.0 · python

Reported evidence · last observed 2026-03-24. 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 = 37 · total_kv = 37 · len_indptr = 2 · bf1610 cases
total_kv
total_q
len_indptr
dtype
112bf16662bf1637372bf1650502bf162002002bf164004009bf165005002bf168008005bf16102410242bf161600160033bf16

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload total_q = 37 · total_kv = 37 · len_indptr = 2 · bf16 · CUDA 12.8 · Framework pytorch 2.9.1+cu128 · Protocol flashinfer-bench · mean · 1 results · last observed 2026-03-24Record history →
Estimated floor 33 ns · record 2538.4× above itestimate, not evidence ›
DRAM 33 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_wrapper_023122baselineFlashInfer-Bench baselines
84.2µs
1.00×
Reported · Apache-2.0 · source
2026-03-24stale

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_wrapper_023122FlashInfer-Bench baselines
python
84.2µs
1.00×
Reported
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 = 4
num_qo_headsconstant = 20
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliasgqa_ragged_prefill_causal_h20_kv4_d128modelqwen3-14bsha2566a60b6d678dc…
Sources: FlashInfer-Bench (2026-03-24) · Apache-2.0last observed 2026-03-24How records are decidedJSON