Skip to content
KernelIndex
Search⌘K
Sign in

MLA ragged prefill causal h16 qk192 vo128

20 eligible runs
mla-ragged

Batched Multi-head Latent Attention prefill with ragged (variable-length) inputs. Uses the absorbed MLA formulation with combined QK dimension (qk_nope=128 + qk_rope=64 = 192) and value output dimension 128. Causal mask is applied. Captured from DeepSeek-V3 during total prefill (no prefix cache) with tensor parallel size 8.

Source baseline · unbeaten
10.2µsmean
flashinfer / wrapperd90b98FlashInfer-Bench baselines · Apache-2.0 · python

Reported evidence · last observed 2026-04-06. 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 = 86 · total_kv = 86 · len_indptr = 3 · bf1610 cases
total_kv
total_q
len_indptr
dtype
112bf1643432bf1686863bf161301304bf161521522bf161371137130bf163157315763bf161294746716bf161807421832bf1619013288564bf16

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload total_q = 86 · total_kv = 86 · len_indptr = 3 · bf16 · CUDA 13.0 · Framework pytorch 2.11.0+cu130 · Protocol flashinfer-bench · mean · 2 results · last observed 2026-04-06Record history →
Estimated floor 176 ns · record 58.08× above itestimate, not evidence ›
DRAM 176 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 / wrapperd90b98baselineFlashInfer-Bench baselines
10.2µs
1.00×
Reported · Apache-2.0 · source
2026-04-06

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 →
2
flashinfer / wrapperd90b98baselineFlashInfer-Bench baselines
10.3µs
1.01×
Reported · Apache-2.0 · source
2026-04-06

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 / wrapperd90b98FlashInfer-Bench baselines
python
10.2µs
1.00×
Reproduction-ready
Apache-2.0 · source

Semantics

Inputs and outputs
qbf16 [total_q, num_qo_heads, qk_dim]
kbf16 [total_kv, num_kv_heads, qk_dim]
vbf16 [total_kv, num_kv_heads, vo_dim]
qo_indptrint32 [len_indptr]
kv_indptrint32 [len_indptr]
sm_scalefp32 scalar
outputbf16 [total_q, num_qo_heads, vo_dim]
lsefp32 [total_q, num_qo_heads]
Axes and behavior
qk_dimconstant = 192
vo_dimconstant = 128
total_qvariable
total_kvvariable
len_indptrvariable
num_kv_headsconstant = 16
num_qo_headsconstant = 16
determinismunspecified
constraintsNo mutation or aliasing
Identity
aliasmla_ragged_prefill_causal_h16_qk192_vo128modeldeepseek-v3modeldeepseek-r1sha256da5710dddb62…
Sources: FlashInfer-Bench (2026-04-06) · Apache-2.0last observed 2026-04-06How records are decidedJSON