List your company

Efficiency Benchmarking on Qwen3-8B (Single-layer forward+backward setup)

36.6Time (ms)DualKV
-77.256691.2721,459.82,228.328May 14, 2026

Evaluation results

MethodLinks
DualKVPrompt length (P)=5K, Micro-batch size (mb)=42026.0536.62.551.984
DualKVPrompt length (P)=8K, Micro-batch size (mb)=42026.0557.22.62.174.9
DualKVPrompt length (P)=5K, Micro-batch size (mb)=82026.0561.42.52.555.5
DualKVPrompt length (P)=8K, Micro-batch size (mb)=82026.0570.33.493.36.5
Prefix Grouper (PG)Prompt length (P)=5K, Micro-batch size (mb)=42026.0572.5——9
FA2Prompt length (P)=5K, Micro-batch size (mb)=42026.0593.2——4
DualKVPrompt length (P)=16K, Micro-batch size (mb)=42026.05108.93.032.267.5
DualKVPrompt length (P)=5K, Micro-batch size (mb)=162026.05110.43.072.859
Prefix Grouper (PG)Prompt length (P)=8K, Micro-batch size (mb)=42026.05124.3——12.8
DualKVPrompt length (P)=16K, mb=8, Hardware=H1002026.051254.413.479
DualKVPrompt length (P)=16K, Micro-batch size (mb)=82026.05125.44.413.479
DualKVPrompt length (P)=8K, mb=16, Hardware=H1002026.051284.283.639.9
DualKVPrompt length (P)=8K, Micro-batch size (mb)=162026.05128.24.283.639.9
FA2Prompt length (P)=8K, Micro-batch size (mb)=42026.05148.9——4.9
FA2Prompt length (P)=5K, Micro-batch size (mb)=82026.05153.3——5.5
Prefix Grouper (PG)Prompt length (P)=5K, Micro-batch size (mb)=82026.05156.9——18.1
DualKVPrompt length (P)=5K, Micro-batch size (mb)=322026.05179.93.323.4813.7
DualKVPrompt length (P)=5K, mb=32, Hardware=H1002026.051803.323.4813.7
DualKVPrompt length (P)=8K, Micro-batch size (mb)=322026.05184.2——14.6
DualKVPrompt length (P)=16K, Micro-batch size (mb)=162026.05205.8——12.5
DualKVPrompt length (P)=32K, Micro-batch size (mb)=22026.05220.31.851.411.8
Prefix Grouper (PG)Prompt length (P)=8K, Micro-batch size (mb)=82026.05232.1——25.8
FA2Prompt length (P)=8K, Micro-batch size (mb)=82026.05245.4——6.5
Prefix Grouper (PG)Prompt length (P)=16K, Micro-batch size (mb)=42026.05245.7——22.9
DualKVPrompt length (P)=32K, mb=4, Hardware=H1002026.052463.292.0812.5
DualKVPrompt length (P)=32K, Micro-batch size (mb)=42026.05246.43.292.0812.5
DualKVPrompt length (P)=32K, Micro-batch size (mb)=82026.05266.8——14.1
Prefix Grouper (PG)Prompt length (P)=32K, Micro-batch size (mb)=22026.05308.1——22.1
Prefix Grouper (PG)Prompt length (P)=5K, Micro-batch size (mb)=162026.05314.2——35.4
FA2Prompt length (P)=16K, Micro-batch size (mb)=42026.05329.9——7.5
FA2Prompt length (P)=5K, Micro-batch size (mb)=162026.05338.9——9
FA2Prompt length (P)=32K, Micro-batch size (mb)=22026.05408——11.8
Prefix Grouper (PG)Prompt length (P)=16K, Micro-batch size (mb)=82026.05434.9——45.9
PGPrompt length (P)=16K, mb=8, Hardware=H1002026.05435——45.9
PGPrompt length (P)=8K, mb=16, Hardware=H1002026.05465——50.7
Prefix Grouper (PG)Prompt length (P)=8K, Micro-batch size (mb)=162026.05465.4——50.7
Prefix Grouper (PG)Prompt length (P)=32K, Micro-batch size (mb)=42026.05513.6——43
PGPrompt length (P)=32K, mb=4, Hardware=H1002026.05514——43
FA2Prompt length (P)=8K, mb=16, Hardware=H1002026.05549———
FA2Prompt length (P)=8K, Micro-batch size (mb)=162026.05549.2——9.9
FA2Prompt length (P)=16K, Micro-batch size (mb)=82026.05553——9
FA2Prompt length (P)=16K, mb=8, Hardware=H1002026.05553———
FA2Prompt length (P)=5K, mb=32, Hardware=H1002026.05597———
FA2Prompt length (P)=5K, Micro-batch size (mb)=322026.05597.2——13.7
DualKVPrompt length (P)=65K, Micro-batch size (mb)=22026.05614.11.911.2621.9
PGPrompt length (P)=5K, mb=32, Hardware=H1002026.05626——69.8
Prefix Grouper (PG)Prompt length (P)=5K, Micro-batch size (mb)=322026.05626.3——69.8
DualKVPrompt length (P)=65K, Micro-batch size (mb)=42026.05653.9——22.7
DualKVPrompt length (P)=65K, mb=4, Hardware=H1002026.05654——22.7
Prefix Grouper (PG)Prompt length (P)=65K, Micro-batch size (mb)=22026.05773.3——42.3
FA2Prompt length (P)=32K, mb=4, Hardware=H1002026.05810———
FA2Prompt length (P)=32K, Micro-batch size (mb)=42026.05810.2——12.5
DualKVPrompt length (P)=65K, Micro-batch size (mb)=162026.05892.6——27.7
FA2Prompt length (P)=65K, Micro-batch size (mb)=22026.051,173.5——21.9
DualKVPrompt length (P)=131K, Micro-batch size (mb)=82026.052,096.7——44.5
DualKVPrompt length (P)=131K, mb=8, Hardware=H1002026.052,097——44.5
DualKVPrompt length (P)=131K, Micro-batch size (mb)=322026.052,883——52.6