AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Forge: Compiling a Unified Abstraction into Scalable Kernels for Linear Attention

Haojie Duanmu, Size Zheng, Ningxin Zheng, Jianqiao Lu, Xuegui Zheng, Xingcheng Zhang, Li-Wen Chang, Xin Liu, Dahua Lin

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Linear attention approximations

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add AOT compilation with static dispatcher (ablation)FlexLA Scalar GLA · seq len 1024dispatch overhead / end-to-end latency: reduces overhead by 46%, 1.6x end-to-end speedupImproved
add AOT compilation with static dispatcher (ablation)FlexLA Scalar GLA · seq len 8192dispatch overhead: 101 us -> 1 usImproved
add tile-level compute-communication fusion instead of serial communication (ablation)FlexLA state propagation · 8xH800 67MB statecommunication time: 873 us -> 560 usImproved
use host-managed pipelined NCCL overlap instead of serial (ablation)state propagation · 8xH800 67MB statecommunication time: 873 -> 902Hurt
replace PyTorch eager with FlexLA compiled kernelscalar GLA · single GPUlatency: 34.6 s -> 9.2 msImproved
scale FlexLA from one GPU to 4 GPUsscalar GLA · Figure 1latency: 9.2 ms -> 2.7 msImproved
replace FLA kernels with FlexLA generated kernelsHGRN · single H100speedup: 1.64-2.02xImproved
replace FLA kernels with FlexLA generated kernelsscalar GLA, vector GLA, vanilla LA · single H100speedup: 1.1-2.0xImproved
replace FLA kernels with FlexLA generated kernelsGDN · single H100 long sequenceslatencyNo clear effect
replace LASP2 all-gather with FlexLA fused communicationGDN · weak scaling to 128 GPUslatency: LASP2 49.2ms on 4 GPUs to 345ms on 128 GPUs; FlexLA flatImproved
replace LASP-2 with FlexLAscalar GLA · H100, 4 GPUs, 16k tokens/GPUlatency (ms): 1.13 -> 0.96Improved
replace Torch-Compile with domain-specific compilerlinear attention · single H100latencyImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy