specific method · filed under inference & serving
FlashKDA
A CUTLASS-based chunkwise KDA kernel that overlaps intra-chunk computation with cross-chunk state propagation.
Also called FlashKDA chunkwise kernel for KDA.
- source
- 1
- model
- 1
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
We therefore develop FlashKDA, a CUTLASS-based chunkwise kernel that overlaps intra-chunk computation with cross-chunk state propagation. The kernel decomposes the work into token-parallel stages and a head-parallel recurrence, each scheduled and tuned independently, and substantially outperforms the Triton reference implementation.
usedunclearin Kimi K3Moonshot AI
Filed alongside
Other methods under inference & serving :: inference kernel.
Batch-invariant deterministic kernelsCUDA GraphCUDA Graph capture size reductionFlashAttention 3Kernel fusionMoE-side chunkingSynchronization-free static-shape MoE executionTensorRT-LLM multi-head attention backendAvoiding split-KDeepGEMM-based batch-invariant matrix multiplicationDistributed shared memory for cross-SM attention data exchangeDual-kernel batch-invariant attention decodingDynamic load balancingExp-free TopK kernelExpert-optimized Triton kernelsFA4 sheared-bias attention kernelFlashAttentionFlashAttention 4FP8 GEMMFused AttnRes merge and RMSNorm kernelFused latent down-projection and MoE-router GEMMFused mHC kernelsFused QSA kernelFused RoPE-attention-RoPE-cast kernel