general family · filed under inference & serving
Batch-invariant deterministic kernels
A family of kernels designed to produce bitwise-reproducible results independent of batch composition.
Also called Deterministic and batch-invariant kernels.
- sources
- 2
- models
- 2
- lab adopt it
- 1
- strongest
- core
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 2core 1
Documented in
Evidence
3 spans quoted from the sources, strongest treatment first.
we implement end-to-end, bitwise batch-invariant, and deterministic kernels with minimal performance overhead
coreunclearin DeepSeek-V4DeepSeek
we implement end-to-end, bitwise batch-invariant, and deterministic kernels with minimal performance overhead
usedsoftware implementationin DeepSeek-V4DeepSeek
Deterministic and batch-invariant kernels ensure that training is bitwise reproducible.
usedsoftware implementationin DeepSeek-V4DeepSeek
Filed alongside
Other methods under inference & serving :: inference kernel.
CUDA 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 4FlashKDAFP8 GEMMFused AttnRes merge and RMSNorm kernelFused latent down-projection and MoE-router GEMMFused mHC kernelsFused QSA kernelFused RoPE-attention-RoPE-cast kernel