Optimizing DeepSeek-V4.1-Flash Decoding on 16 A800 Cards
Two 8x A800-40GB nodes running DeepSeek-V4.1-Flash (552B), how to squeeze out performance? This article records the whole process from excluding cross-machine communication bottlenecks, unloading 203GB knowledge base CPU lock pages, to writing Small-Batch Sparse MLA and Fused Markov sampling kernels using Triton, with a 24.6% decrease in ITL under concurrency.