FlashSampling: Fast and Memory-Efficient Exact Sampling
This work addresses a performance bottleneck in large language model inference for developers and researchers, offering a practical improvement in decoding speed and memory efficiency.
The paper tackles the memory and computational inefficiency of sampling from categorical distributions in large-vocabulary decoding by introducing FlashSampling, an exact sampling primitive that fuses sampling into the LM-head matrix multiplication, eliminating the need to materialize logits in high-bandwidth memory. It achieves up to a 19% reduction in time per output token in end-to-end experiments on various GPU models.
Sampling from a categorical distribution is mathematically simple, but in large-vocabulary decoding, it often triggers extra memory traffic and extra kernels after the LM head. We present FlashSampling, an exact sampling primitive that fuses sampling into the LM-head matmul and never materializes the logits tensor in HBM. The method is simple: compute logits tile-by-tile on chip, add Gumbel noise, keep only one maximizer per row and per vocabulary tile, and finish with a small reduction over tiles. The fused tiled kernel is exact because $\argmax$ decomposes over a partition; grouped variants for online and tensor-parallel settings are exact by hierarchical factorization of the categorical distribution. Across H100, H200, B200, and B300 GPUs, FlashSampling speeds up kernel-level decode workloads, and in end-to-end vLLM experiments, it reduces time per output token by up to $19%$ on the models we test. These results show that exact sampling, with no approximation, can be integrated into the matmul itself, turning a bandwidth-bound postprocessing step into a lightweight epilogue. Project Page: https://github.com/FlashSampling/FlashSampling.