Reviews of machine learning research papers, newest first.
Lianghui Zhu, Bencheng Liao, +4
Follow Vision Mamba from image patches through selective state updates, two aligned scan directions, a constant-resolution backbone, and runnable Python.
Yue Liu, Yunjie Tian, +7
Follow VMamba from a compact running state through four image scan orders, aligned cross-merge, the final VSS block, and runnable Python.
Hanyu Zhao, Zhenhua Han, +9
HiveD reserves groups of GPUs that sit together, not GPU counts, so a tenant never queues longer in a shared cluster than in a private one of the same size.
Mohammad Shoeybi, Mostofa Patwary, +4
Megatron-LM splits each transformer layer across GPUs with two all-reduces forward and two backward, and trains an 8.3B-parameter GPT-2 on 512 V100s.
Min Si, Pavan Balaji, +37
NCCLX is Meta's communication stack for Llama 4: host-driven, zero-copy collectives that start a 96K-GPU job 11x faster and cut decode time by up to 83%.
Yunge Li, Lanyu Xu
Reordering image tokens along a Hilbert curve makes 2D local attention block-sparse, for about 4x faster window attention and 18x faster slide attention.
Leyla Naz Candogan, Arshia Afzal, +2
VIOLIN multiplies ViT attention by decay masks from eight space-filling curves: 1,296 extra parameters on DeiT-B and up to 8.7 points on VTAB-1K spatial tasks.
Siyuan Shen, Anton Korzh, +11
How the Every µs Matters paper brings small GPU AllReduce operations close to the hardware speed-of-light bound: it measures a 1.404 µs floor on GB200, removes memory barriers with LL, sentinel and LL128 atomic signalling, and cuts vLLM inter-token latency by up to 13%.
Kan Zhu, Yufei Gao, +14
How NanoFlow raises LLM serving throughput by overlapping compute, memory and network work inside a single GPU: it shows serving is compute-bound, splits each batch into nano-batches, and shares SMs between concurrent kernels, reaching 1.91x the throughput of TensorRT-LLM.