@ziming_mao on Backlist

mKernel: fused compute and communication kernels for multi-node GPUs (t.co)

mKernel fuses compute and communication into persistent GPU kernels across intra- and inter-node systems, attacking a real bottleneck in distributed training