Songlin Jiang

NeMo-DCR introduces a bit-exact delta refit that transfers only the weights that change per step in canonical coordinates, using fixed affine mappings to project changes directly, residual conversion to cover the rest, mixed XOR/overwrite encoding, in-place application through the serving runtime's native loader, and a recoverable joint commit; in 3% and 5% stress cases across 30B–1T models, refits are 12–40 faster than a transport-only full-checkpoint reference, and a 1T relay-tree refit takes 150 s instead of 87.5 min.

NeMo-DCR introduces a bit-exact delta refit that transfers only the weights that change per step in canonical coordinates, using fixed affine mappings to project changes directly, residual conversion to cover the rest, mixed XOR/overwrite encoding, in-place application through the serving runtime's native loader, and a recoverable joint commit; in 3% and 5% stress cases across 30B–1T models, refits are 12–40 faster than a transport-only full-checkpoint reference, and a 1T relay-tree refit takes 150 s instead of 87.5 min.

Nereus is a cost-aware runtime for reinforcement-learning post-training of large language models that represents each model-stage replica as an Elastic Model Unit and orchestrates Split, Merge, Extend, and Destroy primitives through a global transition graph, adapting the DP/TP/PP execution plan online under a payback rule; on a trace built from real data, online TP/PP adaptation cuts average step latency by 27.7% versus the initial fixed TP/PP layout, six transitions in a 1,000-step run scaling to 1,024 GPUs consume 0.079% of total run time, and end-to-end 8B PPO throughput improves by 2.14-7.27x over OpenRLHF and 1.10-1.47x over Verl.

Nereus is a cost-aware runtime for reinforcement-learning post-training of large language models that represents each model-stage replica as an Elastic Model Unit and orchestrates Split, Merge, Extend, and Destroy primitives through a global transition graph, adapting the DP/TP/PP execution plan online under a payback rule; on a trace built from real data, online TP/PP adaptation cuts average step latency by 27.7% versus the initial fixed TP/PP layout, six transitions in a 1,000-step run scaling to 1,024 GPUs consume 0.079% of total run time, and end-to-end 8B PPO throughput improves by 2.14-7.27x over OpenRLHF and 1.10-1.47x over Verl.