Accelerating Reinforcement Learning Weight Refit with the Reshard Collective
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
NVIDIA's NCCL M2N Reshard, an open-source reference implementation of a new collective, targets RL weight refit, the step where updated parameters move from training GPUs to inference GPUs.