call for contributionenhancementgood first issue
仓库指标
- Star
- (21,533 star)
- PR 合并指标
- (平均合并 3天 6小时) (30 天内合并 125 个 PR)
描述
- Activation offloading (see implementation here)
- Fusing optimizer step into backward pass (see implementation here)
- Utilize
full_shardreshard_after_forward(see here). I wasn't 100% sure if I could see this already implemented in veRL.
These optimizations largely trade off decreased peak memory useage for additional compute, so may only be useful for training larger models, and in GPU-constrained settings.