OpenRLHF · Issues· 381 open

Suboptimal forward dispatch order in make_experience causes unnecessary serialization

LoRA + colocated vLLM: policy→vLLM weight sync fails end-to-end (dtype, then unmapped adapter names)

reward model HTTP service has no authentication, defaults to binding all interfaces, and the training client consumes whatever the endpoint returns

[Bug] Async oversampling checkpoints drop buffered rollouts on resume

[Bug] RewardDataset silently keeps preference pairs that become identical after truncation

[Roadmap] OpenRLHF on Intel XPU

[Bug] seq-mask-tis skips token-level truncation and can produce non-finite PPO updates

assert "expandable_segments:True" not in conf报错

Molt brings an Automodel-powered backend (AutoTP/EP/CP) to OpenRLHF

Droping experiences should also consider standard deviation when using non-binary rewards

Feature request: reward-hacking onset monitoring hooks during PPO/GRPO training

Fix SignalActor concurrency issues and refactor state management based on condition lock

overlong_penalty should exclude tool response tokens from length calculation (use action_mask/action_ranges)

Consider using pre-built Flash Attention kernels via `kernels`

ray error for multi-nodes training