Repository Issue Activity (beta)

#6894 [FSDP2] GRPO reproducibly crashes at step 11 — "ACL stream synchronize failed, error code:507035" (device-side AIV MTE out-of-range) on Atlas 800T A3; FSDP1 stable for 40 steps

#7540 [distillation][bug] OPD multi-teacher distillation loss diverges between v0.9.0 and HEAD despite unchanged numerical code path (reproduced on NVIDIA B300 and Cambricon MLU)

#7509 [RFC] Agent Loop Observability Across VeRL, Uni-Agent and RL-Insight

#6252 Qwen3.5/Qwen3.6 35B-A3B 多轮工具调用 Agent RL 训练中出现工具调用格式异常并导致崩溃

#548 [Bug] A bug in the slurm script example. It will cause "srun: error: Unable to create step for job xxxxx: Requested node configuration is not available"

#7163 filter_overlong_prompts: SIGTERM handling issue during multi-process data filtering

#613 [Feature Request] ARM64 Support Needed: Error Converting Docker Image to Apptainer on Slurm Cluster

#2812 Agentic RL Training Example is not working

#7463 [config][MoE] actor.router_replay.mode is ignored; official docs point at a no-op key

#7269 [RFC] API Refactor: New abstraction of devices and machines allocation to each RL Roles

#7520 [Bug][FSDP2] Fused untied lm_head crashes during gradient accumulation with missing _unsharded_param

#7495 [BUG][LoRA][Async] Disaggregated weight sync never delivers the LoRA adapter — rollout keeps sampling from the initial policy

#7424 Liger fused linear PPO to replace the current FL-PPO implementation.

#7167 [ckpt]: NCCL checkpoint engine broadcast optimization

#7486 [BUG] AgentLoopOutput.as_dict() raises KeyError for default extra_fields

#4604 [RFC][rollout] Should `actor_rollout_ref.rollout.mode` be removed since it should always be `"async"` now?

#1508 [FR] Better metric reduction utility

#5331 Critical Security Vulnerabilities in Mathematical Evaluation and Serialization Code

#7505 [tool] bug: OpenAIFunctionToolSchema silently drops valid JSON Schema constraints

#6949 [docker] verlai/verl:vllm023.dev1 breaks on B300/GB300 (sm_103)

#7460 [trainer] critic.model.model_type is ignored when building TrainingWorkerConfig

#7493 [BUG][FSDP1] fsdp_size=1 on multiple GPUs silently disables gradient synchronization

#5678 Qwen3.5-35B-A3B MTP training rollout 通讯报错

#7489 训练Qwen2.5 32B时总是内存OOM

#7476 ulysses_sequence_parallel_size is a silent no-op on Qwen3.5: per-rank activations stay full length