Repository Issue Activity (beta)

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

#1504 SFT `fsdp_sft_trainer.py` to support FSDP2

#6577 [RFC]Dynamo KV-aware Routing, kv cache migration for Elastic scheduling in async RL, and KVBM Multi-tier KV Cache

#6947 Has anyone noticed that the saved MTP weights remain unchanged after enabling MTP training?

#6439 【Ascend】社区任务发布

#6946 [Feature request] Make fully_async rollouter max_concurrent_samples configurable (follow-up to #6306)

#6810 [OPD] [megatron/fsdp] use forward_kl_topk method occurs oom

#6943 DeepSeek-V4-Flash 全参 SFT 在 DSA attention (csa.py) OOM

#6936 Megatron backend still fails to build models on Megatron-Core >= 0.18 via unpinned `mbridge` (even after #6927)

#6931 [NPU]ModuleNotFoundError: No module named 'megatron.bridge'

#6829 Qwen3.5-4b ppo_kl != 0

#1733 [Bug] verl does not log the final validation stats

#6781 [Auto] Daily issue processing log

#1657 Race condition when creating checkpoint directories causes training failures

#1611 [SGLang Async Rollout] CUDA error: an illegal memory access was encountered

#1569 Unexpected performance degradation on GSM8K when multi-turn/tool use is enabled compared to disabled

#1524 Some doubts about batch size

#1515 [BUG] train_batch_size / val_batch_size has a great influence on the result

#1508 [FR] Better metric reduction utility

#1455 branch-main: where is metrics "actor/entropy_loss"?

#1410 Seamless Rollout

#1655 [Questions]The problem of repeated capture of cudagraph during weight update phase

#1654 Implement Turn-Level Advantage Computation

#1649 [Megatron] reproduce script examples/grpo_trainer/run_qwen3moe-30b_megatron.sh OOM

#1644 why DAO loss mean call token mean? but code is seq mean？