verl · Issues· 1228 open

Google TPU support in verl with Ray

[megatron] seq-mean-token-mean guard rejects CP=1 and appears overly restrictive for reconstructed THD outputs

[Bug] GMPO geo_mean loss changes when a minibatch is split into more microbatches

`log_prob` and `old_log_prob` differ in on-policy setting when rollout and actor micro batch per gpu differ

Multi-LoRA training: resident adapters and mixed-adapter steps on one base model

[rollout][vllm] Garbled multi-language output after weight sync when `free_cache_engine=true` (sleep/resume corrupts rollout weights)

[bug] equal-length batches of 3D mRoPE position_ids produce an inconsistent jagged layout → split_with_sizes crash in V1 trainer

[Tracking] Sharded delta weight sync (delta_sharded): roadmap & known issues

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

[Bug] Qwen3.5 system-only prompt still fails through Continuous Token

[RFC] Tail-aware top-k KL for on-policy distillation

[RFC] Enhanced `verl.single_controller`: a backend-neutral WorkerGroup runtime

[fsdp] Tied embeddings disable rank0-only weight loading, so host RAM scales with rank count; the naive fix unties the model silently

[Bug] KL-Cov token selection changes with the PPO microbatch size

[data] The val_max_samples subsample keys off data.shuffle, so data.validation_shuffle never reaches it