All issues

Search results

choice_reward.compute_score scores empty/malformed responses as correct when ground_truth has no <answer> tag

Status: Open.

#643 In verl-project/verl-omni;

· OnePunchMonk opened on Sep 20, 2026

[Feature] Report per-sample progress inside a diffusion training step

Status: Open.

#638 In verl-project/verl-omni;

· ruiling-smartbear opened on Sep 20, 2026

[RFC] 0-diff train-rollout log-probs for the Qwen3-Omni thinker

enhancementNew feature or requestNew feature or request

Status: Open.

#637 In verl-project/verl-omni;

· cr-gao opened on Sep 20, 2026

[RFC] Add GPU and NPU profiling support to the diffusion V1 trainer

Status: Open.

#632 In verl-project/verl-omni;

· SelfDLer opened on Sep 18, 2026

[RFC] MiniMax-H3 production-level training roadmap

enhancementNew feature or requestNew feature or request

help wantedExtra attention is neededExtra attention is needed

Status: Open.

#631 In verl-project/verl-omni;

· SamitHuang opened on Sep 18, 2026

[RFC]: MiniCPM-o 4.5 thinker GSPO on the omni separate-async V1 trainer (standalone rollout)

enhancementNew feature or requestNew feature or request

Status: Open.

Task

#628 In verl-project/verl-omni;

· zhtmike opened on Sep 18, 2026

[RFC] Save inference-ready LoRA adapters during diffusion training

Status: Open.

#626 In verl-project/verl-omni;

· mengchengTang opened on Sep 17, 2026

[RFC] Support MiniMax-H3 T2VA FlowGRPO on Cambricon MLU

Status: Open.

#619 In verl-project/verl-omni;

· SeptPonts opened on Sep 17, 2026

[Bug] Single-GPU FSDP1 NO_SHARD LoRA weight synchronization fails with CPU offload

Status: Open.

#618 In verl-project/verl-omni;

· 0z5a opened on Sep 17, 2026

[RFC] Fully-async TQ checkpoint recovery and v1 dynamic scheduling (diffusion + omni)

Status: Open.

#615 In verl-project/verl-omni;

· knlnguyen1802 opened on Sep 17, 2026

[RFC] delta_sharded weight sync for diffusion and omni engines

enhancementNew feature or requestNew feature or request

Status: Open.

#612 In verl-project/verl-omni;

· knlnguyen1802 opened on Sep 17, 2026

[bug] diffusion separate_async: the step-end weight sync waits for a running batch and discards results that are already computed

bugSomething isn't workingSomething isn't working

Status: Open.

#610 In verl-project/verl-omni;

· cr-gao opened on Sep 17, 2026