[roadmap] verl 26Q2 roadmap

[wuxibin89]

Model Engine

• Megatron

  • Perf dynamic CP for long-context: [megatron] feat: support dynamic CP #5057
  • Megatron FSDP enabling
  • Megatron FSDP for VLM perf
  • Low precision training: MXFP8/NVFP4 [roadmap] low precision support #5972
  • Qwen 3.5 LoRA support: [megatron] fix: Qwen3.5 LoRA & MTP support (with Megatron-Bridge) #5599
  • Qwen 3.5 MTP support: [megatron] fix: Qwen3.5 LoRA & MTP support (with Megatron-Bridge) #5599 & [megatron] fix: MTP patch for newer mcore #5587

• Veomni:

• Torchtitan: [roadmap] TorchTitan Integration into Verl #5306

• Automodel: [RFC] support Nemo-AutoModel as an alternative training backend #5245

• Model support matrix [roadmap] model support list tracking #5389

• Ascend NPU support: [roadmap][ascend] verl on ascend 2026Q2 roadmap #5526

Rollout Engine

• Performance optimization

  • kv cache awareness load balance
  • workload awareness scheduling, e.g ThunderAgent
  • PD disaggregated
  • mxfp8 rollout [roadmap] low precision support #5972

• Weight refit optimization

  • mbridge only rank 0 export_weights
  • chunk large tensor to fit into smaller update_weights_bucket_megabytes
  • fp8 rollout: quantize weights on trainer side

• Memory optimization

  • NCCL suspend/resume to release communicator memory

• TRT-LLM: [roadmap] verl TensorRT-LLM rollout roadmap #5042

AgentGateway

Make every agent framework seamlessly integrated into verl for agentic rl training.

• [RFC] Agent Abstractions and Trajectory Gateway for VERL #5790
• [rollout, tool] feat: add experimental agent framework and gateway runtime #5931

Trainer

• Sync trainer

  • New sync trainer with TransferQueue: [trainer] feat: add new trainer with TranferQueue #5401
  • Deprecate legacy sync trainer verl/trainer/main_ppo.py

• One-step-off/Fully async trainer

  • Integrate TransferQueue
  • Move from verl/experimental into verl/trainer
  • Elastic rollout and fault tolerance
  • Hybrid arch: colocated+disaggregated
  • Router replay support [rollout] feat: enable R3 in fully async mode #5344
  • Standalone model engine server for log_probs

• Multi-modal Generation RL: [RFC] Multi-modal Generation RL 2026Q2 Roadmap #5755

• Speculative Decoding: target & draft model co-training, weight refiting [rfc]:add speculator training scripts and checkpoint support #4947

• Online policy distillation

  • Multiple teacher support [trainer,cfg,rollout,algo] feat: (MOPD, 1/2) Multi-Teacher Model, Server Managers and Config #5834
  • OPD support for fully async trainer
  • Support sglang/trt-llm teacher server
  • docs

Agentic training recipe

Multi-turn, long context(>=128k), context compression, fully async training: https://github.com/verl-project/uni-agent

• SWE agent: fully async example
• Search agent
• summary agent [algo] feat: supporting agentic rl with context management #5636

Misc

• Remove legacy workers: verl/workers/fsdp_workers.py, verl/workers/megatron_workers.py
• Refactor and clean up examples scripts
  • Unify GPU/NPU scripts
  • Remove outdated examples

[NLPJCL]

I'd like to know if there's a pull request (PR) discussing what the search agent in Verl's Agentic Training recipe should do, so I can see if it can support some of the features.
@wuxibin89

[wuxibin89]

I'd like to know if there's a pull request (PR) discussing what the search agent in Verl's Agentic Training recipe should do, so I can see if it can support some of the features. @wuxibin89

@wdl339 is working on search agent.