Download Latest Version v0.8.0 source code.zip (2.8 MB)
Email in envelope

Get an email when there's a new version of verl

Home / v0.8.0
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-06-01 59.7 kB
0
v0.8.0 source code.tar.gz 2026-06-01 1.8 MB
0
v0.8.0 source code.zip 2026-06-01 2.8 MB
0
Totals: 3 Items   4.6 MB 0

Highlights

Training

Megatron

  • Megatron-FSDP mode for the Megatron backend (#5423).
  • Dynamic context parallel (#5057) and CP for BSHD format (#5826).
  • Checkpoint save as HF PEFT format (#5575).
  • Qwen3.5 MTP SFT/RL (#5898)
  • NVFP4 (W4A16) QAT training via ModelOpt (#5254) with QAT documentation (#5861).

VeOmni

  • VeOmni engine for on-policy distillation (#6072) with native critic support (#6453) and native return_log_probs path (#6184).
  • MoE router replay (R2/R3) support (#6325) and MoE load-balance monitoring (#6470).
  • Qwen3.5 SP, GRPO demos for Qwen3-VL-30B-MOE / Qwen3.5-122b-a10b / Qwen3.5-30b (#6061, [#5275], [#6264], [#6323]).

Hardware / NPU

  • NPU support for Liger-Kernel (#6244).
  • Expandable segments (#5795, [#6346]).
  • MXFP8 rollout on Ascend 950 (#5756).

Rollout

vLLM

  • Split large weights into chunks in NCCL/NIXL checkpoint engine (#6091).
  • MooncakeStoreConnector with hard-reset on weight update (#6373).
  • Upgrade stable version to vllm==0.20.2 (#6393).

SGLang

  • LoRA support for SGLang rollouts (merge + native adapter paths) (#5564).
  • SGLang Prefill-Decode disaggregated rollout (#6117).
  • Upgrade stable version to sglang==0.5.12 (#6435).

TensorRT-LLM

  • Async RL for TRT-LLM rollout, including inter-node support (#5631, [#5992]).

On-Policy Distillation (OPD)

  • End-to-end on-policy distillation support across FSDP, Megatron and VeOmni backends (#5041).
  • Text and multimodal, single/multiple teachers (#6051), sync (#5997) and fully-async (#6056) trainers.
  • Reverse KL loss and forward top-k KL loss, with a fused top-K distillation kernel (#6511) and overlap metrics (#6469).
  • Docs: https://verl.readthedocs.io/en/latest/algo/opd.html

Trainer

Sync Trainer

  • New sync trainer with TransferQueue to decouple control flow and data flow in the single controller (#5401).
  • Up to 2x speedup in multimodal training with 128 GPUs, see TransferQueue blog post.
  • Support multiple trajectories output for each agent loop.
  • TBD: Fully async trainer with TransferQueue will be in next release.

Fully Async Trainer

  • Reuse trainer worker group for hybrid rollout during validation (#6076).
  • Online policy distillation in fully async training (#6056).
  • Qwen3-VL-30B-A3B and Qwen3-VL-8B fully async GRPO training scripts (#6131, [#6006]).

Agentic RL Training

  • New project verl-project/uni-agent for building, running, and training general agents at scale.
  • Support coding, search, GUI agent RL training.

Tools & Reward

  • Simpler function-based tool registration (#6189) with per-sample tool environment routing (#5978) and Gemma4 tool parser (#6406).
  • Multi-step support in skip_rollout v2 (#5556) and improved error messages for malformed tool calls (#6055).

Breaking Changes

  • Deprecate legacy FSDP and Megatron workers, migrate to the unified engine abstraction (#5604, [#6067]).
  • Deprecate verl/interactions (#6074).
  • Move LLMServerManager out of AgentLoopManager (#6129).
  • Migrate Diffusion RL stack to verl-omni (#6200) and experimental/vla to a standalone verl-vla repository (#6162).
  • Remove curriculum sampler + dynamic dataset + tool examples (#6302).
  • main_ppo.py is deprecated with a warning in favor of main_ppo_sync.py (#6384).

What's Changed

New Contributors

Full Changelog: https://github.com/verl-project/verl/compare/v0.7.1...v0.8.0

Source: README.md, updated 2026-06-01