[Roadmap] Relax 2026 Q2

[Yangruipis]

Agentic

• Agentic Rollout — training real agent apps with strict app–RL-framework separation and lightweight API integration
• Recipes
  • DeepEyes v2 — multi-modal visual reasoning agent
  • R2E-Gym (DeepSWE) — repo-level software engineering agent training

High Performance

• Compact Deployment — ref/actor/actor_fwd colocated with async & sync support
• Megatron Dynamic CP — dynamic context parallelism for variable-length sequences
• Low-precision training (FP8 / NVFP4 / INT4) — FP8 and INT4 quantized training for memory and throughput gains
• Dynamo as rollout model router (with @NVIDIA)

Model

• Megatron-Bridge upgrade
• New model support
  • GLM-5, GLM-5.1
  • Qwen3.5-Omni
  • Qwen3.6 dense, Qwen3.6-MoE (30B and larger)

Algorithms

• OPD algorithm exploration — explore more On-Policy Distillation variants (multi-teacher, asynchronous teacher serving, weighted distillation objectives) beyond baseline validation

Ecosystem

• Ascend (Huawei NPU) [Feature] Native Support for Ascend #4
  • Qwen3.5 Dense/MoE DAPO-math，Open-R1
• AMD (ROCm / MI-series GPUs) [Feature] Native Support on AMD GPU #12

Misc

...

[sjmshsh]

Hi, I'd like to contribute to the community. Could you let me know what else I can do or what areas I should focus on?

[sjmshsh]

#16

I raised an issue regarding Agentic Rollout — training real agent apps with strict app–RL-framework separation and lightweight API integration.
@Aurelius84 @Yangruipis @yxyOo @li126com @NINGBENZHE