[Roadmap] Relax 2026 Q3

[NINGBENZHE]

Agentic

@yxyOo @Yangruipis

• Multi-Agent Integration & Customization — support multi-agent systems via lightweight code refactoring and customized advantage.
• Agentic Observability — end-to-end tracing across rollout, reward, and training. @xiaoliang0601
• Agentic RL Recipes
  • Mini-swe-agent — software engineering agent training.
  • ALFWorld — interactive embodied agent training.
  • WebShop — web-shopping agent training.
  • OSWorld — GUI and desktop agent training.
  • Hermes Online RL — end-to-end online learning with trajectory collection, asynchronous training, and live weight updates.

Rollout Performance

@yuanlehome

• SGLang Upgrade — upgrade the rollout backend to SGLang 0.5.12.post1.
• Program-Aware Admission and KV-Cache Lifecycle Scheduling for Agentic Rollouts — build on ThunderAgent's program abstraction to enable training-aware admission, pause/resume, and session lifecycle management (RFC #81).
• Prefill-Decode Disaggregation — validate performance benefits for long-context Agentic workloads.

Training Performance

@NINGBENZHE

• Megatron FSDP — integrate FSDP as an additional training backend for small and medium-sized models. @GUOGUOPOT
• Qwen3.5 GDN CP Performance and Memory Optimization @xiaoliang0601
• Prefix-Tree Training — reuse shared-prefix computation across branched Agentic and multimodal trajectories to reduce training cost and activation memory (RFC #67).
  • Qwen3-VL dense + TP/CP/PP. @mununumu
  • Qwen3.5 MoE + TP/CP/PP/EP (Magi Attention lacks support for GDN CP). @liujia-cc
• Algorithm-specific Optimization — improve OPD teacher utilization and end-to-end training throughput.

Architecture

@Yangruipis

• Unified Inference Service — unify the APIs, engine lifecycle, routing, placement, and recovery of Rollout, GenRM, and Teacher services (RFC #71).
• Role-Device Layout — provide a declarative layout planner for colocated, split, deferred, and fully asynchronous deployments (RFC #73).
• TransferQueue Upgrade — upgrade TransferQueue and integrate and optimize the Mooncake transfer backend.

Model

@li126com

• DeepSeek V4 — end-to-end RL training and rollout support.

Algorithms

@Yangruipis @GUOGUOPOT

• SFT Training — support full-parameter and LoRA fine-tuning with a unified SFT-to-RL workflow.
  • Parameter-efficient RL Training — LoRA.
• PPO — end-to-end support including Critic training, value estimation, and GAE.
• Multi-teacher On-Policy Distillation.
• Off-policy Correction — improve stale-sample handling in asynchronous RL training.

Multi-Modal Generation RL Survey & Design

@NINGBENZHE

• Generative Model RL Survey — investigate RL techniques for image, video, speech, and TTS generation.
• Unified Architecture Design — design rollout, reward, training, and weight synchronization abstractions for token- and latent-based generation.
• Reward System Design — investigate reward modeling for visual quality, video quality, speech naturalness, speaker similarity, and instruction following.

Reliability & Ecosystem

@NINGBENZHE

• Elastic Rollout Enhancement and Faster Failure Recovery. @Yangruipis
• New Hardware Support — NVIDIA B300 and enhanced KLX/NPU capabilities.
• GPU github action @li126com

[gongshaotian]

I am very interested in Program Ware and am working on similar projects. I look forward to communicating with you! @yuanlehome