RoadMap Q2~Q3

[nullnonenilNULL]

Roadmap

This document outlines the upcoming development priorities for LoongForge. Across all model integrations, our goal is to deliver more efficient training performance than publicly available community baselines.

Status legend

• - [ ] Planned / Not started
• - [ ] 🚧 In progress
• - [x] Completed

LLM & VLM

Model coverage

• Add training support for Kimi K2 series, including FP8 recipes.
• Add training support for DeepSeek-V4. ([Question] DeepSeek v4 support plan #9)
• Add training support for MiniCPM. ([FEATURE] Support MiniCPM-V-4.6 model #99)

Training recipes

• 🚧 Introduce INT4 QAT (Quantization-Aware Training) support. ([Feature] INT4 Quantization-Aware Training (QAT) for MoE Expert Linear Layers #12)
• Improve Muon optimizer support and stability.

Performance

• 🚧 DeepSeek-V4 end-to-end optimization: fused kernels, parallelism strategies, and sequence parallel.
• 🚧 MoE elastic expert cloning to mitigate EP load imbalance, targeting better balance and lower algorithmic overhead than existing approaches. ([FEATURE] TAOT: MoE Expert Load Balancing via Online Token Redistribution with Topology-Aware Dispatch Planning #28)
• VLM fully-decoupled parallelism with ViT activation offload (zero memory overhead path), compatible with existing parallel strategies.
• 🚧ChunkPipe for LLM long-context training: extend coverage to SFT and additional model architectures.
• Context Parallel: feature completeness and systematic regression validation.

VLA & WAM

Data Engine

• Ego to lerobot

Model coverage

• Add training support for Pi 0.5.
• Add training support for GR00T N1.6.
• Add training support for X-VLA.
• Add training support for DreamZero.
• Add training support for Lingbot-VA.
• 🚧Add training support for Motus.

Training framework

• Unified data processing pipeline.
• Unified training paradigm abstraction.
• 🚧 Standard model evaluation benchmarks.
• Multi-backend integration: DDP, FSDP2, and beyond.

DiT

Model coverage

• Add training support for Wan2.1.(PR#85)
• Add training support for Wan2.2, support data packing, context parallel.
• Add training support for Qwen-Image. ([FEATURE] Hope to support image generation models #59)

Agent Workflow

Explore agent-driven developer experience to lower the bar of large-scale training:

• Quick-start skill: one-shot environment and recipe bootstrap.
• 🚧 Model development skill: assisted onboarding for new architectures.
• Model performance analysis skill: automated profiling and bottleneck reporting.

---

Feedback, suggestions, and contributions are welcome via GitHub Issues and Pull Requests.