FlyDSL Development Roadmap (v0.4)

Modeled on the SGLang AMD roadmap (sgl-project/sglang#23494). Milestone: v0.4 · due 2026-08-05 · ○ planned.

Contributions and feedback are welcome.

FlyDSL (Flexible LaYout DSL) is the Python front-end of the project. Legend: ✓ done · ▶ in progress · ○ planned.

Focus

• Auto-tuning: Automatic kernel tuning.
• Mega MoE final: Production Mega MoE for Expert Parallel.
• Integration & performance: TVM-FFI, vectorized loads, and torch integration.

Feature and Performance Improvements (v0.4 planned)

• Auto-tuning
  PoC: @coderfeli
  Goal: Automatic kernel tuning.

  • ○ Auto-tune support [Feature]: AUTO TUNE support #625

• Mega MoE final (Expert Parallel)
  PoC: @coderfeli
  Goal: Production Mega MoE for EP.

  • ○ Mega MoE final version for Expert Parallel [Feature]: Mega MOE final version for Expert Parallel #623

• Integration & performance
  Goal: Ecosystem integration and load performance.

  • ○ TVM-FFI
  • ○ Auto vectorized load
  • ○ Evaluate and torch integration

Previous: v0.3 (in progress). All milestones: https://github.com/ROCm/FlyDSL/milestones