MagiCompiler v1.1.0 Release RoadMap (2026 Q2)

[jiahy0825]

Hey everyone! 👋

Here is our roadmap for the upcoming MagiCompiler v1.1.0 release, targeted for June 30, 2026.

In this cycle, we are taking a step back to focus on practical performance gains (better fusions, VRAM optimization) and making the project more friendly for the community (better docs, training/inference examples).

If you want to help out, just drop a comment below and claim an issue! Items marked with Help Wanted are great places to jump in.

⚡ Direction 1: Performance Acceleration

• Robust Artifact Caching: Better load/save mechanism to reuse cache. ([Bugfix] Restore Cache Reuse Mechanism #20 @themistbeforedawn [Fix] Enable AOTAutograd caching for graphs containing autograd_function_apply #16 @wtr0504 )
• Frontend Fixes: Resolve capture speed issues. ([Perf] Reduce magi_compile dispatch overhead to match torch.compile parity #21 @cennn )
• AutoRecompute for Training
  • Code refactoring. ([refactor]: simplify heuristic save-node selection & add Transformer test #11 @wtr0504 )
  • Maximize throughput under given VRAM limits.
• VAE Inference Optimization: Conv layout rearrangement for VAE.
• Matmul Epilogue Fusion
  • Chain fusion: Matmul + element-wise ops. ([Feat] Add CUTLASS matmul-epilogue fusion path #30 @wtr0504 )
  • Fusion: Matmul + swiglu_fwd / swiglu_bwd.

🤝 Direction 2: Community & Ecosystem

• Enhanced Custom Op API (magi_register_custom_op) ([feat] Smart magi-register-op #28 @themistbeforedawn)
  • Support custom dataclasses. ([Feat] Support dataclass in magi_register_custom_op #32 @themistbeforedawn)
  • Auto-infer output meta.
• Training Examples & Benchmarks
  • Llama-3 training recipes. ([Dev] Add Llama3 training example and fix cache save #14 @wtr0504 )
  • Align training loss with standard baselines. (Help Wanted 🆘)
• Inference Framework Integrations
  • Integrate with vllm-omni.
  • Integrate with sglang-diffusion.
• Docs & Official Web
  • Core mechanism blogs (MagiDepyf, Offloading).
  • Better environment setup guides.
• CI/CD
  • Model-level CI & benchmark (both training and inference)

🔮 Future Work (v1.2.0+)

• Deterministic Fusion (Guarantee reproductivity).
• Custom Backend (MegaKernel / Mirage).
• DistCompile (FSDP compilation).

---

Follow the overall progress on our v1.1.0 Milestone.