-
Notifications
You must be signed in to change notification settings - Fork 40
Project Info
Organization: flagos-ai
Main Repository: flagtree
Official Document:docs
FlagTree uses a multi-branch strategy to support different backends, each based on different Triton versions due to vendor-specific constraints. All main branches are protected and have equal status.
| Branch | Vendors | Backends | Triton Version |
|---|---|---|---|
| main | NVIDIA, AMD, x86_64 CPU, ILUVATAR, Moore Threads, KLX, MetaX, HYGON | nvidia, amd, triton-shared, iluvatar, mthreads, xpu, metax, hcu | 3.0, 3.1 |
| triton_v3.2.x | NVIDIA, AMD, Huawei Ascend, Cambricon | nvidia, amd, ascend, cambricon | 3.2 |
| triton_v3.3.x | NVIDIA, AMD, x86_64 CPU, ARM China, Tsingmicro, Enflame | nvidia, amd, triton-shared, aipu, tsingmicro, enflame | 3.3 |
| triton_v3.4.x | NVIDIA, AMD | nvidia, amd | 3.4 |
| triton_v3.5.x | NVIDIA, AMD | nvidia, amd | 3.5 |
Different backends are organized in different branches because each vendor has specific Triton version requirements and constraints. This multi-branch strategy allows FlagTree to:
- Support backends that depend on different Triton versions
- Maintain compatibility with vendor-specific requirements
- Enable parallel development across different Triton versions
- Provide a unified codebase structure while accommodating version differences
FlagTree supports a wide range of hardware backends:
- NVIDIA - GPU acceleration
- AMD - GPU acceleration
- Huawei Ascend - AI accelerator
- AIPU - AI processing unit
- ILUVATAR (天数智芯) - GPU
- Moore Threads (摩尔线程) - GPU
- Cambricon (寒武纪) - AI accelerator
- Tsingmicro (清微智能) - AI accelerator
- Enflame (燧原) - AI accelerator
- MetaX (沐曦股份) - GPU
- HYGON (海光信息) - GPU
- HCU - GPU
- XPU - AI accelerator
- x86_64 CPU - CPU backend via triton-shared
- ARM China - ARM-based processors
FlagTree uses Git tags for releases. Releases are tagged with version numbers following semantic versioning, and may include backend-specific or Triton version-specific variants (e.g., v0.4.0, v0.4.0_triton_v3.2.x, v0.4.0_triton_v3.5.x). All triton_v3.x.x branches are actively maintained and continuously evolving, including experimental updates.
Release notes are located in the reports/ directory:
- Path:
reports/v[version]/release_notes_v[version].md(English) - Path:
reports/v[version]/release_notes_v[version]_cn.md(Chinese)
Each release includes detailed changelog, new features, bug fixes, and backend-specific updates.
See the main README for installation instructions. Different backends may require different installation commands based on their Triton version requirements.
- FlagTree Tutorial: https://github.com/flagos-ai/docs/tree/main/course/tutorial/FlagTree
- FlagTree Documentation: https://docs.flagos.io/projects/FlagTree/en/latest/index.html
All documentation is maintained in this wiki:
- Home: Home - Main entry point and navigation
- Architecture: Architecture - Code structure and extension mechanisms
- Language Extensions: TLE, HINTS, EDSL
- Project Info: This document
- Triton Documentation: triton-lang.org
- MLIR Documentation: mlir.llvm.org
Contributions are welcome! Please refer to CONTRIBUTING.md for guidelines.
FlagTree is licensed under the MIT license.