flagos-ai/FlagTree

FlagTree is a unified compiler supporting multiple AI chip backends for custom Deep Learning operations, which is forked from triton-lang/triton.

What it solves

FlagTree addresses the fragmentation of AI software stacks across different hardware accelerators. It eliminates the need to rewrite AI workloads for every specific chip, reducing the cost of porting and maintaining models while unlocking hardware-specific performance gains.

How it works

FlagTree acts as a unified compiler for multiple AI chips, building upon and extending the Triton ecosystem. It provides a single-repository, multi-backend support system where different hardware backends (such as NVIDIA, AMD, Huawei Ascend, and others) are integrated. To bridge the gap between portable code and hardware-specific tuning, it introduces Triton Language Extensions (TLE), which allow developers to move from portable usage to raw, hardware-oriented tuning through layered abstractions (Lite, Struct, and Raw).

Who it’s for

  • Model users who want to compile and run their AI models across diverse AI accelerators without rewriting code.
  • Chip manufacturers looking for a standardized way to integrate their hardware into the Triton ecosystem.

Highlights

  • Broad Hardware Support: Supports a wide array of backends including NVIDIA, AMD, Huawei Ascend, and various other AI accelerators.
  • Triton Language Extensions (TLE): Provides progressive abstraction levels for deeper performance tuning of memory access and execution patterns.
  • Unified Compilation: Enables a "develop once, run anywhere" workflow for AI workloads.
  • Performance Gains: Achieves speedups in real-world models (e.g., Qwen) without requiring modifications to the original Triton operator code.

Related

  • Project
    flagos-ai/FlagGems

    A high-performance operator library implemented in Triton that enables LLM training and inference to run across diverse AI accelerators using PyTorch APIs.

  • Project
    flagos-ai/FlagScale

    FlagScale is a unified toolkit for the FlagOS ecosystem that provides a single interface for training, reinforcement learning, and inference of large models across diverse AI hardware chips.

  • Project
    ByteDance-Seed/Triton-distributed

    Triton‑distributed is a ByteDance‑Seed open‑source compiler that extends the Triton language with primitives for overlapping GPU compute and communication. It enables high‑performance distributed kernels (GEMM, MoE, flash‑decode, AllToAll) on NVIDIA and AMD GPUs, offering speed‑ups comparable to hand‑tuned libraries. Install via source or a pre‑built pip wheel, write kernels with `@triton_dist.jit`, and use the provided low‑level communication APIs. The project is actively updated (latest news Sep 2026) and released under MIT.

  • Dispatch
    OpenAI Triton 1.0 Release

    OpenAI has released Triton 1.0, an open-source Python-like language and compiler that allows researchers to write highly efficient GPU kernels without deep CUDA expertise.

  • Project
    proliferate-ai/proliferate

    An open-source AI IDE that allows developers to run multiple coding agents in parallel, providing each task with an isolated git worktree and environment.