Skip to content

Releases: flagos-ai/KernelGen

KernelGen V2.1.0 release

Choose a tag to compare

@Dongxu-H Dongxu-H released this 03 Jun 08:48
2d662f0

KernelGen v2.1.0 Release

This is FlagOS KernelGen version 2.1.0


Overview

KernelGen is an AI-powered automatic Triton kernel development platform built on the FlagOS ecosystem. It provides a fully automated workflow for operator (kernel) generation, optimization, testing, and deployment across diverse hardware platforms.

With the 2.1 release, KernelGen further expands its hardware ecosystem and introduces experimental support for Triton Language Extensions (TLE) across the full development workflow — including Web, MCP, and Skills integration.

Experience it from: https://kernelgen.flagos.io
MCP Service (ModelScope): https://www.modelscope.cn/mcp/servers/flagos-ai/FlagOS_KernelGen


Core Updates in v2.1

1. New Hardware Backend Support — Sunrise

KernelGen now officially supports the Sunrise AI accelerator platform, expanding hardware coverage from 6 supported chips to 7.

The Sunrise backend is fully integrated into the existing KernelGen ecosystem and supports the complete feature set already available on other hardware platforms, including:

  • Web Platform support
  • MCP automated workflows
  • Skills integration
  • Kernel generation & optimization
  • Auto testing & benchmarking
  • Multi-hardware adaptation workflows

This enables developers targeting Sunrise hardware to use the same AI-native Triton development experience already available across the FlagOS ecosystem.


2. Experimental Triton Language Extensions (TLE) Support

KernelGen v2.1 introduces experimental support for Triton Language Extensions (TLE) across:

  • Web Platform
  • MCP workflows
  • Skills-based development

TLE support is currently available as a laboratory / experimental feature and will continue evolving in future releases.

TLE documentation:
https://github.com/flagos-ai/FlagTree/wiki/TLE


What is TLE?

TLE (Triton Language Extensions) is an extension architecture for Triton designed to address limitations in:

  • Hardware adaptation
  • Memory hierarchy abstraction
  • Tile programming
  • Parallelism abstraction
  • Distributed kernel programming

As modern AI accelerators evolve rapidly, traditional Triton workflows increasingly face challenges in portability, fine-grained optimization, and heterogeneous hardware adaptation.

TLE extends Triton across three abstraction layers:

Layer Description
TLE-Lite Lightweight Triton-compatible extensions with minimal code changes
TLE-Struct Architecture-clustered abstractions for deeper performance tuning
TLE-Raw Hardware-native programming interfaces for maximum performance

Key goals include:

  • Better portability across diverse AI chips
  • Improved tile-level programming abstractions
  • Flexible memory hierarchy optimization
  • Enhanced performance tuning capabilities
  • Future distributed programming support

KernelGen v2.1 now enables developers to experiment with TLE-powered kernel generation and optimization directly through AI-native workflows.


Core Features

Fully Automated Workflow

End-to-end kernel lifecycle automation with MCP + AI agents.

Multi-Backend Support

Broad compatibility across AI frameworks and hardware platforms.

AI-Native Development Experience

Deep integration with IDEs, agents, and developer workflows.

Standardized Verification

Automatic correctness and performance validation.

Deep Ecosystem Integration

Seamless collaboration with:

  • FlagGems
  • FlagTree
  • FlagOS infrastructure

Hardware Platform Coverage

KernelGen v2.1 now supports 7 hardware platforms, including newly added Sunrise support.

Supported capabilities across platforms include:

  • Triton kernel generation
  • Auto optimization
  • Auto tuning
  • Testing & benchmarking
  • Web workflows
  • MCP integration
  • Skills integration
  • TLE experimental workflows

AI Skills Integration

KernelGen continues providing unified AI skill integration through:

  • Claude Code
  • VS Code (with Copilot)
  • OpenClaw
  • Other MCP-compatible agents

Skills repository:
https://github.com/flagos-ai/skills

Available workflows include:

  • Kernel generation
  • Code adaptation
  • Testing & benchmarking
  • Auto PR integration
  • TLE experimental workflows

Web Platform Enhancements

The KernelGen web platform continues providing:

  • Kernel history tracking
  • Enhanced workflow visibility
  • Generation & optimization management
  • TLE-enabled experimental workflows
  • Multi-backend kernel management

Access:
https://kernelgen.flagos.io


Quick Start (v2.1)

Step 1: Get Bearer Token

Get your Bearer Token from the MCP Service section on:

https://kernelgen.flagos.io

Step 2: Install Skill

npx skills add flagos-ai/skills --skill kernelgen -a claude-code

Step 3: Run

/kernelgen-flagos relu_plus_exp

Experimental Feature Notice

TLE Experimental Status

Current TLE support in KernelGen is still in an experimental laboratory stage.

Users may encounter:

  • Unstable optimization results
  • Inconsistent performance across hardware
  • Incomplete backend feature coverage
  • Ongoing API and workflow changes

The TLE ecosystem will continue evolving with future KernelGen and FlagTree updates.


Known Issues

  • Some complex kernels may require multiple optimization iterations for optimal performance
  • Auto-tuning may incur longer execution time on certain hardware platforms
  • MCP workflows may depend on external agent/IDE integration stability
  • Experimental TLE workflows may produce unstable or non-deterministic optimization results

Future Roadmap

  • Broader hardware backend support
  • Enhanced TLE capabilities
  • Improved TLE compiler/runtime integration
  • More intelligent AI-driven optimization
  • Expanded distributed programming support
  • Deeper AI agent integration
  • Improved debugging & profiling tools

Contact Information

contact@flagos.io


Powered by FlagOS — Building the unified open-source system software stack for diverse AI chips.

KernelGen V2.0.0 release

Choose a tag to compare

@Dongxu-H Dongxu-H released this 26 Mar 05:00
ddeeedd

KernelGen v2.0.0

This is FlagOS KernelGen version 2.0.0


‍Overview

KernelGen is an AI-powered automatic Triton kernel development platform built on the FlagOS ecosystem. It provides a fully automated workflow for operator (kernel) generation, optimization, testing, and deployment across diverse hardware platforms.

With the 2.0 release, KernelGen evolves into a complete AI-native kernel engineering system, introducing MCP-based automation, IDE-integrated skills, enhanced web capabilities, and advanced Triton language extensions.

Experience it from: https://kernelgen.flagos.io
MCP Service (ModelScope): https://www.modelscope.cn/mcp/servers/flagos-ai/FlagOS_KernelGen


Core Features (Enhanced)

  • Fully Automated Workflow
    End-to-end kernel lifecycle automation with MCP + AI agents
  • Multi-Backend Support
    Broad compatibility across AI frameworks and hardware platforms
  • AI-Native Development Experience
    Deep integration with IDEs, agents, and developer workflows
  • Standardized Verification
    Automatic correctness and performance validation
  • Deep Ecosystem Integration
    Seamless collaboration with:
    • FlagGems
    • FlagTree
    • FlagOS infrastructure

Core Capabilities Comparison

KernelGen 2.0 transforms Triton kernel development from a fixed pipeline into a fully AI-native, agent-driven system — enabling automatic generation, optimization, and integration across hardware and repositories.


Kernel Development & Optimization

Feature Version 1.0 Version 2.0
Workflow Type Fixed step (Linear pipeline) Agentic (Iterative & Adaptive)
Error Handling Manual debugging Automatic error fixing (log-driven)
Optimization Basic performance test Auto-tuning + AI-driven optimization
Testing Basic correctness & performance tests Fully automated test generation (correctness + benchmark)
Kernel Lifecycle Management Partial Full lifecycle (generate → optimize → test → integrate)

Hardware & Performance Capabilities

Feature Version 1.0 Version 2.0
Multi-Hardware Adaptation Supported Intelligent auto-adaptation & specialization

Developer Experience

Feature Version 1.0 Version 2.0
Interface Web Browser only Web + IDE + CLI (MCP)
Development Entry Web UI only Natural language + CLI + AI agents
IDE / Agent Integration Not supported Claude Code / VS Code / OpenClaw / MCP agents
User Productivity Assisted development Fully automated development

Integration & Ecosystem

Feature Version 1.0 Version 2.0
Repository Integration Manual download & integration Automatic PR generation via Skills
Web Platform Features Basic UI Operator history tracking + enhanced UX
Ecosystem Integration FlagOS basic integration Deep integration with FlagGems / FlagTree / Skills
Target Users Triton developers Triton developers + AI-native developers

Major Updates in v2.0

MCP-Powered Kernel Automation

KernelGen now introduces the MCP (Model Context Protocol) Server, enabling fully automated kernel development workflows through AI agents.

Key Capabilities

  • Automatic Operator Generation
    Generate Triton kernels directly from natural language descriptions

  • Automatic Error Fixing
    Fix kernel code based on compilation and test logs

  • Automatic Performance Optimization
    Optimize kernels using runtime feedback and performance logs

  • Auto Tuning & Specialization
    Automatically tune and specialize kernels for target hardware

  • Multi-Hardware Adaptation
    Seamless support for diverse hardware platforms (including domestic AI chips)

  • Automated Testing Pipeline
    Auto-generated correctness test cases
    Auto-generated performance benchmarks (vs CUDA)

  • One-Click Workflow
    End-to-end kernel development (generation → optimization → testing) in a single pipeline


AI Skills Integration (IDE & Agent Native)

KernelGen 2.0 introduces a unified AI skill: kernelgen-flagos, enabling deep integration with modern AI coding environments:

  • Claude Code
  • VS Code (with Copilot)
  • OpenClaw
  • Other MCP-compatible agents

Highlights

  • One-command kernel generation
/kernelgen-flagos relu  
  • Automatic repository detection

    • FlagGems → specialized workflow
    • vLLM → specialized workflow
    • Generic Triton repo → adaptive workflow
  • Built-in workflows

    • Kernel generation
    • Code adaptation
    • Testing & benchmarking
    • Auto PR integration
  • Unified Skill Architecture

    • kernelgen-general
    • kernelgen-for-flaggems
    • kernelgen-for-vllm
    • kernelgen-submit-feedback

Skills repository: https://github.com/flagos-ai/skills


Web Platform Enhancements

  • Operator History Tracking
    Users can now view and manage previously generated operators directly from the web interface

  • Improved usability and workflow transparency
    Better visibility into generation, testing, and optimization processes

Access: https://kernelgen.flagos.io/


Quick Start (v2.0)

Step 1: Get Bearer Token

Get your Bearer Token from the MCP Service section on:
https://kernelgen.flagos.io/

Step 2: Install Skill

npx skills add flagos-ai/skills --skill kernelgen -a claude-code

Step 3: Run

/kernelgen-flagos relu_plus_exp


Known Issues

  • Some complex operators may require multiple optimization iterations for optimal performance
  • Auto-tuning may incur longer execution time on certain hardware platforms
  • MCP workflows may depend on external agent/IDE integration stability

Future Roadmap

  • Broader hardware backend support
  • Enhanced MCP intelligence
  • Deeper AI agent integration
  • Improved debugging tools

Contact Information

contact@flagos.io


Powered by FlagOS — Building the unified open-source system software stack for diverse AI chips.

KernelGen V1.0.0 release

Choose a tag to compare

@Dongxu-H Dongxu-H released this 31 Dec 08:05
6d82b9f

This is FlagOS KernelGen version 1.0.0

‍Overview

KernelGen is an AI-powered automatic Triton kernel development tool built on the FlagOS ecosystem. It offers a fully automated workflow for operator (kernel) development, tuning, testing, and deployment, significantly boosting efficiency and speeding up multi-hardware adaptation.

Experience it from: https://kernelgen.flagos.io

‍Initial Release

  • Basic Workflow : Supports single, fixed - step code generation including GroundTruth → TritonKernel → Correctness Test → Performance Test.
  • User Registration : Supports self - service registration and application, with platform approval required for trial use.
  • Web Interface : Online access at https://kernelgen.flagos.io/.
  • Core Features :
    • Fully Automated Workflow : Automatically generates, tests, and optimizes complete AI operator sets.
    • Multi - Backend Support : Seamlessly supports multiple AI libraries and chips with automatic adaptation and debugging.
    • Easy - to - Use : Browser - based interface requiring no setup or prior experience.
    • Standardized Verification : Automatically generates test cases to ensure operator correctness.
    • Deep Ecosystem Integration : Collaborates with FlagGems and FlagTree to accelerate operator library development.

Known Issues

  • Some complex operators may require multiple iterations to achieve optimal performance.
  • Performance test timeouts may occur for certain operators.
  • Limited support for edge computing platforms (work in progress).

Future Roadmap

  • Integration with more AI frameworks and hardware platforms
  • Enhanced operator optimization capabilities

Contact Information

For questions or feedback, please contact the FlagOS team at contact@flagos.io or record issues in this repository.

Powered by FlagOS - Building the unified open - source system software stack for diverse AI chips.