flagos-ai/FlagGems

Health v2 · maintenance only

Adoption not part of the score

1089 stars · 509 forks observed · 2026-09-03

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

FlagGems is a high-performance operator library for large language models written in the Triton language, providing backend-neutral GPU kernels. It registers with PyTorch's ATen backend so developers can accelerate training and inference across diverse hardware platforms without changing their PyTorch APIs.

Use cases

• accelerate LLM training and inference on non-NVIDIA accelerators
• replace CUDA kernels with portable Triton kernels in PyTorch models
• run PyTorch workloads across multiple AI chip backends without code changes
• generate pointwise operators automatically for arbitrary input types and layouts
• write and contribute GPU kernels in Triton instead of CUDA

When to choose

• you need hardware-portable PyTorch operator acceleration across diverse AI accelerators
• you want Triton-based kernels with PyTorch API compatibility via ATen registration
• you are building LLM training or inference pipelines that must run on multiple chip vendors

When to avoid

• you need highly specialized hand-tuned CUDA kernels for a single NVIDIA GPU
• your project does not use PyTorch or Triton-compatible hardware
• you need operators not yet covered by the library's kernel collection

Facets

library · maturity active

machine-learning llm-inference llm-training gpu-computing compiler machine-learning deep-learning large-language-models gpu-computing developer-tools python cross-platform triton pytorch kernels operator-library hardware-acceleration aten-backend gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="flagos-ai/FlagGems")