# bytedance/flux

A fast communication-overlapping library for tensor/expert parallelism on GPUs.

Repository: https://github.com/bytedance/flux
Canonical: https://ross.abutalabs.com/products/bytedance-flux
Language: C++
License: Apache-2.0
License Family: permissive
Topics: cutlass, pytorch, cuda, gpu
Last push: 2025-08-28T06:26:38+00:00

## Health v2 (maintenance only)
Score: 33/100 (v2, computed 2026-09-03T02:39:23.370411+00:00)
- activity 39, release rhythm 8, longevity 65
- inputs: {"age_days": 915, "days_push": 370, "days_rel": 673, "gap_med": null, "n_releases_24m": 1}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1354, forks 115 (observed 2026-08-28T04:04:29.043188+00:00)

## What it is
Flux is a GPU kernel library from ByteDance that overlaps computation with communication for tensor and expert parallelism in dense and MoE model training and inference. It provides high-performance, pluggable CUDA kernels compatible with PyTorch across multiple Nvidia GPU architectures.

## Use cases
- overlap allreduce communication with gemm computation during tensor parallel training
- speed up MoE expert parallelism by hiding dispatch/combine communication
- integrate fast communication-overlapping kernels into an existing PyTorch training framework
- reduce GPU idle time in multi-node LLM training
- train dense or mixture-of-experts models more efficiently on Hopper or Ampere GPUs

## When to choose
- you are training or serving large dense or MoE models with tensor/expert parallelism on Nvidia GPUs
- you need pluggable PyTorch-compatible kernels to overlap communication with computation
- you want to squeeze more performance out of multi-GPU or multi-node training

## When to avoid
- you are doing single-GPU workloads with no distributed communication to overlap
- you use non-Nvidia GPUs or unsupported CUDA architectures
- you need a full training framework rather than a kernel library

## Facets
- artifact type: library
- maturity: active
- function: gpu-computing, machine-learning, llm-training, concurrency
- domain: deep-learning, large-language-models, gpu-computing, microservices, machine-learning
- platform: python, cpp
- tags: communication-overlapping, tensor-parallelism, expert-parallelism, moe, nvshmem, cutlass, pytorch-kernels, gpu, cuda, linux

## Member repositories
- bytedance/flux (main) score 33

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:29.043188+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:42:01.176215+00:00, confidence not recorded.
  - readme: https://github.com/bytedance/flux (fetched 2026-08-28T04:04:29.043188+00:00, sha 44bada969f69)
- Data as of 2026-08-30T08:39:29.467469+00:00.