# bytedance/lightseq

LightSeq: A High Performance Library for Sequence Processing and Generation

Repository: https://github.com/bytedance/lightseq
Canonical: https://ross.abutalabs.com/products/lightseq
Language: C++
License: NOASSERTION
License Family: other
Topics: inference, transformer, beam-search, bert, cuda, sampling, diverse-decoding, multilingual-nmt, training, bart, gpt, accelerate
Archived: true
Last push: 2023-05-16T10:47:48+00:00

## Health v2 (maintenance only)
Score: 10/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 0, release rhythm 8, longevity 100
- inputs: {"age_days": 2462, "days_push": 1205, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: archived, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 3295, forks 334 (observed 2026-08-28T04:07:54.091769+00:00)

## What it is
LightSeq is a high-performance CUDA-based library for training and inference of sequence models like Transformer, BERT, GPT, and BART, with fused custom kernels and fp16/int8 mixed-precision support. It also provides a serving backend built on TensorRT Inference Server for easy deployment.

## Use cases
- accelerate transformer model training on gpu
- run fast transformer inference with beam search
- fine-tune bert with fp16 mixed precision
- deploy gpt or bart models with an inference server
- speed up machine translation training
- int8 quantized inference for transformer models

## When to choose
- you need maximum GPU throughput for Transformer-family training or inference
- you want a deployable serving backend for sequence generation models
- you work with fairseq or Hugging Face models and want drop-in acceleration

## When to avoid
- you need a general-purpose framework for non-Transformer architectures
- you require active maintenance and recent updates, as the last release was in 2023
- you need CPU-only inference

## Facets
- artifact type: library
- maturity: maintenance
- function: llm-inference, llm-training, machine-learning, deep-learning, nlp, gpu-computing
- domain: deep-learning, machine-learning, gpu-computing
- platform: python, cpp
- tags: transformer, cuda-kernels, beam-search, bert, gpt, mixed-precision, inference-server, machine-translation, natural-language-processing, linux, gpu, docker

## Member repositories
- bytedance/lightseq (main) score 10

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:54.091769+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 2026-08-30T07:22:43.168510+00:00, confidence not recorded.
  - readme: https://github.com/bytedance/lightseq (fetched 2026-08-28T04:07:54.091769+00:00, sha 98121c669f66)
- Data as of 2026-08-30T08:39:29.467469+00:00.