# ByteDance-Seed/VeOmni

VeOmni: Scaling Any Modality Model Training with Model-Centric Distributed Recipe Zoo

Repository: https://github.com/ByteDance-Seed/VeOmni
Canonical: https://ross.abutalabs.com/products/veomni
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-08-26T12:11:31+00:00

## Health v2 (maintenance only)
Score: 82/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 86, longevity 37
- inputs: {"age_days": 523, "days_push": 7, "days_rel": 99, "gap_med": 5, "n_releases_24m": 2}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2173, forks 259 (observed 2026-08-28T04:06:21.980970+00:00)

## What it is
VeOmni is a PyTorch-native framework for single- and multi-modal model pre-training and post-training, with a modular, trainer-free design exposing full training logic. It scales any-modality models (including omni and MoE models) across accelerators using distributed recipes like FSDP2, sequence parallelism, and experts parallelism.

## Use cases
- pretrain a large language model on multiple GPUs
- train a vision-language or omni-modal model at scale
- fine-tune a Qwen3-MoE model with experts parallelism
- run post-training or RL training for LLMs
- scale training across different accelerator types with PyTorch native tools

## When to choose
- you need transparent, trainer-free training scripts with full control over training logic
- you are training large MoE or multimodal models requiring FSDP2, sequence parallelism, or experts parallelism
- you want a PyTorch-native framework without heavy trainer abstractions

## When to avoid
- you prefer high-level trainer abstractions like HuggingFace Trainer or PyTorch Lightning
- you only need simple single-GPU fine-tuning of small models
- you need a framework with broad non-PyTorch backend support

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, machine-learning, deep-learning
- domain: large-language-models, machine-learning, deep-learning, gpu-computing
- platform: python
- tags: distributed-training, multimodal, fsdp2, sequence-parallelism, mixture-of-experts, pytorch, post-training, reinforcement-learning, gpu, linux

## Member repositories
- ByteDance-Seed/VeOmni (main) score 82

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:21.980970+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 2026-08-30T02:49:01.796763+00:00, confidence not recorded.
  - readme: https://github.com/ByteDance-Seed/VeOmni (fetched 2026-08-28T04:06:21.980970+00:00, sha ee04896f9b42)
  - registry_pypi: https://pypi.org/pypi/veomni/json (fetched 2026-08-29T10:29:09.588709+00:00, sha 791f454743f9)
- Data as of 2026-08-30T08:39:29.467469+00:00.