# TRI-ML/prismatic-vlms

A flexible and efficient codebase for training visually-conditioned language models (VLMs)

Repository: https://github.com/TRI-ML/prismatic-vlms
Canonical: https://ross.abutalabs.com/products/prismatic-vlms
Language: Python
License: MIT
License Family: permissive
Last push: 2024-07-04T13:08:39+00:00

## Health v2 (maintenance only)
Score: 25/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 0, release rhythm 35, longevity 66
- inputs: {"age_days": 932, "days_push": 790, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1009, forks 1206 (observed 2026-08-28T04:03:12.791738+00:00)

## What it is
Prismatic VLMs is a PyTorch-based codebase for training visually-conditioned language models (VLMs) with flexible vision backbones like CLIP, SigLIP, and DINOv2. It supports efficient scaling from 1B to 34B parameters using FSDP and Flash-Attention.

## Use cases
- train a vision-language model from scratch
- fine-tune a VLM with custom vision backbones
- train multimodal LLMs on image-text data
- experiment with fused visual representations like DINOv2 + SigLIP
- scale VLM training to 34B parameters with FSDP
- instruct-tune a vision-language assistant

## When to choose
- you need a research-grade, configurable codebase for training VLMs
- you want to swap or fuse different vision encoders (CLIP, SigLIP, DINOv2) via timm
- you need efficient large-scale training with FSDP and Flash-Attention
- you want to train on custom multimodal dataset mixtures

## When to avoid
- you only need to run inference with an existing VLM
- you want a plug-and-play hosted API for multimodal models
- you lack multi-GPU resources for large-scale training
- you need a full evaluation suite rather than training (use their vlm-evaluation repo instead)

## Facets
- artifact type: library
- maturity: active
- function: llm-training, machine-learning, deep-learning
- domain: large-language-models, machine-learning, computer-vision, deep-learning
- platform: python
- tags: vision-language-models, multimodal, pytorch, fsdp, flash-attention, model-training, gpu, linux

## Member repositories
- TRI-ML/prismatic-vlms (main) score 25

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:12.791738+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 2026-08-30T07:12:00.961786+00:00, confidence not recorded.
  - readme: https://github.com/TRI-ML/prismatic-vlms (fetched 2026-08-28T04:03:12.791738+00:00, sha b08831ac0ba0)
- Data as of 2026-08-30T08:39:29.467469+00:00.