When our team first started working on machine learning, we hit a wall. CUDA errors, broken Conda environments, and constantly breaking dependencies with PyTorch versions. We spent days just getting code to run, long before we could do anything interesting or innovative. Over time, I've learned this experience is incredibly common. Researchers have ideas but spend huge portions of their time just trying to make their work run.

And then comes the time to scale up. Moving a project to multiple GPUs can feel nearly impossible without an entire infrastructure team to help. This was our experience, and it's the core reason we started building this platform. We wanted to build the tool we wished we'd had from day one.

ML Research is Messy​

The dream of ML research is simple: Hypothesis → Experiment → Results. But the reality is often messier:

• Getting the machine: Find the right hardware at the right price.
• Setting it up: Provision, configure environments, wrangle CUDA versions.
• Storing data: Manage gigabytes of outputs, artifacts, logs, and checkpoints across nodes.
• Running the job: Launch training, monitor logs, ensure proper networking.
• Tracking everything: Random logs, graphs, and spreadsheets to track runs.
• Oops, it died: Debug failures, find checkpoints, restart.

And that's just one run. Then you need to manage experiments, version datasets, compare models, and share findings—without resorting to screenshots on Discord.

A New Category: Machine Learning Research Platform​

Frontier labs like OpenAI have spent years building and maintaining their own tooling and infrastructure. We're aiming to make this category of advanced tooling open and accessible to everyone. Users have already told us that Transformer Lab is the "essential open source stack for serious ML teams".

Our goal is to build a unified, open-source platform that handles infrastructure so researchers can focus on research. Better tools accelerate progress for everyone—when the ML community has access to best-in-class platforms, discovery moves faster. For this to work, it must follow core principles that empower researchers, not constrain them.

Our Guiding Principles:​

☁️ Compute Agnostic: You should be able to run your workloads on any cloud, any cluster, or even on-prem hardware without changing your code.

• 🏃‍♂️ Get Out of the Way: The platform should serve you, not the other way around. You shouldn’t have to learn “our way” of training, the system should be able to run any workload. If you want to use PyTorch, JAX, or your own custom framework, there should be zero barriers.
• ❤️ Open Source: Foundational tools for science should be open. This fosters trust, encourages community contribution, and prevents vendor lock-in.
• 🧩 Composable: You shouldn't have to adopt a monolithic, all-or-nothing system. Just need workload orchestration? Use that part. Just want experiment tracking? Great. Use only what you need.
• 😌 A Great Developer Experience: At the end of the day, it just has to be easy and enjoyable to use. An API that makes sense. A CLI that's intuitive. Documentation that's actually helpful.

How We're Building It: Standing on the Shoulders of Giants​

We're building this platform in layers, leveraging the best open-source tools:

Layer 1: Workload Orchestration​

We use SkyPilot for GPU and workload orchestration, adding team and user isolation for shared control nodes.

Layer 2: Scaffolding (Management & Persistence)​

We added team-based features like user/quota management and a unified SDK for consistent artifact, model, and checkpoint storage.

Layer 3: User Experience (Experiment Management)​

Our Experiment Management layer tracks everything automatically, making models, evaluations, and datasets versioned, shareable, and easy to compare.

From Frustration to a Framework​

We built this platform because we remember losing days to broken environments and CUDA errors. We wanted a tool that lets researchers focus on the exciting "what if" questions, not infrastructure headaches.

Try us out, and give us feedback; or better yet sign your org up to be a design partner. We'd love to hear about your pain points and build indispensable tools for you and your team.

Let's build something better, together.

Today, we’re launching Transformer Lab GPU Orchestration, a flexible open-source platform for AI/ML teams to manage large-scale training across clusters of GPUs.

🎉 Transformer Lab just got a voice! We’re thrilled to announce audio modality support so you can generate, clone, and train voices directly in Transformer Lab.

What’s included in this release​

• 🎙️ Turn text into speech (TTS) with CUDA, AMD and MLX
• 🛠️ Train your own TTS models on CUDA and AMD
• 🧬 Clone a voice in one shot for lightning-fast replication on CUDA and AMD

kk, Transformer Lab now supports Diffusion image generation and training!

Out of the box we support major open weight base models including:

• Stable Diffusion (1.5, XL, 3)
• Flux
• And a lot more! (see the full list here)
• (We also support thousands of LoRAs trained on these base models)

What can you do? Well...

We're excited to announce that Transformer Lab now supports AMD GPUs! Whether you're on Linux or Windows, you can now harness the power of your AMD hardware to run and train models with Transformer Lab.
👉 Read the full installation guide here

TL;DR​

If you have an AMD GPU and want to do ML work, just follow our guide above and skip a lot of stress.

The journey for us to figure out how to build a reliable PyTorch workspace on AMD was... messy. And we've documented everything below.

We're excited to announce a significant enhancement to Transformer Lab - integration with the open-source Markitdown library from Microsoft! This update dramatically expands the types of documents you can work with in Transformer Lab, making it more versatile and powerful for your AI projects.

Retrieval-Augmented Generation (RAG) combines the power of retrieval systems with generative AI to create more accurate, factual, and contextually relevant responses. In this hands-on tutorial, we'll walk through building and evaluating a complete RAG pipeline in Transformer Lab using documentation files as our knowledge base.

Transformer Lab is excited to announce robust multi-GPU support for fine-tuning large language models. This update allows users to leverage all available GPUs in their system, dramatically reducing training times and enabling work with larger models and datasets.

Transformer Lab has recently added an Ollama Server plugin which allows users to run inference through Ollama on their local machine.

In this guide, we'll walk through creating an evaluator plugin within Transformer Lab named sample-data-print. This plugin will load a dataset and print its contents, along with some sample parameters, using the new tlab_trainer decorator approach.