Runhouse

Many ML teams are becoming #Ray-curious; we've had many conversations about how to navigate Ray adoption. The richness of Ray's libraries, the simplicity of distributed programming with Ray, and portability of Ray programs are all extremely attractive features. But for teams just starting to experiment with Ray, there's often significant infra challenges. From the Platforms perspective, standardizing the path for researchers to launch Ray clusters is a headache unless your team is ready to universally adopt Ray. There's also other subtle sharp edges -- everything from making logging flow correctly to avoiding code serialization issues -- that make it incrementally more challenging to "gently" or "lightly" adopt Ray. In this blog post, we praise Ray for its strengths, but raise some common misconceptions or challenges that teams have. Finally, we propose that using Ray shouldn't be more complex than any other library that you can reach for. With Runhouse, you can write regular Ray programs without worrying about the infrastructure. You simply request compute from Runhouse (running over your Kubernetes clusters or elastic cloud compute), and call`.distribute('ray')` -- we wire up the Ray cluster for you and make it available for "serverless" execution of your Ray code. You can take your existing code and dispatch it to Ray clusters in an afternoon (rather than after a quarters-long Platforms team effort) https://lnkd.in/eYEtukaJ #ML #AI #Training #RayData #RayTrain

Many ML teams are becoming #Ray-curious; we've had many conversations about how to navigate Ray adoption. The richness of Ray's libraries, the simplicity of distributed programming with Ray, and portability of Ray programs are all extremely attractive features. But for teams just starting to experiment with Ray, there's often significant infra challenges. From the Platforms perspective, standardizing the path for researchers to launch Ray clusters is a headache unless your team is ready to universally adopt Ray. There's also other subtle sharp edges -- everything from making logging flow correctly to avoiding code serialization issues -- that make it incrementally more challenging to "gently" or "lightly" adopt Ray. In this blog post, we praise Ray for its strengths, but raise some common misconceptions or challenges that teams have. Finally, we propose that using Ray shouldn't be more complex than any other library that you can reach for. With Runhouse, you can write regular Ray programs without worrying about the infrastructure. You simply request compute from Runhouse (running over your Kubernetes clusters or elastic cloud compute), and call`.distribute('ray')` -- we wire up the Ray cluster for you and make it available for "serverless" execution of your Ray code. You can take your existing code and dispatch it to Ray clusters in an afternoon (rather than after a quarters-long Platforms team effort) https://lnkd.in/eYEtukaJ #ML #AI #Training #RayData #RayTrain

Donny Greenberg went on the AI Engineering podcast with Tobias Macey to discuss the state of #ML infrastructure. A lot of great conversation, but give it a listen if you're interested in: -Standardizing the path to launch ML experiments across your org -#GenAI vs. enterprise ML -Avoiding ML platform opinionation, which creates barriers to adopt new frameworks like #Ray or #PyTorch distributed as you scale up. https://lnkd.in/gnCC8PvJ

Hear a replay of our presentation at the Anyscale Ray Summit and hear how Runhouse enables easy adoption of #Ray without disrupting your existing infrastructure while giving you fault-tolerant, flexible, and cost-efficient execution. https://lnkd.in/exc_WR4i #ML #MLOps

While LLMs dominate the news cycle, "boring" ML like recommender systems or fraud detection have quietly continued to be the main driver of realized value from applied AI in the enterprise. Pre-trained language models cannot conquer every use case, and deploying (useful) LLM pilots should not distract from enabling an ML platform that enables the creation of first-party models from your first-party data. https://lnkd.in/eezae4b7 #LLMs #genAI #datascience #ml #mlops

While LLMs dominate the news cycle, "boring" ML like recommender systems or fraud detection have quietly continued to be the main driver of realized value from applied AI in the enterprise. Pre-trained language models cannot conquer every use case, and deploying (useful) LLM pilots should not distract from enabling an ML platform that enables the creation of first-party models from your first-party data. https://lnkd.in/eezae4b7 #LLMs #genAI #datascience #ml #mlops

What is an ML platform? We have found no clear reference architecture and in most organizations, the ML platform means the hard work of multiple smart people rather than a true platform enabling software-like development lifecycles. This ambiguity is exacerbated by the fact that the definition of an ML platform has changed every 2-3 years, while vendors / cloud providers / generative AI solutions continuously add noise. In this article, we trace the history of ML platforms to understand their evolution over the past decade. Then, we attempt to define the requirements of a modern platform that lets engineers create, test, and deploy ML like software. https://lnkd.in/eDJ2ngxT #mlplatform #mlops #ml #ai #platformengineering #mlengineering

Excited to be giving a talk among great company at the #RaySummit. Let us know if you're going to be around!

We are excited to announce that Donny Greenberg, the co-founder and CEO of Runhouse ??, will speak at #OSACON 2024! Explore a novel approach to ML orchestration that allows seamless execution across diverse environments, overcoming the challenges of constrained resources, cloud commitments, and data residency. ?? Register for the conference: osacon.io

Last night, I teamed up the Runhouse team (cc Donny Greenberg) to co-host an ML Engineering mixer in the city ???? Stunning views and fantastic turnout from ML practitioners at The Home Depot Glean Datadog ASAPP & MORE. We dove deep into all things ML Infra ????? Building AI on your infra? Give Runhouse a spin - they’re game changers!