Why Pytorch Won

During this week’s PyTorch Conference 2025 successful San Francisco, I sat down pinch Luca Antiga, nan caput of nan Technical Advisory Council for nan PyTorch Foundation and nan CTO of end-to-end AI level Lightning AI. Antiga was portion of nan squad that wrote nan original PyTorch paper and co-authored nan “Deep Learning pinch PyTorch” book. Who amended to talk to astir nan authorities of PyTorch itself and get an update connected nan PyTorch Foundation?

Antiga, whose world inheritance is really successful biomedical engineering, noted that 1 of nan reasons why PyTorch became truthful popular is that it is very friends to researchers. Those researchers yet graduated aliases went to activity successful nan tech manufacture and brought PyTorch pinch them.

“It’s very ‘pythonic,'” Antiga explained. “Whereas backmost successful nan day, location were galore frameworks that were successful Python, but you needed to codification successful a benignant of meta connection that gets betwixt nan problem successful yourself, and it makes debugging a batch harder, and truthful on. Pytorch was benignant of revolutionary successful that consciousness successful that it brought nan easiness and loop velocity of Python, and, you know, nan bias toward action that Python has into nan world of for illustration neural networks and backmost propagation and GPU computing.”

When PyTorch was created, nan manufacture still focused connected neural networks — often for image aliases sentiment analysis. Then ChatGPT brought generative AI (GenAI) to a wider audience, but Antiga argues that PyTorch ever remained relevant.

“Throughout each these revolutions that came, you ever spot PyTorch there,” he said. “And location are, of course, others for illustration JAX and truthful connected — they’re very strong. But comparatively, PyTorch is simply a full manufacture and empowers nan full industry. And together pinch that came a awesome ecosystem.”

And while PyTorch is astir often associated pinch training models, it has besides go cardinal to moving models successful accumulation now.

“If you spell and spot nan conclusion frameworks that are astir celebrated — vLLM, SGLang — and they tally PyTorch, right? And they tally PyTorch successful production,” Antiga noted. “If you interact pinch immoderate chatbot today, chances are that you’re moving PyTorch astatine that time.”

Another logic PyTorch has been gaining successful fame precocious is that reinforcement learning — which rewards models for nan correct action and provides antagonistic feedback erstwhile they return nan incorrect move — is now routinely being utilized to fine-tune pre-trained ample connection models (LLMs). As it turns out, PyTorch is well-suited for this benignant of usage case, too.

“Reinforcement learning will promote a exemplary to do much of nan things that activity erstwhile confronted pinch an situation wherever you tin spell successful pinch an action, nan situation authorities changes, and you get a reward for your actions,” Antiga explained.

The State of nan PyTorch Foundation

As for nan PyTorch Foundation itself, it’s worthy noting that it was only a fewer months agone that nan instauration started bringing successful further projects, starting pinch vLLM and DeepSpeed. With Ray, nan celebrated unfastened root distributed computing model for scaling AI and Python applications, nan instauration has now added its 4th project.

But nan scheme isn’t to go a ample umbrella foundation, Antiga stressed.

“What I’m mostly concerned astir successful nan ecosystem is nan users. What travel do nan users acquisition erstwhile they attack an ecosystem that is benignant of sanctioned by nan PyTorch Foundation,” he explained. “My extremity is to person them succeed.”

For much of our conversation, subscribe to our podcast aliases watch nan afloat question and reply connected YouTube.

Group Created pinch Sketch.