Luca Antiga

Luca Antiga is an early PyTorch core contributor, coauthor of Deep Learning with PyTorch, and CTO of Lightning AI. His work includes contributions to Thunder, a compiler for optimizing PyTorch models, alongside public arguments for giving developers control over those optimizations. In February 2026, he also chaired the PyTorch Foundation’s Technical Advisory Council; he served as a program chair for PyTorch Conference Europe that April.[4][5][7][8]

Antiga’s case for Thunder centers on the dependence of performance on a particular model, hardware, and workload. In his October 2024 interview with Jon Krohn, he argued that practitioners should be able to inspect and transform a computation before executing it: choose which operations to fuse, change numerical precision, or discard intermediate results and recompute them to save memory. He described Thunder as work in progress and a complement to torch.compile, which he called the natural starting point for trying compilation.[8]

Thunder interprets Python bytecode to produce a straight-line Python program, transforms the model and computation trace, and routes parts of that trace to execution backends. Those backends include specialized libraries, custom Triton or CUDA kernels, PyTorch operations, and torch.compile. This source-to-source approach exposes the computation for modification rather than requiring the compiler to generate every kernel itself. Thunder is a collaborative open-source project with significant NVIDIA contributions.[6]

Antiga’s own contributions include extending thunder.jit coverage on Hugging Face models, improving coverage reporting in continuous integration, and adding dependencies to reduce skipped cases. The Thunder 0.2.5 release record credits these changes to lantiga, the GitHub account linked from his personal website.[7][3]

In October 2024, Antiga described a practical feedback loop with Thomas Viehmann: their Friday Thunder Sessions podcast involved explaining or using the compiler while developing it. Working with the tool in public exposed shortcomings they needed to address—“a great way to put ourselves in front of what we’re creating,” as he put it.[8]

Antiga encountered PyTorch Lightning through use before meeting its creator, William Falcon: Orobix, the Italian AI company Antiga cofounded, adopted it to standardize training code. He explained that models remain in pure PyTorch, while lifecycle hooks organize tasks such as creating optimizers and running training steps. The trainer calls that code at the appropriate time and handles engineering details, allowing practitioners to work without first becoming experts in distributed training.[8]

As Lightning AI CTO, Antiga emphasized minimal tools that remove burdens while leaving users room to make their own choices. He credited Falcon with pushing the team toward polished releases, including substantial work on READMEs and documentation. His caution extended to prospective tools for controlling model behavior: “But releasing something is a big commitment.” Interest in an idea, he argued in 2024, was insufficient reason to ship an open-source product before its design felt ready.[8]

In 2024, Antiga argued that distillation and more efficient use of model capacity left considerable room for smaller language models to improve. He distinguished factual knowledge from reasoning—“reasoning and knowing everything are two different things”—and treated scaling laws as relationships formulated under particular conditions. His favorable experience with Llama-3-8B was explicitly limited to his private test cases. He also anticipated more visible inference-time controls for directing model behavior and improving compound AI systems.[8]

Antiga described using language models as a partner for exploring ideas: “You need to use it as a wall to bounce ideas back and forth.” He expected repetitive coding tasks to become easier to delegate and personal automation to become more accessible through expressing intent. He nevertheless maintained that developers still needed to understand and connect the layers of a system; in his assessment, AI could accelerate architectural work without replacing that understanding.[8]

Before his PyTorch and Lightning work, Antiga trained in bioengineering and developed medical-imaging software. He holds a PhD in Bioengineering and developed VMTK, a toolkit for modeling blood vessels from medical images, as well as contributing algorithms to ITK and 3D Slicer. Recalling this period in 2024, he dated VMTK’s development to 2003–2004 and its open-source release to 2004, while working with David Steinman at Robarts Research Institute in London, Ontario.[3][8]

Antiga recalled that the open-source culture around him at Robarts encouraged him to release his computational-geometry algorithms. Supporting other researchers subsequently showed him the reach of that decision: he remembered entering a Norwegian laboratory and seeing students using his command-line tools. By the 2024 interview, he said he no longer developed VMTK himself.[8]

Antiga recalled founding Orobix in 2009 and later facing a medical-image segmentation challenge that his existing algorithms could not adequately solve. Random forests provided only a partial answer; around 2013–2014, the team tried Torch7 because its 3D convolution kernels suited their volumetric data. He remembered useful results arriving quickly. After a small contribution to Torch7, he began contributing to PyTorch when it appeared in early 2017.[8]

With Eli Stevens and Thomas Viehmann, Antiga coauthored Deep Learning with PyTorch, published in July 2020. The book teaches through a tumor-image classifier, following the work from data loading and training through diagnosing unreliable initial results, augmenting data, and improving the architecture. Manning lists him under the fuller name Luca Pietro Giovanni Antiga.[2][8]

Antiga appeared as Lightning.ai CTO in Latent Space’s NeurIPS 2023 startup recap, published December 30, 2023. His segment begins at 02:12:39, with show notes covering PyTorch Lightning, Lightning Studios, and LitGPT.[1]