The PyTorch Foundation is expanding what PyTorch Conference North America offers to newcomers. Alongside its usual program, the conference will run a dedicated Introduction Track and an official, full-day PyTorch Associate Training session on Monday, October 19, 2026, from 9am to 5pm in San Jose, California.
The foundation frames the addition around a shift it sees in its own user base: deep learning models are moving quickly from research prototypes into core enterprise infrastructure, and the practical skills needed to train, optimize, and deploy them are not the same as a high-level familiarity with the library.
The Introduction Track is open to everyone attending the conference, from students and first-time attendees to experienced AI engineers who want to pick up an unfamiliar corner of the PyTorch library, ecosystem, or AI stack. Sessions cover foundational concepts, common workflows, project orientation, and approachable starting points for building with PyTorch.
What the Introduction Track covers
The track lists six sessions, each aimed at a different entry point into the stack.
- From Math Panic to PyTorch Confidence — Yashasvi Misra, 3:25 PM PDT, Room LL20AB. Visual explanations and interactive exercises for Python developers, students, and educators new to machine learning, covering tensors, gradients, and dynamic computation graphs before attendees implement a foundational neural network and training loop.
- Practical GPU Programming with Triton for PyTorch Developers — Suman Debnath and JanakiRam Goteti, 11:45 AM PDT, Room LL20CD. An introduction to Triton, the open-source Python-embedded programming model for writing high-performance GPU code without CUDA or C++, including GPU memory movement fundamentals, thread parallelization, and custom vector addition and matrix multiplication kernels.
- What You Cannot Profile, You Cannot Optimize: Learning to Read PyTorch Traces — Aritra Roy Gosthipaty and Suvaditya Mukherjee, 2:50 PM PDT, Room LL20AB. Frameworks for understanding dispatch from CPU to GPU, distinguishing compute-bound from overhead-bound tasks, and analyzing kernel timing, with workload comparisons using torch.compile, fused Triton kernels, and Liger kernels.
- From Scratch to PyTorch: Demystifying ML Frameworks by Building Your Own — Andrea Mattia Garavagno and Vijay Janapa Reddi, 2:15 PM PDT, Room LL20AB. A walkthrough of TinyTorch, an open-source CLI project that guides developers through rebuilding tensors, autograd, optimizers, and transformers in pure Python.
- Understanding Modern Vision Language Models — Aastha Jhunjhunwala and Mark Moyou, 12:20 PM PDT, Room LL20AB. A deconstruction of five open source architectures covering image-to-token encoding, vision-language fusion strategies, multi-GPU fine-tuning, and deployment bottlenecks like KV cache pressure and image token expansion.
- A Developer's Guide to Attention in vLLM — Lucas Wilkinson and Matthew Bonanni, 11:45 AM PDT, Room LL20AB. How the vLLM engine represents hybrid, sliding-window, sparse, and linear attention patterns through attention backends, KV-cache connectors, and hybrid memory allocators.
The Triton and vLLM attention sessions share an 11:45 AM PDT start time in different rooms; the announcement does not say which day each session runs.
The Associate Training day
The training is a separate, in-person, instructor-led program running the full day on October 19, 2026. Its curriculum is organized as a sequence of modules mapped to the key domains of the PyTorch Certified Associate exam, and the foundation describes it as bridging core concepts and production-ready implementation.
Participants work through interactive lectures and live demos of model construction and training, guided Jupyter Notebook labs, comprehension checkpoints in the form of quizzes and reviews, and practical industry projects modeled on production scenarios and deployment challenges. The format is deliberately small: pre-registration is required, and the foundation says space is limited to maintain an optimal student-to-instructor ratio. Attendees can add the training to their conference registration through Tuesday, October 13.
Prerequisites are modest but specific. Attendees should be proficient in Python, familiar with Jupyter Notebooks and basic Google Colab workflows, have a basic understanding of machine learning concepts, and hold an active Google Account for the practical exercises.
Faradawn Yang, an engineer on the AI platform software team at NVIDIA focused on AI inference products, will lead the session. Yang earned both a master's degree in computer science and a bachelor's degree in mathematics and computer science from the University of Chicago, and previously worked as a data engineer at a marketing measurement company.
Everyone who completes the training receives a voucher valued at $250 for the official PyTorch Certified Associate (PTCA) certification exam. The foundation positions the exam as an industry-recognized credential validating practical deep learning workflow skills in PyTorch.
Why it matters
Conference training tracks are usually a side attraction. Here the foundation is treating onboarding as a product problem, pairing an introduction track open to every conference attendee with a separate pre-registration training day and a certification voucher. For teams hiring into PyTorch roles, the PTCA exam gives them a named credential to screen against — a small but real shift in how the ecosystem signals competence.
The session list also reads as a map of where PyTorch practitioners currently get stuck: kernel authoring without CUDA, reading profiler traces, serving vision-language models, and managing KV-cache memory across long contexts. Those are the operational problems that show up after a model works, not before.
Our take
The interesting detail is the prerequisite list. Requiring Jupyter, Colab, and a Google Account for the labs tells you the training is aimed at people who have never shipped a model, not at teams already running inference in production. That is a sensible funnel for a foundation that needs its user base to keep growing, but it also means the Introduction Track sessions on profiling and vLLM attention are doing double duty — recruiting newcomers while giving existing engineers a reason to show up.
