The PyTorch Foundation has decided to take the next North America conference beyond research presentations. On October 19, 2026, in San Jose, it will run a full day of official training for the PyTorch Associate certification, while the main conference days will feature a new introductory track aimed at developers, students, and engineers at different experience levels. The news is not just about the calendar: the chosen topics show how the industry is shifting its center of gravity from model theory to inference engineering.

Two talks on GPU programming and trace analysis point in a clear direction. 'Practical GPU Programming with Triton' teaches how to write custom kernels without CUDA or C++, while 'What You Cannot Profile, You Cannot Optimize' shows how to distinguish compute-bound from overhead-bound workloads using torch.compile, fused Triton kernels, and Liger kernels. This is not generic outreach: it is the kind of skill needed to understand why a GPU is not reaching expected throughput.

At the same time, 'From Scratch to PyTorch' proposes rebuilding core components such as tensors, autograd, optimizers, and transformers in pure Python. The TinyTorch project becomes a path to reading framework internals, a capability that proves decisive when a production error emerges and there is no managed layer to rely on. For those evaluating self-hosted deployments, this attention to the subsystem has direct value: cache bottlenecks and VRAM allocation choices are not solved by changing endpoints, but by working on the engine itself.

The vLLM session is the clearest symptom. Recent models use hybrid, sliding-window, sparse, or linear attention to contain KV-cache occupancy over long context windows. The vLLM engine exposes attention backends, KV-cache connectors, and hybrid memory allocators. For teams running self-hosted deployments, these are not distant abstractions: they are direct levers to reduce required VRAM and increase throughput without changing hardware. The vision-language models session adds another piece: image-token expansion can silently multiply cache pressure, a problem that only emerges when moving from prototype to real service.

The full-day certified training, led by Faradawn Yang from NVIDIA, takes a lab-based approach: interactive lectures, guided Jupyter notebooks, comprehension checkpoints, and projects on realistic use cases. Participants need Python proficiency, familiarity with Jupyter and Google Colab, and a basic understanding of machine learning. Those who complete the course receive a $250 voucher for the PyTorch Certified Associate exam, a credential designed to validate practical skills in deep learning workflows with PyTorch.

Pre-registration is required and must be added to the conference registration by Tuesday, October 13. Space is limited to maintain an optimal student-to-instructor ratio.

Overall, the PyTorch Foundation's move signals a maturation: it is no longer enough to know how to train a model in a notebook. Those building inference services, especially outside managed clouds, need to control the chain from kernel writing to cache management. The fact that a major conference dedicates an introductory track to these topics is not educational charity: it is a response to concrete demand for system-level skills. For those evaluating on-premise deployments, AI-RADAR offers analytical frameworks at /llm-onpremise to weigh these trade-offs.