wal.sh
home > events > pytorch-conf > 2026 · Oct 20, 2026
pytorchconferencesan-joselinux-foundationagentic-airlkernel-agentsexecutorchvllmroboticsagntconmcpcon

PyTorch Conference North America 2026

Table of Contents

PyTorch Foundation's flagship, San Jose, October 20-21 2026. What's new for 2026: agents are the through-line across every track, not a single sidebar. Post-training (RL harnesses, on-policy vs off-policy trade-offs), serving (long-lived agent sessions break inference-stack assumptions), kernel search (KernelAgent, Primus Tuning, Helion), robotics (Cosmos + GR00T + ExecuTorch), and PyTorch's own project-level use of agents for CI, PR review, and release triage all land in the same two days.

Co-located with AGNTCON + MCPCon North America 2026 (Oct 21–23) — the day-two AI Community Bash unites all three communities (De La Soul live set).

1. Facts

  • Dates: October 20-21, 2026
  • Location: San Jose, California (venue: San Jose Convention Center; Grand Ballroom for keynotes)
  • Organizer: PyTorch Foundation (Linux Foundation)
  • URL: pytorch.org agentic-AI session preview
  • Registration: open now (see PyTorch Foundation conference page)
  • Co-located: AGNTCON + MCPCon NA 2026 (Oct 21–23, same city)

2. Keynotes

  • Sara Hooker (Adaption) — "Beyond Brute Force: The Era of Adaptive Intelligence". Continual, gradient-free learning; updating deployed models without full retraining or catastrophic forgetting
  • Mazin Gilbert (Agentic AI Foundation) — keynote on the LF-hosted neutral governance home for agentic infrastructure
  • Bill Jia (Google Cloud) — "Workload Fungibility in the Age of Agents". Long-horizon agentic workflows migrating models GPU→TPU with minimal human intervention; TorchTPU going open source

3. Why to attend

Five thematic clusters worth the trip:

  • RL post-training moved to production — Meta's Torchtitan RL (single model definition for training + generation, bitwise-reproducible on-policy mode), Hugging Face's OpenEnv (co-owned interoperability layer for RL environments — Meta, Unsloth, Prime Intellect, Modal, NVIDIA, Mercor), and a sharp lightning talk from ByteDance + UVa on training-inference numerical mismatch quietly reshaping PPO/GRPO objectives.
  • Kernel agents ship real numbers — KernelAgent (Meta): 1.56x avg speedup over torch.compile, 89% hardware roofline on H100. Primus Tuning (AMD): +27% Mixtral 8x22B throughput in under 30min without full-cluster profiling. Helion DSL LFBO+LLM-guided autotune: 10x faster tuning. Intel: vLLM kernel bring-up weeks → overnight.
  • Agents that build PyTorch — Meta's "Fighting Agents with Agents" (Driss Guessous): Claude wired into PyTorch CI, PR review, autorevert investigation, triage — with adoption-curve data. Edward Yang's BoF on how agents should be contributing to PyTorch.
  • Physical AI ships this year — NVIDIA Cosmos + GR00T → PyTorch → ExecuTorch runtime portable across NVIDIA/Intel/ARM/microcontrollers. Rerun on multimodal dataloaders avoiding GPU starvation. Arm's agentic runtime synthesis (bespoke per-target runtime instead of one hand-tuned general runtime).
  • Governance actually engineered — Apple's Prakshal Doshi + Aditi Mewada on 3am agent-driven prod outages (what agents fix vs flag vs never touch), Amazon's Purva Chiniya on SLM guardrails for prompt-injection detection at multi-agent throughput, Capital One on four-layer AI-output validation.

4. Selected sessions

4.1. Training

  • Meta: Agentic RL Training in PyTorch (Yichuan Wang, Shuhua Yu, Oct 20)
  • Hugging Face: Open Source RL with Agent Harnesses (Ben Burtenshaw, Oct 20 – OpenEnv demo forkable on one GPU)
  • Hugging Face: Train the Agent, Not Just the Model (Sergio Paniego Blanco, Oct 21 – SFT + GRPO in agentic environments)
  • Meta: Torchtitan RL (Felipe Mello, Jiani Wang, Oct 21)
  • RadixArk: Miles: Enterprise-facing Agentic RL Framework (Mao Cheng, Oct 21 – MoE RL + speculative rollouts)
  • ByteDance + UVa: When Rollout and Training Disagree (Neiwen Ling, Tianle Zhong, Oct 21 – lightning talk on numerical mismatch)

4.2. Serving

  • Red Hat: Making Enterprise Agentic Inference Production-Ready with PyTorch and vLLM (Joseph Groenenboom, Tyler Smith, Oct 20)
  • IBM: Native Tiered KV Cache Offloading in vLLM (Or Ozeri, Oct 20 – CPU memory as universal transport hub)
  • Tensormesh: LMCache (Kuntai Du, Oct 20 – cluster-wide open source prompt caching)

4.3. Kernel + performance

  • Meta: KernelAgent: Hardware-Guided GPU Kernel Optimization (Kaiming Cheng, Laura Wang, Oct 20)
  • Meta: Helion DSL LLM-guided autotune (Jongsok Choi, Ethan Che, Oct 20)
  • Intel: From Weeks to Overnight: Autonomous Day-0 Kernel Bring-Up (Xiaogang Gu, Qun Yang, Oct 20)
  • AMD: Primus Tuning: Hybrid Projection and Agentic Search for Distributed Training (Anshu Raina, Peyman Razaghi, Oct 21)
  • Meta: TorchInsights: Zero-GPU Memory & Runtime Estimation (published as TorchSim at ICML 2025)

4.4. Physical AI

  • NVIDIA: From Pixels to Physical Motion: World Action Models with PyTorch & NVIDIA Cosmos (Oct 21)
  • Meta: Deploying Robot Policies Across Many Targets with ExecuTorch (Jacob Szwejbka, Oct 21 – LeRobot/OpenVLA → NVIDIA/Intel/ARM/MCU without per-target rebuilds)
  • Meta: Building Portable, Composable Local Agents with ExecuTorch (Mergen Nachin, Digant Desai, Oct 20)
  • Arm: From PyTorch to the Edge: Agentic Synthesis of Inference Runtimes (Thomas Cottenier, Oct 21 – validated on Apple silicon then pointed at zero-baseline hardware)
  • Boson AI: Scaling Audio AI Infrastructure (Mu Li, Huapeng Zhou, Lindsey Allen, Oct 21 – extending SGLang for audio)

4.5. Governance

  • Apple: Who Owns Production When the Agent Does the Fixing? (Prakshal Doshi, Aditi Mewada, Oct 20)
  • Amazon: Compute, Latency, and Safety: Stateful Guardrails for Multi-Agent Workflows (Purva Chiniya, Oct 20 – SLMs as deterministic guardrails)
  • AWS: Guardrails at the Infrastructure Layer BoF (Sai Charan Teja Gopaluni, Aaresh Sharma, Oct 20 – container sandboxing + GPU resource boundaries + token-level observability)
  • Capital One: Closing the Confidence Gap: Four-Layer Validation Architecture (Ravi Teja Prabhala Venkata, Oct 20)

4.6. Agents that build PyTorch

  • Meta: Contributing to PyTorch with AI agents BoF (Edward Yang, Oct 20)
  • Meta: Fighting Agents with Agents: Bringing Claude to PyTorch CI, triage, and PR review (Driss Guessous, Oct 20)
  • Meta: Shipping PyTorch and Its Ecosystem: A Modern Release Story (Andrey Talman, Oct 20)

5. Related

6. Placeholder note

Scaffolded 2026-09-09 from the PyTorch Foundation blog post (source). Fill in with attendance notes if attended; leave as a directory pointer to the primary source otherwise.