PyTorchCon Demo Theater 13 场演示：从 Trainium 强化学习到 MoE 推理
PyTorchCon 一天 13 场硬核演示，Wan 2.2 视频生成提速 53.6 倍、万节点 Ray 后训练都在列，做基建的值得扫一遍日程。
PyTorchCon 周三 Demo Theater 安排 13 场演示，覆盖训练、推理、硬件适配和智能体工作流。Baseten 展示 Wan 2.2 生成五秒视频用时不到三秒，较基线提速 53.6 倍，靠时间步蒸馏、定制内核和 4-bit 量化实现。Verda Cloud 演示在 NVIDIA NVL72 上跑 WideEP 专家并行 MoE 推理，Anyscale 讲 Ray Core 扩展到 10,000 节点集群的后训练方案。其他演示包括华为的 CRCR 跨仓库 CI 测试、Clockwork 的训练断链恢复，以及 Lightning AI 上自动复现 ML 研究的智能体。
Wednesday's Demo Theater at #PyTorchCon features 13 presentations on training, inference, hardware enablement, and agentic workflows. All sessions are in the Community Expo on the Concourse Level.
10:40–10:50 AM: Maen Suleiman (@awscloud) demos reinforcement learning on Trainium through TorchNeuron, the native PyTorch backend for AWS Trainium, with TorchTitan training, vLLM-Neuron generating rollouts, Monarch orchestrating, and TorchStore syncing weights.
10:55–11:05 AM: Ali Taha (@baseten) breaks down how Wan 2.2 generates a five-second video in under three seconds, a 53.6x improvement over baseline, using timestep distillation, custom kernel engineering, and four-bit quantization.
12:20–12:30 PM: Rain Jiang and Kevin Li (@verdacloud) demo WideEP for expert-parallel MoE inference on an NVIDIA NVL72 system, using a Rust-based frontend to run and monitor the workload.
12:35–12:45 PM: Tristan Webb (@dMatrix_AI) presents dmx-compressor, a PyTorch 2.0 based quantization framework that gives ML engineers numerical reference output to validate models running on ASIC hardware.
12:45–12:55 PM: Lin She (@clockworkio) demos surviving link and node failures in PyTorch training live, with LinkPass and TorchPass keeping the job running on the same training step.
1:00–1:10 PM: Ashok Emani and Qiacheng Li (@intel) show AI Playground Home Agent, which runs PyTorch AI workflows on an Intel AI PC and lets you use them from your phone.
1:15–1:25 PM: Zesheng Zong (@Huawei) presents CRCR (Cross-Repository CI Relay) for joint testing between PyTorch pull requests and accelerator backends, catching regressions at the PR phase.
1:30–1:40 PM: FNU Harsh and Sarah Ahmed (@LightningAI) demo an agent reproducing ML research on Lightning, showing where it succeeds autonomously and where it needs intervention.
1:45–1:55 PM: George Novack (@inferact) introduces TPU megakernels, with Kimi K3 megakernels running inside vLLM-TorchTPU.
2:00–2:10 PM: Jay Dawani (Lemurian Labs) examines why heterogeneous computing is so hard and what a hardware-agnostic stack has to get right.
2:15–2:25 PM: Minwook Ahn (@RebellionsAI) shares how extending PyTorch's device abstraction to support device-persistent tensors minimizes data movement between host and device.
2:30–2:40 PM: Jesse Treger (HPE) introduces OpenCLIC, an open inference communication framework for connectionless networks that is being integrated with vLLM.
3:55–4:05 PM: Mengjin Yan and Joshua Lee (@anyscalecompute) cover evolving Ray Core for post-training at scale, from scaling Ray's scheduling to 10,000-node clusters to moving PyTorch tensors directly between GPUs with Ray Direct Transport.
View the Demo Theater schedule: https://t.co/Sn5gpWxGHO
