NVIDIA 将携 12 场演讲与 25 个海报展示亮相 PyTorch 大会
去 PyTorch 大会优先排 NVIDIA 这 12 场：FP4 训练、vLLM 弹性专家并行、WAM 都有工程师现场答疑。
NVIDIA 公布了在 PyTorch Conference 的议程：12 场演讲和 25 个海报展示，时间为 10 月 20-21 日，地点在 San Jose Convention Center。低精度方向覆盖 FP4 量化、FP8 多模态训练、FP8-RL 和 NVFP4 预训练。编译器与 CUDA Graphs 议题包含动态 shape 支持、MoE 训练以及基于 Inductor-TV 的形式化方法。分布式系统部分涉及 Symmetric Memory 的计算通信融合、NCCL 扩展和 vLLM 中的弹性专家并行。物理与智能体 AI 方面会介绍用 PyTorch 构建 World Action Model（WAM）以及面向智能体模型的多教师蒸馏。
Sponsor news from @NVIDIAAI: PyTorch Conference is where our engineers share the work — and this year, we have a lot to share.
12 talks and 25 poster sessions across numerous areas: - Compiler & CUDA Graphs — dynamic shape support, MoE training, shape-stable control flow, formal methods with Inductor-TV - Low-precision training & inference — FP4 quantization, FP8 multimodal training, FP8-RL, NVFP4 pretraining - Distributed systems — fused compute-comms with Symmetric Memory, NCCL extensions, elastic expert parallelism in vLLM, resilient distributed with NCCL Shrink & Revoke - Physical & agentic AI — building the World Action Model (WAM) with PyTorch, multi-teacher distillation for agentic models, closed-loop robotics evaluation
The developers behind this work will be presenting and taking questions. If you're attending, these are sessions worth prioritizing.
📅 Oct 20-21 | San Jose Convention Center
🔗Register https://t.co/1z0jDhdUZm