PyTorch Day France 2025
2025-05-07
Paris, France
who it's for
machine learning engineers, deep learning researchers, infrastructure engineers, mlops engineers, research scientists, model inferencing engineers
topics
video and audio decoding for PyTorch; multi-accelerator LLM inference; mixture of experts architectures; FP8 and int8 quantization; sim-to-real robotics with RL; training at thousands of GPUs; FSDP and Zero optimizer sharding; recurrent-state Mamba attention; PyTorch compiler and kernel fusion; open multilingual LLM evaluation
| title | author | length |
|---|---|---|
| Welcome & Opening Remarks | Matt White | 11m |
PyTorch Foundation shifts into an umbrella foundation hosting multiple projects, now with over 30 members and an ecosystem of over 70 projects. Two… | ||
| Real-World Robotics as the Next Frontier for AI? | Pierre Rouanet | 16m |
Argues that robotics advances owe real gains to deep reinforcement learning plus high-fidelity simulation and to LMs/VLMs for perception and instruction… | ||
| TorchCodec: The Media Decoding Library for PyTorch | Nicolas Hug | 13m |
A PyTorch library that decodes video and audio into tensors, available via pip install torchcodec, wraps FFmpeg to provide CPU and CUDA decoding and returns… | ||
| Scaling LLM Inference with vLLM | Erwan Gallen, Eldar Kurtic | 21m |
vLLM provides multi‑accelerator inference with tensor parallelism, pipeline parallelism via Ray and an expert (Mixture of Experts) routing option, and supports… | ||
| Llama 4 | Christian Keller | 16m |
Llama 4 launches two models, Scout and Maverick, that use mixture‑of‑experts with about 17 billion active parameters (Scout uses 16 experts, Maverick 128… | ||
| Harnessing Common Crawl for AI and ML Applications | Pedro Ortiz Suarez | 14m |
Common Crawl provides 275 billion pages and about 9 PB of data (monthly snapshots) hosted on an open S3 bucket with an official client for easy download… | ||
| The Ultra-Scale Talk: Scaling Training to Thousands of GPUs | Nouamane Tazi | 12m |
Demonstrates a practical 5D recipe—data, tensor, context, pipeline and expert parallelism—and a running stack that uses nanotron plus PyTorch DDP/FSDP patterns… | ||
| Teaching Mistral to Reason: Post-Training with PyTorch and NVIDIA | Meriem Bendris | 24m |
Shows a concrete pipeline that turns base models into stepwise reasoners by distillation, continual pretraining, supervised fine-tuning on synthetic traces and… | ||
| DeepSpeed – Efficient Training Scalability for Deep Learning Models | Olatunji Ruwase | 19m |
DeepSpeed presents concrete systems: Zero optimizer shards model state to remove GPU memory limits and enables training from a single GPU up to 100k GPUs… | ||
| Advancing Mamba in PyTorch | Garrett Goon | 18m |
Mamba 2 is a drop-in replacement for attention that keeps a fixed-size recurrent state so inference is constant-time token generation and the cache does not… | ||
| Lightning Thunder: Supercharged PyTorch for Modern Hardware | Luca Antiga | 17m |
Lightning Thunder is an open source compiler that traces a PyTorch NNModule on the meta device and emits a prologue, a purely functional trace, and an epilogue… | ||
| Beat GPT 4 & Build the Best AI For You | Paul Yang | 19m |
Shows concrete gains: GPT‑4o scores 81% on FashionMNIST vs a tree model 90% and a ResNet 95%, and a LLaMA 8B fine‑tuned to generate code diffs is ~40% more… | ||
| Best Practices for Open Multilingual LLM Evaluation | Catherine Arnett | 16m |
Advice: pick benchmarks whose difficulty still separates current models — CIV 200 (200 languages) and MultiBlim (100 languages) are too easy, MultiBlim already… | ||
| Multimodal Open Source at Kyutai, From Online Demos to On-Device | Alexandre Défossez | 21m |
Nonprofit lab open sources multimodal tools: full-duplex 7B Moshi (cloud on L4 or quantized on MacBook Pro), a neural audio codec Mimi to tokenise speech, and… | ||
| PyTorch x Transformers Journey | Pablo Montalvo | 20m |
Transforms the library story: claims that one model one file plus strict backwards compatibility cut review and reproduction time from weeks to hours and lets… | ||
| nothing matches | ||