PyTorch Conference 2025
2025-10-22..2025-10-23
San Francisco, United States
[PyTorch Conference 2022, PyTorch Conference 2023, PyTorch Conference 2024, PyTorch Conference 2025]
| title | author | length |
|---|---|---|
| Welcome & Opening Remarks | Matt White | 17m |
Announces the PyTorch Foundation moving to an umbrella model and hosting core projects VLM, DeepSpeed, Ray to cover inference, training, and distributed… | ||
| Recognizing the PyTorch Foundation CICD Infrastructure Community | Eli Uriegas | 04m |
PyTorch CI infrastructure now runs 24 million CI/CD jobs year to date across MacOS Windows Linux multiple CPUs and accelerators, consuming 794 million minutes… | ||
| vLLM & Deepspeed Updates | Simon Mo, Tunji Ruwase | 10m |
vLLM presents an inference engine that runs open models on heterogeneous datacenter hardware, integrates PyTorch features like torch.compile and flex… | ||
| Ray: A Distributed Compute Engine for AI | Robert Nishihara, Ion Stoica | 15m |
Ray claims to be a general distributed compute engine built for modern AI, reporting 7x downloads last year and adoption by thousands of companies. It exposes… | ||
| The Physical Turing Test: Solving General Purpose Robotics | Jim Fan | 19m |
Claims concrete engineering wins: GPU simulation runs up to 10,000x real time in Isaac Lab/CUDA, teleoperation yields ~24 hours max per robot (practically ~4)… | ||
| From Transactions to Transformers: Hardware Built for... | Christian Jacobi | 05m |
Claims enterprise mainframes run 33 billion encrypted transactions per day with eight nines availability and handle 70% of the world’s financial transactions… | ||
| Olmo-Thinking: Training a Fully Open Reasoning Model | Nathan Lambert | 15m |
7B Olmo-Thinking claims parity with Qwen 3 by combining midtraining, targeted SFT distillation from high-quality “thinking” data, and reinforcement learning… | ||
| mxfp8, mxfp4, nvfp4 formats and applications in PyTorch | Vasily Kuznetsov, Driss Guessous | 25m |
MicroScaling formats MXFP8, MXFP4 and NVFP4 appear as concrete PyTorch features: MXFP8 stores tensors ~2x smaller and achieves up to 2x speed on matrix… | ||
| ExecuTorch 1.0: General Availability Status for Mobile and Embedded... | Mergen Nachin, Cemal Bilgin | 25m |
ExecuTorch 1.0 is GA: a PyTorch-first export-to-edge stack using torch.export and progressive lowering to produce a flatbuffer runtime that runs without a… | ||
| Behind Every AI SBOM Lies a Minefield: Here's How to Not Step... | Matthew Crawford, Jimmy Ahlberg | 24m |
OpenChain launches an AI SBOM guide launched (6 pages) that prescribes process controls (policy, roles, competency, artifact retention) and urges documenting… | ||
| Open Source Model Performance Optimization With SGLang | Yineng Zhang | 25m |
SGLang presents concrete performance wins: flash attention integration, dynamic MHA switching, kernel fusion and DP attention deliver ~2x speed on DeepSE F3… | ||
| PyTorch Symmetric Memory: A New Programming Paradigm for Distributed AI | Ke Wen, Chien-Chin Huang | 27m |
PyTorch adds Symmetric Memory: a global GPU address space that lets device kernels read/write peer GPU memory directly and author custom communication logic in… | ||
| Generating State-of-the-Art GEMMs for Heterogeneous Hardware with... | Michael Lazos, Henry Tsang | 28m |
Integration of CUTLASS and Composable Kernel into TorchInductor generates and benchmarks thousands of templated GEMM kernels at JIT time, precompiles them… | ||
| Lightning Talk: Accelerating Experimentation... | Niall Lyons, Luca Antiga | 11m |
Lightning’s platform plus Infineon toolchain trains and deploys tiny language models (25M, 13M, 8M parameters) to Cortex‑M55/U55 microcontrollers with 6.5 MB… | ||
| Mojo + PyTorch: A Simpler, Faster Path To Custom Kernels | Spenser Bauman | 25m |
A new workflow uses Mojo to write PyTorch custom kernels without C++, CUDA, or custom build systems and exposes them via a per-op custom op loader or a Max… | ||
| Deploying GenAI for Audio Generation on Mobile CPUs With ExecuTorch | Gian Marco Iodice | 27m |
ExecuTorch plus Clyde AI runs Stable Audio on Arm CPUs generating 11s stereo 44 kHz clips with local inference under 10s (about 7s on one core, <4s on four… | ||
| SPDX: Tackling System Risk in Modern Supply Chains | Kate Stewart, Gary O'Neall | 28m |
SPDX 3.0 encodes software, models, datasets and hardware as linked artifacts so system-level risk can be reasoned about using a knowledge graph. The talk gives… | ||
| Coding with AI | Chip Huyen | 26m |
Overview of current AI coding tooling, interfaces and measurable claims: compares IDE, CLI, GitHub and web UIs and names Cloud Code, OpenAI Codex, Gemini CLI… | ||
| Beyond the Node: Scaling Inference with Cluster-Wide KVCache... | Alon Yariv | 19m |
Claims a cluster-wide KVCache that turns prefill recomputation into cheap fetches, demonstrated on Llama 8B with KVCache offloaded to shared DRAM and served… | ||
| Hardware-Aware Python Packages ~ PyTorch and... | Jonathan Dekhtiar, Eli Uriegas | 11m |
Goal: make pip install torch pick the right binary by reading machine hardware (CUDA version, GPU SMs, CPU ISA like AVX/ARMv8/ARMv9). Implementation: new Wheel… | ||
| The Future Is Tiled: Using CuTile & TileIR To Write Portable, High-performance GPU... | Jared Roesch | 28m |
NVIDIA introduces a new stack: cuTile Python DSL atop a virtual ISA called TileIR exposed as a TileIR MLIR dialect, with a TileIR backend for Torch Inductor… | ||
| Leveraging PyTorch for Generative AI in Distributed Edge Clouds | Tina Tsou | 25m |
PyTorch Edge and SQTorch enable exporting PyTorch 2.x models to compact on-device runtimes and use compiler-backed backends for Metal and NNAPI while combining… | ||
| The OpenMDW License Agreement: Simple, Permissive Terms for AI Model Materials | Steve Winslow | 21m |
OpenMDW is a short permissive license (openmdw.ai) explicitly written for AI model materials — architectures, parameters and related artifacts — that borrows… | ||
| Architecting AI: Efficient Models, Everywhere | Sharbani Roy | 20m |
Arm argues that small models let AI run on device to cut token fees, lower latency and keep private data local, citing a 300 ms UX ceiling and that training… | ||
| Efficient MoE Pre-training at Scale on AMD GPUs With TorchTitan | Liz Li, Yanyuan Qin, Matthias Reso | 26m |
Claims an end-to-end PyTorch pipeline (TorchTitan + Primus Turbo) that uses FP8 blockwise quantization and grouped GEMM via torch._scaled_mm plus a custom… | ||
| Our Journey With TorchTitan | Linsong Chu, Garrett Goon | 30m |
IBM Research moves from throughput benchmarks to production using TorchTitan and trains an in-house 70B model matching LLaMA 3 while consuming roughly… | ||
| Welcome Back | Matt White, Joe Spisak | 08m |
Meta announces a PyTorch stack with concrete tools and partnerships: Helium lets engineers write GPU kernels in ~30 lines with autotuning and… | ||
| The Building Blocks of Agentic Al | Joe Spisak | 06m |
Company-scale stack for agentic AI that targets hardware heterogeneity and on‑device inference, with concrete projects and claims: LLMs generating kernels and… | ||
| Foundations for AI+Science | Animashree Anandkumar | 20m |
Introduces neural operators that map between continuous function spaces so models run at any resolution and learn from discretized samples, claims the 2021 AI… | ||
| Build AI Anywhere with ROCm™ Software: AMD and PyTorch Bring... | Anush Elangovan | 05m |
ROCm is presented as an open software stack tightly integrated with PyTorch, claiming Day-0 framework and model support, sweeping performance gains in recent… | ||
| PyTorch Technical Deep Dive | Alban Desmaison, Peng Wu, Mark Saroufim, Edward Yang | 43m |
PyTorch reports concrete growth and new tools: over 16,000 research papers and ~14,000 commits last year, 1,000+ contributors and 400,000 new GitHub projects… | ||
| Unlocking AI at Scale for PyTorch Developers | Jeff Gehlhaar | 08m |
Qualcomm outlines a full edge stack: hardware (CPU GPU NPU), Qualcomm AIHub for web-enabled model testing and deployment, and ExecutorTorch reaching GA on the… | ||
| Agentic AI | A. Dhinakaran, J. Liu, A. Jones, S. Colvin, B. Lorica | 28m |
Coding agents outperform expectations thanks to strong feedback loops like type safety and unit tests, while teams that succeed treat evals as… | ||
| Build and Deploy AI Flows with an Agent Factory | Arjun Upadhyay | 09m |
Oracle shows a no-code Private AI Agent Factory that builds visual agent workflows, covers drag-and-drop components, and exports/imports a single JSON open… | ||
| Everything Everywhere all at Once: vLLM... | Brittany Rockwell, Shireen Kheradpey | 26m |
vLLM TPU backend now supports both PyTorch and JAX under one code path and ships a new Ragged Page Attention kernel v3 (Palace) and XLA lowering that yields… | ||
| Beyond TorchScript: The Future of C++ Inference in PyTorch | Sherlock Huang | 27m |
TorchScript is being deprecated in PyTorch 2.9 with a four‑cycle code deletion target around version 24; migration moves models to torch.export (FX graphs with… | ||
| An Open Source Post-Training Stack: Kubernetes + Ray + PyTorch + vLLM | Robert Nishihara | 23m |
An open source stack using Kubernetes + Ray + PyTorch + vLLM runs post‑training workloads by letting Kubernetes provision nodes, Ray orchestrate heterogeneous… | ||
| PyTorch-Native Stack for Agents | Allen Wang, Davide Testuggine | 27m |
A PyTorch-native RL stack that actually ships orchestration, RDMA weight sync, and environment tooling: TorchForge for actor-based rollout/trainer… | ||
| Low-Precision Inference without Quality Loss... | Pankaj Gupta, Philip Kiely | 26m |
Claims that FP8 and NVFP4 microscaling cut latency and cost while preserving quality, with NVFP4 yielding ~30% model-size/accuracy efficiency and GEM kernels… | ||
| Optimizing Model Inference with PyTorch 2.0 | Devansh Ghatak | 11m |
Cloud-focused walkthrough of PyTorch 2.0 techniques: int4 quantization cuts memory enough to run a 12B model on smaller GPUs with only ~0.01% accuracy loss… | ||
| Accelerating PyTorch FSDP Via Overlapping Collectives With In... | Nariaki Tateiwa | 14m |
Extension of the UCC library and a prototype chain of Nvidia SHArP plus a zerocopy SPIN backend enables overlapping ReduceScatter and Allgather for PyTorch… | ||
| Thunder: Distribute and Optimize Your PyTorch Models With Zero... | Luca Antiga, Thomas Viehmann | 26m |
Thunder compiles PyTorch by applying composable, full-graph Python transforms that produce debuggable code and enable Megatron-style sharding, kernel fusion… | ||
| Bridging the Gap: Embedded Deployment of PyTorch Deep Learning Models in Complex Engineered Systems | Ashvant Ram Selvam, Sarah Mohamed | 26m |
Workflow trains a PyTorch model, imports it into Simulink/MathWorks model‑based design to replace the state‑of‑charge estimator with an AI virtual sensor, adds… | ||
| vLLM: Easy, Fast, and Cheap LLM Serving for Everyone | Simon Mo | 25m |
vLLM serves open‑source LLMs with claims of production scale on NVIDIA/AMD GPUs, TPUs, Trainium and CPUs, plus OpenAI and Anthropic API compatibility and… | ||
| Amazingly Fast and Incredibly Scalable Inference... | Harry Kim, Laikh Tewari | 26m |
NVIDIA shows concrete system and model-level tricks to cut LLM inference cost and raise interactivity: NVFP4 4-bit quantization with 16-value block scaling and… | ||
| Mapping the Open Model Landscape in 2025 | Nathan Lambert | 34m |
2025 is an inflection year: DeepSeek R1 drops in January and open Chinese labs (Qwen, Z.ai, Moonshot AI, Alibaba) release frequent models at scale (DeepSeek at… | ||
| Unlock the Future of Generative AI: TorchTitan's Latest... | Tianyu Liu, Jiani Wang | 11m |
TorchTitan adds concrete systems: 5D parallelism (context-sharded sequence dimension plus expert parallel for Mixture of Experts), compiler-driven parallelism… | ||
| Per-iteration PyTorch Visibility... | Shruti Verma, Yi-Ting Wu | 11m |
Claims per-iteration visibility into PyTorch step timing with no code change, using a Python site package to emit job/iteration indices and… | ||
| FlexCP: FlexAttention for Long Sequence Training... | Xilun Wu, Chien-Chin Huang | 11m |
FlexAttention integrates with context parallelism to shard QKV plus the block mask and rewrite mask functions so local QKV indexes map to global ones; an… | ||
| Unlocking Performance: Harnessing LLMs To Streamline GPU Kernel Development in... | Jiannan Wang | 20m |
KernelBook provides 25,000+ PyTorch→Triton pairs as training data and KernelLLM yields surprising efficiency: an 8B-parameter model matches 671B DeepSeek‑R1 on… | ||
| Breaking Heterogeneity Barriers: Unified Cloud-to-Robot AI System SW Stack for... | Yonghua Lin | 30m |
A PyTorch-native stack called FlagOS stitches cloud VLMs to edge VLAs using a Triton-based operator library FlagGEM (200+ Triton kernels) plus compiler… | ||
| Distributed Training in PyTorch: Zero to Hero | Corey Lowman | 25m |
A step-by-step GitHub guide shows how to scale a single-GPU script to multi-node production, with runnable examples and diagnostic tips. Practical claims… | ||
| Fine Tune Your AI Toolbox with GitHub Copilot, Azure AI Foundry... | Sarah Kaiser | 23m |
Live demo shows GitHub Copilot in VS Code generating a Jupyter notebook, using GPT5 Mini to scaffold pandas data loading and plots, auto-installing missing… | ||
| Flex Attention for Inference | Boyuan Feng, Driss Guessous | 11m |
FlexAttention ships in PyTorch 2.5.0 and implements most attention variants in a few lines of PyTorch; it adds a compiler-driven FlexDecoding backend that uses… | ||
| Improving Drug Discovery with Machine Learning... | Simon Axelrod | 11m |
D.E. Shaw Research pairs atomistic molecular dynamics and PyTorch models to speed and improve drug design: Anton 3: 1 ms of chemical time in 6–9 days versus… | ||
| KernelBot: The World's First Competitive GPU Programming Platform | Mark Saroufim | 13m |
KernelBot runs a time‑shared GPU service with a CLI and Discord integration that yields sub-15s feedback (modal hits sub‑10s) and uses PyTorch tensors as the… | ||
| What's New in Torch.onnx | Maanav Dalal, Justin Chu | 11m |
New exporter path uses dynamo=True with torch.export/ExportedProgram to capture functional FX graphs, producing ONNX IR models that share tensors with the… | ||
| Accelerated Software for a Post-Moore World | Jay Dawani | 11m |
Argues for replacing kernel‑first tuning with a system‑level optimizing compiler that treats a full workload as the unit of performance, integrates with… | ||
| Building the Next Generation of Open Source AI Tooling | Travis E. Oliphant | 26m |
OpenTeams describes an open-source platform called Nibari that installs a Jupyter-based AI stack in about 30 minutes and evolves into a modular AI deployment… | ||
| No GPU Left Behind: Scaling Online LLM Training With... | Mert Toslali, Yu Chin Fabian Lim | 19m |
Implements co-located vLLM inside TRL so inference runs in-process (no HTTP server), removes ping-pong GPU idle time, and supports torchrun SPMD with DP and TP… | ||
| Enabling Lightweight, High-Performance FSDP With NVIDIA GPU | J. Chang CN, C. Ye, X. Chen, S. Lym | 29m |
Megatron‑FSDP is a PyTorch‑compatible FSDP implementation that reaches ~2020 TFLOPS (versus 1340 TFLOPS baseline) on Llama 3 4.5B across 64 B200/GB200 GPUs by… | ||
| Lightning Talk: Challenges and Standardization in PyTorch Ecosystem... | Zesheng Zong, Ashok Emani | 13m |
Sets a concrete path for out‑of‑tree accelerators: a public testing repo pytorch-fdn/oota provides daily CI and torchbench validation, unit tests merge into… | ||
| Keynote: Welcome Back to Day 2 | Matt White | 01m |
Day two opens with logistics and community notes: attendees report posters and a busy Flare Party, and the exhibit floor hosts a PyTorch gear store with… | ||
| Keynote: Building the Open Agent Ecosystem Together: Introducing OpenEnv | Joe Spisak, Lysandre Debut | 09m |
Announces OpenEnv as an end-to-end framework for creating, deploying, hosting and using agentic RL environments with simple APIs and a Hub integration on… | ||
| Keynote: Enabling PyTorch’s Thousand Ops for Software First Silicon Design | Andrew Ling | 13m |
Claims a hardware first redesign that makes compilation simple by making the architecture deterministic and software controlled and shows it works on inference… | ||
| Keynote: Marin: An Open Lab for Frontier AI | Percy Liang | 16m |
Marin runs model development in the open using GitHub issues to pre-register experiments and pull requests that contain full experiment pipelines in Python so… | ||
| Keynote Panel Discussion: Signal vs. Noise: Betting on AI’s Next Wave | Sarah Catanzaro, Sarah Guo, Brian Zhan, Peter Deng, Joe Spisak | 34m |
Five VCs debate concrete signals: forks over stars and founders who show grit by doing 100 customer interviews, not vanity metrics; one firm reports one in… | ||
| Keynote: Robot Foundation Models | Sergey Levine | 16m |
Demo shows a robot completing "put the eggplant in the pot" using RT2X, and a cross-embodiment dataset RTX: 34 labs, 22 robots trains a single model that… | ||
| Enabling VLLM V1 on AMD GPUs With Triton | Thomas Parnell, Aleksandr Malyshev | 22m |
vLLM V1 runs on AMD MI300-series GPUs by replacing legacy HIP primitives with Triton kernels and by splitting heterogeneous batches into specialized kernels… | ||
| Transformers: Standardizing Model Definitions Across the PyTorch Ecosystem | L. Debut, A. Zucker | 25m |
Transformers V5 breaks compatibility in specific ways: it removes TensorFlow and JAX model implementations, standardizes tokenizers on the Rust tokenizers… | ||
| The Building Blocks of Today’s and Tomorrow’s Language Models | Sebastian Raschka | 49m |
Survey of 2025 LLM architectures highlights concrete tradeoffs and wins: models like DeepSeek v3 reach 671 billion parameters while using mixture of experts… | ||
| “Many-model” Time Series Forecasting: Scaling PyTorch Training Across 1000s... | V. Sridhar, S. Chen | 26m |
A pipeline that trains thousands of PyTorch models in parallel by partitioning Snowflake tables, converting partitions to Ray datasets, and mapping each to a… | ||
| How Modern PyTorch Supercharges Multimodal Training and Inference at Luma AI | Thomas Neff | 25m |
Luma AI ships Ray 3 and Dream Machine (built in five months) and runs multimodal training using a unified forge config system that applies FSDP2/FSTP2… | ||
| Optimizing Long-Tail and MoE Challenges in Reinforcement Learning with SGLang | Chenyang Zhao | 28m |
SGLang presents concrete fixes: partial rollouts use oversampling so the first ~80% of fast responses (arriving in 50% of time) train immediately while slow… | ||
| Helion: A High-level DSL for Kernel Authoring | Jason Ansel | 28m |
A Python-embedded DSL that compiles to Triton, Helion exposes a single critical construct hl.tile tiling construct and maps inner tile bodies to device kernels… | ||
| Sponsored Session: PyTorch Distributed and Fault Tolerance | Tristan Rice | 24m |
Open source torchcoms provides a thin, object oriented communicator abstraction over new backends like nickel x and Catran transport, claiming fast startup and… | ||
| Scaling Inference of O(10K)-length Sequence Recommendation Models Using... | S. Joshi, K. Rajesh | 25m |
Pinterest scales sequence length from 500 to 16,000 for a 10M+ inferences/sec ranking fleet while keeping P90 <50 ms by combining TorchScript + a custom… | ||
| Serving PyTorch LLMs at Scale: Disaggregated Inference With Kubernetes and Llm-d | M. Ayoub, C. Liu | 27m |
llm-d implements disaggregated prefill/decode on Kubernetes, running vLLM-backed model servers with a routing sidecar and NVIDIA NIXL for GPU-to-GPU KV… | ||
| Verl: A Flexible and Efficient RL Framework for LLMs | Hongpeng Guo, Ziheng Jiang | 25m |
verl is an open source RL pipeline that combines a hybrid-controller (central logic + many SPMD controllers) with WorkerGroup and ResourcePool for cluster… | ||
| Sponsored Session: Lightning Talk: PyTorch and Democratization of AI Accelerators | Hong-Seok Kim | 11m |
Rebellions describes native PyTorch integration via a 2024-launched torch-rlbn package that targets a chimnet-based flagship chip claiming 32x compute over… | ||
| Multi-Accelerator PyTorch Serving With NxD Inference and vLLM | Yahav Biran, Liangfu Chen | 29m |
NxD Inference packages PyTorch models into neuron executables and runs alongside TorchInductor CUDA kernels inside a vLLM-based Kubernetes cluster so Llama 8B… | ||
| Monarch: A Distributed Execution Engine for PyTorch | Colin Taylor, Zachary DeVito | 24m |
Monarch provides a single-controller Python API that treats a cluster like one process, exposing hosts/processes/actors as typed meshes and supporting scalable… | ||
| Scaling KV Caches for LLMs: How LMCache + NIXL Handle Network and Storage... | J. Jiang, M. Khazraee | 33m |
LMCache integrates with NVIDIA’s NIXL transfer library to move big K and V tensors between GPUs, CPUs, and storage using a buffer-list primitive and… | ||
| Maximizing Luck in Reinforcement Learning | Daniel Han | 18m |
Unsloth claims RL training speeds and memory drop by specific tweaks: float8 quantization and keeping the LM head in float32 for accuracy, plus algorithm swaps… | ||
| Sponsored Session: Lightning Talk: PyTorch in Production: Boosting LLM Training and... | F. Hua | 08m |
Shows concrete PyTorch changes for Ascend NPU: adds privateuse1 plugin loading in PyTorch 2.1 so backends auto-import via Python entry points, provides an… | ||
| Keynote: Towards Building Safe and Secure Agentic AI | Dawn Song | 19m |
Describes an agent visual blackbox red‑teaming method that uses iterative optimization to craft environment payloads (example: a malicious review that triggers… | ||
| Keynote | Jeremy Howard, Anna Tong | 20m |
Argues that transfer learning fine‑tuning is the core breakthrough for language models and that ULMFiT, built in a tiny Jupyter notebook on PyTorch, beats… | ||
| Keynote Panel: Hardwares & Accelerators | D Patel, S Zhou, P Salanki, N Perumbeti, M Saroufim | 30m |
Panel argues big deals close by repeatedly delivering scale and reliability, not magic: negotiating aligns three trains—data center & energy, hardware… | ||
| Keynote: An OPEN Path to Super Intelligence | Eric Xing | 17m |
An academic-led institute is building open foundation models that aim to match much larger closed models: K2 (32B parameters) launches with reported parity to… | ||
| Keynote: Contributor Awards | Chris Scottrath, Greg Channon | 09m |
Eleven contributor awards honor specific community roles: Review Powerhouse goes to Zingo Anderson, Innovator to Jonathan Decar, Trailblazer (docs) to Justin… | ||
| Keynote: Closing Remarks | Matt White | 06m |
Announces conference giveaways and contest winners, reading names and ticket numbers live and handing out prizes: top scorer on the Hoova app leaderboard Ara… | ||
| From Model To Trust: Building Upon Tamper-proof ML Metadata Records | Mihai Maruseac | 25m |
Demonstrates a working system that cryptographically signs bundles of ML artifacts (models, datasets, model cards) using OpenSSF Model Signing, produces… | ||
| Blazing Fast GenAI Inference With Torch.compile | Richard Zou | 23m |
Torch.compile captures PyTorch graphs with Dynamo and generates fused kernels via inductor, delivering measured speedups like ~60% on Gemma 3, benchmarks from… | ||
| Sponsored Session: Empowering AI Everywhere: Democratizing PyTorch with Intel... | F. Zhao, E. Wang | 26m |
PyTorch runs natively on Intel AIPCs using an XPU device with >90% ATen ops implemented (via SYCL-like kernels and oneDNN/oneCCL backends), a Triton-enabled… | ||
| Designing and Building Custom Reinforcement Learning Environments for Fine-tuning LLMs | N. Bantilan | 25m |
Built a token-level RL environment called Wikipedia Maze that casts pages as states and link clicks as actions, with deterministic terminals and a shaped… | ||
| Sponsored Session: Accelerating GenAI Inference: From AWS Deep Learning... | P. Nguyen, A. Zhao | 25m |
AWS Deep Learning Containers plus VLM and PyTorch power a multi‑node inference cluster that shards models across GPUs and Trainium using Neuron and EFA over… | ||
| AI On-board of Satellites, Towards Autonomous Scientific Instruments | Vit Ruzicka | 30m |
Deployed a tiny foundational model RaVAEn on a D‑Orbit satellite in 2023 and runs the world’s first PyTorch on-board training for a few‑shot cloud classifier… | ||
| Lightning Talk: Vllm-triton-backend: How To Get State-of-the-art Performance on... | B. Ringlein | 09m |
vLLM gains a single Triton-only code base that runs unchanged on NVIDIA H100 and AMD MI250, using 3D launch tiles and parallel softmax plus persistent-launch… | ||
| Sponsored Session: Lightning Talk: Efficient Inference Serving with Kubernetes Gateway... | Lin Sun | 14m |
Demonstrates a working stack: vLLM simulator plus the Kubernetes Gateway API Inference Extension and an agent gateway implementation that routes by header and… | ||
| Lightning Talk: FTorch: A Library To Couple PyTorch ML Models With Fortran Climate Models | M Archer | 12m |
FTorch is an open-source library that directly loads torchscript models into Fortran via LibTorch via Fortran 2003 bindings, bypassing Python and supporting… | ||
| PyTorch APIs for High Performance MoE Training and Inference | D. Vega-Myhre, Ke Wen, N. Gimelshein | 25m |
PyTorch adds GPU-native all_to_all_2d dispatch/combine that returns device-side offsets to avoid host-device syncs and to place per-expert tokens contiguously… | ||
| Rethinking the Transformer: Toward Native Multimodal Architectures | Bowen Peng | 26m |
Argues for training on raw information rather than language proxies, proposing tokenizer-free transformer designs and hierarchical attention to handle wildly… | ||
| Sponsored Session: Powering PyTorch: Decentralized Training for an Energy-Hungry Future | Greg Osuri | 23m |
Claims that distributing PyTorch training across modular datacenters and a decentralized GPU marketplace cuts energy and expands access, demonstrates concrete… | ||
| Lightning Talk: Improved GEMM and SDPA Performance on ROCm With Composable Kernel | Andres Lugo | 10m |
Composable Kernel (CK) is now a ROCm backend for GEMMs and SDPA, reachable by build-time env vars like use_rocm_ck_gemm and a single Python line to select the… | ||
| Lightning Talk: But How Do You Autograd Through Your Simulator? Making a... | R. Dagli, S. Lamba | 11m |
Practical recipe for turning PyTorch physics code into a differentiable simulator: replace non‑smooth collision and Coulomb friction with smooth proxies… | ||
| Lightning Talk: Dynamic Shape Recompilations | Bob Ren | 09m |
Explains when recompilations from dynamic shapes are worth the cost with concrete numbers: a 20 ms kernel win invoked 100 times nets 2 s but loses to a 5 s… | ||
| Lightning Talk: Model Checkpoint Compression With OpenZL | Nick Terrell, Teja Rao | 09m |
OpenZL is an open source lossless compressor that uses tensor type information and a DAG of codecs to transform and entropy‑encode numeric data rather than… | ||
| Addressing the Challenges of Public Web Data | Greg Lindahl | 41m |
Common Crawl maintains a 10 PB archive of about 30 billion pages, adding ~2.5 billion pages a month and serving nearly 1 exabyte of downloads last year, and… | ||
| DeepSpeed: Efficient Training Scalability for Deep Learning | Tunji Ruwase | 26m |
DeepSpeed presents concrete systems: super offload overlaps optimizer work across GPU and Grace CPU with speculative validation, fine-tunes GPTOS 20B and Qwen… | ||
| Kimi K2 and Our Contributions to Open Source | Yuxin Wu | 26m |
Kimi K2 is an open-weight 1 trillion parameter foundation model that activates about 32 billion parameters per token, trains on 15 trillion tokens and uses 384… | ||
| Controllable Generation of Synthetic Video Data With Human Motion | Vaclav Knapp | 18m |
A PyTorch-compatible toolkit generates controllable human-action videos by combining pose transfer and 3D-grounded animation, releases 100 controllable… | ||
| Lightning Talk: A Stable Limited LibTorch ABI? How?! (and Why?) | Jane Xu | 12m |
Proposal to add a limited stable libtorch ABI that lets C++/CUDA extensions survive PyTorch upgrades by building against a promised surface rather than… | ||
| Lightning Talk: In-Cluster Distributed Checkpointing: Optimizing Training... | G. Kroiz, S. Mishra | 10m |
PyTorch Distributed Checkpointing adds an In-Cluster Checkpointing mode that saves and restores each GPU node’s state to node-local storage, replicates those… | ||
| Sponsored Session: Arctic Inference: Breaking the Speed-Cost Tradeoff in LLM Serving | Aurick Qiao | 28m |
Arctic Inference is an open‑source vLLM extension that ships specific optimizations: Shift Parallelism (combines tensor and sequence parallelism to match… | ||
| Lifecycle of a Parameter | Philip Bontrager | 25m |
Follows one parameter through an LLM RL post‑training pipeline: reading 100s of GB checkpoints, doing FP8→BF16 dequant on load, splitting the tensor across… | ||
| Fault Tolerance Training at Pinterest | Chia-Wei Chen, Karthik Anantha Padmanabhan | 28m |
Pinterest implements iteration-level checkpointing using StatefulDataloader and stores model, optimizer and LR scheduler state plus the exact dataloader… | ||
| PyTorch Conference 2025 Recap | 02m | |
Recap captures collaboration and energy at PyTorch Conference 2025 where 3,000 developers convene over two days in San Francisco; the video compiles keynotes… | ||
| nothing matches | ||