PyTorch Conference Europe 2026
2026-04-07..2026-04-08
Paris, France
as the organisers put it
Join top-tier researchers, developers, and academics for a deep dive into PyTorch, the cutting-edge open-source machine learning framework.
| title | author | length |
|---|---|---|
| Co-Evolution: How the Open Source Intelligence Stack Compounds | Mark Collier | 12m |
Argues that AI progress comes from synchronized updates across models, training frameworks, inference engines, RL systems, hardware and cloud, and that open… | ||
| PyTorch Updates | Edward Yang | 21m |
PyTorch reports 15,000 contributions this year and 1,250 new contributors while remaining the backend for 9/10 top research projects and 90% of open‑source ML… | ||
| Community Led Open Source RL | Joe Spisak | 12m |
Open Enve is a GitHub spec under Meta PyTorch plus a Hugging Face hub with a CLI that lets developers push environments; the initial project surfaces a… | ||
| From One Node to Distributed Training and Inference. How the PyTo... | Ramine Roane | 05m |
Claims that single‑GPU or single‑node inference is obsolete as open models exceed 500 billion parameters — DeepSeek 671B and MyMovie 2 Pro 1T — and that these… | ||
| Stream Everything - Moving from Request input to Streaming input | Patrick von Platen | 15m |
Describes a working streaming-input system called Vox Real-Time that feeds small audio frames plus autoregressive text embeddings into a causal LM, trains the… | ||
| Any [ Agent | Model | Accelerator | Cloud ]... | Maryam Tahhan, Nicolò Lucchesi | 04m |
Red Hat claims to be the leading contributor to vLLM and to have integrated community projects from the PyTorch Foundation (vLLM, LLMD, Smart Router, got LLM)… | ||
| The Unbearable Lightness of (Agentic) Evaluations | Besmira Nushi | 11m |
Claims that agentic evaluation is fragile and nonreproducible, and that open tooling is required to own scores. A simple proxy called Kodak contamination proxy… | ||
| Why Your Forecasting Transformer Isn’t Working (And How To Fix It... | Rosheen Naeem | 10m |
Implements a Temporal Fusion Transformer via PyTorch Forecasting and PyTorch Lightning, training on hourly PV data from Hugging Face, site metadata (capacity… | ||
| Training Embedding Model Resiliently for Multimodal M... | Huamin Chen, Haichen Zhang | 13m |
Describes training production-grade embedding and classifier models for a vLLM Semantic Router, delivering a multilingual text embedding with 32K context… | ||
| Helion 1.0: A High-Level DSL for Performance Portable Kernels | Oguz Ulgen | 24m |
Helion is a Python-embedded DSL that preserves PyTorch APIs, lowers through TorchInductor to multiple backends (Triton today, QTESL/Palace/Metal planned), and… | ||
| Lights, Camera, Inference! Video Generation as a Service With VLLM-O... | Ricardo Noriega, Doug Smith | 25m |
Prototype stacks vLLM-Omni + LTX-2 run live text-to-video and image-to-video endpoints with ComfyUI nodes and an async video serving path, and the repo already… | ||
| Deep Learning in the Wild: Embedded PyTorch for... | Taraqur Rahman, Owen O'Donnell | 10m |
Owl IoT “ducks” form a LoRa mesh that hops detections to a single gateway and then to satellite; initial deployment runs BirdNet on Raspberry Pi, gets… | ||
| Flexible Deployment of PyTorch Models on MCU-Class... | Robert Kalmar, Martin Pavella | 14m |
ExecuTorch compiles PyTorch for Cortex‑M, Cadence DSPs, Ethos‑U and eIQ Neutron using portable C++ kernels, core‑specific backends and NPU delegates… | ||
| Coding Agents for Compiler Construction: Beyond the... | Reza Rahimi, Stefan Krassin | 12m |
An agentic compiler prototype is implemented as a PyTorch add-on that ingests nn.Module or FX graphs, runs an LLM-based code generator plus an eval tool that… | ||
| How DeepInverse Is Solving Imaging in Science and H... | Andrew Wang, Minh Hai Nguyen | 10m |
Claim that imaging is the bottleneck in computer vision and present a usable stack: DeepInverse PyTorch library ties physics models, differentiable ops and… | ||
| LLM Inference Optimization From Simple To Sophisticated | Christin Pohl | 24m |
Practical roadmap: choose a model just “smart enough” and quantize to get ~1.3–1.7x speedups, often delivering ~50% gains from model choice; use vLLM, SGLang… | ||
| Why Classic IAM Collapses for Agents: Rethinking IAM for Agentic Systems | Parul Singh | 25m |
Classic IAM fails when autonomous pipelines contain multiple agents with shared long‑lived service accounts — an ML example has five agents (orchestrator… | ||
| ExecuTorch on Microcontrollers: Deploying PyTorch To... | RJ Ascani, Matthias Cremon | 10m |
ExecuTorch brings PyTorch models to microcontrollers and tiny NPUs with an ahead‑of‑time pipeline: torch.xport graph tracing, node‑level quantization from… | ||
| TorchJD: Jacobian Descent in PyTorch | Pierre Quinton, Valérian Rey | 10m |
TorchJD provides a PyTorch library that computes per-objective Jacobians and aggregates them into update directions, implementing Jacobian descent with… | ||
| Ethical, Privacy and Sustainability Considerations in PyTorch S... | Paula Mesa Macias | 10m |
PyTorch models sit inside data pipelines, APIs, logs and infrastructure so risks appear beyond weights and accuracy. The Ethical Software Framework and an… | ||
| Bringing Google’s Colossus to PyTorch: Rapid Stor... | Ankita Luthra, Trinadh Kotturu | 10m |
Google brings Colossus’ stateful gRPC protocol into the Python stack by adding Rapid buckets to GCSFS and fsspec, bypassing REST and front-end proxies… | ||
| Parameterized CUDA Graph Launch in PyTorch: CUDA Graphs Without the Pain | Daniel Galvez | 26m |
Explains a new Parameterized CUDA Graph launch that lets CUDA Graphs replay with new tensor pointers by patching kernel pointer args and borrowing memory from… | ||
| Write Once, Run Everywhere with Pytorch Transformers | Pedro Cuenca | 19m |
Hugging Face transformers is a pure PyTorch model-definition framework that now targets immediate downstream availability by going PyTorch-only, using a… | ||
| FlexAttention + FlashAttention-4: Fast and Flexible | Driss Guessous | 09m |
FlexAttention integrates with FlashAttention-4 to run Flex-style score and mask hooks as JIT-instantiated CuTeDSL kernels via PyTorch Inductor, converting… | ||
| From Pretrained To Personal: Privacy-First... | Daniel Holanda Noronha, Iswarya Alex | 10m |
Claim: local AI PCs now run meaningful fine-tuning so sensitive data never leaves the device. Demonstrated setup uses PyTorch 2.10 + ROCm 7.12 on an AMD Strix… | ||
| Bringing ExecuTorch To the Next Frontiers of Edge AI | Mergen Nachin | 23m |
ExecuTorch 1.0 GA (Oct 2025) powers an ahead-of-time PyTorch export -> C++ runtime that runs on bare-metal ~50 KB runtimes, mobiles, M1 laptops and… | ||
| Teaching PyTorch To Read Your Worst PDFs With Docling | Mingxuan Zhao, Peter Willem Jan Staar, Carol Chen | 24m |
Docling is an MIT‑licensed PyTorch-native library that detects file formats, runs optimized pipelines, and returns a single pydantic canonical document… | ||
| Why WideEP Inference Needs Data-Parallel-Aware Scheduling | Maroon Ayoub, Tyler Michael Smith | 26m |
WideEP fails not from expert cost but because routing ignores where KV state lives. Using vLLM plus LMD and a router extension called llm-d that reads KV-cache… | ||
| Accelerating On-Device ML Inference With ExecuTorch and Arm SME2 | Jason Zhu | 12m |
ExecuTorch plus Arm SME2 runs SqueezeSAM on a single CPU core and cuts FP16 end to end latency from >1s to ~300ms yielding 3.9x end-to-end speedup. The… | ||
| TorchTPU: Expanding TPU Programmabil... | Kat Ko, Claudio Basile, Jana van Greunen | 23m |
TorchTPU brings PyTorch to Google TPUs with native eager-first and compiled paths, lowering Aten ops to StableHLO via OpenXLA and integrating with… | ||
| The Token Slice: Implementing Preemptive Scheduling Via Chunked Decod... | Maroon Ayoub, Kellen Swain | 22m |
An experimental Chunked Decoding system slices generation into fixed-size token time slices and runs as a sidecar implementation for PyTorch servers (example… | ||
| Combo Kernels: Horizontal Fusion Optimizat... | Karthick Panner Selvam, Elias Ellison | 10m |
Combo kernels in PyTorch Inductor fuse many small independent Triton launches into a single kernel that uses a PID-based dispatch so one launch replaces N… | ||
| Implementing Single-Dim Strategies With Sharding Validator | Anshul Sinha | 09m |
Prototype separates semantic correctness from pruning by requiring each operator to declare Single Mesh Dim Strategies for one mesh dimension while infra… | ||
| Brevitas Quantization Library | Pablo Monteagudo Lago | 30m |
Brevitas is an AMD PyTorch library that composes PTQ stages—rotation transforms, scale/zero choices, and greedy error‑correction—into reproducible pipelines… | ||
| Model-Changing Transforms With Torch.compile | Thomas Viehmann | 22m |
Proposes using torch.compile’s three outputs — graph, guards, and wrapper — as a transformable contract so quantization, distribution and other changes act on… | ||
| The Science and Practice of Open and Scalable LLM Evaluations | Grzegorz Chlebus | 27m |
Describes an open evaluation stack that versions configs and hashes results, runs hundreds of benchmarks reproducibly with repeats and confidence intervals… | ||
| Jigsaw: Domain and Tensor Parallelism for High-Resolution Inp... | Deifilia Kieckhefen | 10m |
Jigsaw shards both model weights and input data so GPUs hold disjoint weight and data slices and avoid redundant I O, implementing distributed matrix… | ||
| Graph Based Pipeline Parallelism | Sanket Purandare, Simon Fan | 10m |
Compiles stage code to reusable FX graphs and runs explicit schedules via a schedule IR now upstreamed in torch.distributed, integrated with TorchTitan and… | ||
| Cross-Region Model Serving: PyTorch Inference, Observability... | Suraj Muraleedharan | 16m |
Production-tested architecture for running PyTorch inference across three regions handling 10K+ RPS per region with a P99 time-to-first-token target of <200… | ||
| Enabling State-of-the-art Asynchronous Execution in Torch.compile With CUDA Streams | Michael Lazos | 29m |
Adds native CUDA stream support to torch.compile (Dynamo → AOT Autograd → TorchInductor) so eager-style stream context managers become symbolic stack state… | ||
| torch.compile and Diffusers: A Hands-On Guide to Peak Performance | Sayak Paul | 25m |
Shows how to use torch.compile with Diffusers to speed up diffusion models like Flux‑1‑Dev, including concrete configs: compile the Diffusion Transformer for… | ||
| Running ExecuTorch Applications With Silicon Accelera... | George Gekov, Aki Makkonen | 13m |
Shows a full pipeline that trains a 11 million‑parameter Conformer in PyTorch, exports a trace, applies post‑training quantization to 8‑bit weights (8/8 or… | ||
| Beyond Generic Spans: Distributed Tracing for Actio... | Sally Ann O'Malley, Greg Pereira | 12m |
Implemented end-to-end tracing for llm-d with manual OpenTelemetry instrumentation that adds targeted spans (gateway/EPP scoring, KV cache lookups, PD proxy… | ||
| Build PyTorch to Understand PyTorch | Vijay Janapa Reddi, Andrea Mattia Garavagno | 21m |
TinyTorch is a 20-module curriculum that has learners implement a PyTorch-like framework in pure Python tensor + autograd with unit tests that export working… | ||
| On-Device LLM Inference on Android With ExecuTorch and Qualcomm QNN | Shivay Lamba, Kartikey Rawat | 25m |
Shows how to run a full ViT‑B/32 CLIP pipeline ExecuTorch + Qualcomm QNN on Android, with model export via torch.export, AOT graph transforms, and on‑device… | ||
| Optimizing Reinforcement Learning at Trillion-Parameter Scale | Songlin Jiang | 25m |
Trillion-parameter MoE reasoning model with 384 experts and ~32.6B active parameters per token trains using LoRA adapters (rank 32 across all linear layers) so… | ||
| TorchStore: What We Learned Building Distributed Storage Sol... | Lucas P, Danielle P, Allen W, Amir Afzail | 26m |
TorchStore is an open source distributed tensor store built on Monarch actors that combines save-anywhere semantics with fast direct transfers using pluggable… | ||
| TerraKit: Standardising AI-Ready Geospatial Data... | Rosie Lickorish, Romeo Kienzler | 11m |
An open‑source Python library called TerraKit automates retrieval, processing, alignment and spatial train/val splitting of EO raster/vector labels into… | ||
| Optimizing PyTorch on CPU-GPU Coherent Platforms | Matthias Jouanneaux | 27m |
Covers practical steps for GB200 and coherent platforms: avoid needless CPU→GPU duplication by using a custom transfer tensor class that solves dispatch… | ||
| Securing Agentic AI With PyTorch: Threat Modeling & LLM Red Teaming in Practice | Valeri Milke | 33m |
Practical guide that applies MAESTRO-based AI Threat Modeling to PyTorch agent architectures, mapping specific risks (model hallucination rates, vector DB… | ||
| Bayesian Neural Networks With Variational Inference in PyTorch | Lars Heyen | 12m |
Claims a light‑weight PyTorch library torch_blue that turns standard models into Bayesian neural networks with variational inference by changing only four-line… | ||
| PyTorch CTO | Matt White | 10m |
PyTorch Foundation reports 73 projects total with five hosted in the foundation and 68 ecosystem projects, adding nine new projects in the last three months… | ||
| vLLM & Ray Updates | Tyler Michael Smith, Artur Niederfahrenhorst | 11m |
vLLM reimplements its GPU runner as Model Runner V2, splitting a 6,000‑line file into several ~1,000‑line modules, decoupling per‑request state, moving input… | ||
| The Hub as Infrastructure. From Open PyTorch Models, to a Safe and Perfor... | Lysandre Debut | 10m |
Hugging Face runs a distribution hub with 2.74M public models and 1.44M PyTorch-compatible checkpoints, where the top 200 models drive 50% of downloads and the… | ||
| Open Source Infrastructure for the AI Native Era | Jonathan Bryce | 05m |
Cloud Native Computing Foundation runs Kubernetes plus 230 projects and focuses on operationalizing AI: deploy, secure, observe, scale. Industry compute flips… | ||
| Gemma 4: Compacting Intelligence for the Edge | Léonard Hussenot | 16m |
Claims that Gemma 4 maximizes “intelligence per byte” with concrete wins: a 31B dense model and a 24B MoE with 4B activated parameters outperform models… | ||
| Monarch: An API To Your Supercomputer | Marius Eriksen | 12m |
Monarch presents a Pythonic API that programs GPU clusters as meshes of hosts, processes and actors, exposing precise placement, lifecycle and one-sided memory… | ||
| Live Migration of PyTorch GPU Nodes From Azure To European Clouds | Mike Krom | 07m |
Live migration of a full GPU PyTorch VM including CUDA drivers, kernels and Jupyter notebooks runs by exporting disks, transforming the OS image and importing… | ||
| Beyond JSON-RPC: Scaling Model Context Protocols With gRPC in the Py... | Ashesh Vidyut, Madhav Bissa | 19m |
Proposal adds SEP-1352 to make MCP use Protobuf over gRPC as a pluggable transport, claiming binary encoding cuts an 8 MB context from ~15–20 MB in JSON down… | ||
| How To Write C++ Extensions in 2026 | Jane Xu, Mikayla Gawarecki | 24m |
Shows how to build a real C++ custom op (a fused multiply‑add with a CUDA kernel) that compiles against the stable ABI so one binary replaces an M×N wheel… | ||
| Lightning Talk: Achieving SOTA GEMM Performance: A CuTeDSL Backend for PyTorch Induc... | Nikhil Patel | 08m |
Adds a CuTeDSL GEMM backend to PyTorch Inductor that queries NVIDIA’s Cutlass kernel library, prunes candidates with the NVMAL heuristic to about 5 candidate… | ||
| Lightning Talk: Step-Aligned Telemetry for Distributed PyTorch Training (Time... | Abhinav Srivastav | 10m |
Step-aligned telemetry library TraceML captures per-step timing and memory, not time-sampled averages, by wrapping the loop with a trace_step context manager… | ||
| Lightning Talk: Accelerating PyTorch Models With Torch.compile's C++ Wrapper Mode | Bin Bao | 14m |
torch.compile replaces the Python wrapper with C++ wrapper mode to cut CPU orchestration overhead, claiming up to 39% speedup over default torch.compile on OSS… | ||
| Lightning Talk: KV-Cache Centric Inference: Building a State-Aware... | Maroon Ayoub, Martin Hickey | 10m |
KV-cache centric inference reframes serving so KV state becomes the primary resource, implemented with llm-d plus vLLM, LLMCache and Envoy on Kubernetes; it… | ||
| Bringing PyTorch Monarch to AMD GPUs: Single-Controller Distributed Tra... | Liz Li, Zachary Streeter | 20m |
Ports Monarch’s runtime and collectives to ROCm so a single-Python controller runs on MI300-class clusters, compiling the same Rust and Python code for CUDA or… | ||
| Fp8 Training From Hopper To Blackwell | Luca Wehrstedt | 25m |
FP8 delivers roughly ~4x speed versus FP32 on modern NVIDIA GPUs but only if tensors are scaled to avoid overflows; scaling is applied right before matmuls and… | ||
| Lightning Talk: Building AI That Ops Teams Actually Trust | Robert King | 10m |
Build AI that ops teams actually trust by replacing monolithic agents with modular subagents that separate context gathering, hypothesis formation, telemetry… | ||
| Lightning Talk: Not All Tokens Are Equal: Semantic KV-Cache for Agen... | Maroon Ayoub, Hyunkyun Moon | 10m |
A runtime called Semantic KV-Cache replaces anonymous KV blocks in llm-d and vLLM with typed state that tags tokens as SystemPrompt, ToolDefinition… | ||
| Lightning Talk: Enabling the Audio Modality for Language Models | Eustache Le Bihan | 14m |
Hugging Face observes 2.7M models on the hub with ASR dominating and TTS rising since Feb 2025. Integration uses pretrained audio encoders (commonly Whisper… | ||
| Accelerating Complex-Valued Tensors With Torch.compile | Hameer Abbasi | 13m |
Provides a working subclass storing real and imaginary parts that lets torch.compile succeed on complex dtypes today on PyTorch nightly, plus an open PR to… | ||
| Optimizing Large MoE Inference on NVIDIA Blackwell: NVFP4, ADP, and DualPipe Strat... | Julien Demouth | 22m |
Describes moving MoE weights to NVFP4 to cut memory and keep accuracy on DeepSeek‑V3/R1, with FP4/FP8 KV caching to shrink attention state; NVFP4 uses… | ||
| Portable High‑Performance LLM Serving: A Triton Backend for... | Burkhard Ringlein, Jan van Lunteren | 24m |
Introduces a Triton-only attention backend for vLLM that replaces hand‑written CUDA/HIP kernels and achieves cross‑vendor performance portability by autotuning… | ||
| Lightning Talk: From Hugging Face To Handheld: Scaling LLM Deployment W... | Cormac Brick, Weiyi Wang | 13m |
Converts a custom Hugging Face checkpoint into a cross‑platform LiteRT IR in under a minute, runs quantization in ~40 seconds, and pushes the packaged model to… | ||
| PyTorch on RISC-V: From Cross-Compilation To Native CI | Ludovic Henry | 25m |
Porting PyTorch to RISC‑V moves from cross‑compilation to native CI by resolving concrete gaps: GCC/LLVM, Ubuntu 24.04 builds, OpenBLAS RVV support and… | ||
| PyTorch Symmetric Memory + NCCL Device APIs: A New Path Towards Multi-GP... | Ke Wen, Sylvain Jeaugey | 25m |
A new stack enables direct GPU-to-GPU memory access without CPU involvement using PyTorch Symmetric Memory for symmetric tensors and NCCL device APIs (LSA… | ||
| Optimizing CPU LLM Inference in PyTorch: Lessons From VLLM | Crefeda Rodrigues, Fadi Arafeh | 24m |
vLLM on Arm moves from barely running to about sixfold faster by combining system and kernel work: replacing glibc malloc with MiMalloc yields +133%… | ||
| Lightning Talk: Slash LLM Cold-Start Times by Pre-distributing GPU... | Billy McFall, Maryam Tahhan | 07m |
Pre-distribute compiled GPU kernel caches as OCI images so pods skip repeated JIT compilation and start minutes faster. A utility called Model Cache Vault… | ||
| Lightning Talk: Pluggable PyTorch LLM Inference Architecture With VLL... | Yahav Biran, Maen Suleiman | 10m |
Modular plugin code lets PyTorch models run on Trainium by registering a torch.neuron device backend so devices appear alongside CUDA and MPU, with automatic… | ||
| Lightning Talk: Backpropagation-Free Optimization in PyTorch | Andrii Krutsylo | 12m |
Compares four concrete families — difference target propagation, direct feedback alignment, local layerwise losses, and forward-forward — and shows… | ||
| Lightning Talk: Debugging the Undebuggable: Introducing Torch.distributed.debug | Tristan Rice | 10m |
A new PyTorch Distributed Debug Server called torch.distributed.debug runs an HTTP front end plus lightweight worker handlers to fetch C++ and Python stack… | ||
| Deploying PyTorch Models To the Browser and Beyond With Transformers.js | Joshua Lochner | 26m |
Transformers.js exports PyTorch models to ONNX using torch.export (Dynamo), applies quantization and graph rewrites with ONNX Runtime GenAI and Onnx‑Slim, and… | ||
| Lightning Talk: Distributed AI Without the Infrastructure Tax | Yahav Biran, Maen Suleiman | 12m |
Provides an actionable stack: open-source Neuron Deep Learning Containers that lock PyTorch, Neuron SDK, XLA and PrivateUse1 as a tested contract to avoid… | ||
| Lightning Talk: Inside VLLM's KV Offloading Connector: Async Memory Transfers for... | Nicolò Lucchesi | 11m |
vLLM adds a pluggable KV connector that offloads KV-cache from GPU HBM to CPU DRAM using async GPU DMA transfers, exposing a stateless, block-based API enabled… | ||
| Lightning Talk: Scaling Recommendation Systems To 2K GPUs and Beyond | Zain Huda | 10m |
Meta describes a production system that implements 2D sparse parallelism to shard massive embedding tables across hierarchical model plus data parallel groups… | ||
| Lightning Talk: Torch-Spyre: Compiling To a Multi-core Dataflow Accelerator... | D. Grove, O. Tardieu | 11m |
Torch-Spyre compiles PyTorch models to the IBM Spyre accelerator using Inductor, replacing a significant fraction of proprietary compiler code and improving… | ||
| Lightning Talk: Every Millisecond Counts: The Fine-tuning Journey of an Ultra-Eff... | Pavel Macenauer | 14m |
NXP’s flow shows an end-to-end path from PyTorch or MATLAB exports into ONNX, two open source converters (ONNX-to-TFLite, ONNX-to-Quant) that map ONNX directly… | ||
| Seamless Integration: Custom Kernels in the Torch.compile Stack Wi... | Kshiteej K, Masaki K, Pawel Gadzinski | 23m |
Transformer Engine integration with torch.compile focuses on eliminating graph breaks by registering kernels as first-class ops so Dynamo sees output metadata… | ||
| From Responses To Trajectories: Multi-Turn and Multi-Environ... | Kashif Rasul, Sergio Paniego Blanco | 23m |
Claims that post-training moves from single-response optimization to GRPO on trajectories where the training unit is multi-turn rollouts, not completions… | ||
| Lightning Talk: Building a PyTorch‑native VLLM Plugin for IBM Spyre | Thomas Parnell, Thomas Ortner | 10m |
IBM Spyre plugin is redesigned to be PyTorch‑native, using torch‑spyre PrivateUse1 device to place tensors and leverage torch.compile; the repo has ~550… | ||
| Lightning Talk: Full-Stack PyTorch Robotics VLA: From Data To... | Samet Akcay, Dmitriy Pastushenkov | 14m |
A PyTorch-native pipeline that takes teleoperation data through training to optimized edge inference, claiming an end-to-end Physical AI flow that runs on… | ||
| Lightning Talk: Trinity Large - Torchtitan on 2000+ B300s | Matej Sirovatka | 11m |
Describes training a 400B MoE with 13B active params from scratch using torchtitan, achieving a smooth loss after 17 trillion tokens and one month on 2,000… | ||
| Bridging the Hardware Gap With Code Harnesses on the Hugging Face Kernels Hub | Ben Burtenshaw | 22m |
Benchmarks compare six harnesses across ten models on CUDA and Metal, measuring agent cost, kernel latency, VRAM use and end-to-end inference performance; one… | ||
| Beyond the Theory: What Actually Breaks When You Scale Your Disaggregat... | Ekin Karabulut, Ron Kahn | 21m |
Disaggregated inference runs vLLM and SGLang on Kubernetes with Dynamo or LLMD, Kai for scheduling, and Grove or leader-worker-set for workload APIs, and… | ||
| Building Trust for Users and Regulators Alike: A Cost-Efficient PyTorch Pat... | Raja Gopal Hari Vijay | 19m |
Claims a reproducible PyTorch pattern that embeds executable regulatory controls into training and inference using existing primitives, with zero new… | ||
| Lightning Talk: Faster Than SOTA Kernels in Torch.compile With Subgrap... | Elias Ellison, Paul Zhang | 10m |
Torch.compile extensions (Inductor + Triton) implement DecomposeK for large-K matmuls and a Custom Op Autotuning pipeline that benchmarks Triton vs ATen… | ||
| Lightning Talk: Why Logging Isn’t Enough: Making PyTorch Training Regressions Vi... | Sahana Venkatesh | 12m |
PyTorch training regressions become visible by turning logs into operational checks: scheduled short and long baseline runs compared to a 30‑day average… | ||
| From Gradients To Governance: Making PyTorch Lineage-Aware | Kateryna Romashko, Clodagh Walsh | 21m |
Argues that lineage must live inside PyTorch as a runtime constraint, not just post hoc logs, and shows an integration path using Open Lineage + Marquez with… | ||
| DualPipe from Scratch: Implementing DeepSeek's 5D Parallelism in PyTorch | Dev Jadhav | 20m |
Implementation-first walkthrough of a PyTorch DualPipe reference that reproduces DeepSeek V3 at 2,048+ GPUs and explains the missing details: exact warm-up… | ||
| Sponsored Session: Fault-Tolerant Training: How We Build Rel... | Cyril Konkratenko, Maurits de Groot | 21m |
Nebius describes concrete fault‑tolerance choices and measured results: multi‑stage acceptance tests and continuous active and passive health checks (DCGM… | ||
| Lightning Talk: Ball Tracking and Detection in Soccer Videos - Comparison of... | Maciej Szymkowski | 10m |
Compares fine‑tuned YOLOv11 (250‑epoch schedule, early stop at 155) and a traditional image‑processing pipeline against on‑prem VLMs (Qwen 3, InternVL‑3.5 with… | ||
| Lightning Talk: Bridging the Gap: Engineering Compliant... | Muhammad Saqib Hussain, Mohaddisa Maryam | 09m |
A PyTorch-based Clinical Auditing System backpropagates gradients from 4D fMRI outputs into EEG bands to expose what drives predictions, aiming to satisfy the… | ||
| De-mystifying PyTorch for ASICs: When (and Why) To Move Your Development To AI A... | Alpha Romer Coma | 16m |
Benchmarks compare PyTorch/XLA (TPU) and TorchNeuron (Trainium) with side‑by‑side code diffs and real runs: fine‑tuning Llama 4, Gemma 3, Qwen 3 plus ResNet50… | ||
| nothing matches | ||