Call for Papers
Call for Papers is closed. Submissions are no longer possible. Sorry.
in 3 months
PyTorch Conference North America 2026 - POSTER SESSIONS
event starts
20 Oct 2026
event ends
21 Oct 2026
location
San Jose Convention Center San Jose, California, United States
Join us in San Jose, CA, October 20-21 for PyTorch Conference North America 2026. This two-day event hosted by the PyTorch Foundation gathers top-tier AI pioneers, researchers, and developers to explore the future of open source AI and the impact of PyTorch Foundation projects like PyTorch, vLLM, DeepSpeed, and Ray.
PyTorch Conference features in-depth technical talks, hands-on workshops, and candid conversations spanning the full AI stack, from bare metal infrastructure to applications and agent-based systems. The program features keynote sessions from leading voices in AI and practical deep dives on training, inference, GenAI, and focused tracks on responsible AI & compliance, security & privacy, frameworks & compilers, and more.
finished 7 days ago
Call for Papers
Call opens at 12:00 AM
31 Mar 2026
Call closes at 11:59 PM
26 Jul 2026
Call closes in Pacific Daylight Time (UTC-07:00) timezone.
Closing time in your timezone () is .
CFP GUIDE
Please review our CFP Guide to answer many common questions before submitting.
PLEASE NOTE - Only an abstract is required with your Poster submission. If selected to present a poster, you’ll receive additional details about printing requirements and event logistics by email.
DATES TO REMEMBER:
- Poster Session CFP Close: Sunday, July 26 at 11:59 pm PDT (UTC +-7)
- Poster Session Notifications: Monday, August 17, 2026
- Poster Sessions Announced: Tuesday, August 18, 2026
- Event Dates: Tuesday, October 20 - Wednesday, October 21, 2026
Reminder: This is a community event — so no product and/or vendor sales pitches.
CODE OF CONDUCT
By submitting, you agree to The Linux Foundation's Code of Conduct.
COMMITMENT TO INCLUSIVITY
Please review The Linux Foundation's Inclusive Speaker Orientation and Inclusive Language Initiative.
Please be mindful that The Linux Foundation does not allow any talk with 3 or more speakers to only have men participating in an effort to increase equity & inclusion.
PRIVACY POLICY
At The Linux Foundation, we are committed to safeguarding your privacy. Sensitive speaker information will only be accessible to event organizers and program committee members who adhere to the highest confidentiality standards. Rest assured that your information will never be sold or shared beyond these parties.
Speaker personal information including name, company, job title, biography and photo will appear on the public schedule. For information on our privacy practices and commitment to protecting your privacy, please review our Privacy Policy.
You can update or delete your account information at any time through your Sessionize profile account settings. Please contact support@sessionize.com directly for any questions or issues.
QUESTIONS?
Question about submitting a proposal? Contact us at cfp@linuxfoundation.org.
all submitted sessions
publicly listed on this page
292 submissions
Submitted sessions
Angel Li
- Resolving Numerical Differences in Attention for RL
- Agentic Triton Kernel Development: Using AI Agents to Accelerate GPU Kernel Optimization
- Scaling LLM Inference with Megakernels
- Network fault tolerance for AI workloads
- Fractional-Order Memory in Single-Cell Habituation
- Neptune Serving Engine: General Purpose Inference
- KernelFoundry: Hardware-Aware Evolutionary GPU Kernel Optimization
- Trident: Compiling Triton Kernel Launch Paths from Python to LLVM via MLIR-Based JIT Compilation
- Agentic Kernel Optimization at the Scale of the Web
- Koi: A Self-Learning Planner for LLM Inference on Heterogeneous Multi-Cloud GPUs
Pratinav Seth, Zera Lyngkhoi
- AlignTune: One Workflow for LLM Post-Training
- Rebuilding the Build: PyTorch's Move to a Standards-Based Packaging Backend
- Building On-Device AI Agents on Exynos with ExecuTorch
- The Foundation Effect: How Umbrella Governance Reshapes vLLM and DeepSpeed
- Fenceline: a self-hosted prompt firewall for locally served open-weights models
- Understanding Distributed Checkpointing
- Don't Tune It Yourself: Autotuning GPU Kernels with Helion
- Safe Checkpoint Provenance using Sigstore + Safetensors
- Profiling CUDA Memory Fragmentation in Dynamic Shape PyTorch Workloads
- Characterizing Graph Breaks in torch.compile Across Production Vision Models
- Understanding KV Cache Memory Behavior in vLLM for Long Context
- Static Security Analysis of PyTorch Checkpoints for Safe Model Deployment
- Firetap: Model Internal Telemetry for Large-Scale ML Training
- Physical AI on the Edge: Adapting and Deploying VLMs with ExecuTorch on Qualcomm HTP
- Spatial Reasoning for Edge Robotics Using Parameter-Efficient Vision-Language Models on Ventuno Q
- From Simulation to Edge AI: A PyTorch Sim2Real Deployment Workflow for the Arduino Ventuno Q
- From PyTorch to Edge: Accelerating GenAI Inference with ExecuTorch on Qualcomm NPUs
- On-Device LLM Inference on Android with ExecuTorch and Qualcomm QNN
- Memory Management in CUDA Graphs
- From Hugging Face to Edge: Enabling Native PyTorch Model Onboarding with Optimum–ExecuTorch
- MoRI + vLLM: Wide Expert Parallelism and RDMA KV-Cache Transfer for Disaggregated MoE Serving on AMD
- Performance and optimization of distributed MoE on Arm-CPU clusters
- ExecuTorch and Monarch powered Snowflake CoCo applications
- Smart routing for packing many model lifecycle functions such as training, RL, inference,serving
- Red-Teaming Self-Hosted Agents: What Each Injection Defence Stops, Breaks, and Costs
- TorchTPU Agent Suite Automating PyTorch Model Migration and Benchmarking on Cloud TPU
- Bringing vime RL Post-Training to ROCm: End-to-End Training and Rollout on AMD Instinct
- Seal Forward, Diagnose Backward: Signed Evidence Around a PyTorch Inference Step
- Reliable and High-Availability LLM Serving with In-place Failover and Elastic Recovery
- Accelerating PyTorch LLM Inference by Reducing Tensor Layout Overhead
- TCCL: Low-latency Communication Backend for Apple Silicon
- Composable Kernel HSTU Attention Kernel for Generative Recommenders
- Retrieval-Augmented Autotuning: Navigating the Quality–Readiness Frontier with Historical Evidence
- Integrating RCCL GPU-Initiated Networking into TorchTitan’s MoE Communication
- GraphTrainer on AMD Instinct: Bringing a New Distributed-Training Paradigm to Eager Parity
- Extending OpenSHMEM for AI and MoE Communication Patterns
- Closing the Source-to-Kernel Provenance Gap for Out-of-Tree PyTorch Accelerators
- From Bottlenecks to Autonomy: Profiling RLHF Pipelines with PyTorch TorchTItan
- Scaling PyTorch Validation with Neuron as the success story
- From FlashAttention to DeepGEMM: IO-Aware Kernel Design for Inference workloads
- When Acceptance Rates Are Not Enough: An Inference Systems Perspective on Speculative Decoding
- Integrating Out-of-Tree Accelerators with PyTorch Profiler
- "Are We There Yet?" (AWTY): A Framework for Assessing Small Model Viability on Local Hardware
- Five Bits on Metal: Adding Q5_K Quantization to ExecuTorch's MLX Backend
- TorchTitan Performance Optimizations
- Beyond Registration: What It Actually Takes to Be a First-Class PyTorch Backend
- The Graph Break You Didn't Expect: Making torch.compile and FSDP2 Work Together
- Building Reproducible ML Infrastructure for Digital Pathology with PyTorch
- One Policy, Two Engines: Closing Training–Rollout Drift in DeepSeek-V4 Flash RL
- From MP4 to Tensor: Optimizing Multi-Camera Video Data Loading for PyTorch Training
- TorchTitan GraphTrainer: LLM Distributed Training with PyTorch’s Compiler Toolkit
- Scaling PyTorch Validation for a Multi-Accelerator Future
- Connectionless Scale-Out EP: GPU-Initiated MoE Dispatch/Combine for vLLM
- Getting Cozy in the Scratchpad: Compile Time Memory Planning for the Spyre Dataflow Accelerator
- Zero-Code Model Portability: Running HuggingFace Transformers & Diffusers on Any Chip via Torch-FL
- One PrivateUse1, Many Chips: A Virtual Device Layer for PyTorch
- From MCP to On-Device AI: Building Multimodal Agents with SafeTensors and ExecuTorch
- Trust Before Inference: Verifiable Multimodal Model Deployment with SafeTensors and ExecuTorch
- From Model Integrity to Runtime Integrity: Trusted MCP Agents on ARM Edge Devices
- vLLM IR: Graph Compiler Intermediate Representation for Fast LLM Inference
- Upgrading CPU Preprocessing Modules from TorchScript to PT2: IR Design, Developer Experience, and Pr
- Triton-Based Operator and Compiler Optimizations for LLM Inference Across AI Accelerators
- Accelerating Modern Attention Kernels with Triton-TLE
- Testing the Edge: Scalable Validation for ExecuTorch
- Neuron - PyTorch Native Backend for Trainium
- Neuron Native PyTorch Backend for Trainium
- Portable Semantics, Specialized Kernels: PyTorch FlexAttention in ONNX
- Evaluating Local INT4 SLM Judges on Apple Silicon
- Production FP8 Training for Multimodal Transformers in PyTorch
- FlexBudget: Interchangeable KV Importance Signals under One Compiled FlexAttention Budget
- The Agent Roofline: When GPU Offloading Helps (and Hurts) in LLM Agent Runtimes
- Scaling TorchTitan on Intel XPU Architecture : Very Large MoE LLM Training at 10K Scale.
- AutoBalance: A Self-Improving Agent Architecture for Imbalance Mitigation in Distributed Execution
- TorchTitan on Intel XPU: From Desktop to Supercomputer - A Guide to Large-Scale LLM Training
- From Client Communications to Competing Risks: A PyTorch Transformer for Client Attrition Model
- Fidelity Over Capability: Distilling a Paid LLM Judge from Its Public Verdict Archive
- Reward Hacking in Long-Horizon LLM Agents: A Health-Gated Reward for an OpenEnv Benchmark
- The Grader Looked Fine Until the Model Learned to Hack It
- Serving Interpretability Probes on Open Models
- AugVein: Vein Detection through Augmented Reality with PyTorch
- Phalanx: Adaptive Kernels for Sparse Retrieval Encoding
- When Does Offloading FSDP's Reduce-Scatter to a DPU Actually Pay Off?
- Hardware-Accelerated Strided KV-Cache Transfers for Asymmetric LLM Inference
- MAKCI: A Multi-Abstraction AI Kernel Cost Instrumentor for Performance Modeling in AI Compiler
- Semantic Tensor Contracts: Detecting and Localizing Train–Serve Corruption in PyTorch
- Evaluating TorchComms Adoption in vLLM on XPU
- Tuning the Compiler, Not Just the Kernel: Bespoke -O3 for your PyTorch Helion GPU kernels
- When Should I Disaggregate? Portable Diagnostics for LLM Inference on vLLM and llm-d
- A Remote Autotuning Service for Helion: Async, Parallel Tuning on Production Hardware
- 85% of the Time Was Preprocessing: A Vendor-Portable GPU Kernel for Real-Time ICU Monitoring
- The Same Kernel Bug Costs 10x More on AMD: Vendor Asymmetry in Protein-Binding Code
- From 129 Seconds to 79: Fused GPU Kernels for Whole-Slide Pathology Foundation Models
- ERRC: Entropy-Reinvested Residual Correction for Tensor-Parallel LLM Communication
- Serving Video Diffusion on Trainium by fusing Communication, Attention, and Quantization
- Layout-Aware Compilation Cache Validation for Custom AI Accelerators in PyTorch
- No Hardware? No Problem : A Compiled-Artifact Execution Layer for Out-of-Tree PyTorch Backends
- Batch Cost Lookahead: Coordination-Free Workload Regularization for Packed Variable-Length Training
- Step-Aligned Telemetry for Diagnosing Distributed Training Slowdowns
- Pytorch Legate Interoperability through DLPack
- Inference Performance Analysis using TraceLens
- What actually governs edge inference on a Raspberry Pi 5
- A 3D U-Net Surrogate for CFD Prediction in Industrial Enclosures
- TraceWeave: Eliminating Weight Reloads for Faster LLM Inference
- Bringing High-Performance Inference Serving to Trainium with an Out-of-Tree vLLM Plugin
- ExecuTorch WebGPU: PyTorch's First GPU-Accelerated Inference in the Browser
- Improving the torch.compile user experience
- Telemetry-Complexity Aware Adaptive Routing for Agentic Workloads on Heterogeneous Hardware Systems
- Neuron Native PyTorch Runtime - An Asynchronous Execution Runtime
- Efficient EPD Architecture to Accelerate VLM Performance on AWS Trainium
- Agents as Distributed Processes: A Native PyTorch Design Pattern for Multi-Agent Orchestration
- Gradient Ghosts: Real-Time Numerical Fault Detection for PyTorch-Native Training
- AI Edge Quantizer: Streamlining PyTorch Model Optimization for LiteRT On-Device Inference
- Race-Free TorchInductor Artifact Reuse for Elastic Training
- GenieX: A Dual-Runtime PyTorch Inference Backend for NPU/GPU/CPU Deployment on Snapdragon
- Day-0 RL at Scale for Emerging LLM and VLM Architectures
- From Domain Documents to Custom Embeddings: A PyTorch Recipe for Better Retrieval
- Regional-AOTI: A PyTorch-Native Partial Compilation Framework For High-Performance Inference
- Multi-Teacher On-Policy Distillation for SOTA Agentic AI models
- Accelerating Whisper Inference on Arm CPUs through PyTorch Ecosystem Optimizations
- Flint: Compiler Enabled Cluster-Free Design Space Exploration for Distributed ML
- Day-1 Cooperative Matrix Support in ExecuTorch: Accelerating On-Device Inference on Mobile GPUs
- Beyond export: building Raspberry Pi camera apps from PyTorch vision models with ExecuTorch
- From Cloud to Device: A Guide for Deploying AI Models on Embedded Systems Using Quantization
- Bridging the Edge Deployment Gap: High-Performance PyTorch with Torch-TensorRT and ExecuTorch
- Hypothesis-to-Skill Loop for Agentic SYCL Kernel Optimization
- Empirically-guided Tensor Placement in AutoParallel
- Hospital Bed Assignment: A PyTorch-Forecast and LLM Decision Support Architecture
- Turning Device Backend Bring-Up into Upstream-Tested Confidence: A PrivateUse1 Journey with Spyre
- Reinforcement Learning for LLMs with Intel XPU and PyTorch on ALCF Aurora
- Automatic Graph-Based Activation Checkpointing and Offloading for TorchTitan Frontier Model Training
- Beyond the Python Runtime: AOT-Compiled Expert-Parallel MoE for Edge Deployment in TensorRT
- The Power of Graph-level Optimizations for LLM Inference on Edge CPUs: Efficient SDPA with YNNPACK
- RAFT for API Code Generation Under Distractor Noise
- Beyond the Execution Layer: Integrating Custom Accelerators with vLLM
- Efficient Segmented LoRA Inferencing on Intel GPUs
- DLBacktrace: Tracing How AI Models Reach Their Decisions
- CuratorKIT: Building Better LLM Training Data from Raw Documents
- TabTune: One Interface for the Tabular Foundation Model Lifecycle
- SafeTune: Auditing and Repairing Safety Drift After Fine-Tuning
- CircuitKIT: From Circuit Discovery to Model Intervention
- Helion Beyond the GPU: A Qualcomm Hexagon Backend for PyTorch
- Cross-Language Compilation in Triton: From CUDA Source to Distributed GPU Kernels
- Fail Fast or Restart? Automatically Triaging Distributed PyTorch Training Failures
- Efficient Data Representations for Large-Scale PyTorch Recommendation Training
- Scaling PyTorch Recommendation Models with Request-Level Training
- Grouped GEMM on AMD ROCm in PyTorch
- An Empirical Study Of torch.compile across Inference Engines
- veRL: Extreme Optimization Practices for Ultra-Large-Scale MoE Models
- vLLM Ascend: Production-Grade Inference on Huawei Ascend NPUs
- Support Helion Flow to MLIR Linalg for RISC-V Vector Extension
- Beyond SPMD: Compiling Triton Kernels for Dataflow AI Architectures
- Multi-Chip KV-Cache Transfer for Disaggregated vLLM Serving with FlagCX
- FlagCX: Portable Device-Side Communication Primitives for Triton Kernels on Any Accelerator
- Solving Tail-Latency Inflation in Production Vision-Language Serving with Virtualized GPUs
- Resource-Efficient Secure and Privacy-Preserving LLM Inference via GPU Virtualization
- Triton-Ascend: Simplifying NPU Kernel Programming and Enabling the Open-Source Kernel Ecosystem
- Session-Aware Routing for Local Agentic Workflows: Mixture-of-Models on Intel Workstation GPU
- Scaling Satellite Object Detection: A PyTorch Geospatial Pipeline on Intel XPU
- Scaling LLM Inference Beyond Accelerators with CPU Clusters
- StageFrontier: Always-On Stage Accounting for Distributed PyTorch Training
- Which Quantization Scheme Should You Actually Use? A Decision Tree for INT8, FP8, GPTQ, and AWQ
- From Training to Your Pocket: Mapping the PyTorch Ecosystem End-to-End
- What Privacy Actually Costs: An Accuracy-vs-Epsilon Curve for LoRA Fine-Tuning
- Sandboxing Agentic Tool Calls: An Isolation Architecture for Untrusted Execution in Production
- Your Agent Passed the Benchmark. Then Production Broke It.
- Offloading KV Cache to Secondary Memory: A Three-Tier Hierarchy Emulator using vLLM
- Resilient LLM Serving with Shared Secondary Memory: Fast KV-Cache Failover Across Deployment
- Where the Blocks Live: Tier-Aware KV-Cache Routing for vLLM at Scale
- Using HugePages with PyTorch for Faster LLM Inference
- EvoTuner: Evolutionary Search with a Learned Cost Model for Triton Kernel Autotuning
- TorchInductor for SGLang Inference on RISC-V CPUs
- Enabling APC for hybrid models in vLLM
- Transparent Pipelined Compression for PyTorch Inference via a NIXL Utility
- FlagRelease: A Self-Verifying Agent Platform for Model Migration Across Heterogeneous Accelerators
- Fault-Tolerant PyTorch FSDP at Scale: Benchmarking NVRx
- KernelGen: A Self-Evolving AI Kernel Engineer for PyTorch Across Heterogeneous Accelerators
- TLE: Compiling End-to-End LLM Inference into Triton Megakernels with E-Graphs
- SGLang-Plugin-FL: A Unified Multi-Chip Backend for SGLang Inference
- What Should SRE Watch? Calibration-Gated Decisions and the Missing Golden Signals for GPU Clusters
- New Activation Memory APIs in PyTorch
- Regional Inductor: Selective Inductor Compilation within torch.compile
- More Than Meets the Eyes: Mirroring CPython's Object Protocol in Dynamo
- Warm Standby, Cold Start: Multi-Region Resilience for PyTorch Inference on Kubernetes
- Train More with Less: Fewer GPUs, Less Data Movement, Better Cluster Utilization
- Choosing BF16, FP8, or TurboQuant: A PyTorch KV Cache Guide for Memory‑Constrained GPUs
- Private Pool Memory Visualization for CUDA Graphs
- Efficient, Large-Scale LoRA and Fine-Tuning for Diffusion Models in PyTorch
- Beyond Autoregression: Fine-Tuning Diffusion LLMs at Scale in PyTorch
- Incremental Neighbor Sampling for GNN Training on Continuously Growing Graphs
- Power-Aware Profiling: Bringing Arm Telemetry into PyTorch for Efficient Cloud–Edge Inference
- Real Shapes, Real Bugs: Model-Derived Operator Tests for Quick Out-of-Tree Accelerator Bring-Up
- Bridging PyTorch Export to Edge GPU Backends: Enabling High-Performance grid_sample with ExecuTorch,
- FP8-RL: A Practical and Stable Low-Precision Stack for LLM Reinforcement Learning
- High-Performance CPU Inference in PyTorch: TorchInductor, oneDNN, and vLLM Ecosystem Integration
- From FX Graph to Silicon: Building a Full-Stack PyTorch Backend for Custom AI Accelerators
- Production Stability for vLLM: Elastic Scaling and Fault Tolerance Mechanisms on Ascend Platform
- FlagQuantum: A PyTorch-Native Runtime for Differentiable and Distributed Quantum Computing
- OSU HPC-AI: Distributed Pre-training, Reinforcement Learning, and Inference with MPI Communication
- Graphing the Ungraphable: Making MoE Training CUDA-Graphable in PyTorch
- Dynamo Snapshot: Autoscaling and Failure Recovery in Seconds Using GMS
- Heterogeneous CPU + GPU EPD Disaggregation to Boost VLM Serving
- Better Together: Get More Concurrent Users with Heterogeneous Multi-Agent Workload optimization
- Beyond the Learner: The CPU’s Role in Scaling PyTorch Multi-Agent Reinforcement Learning
- Beyond the Learner: The CPU’s Role in Scaling PyTorch Multi-Agent Reinforcement Learning
- Beyond the Learner: The CPU’s Role in Scaling PyTorch Multi-Agent Reinforcement Learning
- Inductor-TV: Formal Methods for the Pytorch Compiler
- Optimizing Streaming ASR on Arm Edge Devices with ExecuTorch
- Elastic Endpoint P2P Transfers for Inference with NIXL
- From PT2E to Arm/TOSA: Dynamic W8A8 Linear Quantization in ExecuTorch
- X-Raying a Delegate: A Debugging Toolkit for Delegated Inference in ExecuTorch
- Two GPU Delegates, One Vulkan Context: Mixed VGF–Vulkan Execution in ExecuTorch
- Shape-Stable Dynamic Control Flow in PyTorch CUDA Graphs
- Intel Client GPU Expands PyTorch Use Cases: Enabling Robotics AI on iGPU and dGPU
- Debuggable and Profileable Intel GPU Software Stack for AI Computing with PyTorch
- AOTI with CUDA Graph
- Tensor KV Cache: A PyTorch-Native Approach to KV Cache Offload
- Inside TokenSpeed-Kernel: Portable, High-Performance LLM Inference Kernels Across GPUs
- Accelerating Helion Autotuning with Warm-Start and Shared Caches
- VLAs for Autonomous Driving: Inference Optimization and Production Lessons
- Resilient PyTorch Distributed with NCCL Shrink and Revoke
- Bringing SmolLM2-135M to Resource-Constrained Arm Edge Devices with ExecuTorch
- ASystem: An Open-Source Reinforcement Learning System for Foundation Model Post-Training
- Retrieval as Inference: Building a Production-Scale Torch-Native Retrieval Engine
- Scaling PyTorch Knowledge Through Community: Lessons from PyTorch Korea
- Encryption Hides the Data, Not the Agent: Trace-Privacy for PyTorch LLM Agents
- 2 in 1: Deep Dive into PyTorch Certified Associate (PTCA) and PyTorch Ambassador Program
- From PyTorch to Edge: An Open-Source Workflow for Converting, and Optimizing On-Device Models
- AI Power Efficiency Index (APEI): Framework for measuring Per-Token CO2 Emissions in Frontier LLMs
- TraceLens Agent: Automated Performance Bottleneck Identification from PyTorch Profiler Traces
- Efficient Inference of Masked Diffusion Language Models on AMD XDNA2 NPU
- Multimodal Sensor Fusion for AI-Powered Robotics: Advancing Real-World Perception
- Human-in-the-Loop Requirement Classification for Enterprise API Workflow Generation with PyTorch
- Perch 2.0 and Perch-Hoplite in Pure PyTorch: An End-to-End TensorFlow-Free Bioacoustics Pipeline
- Scaling Multi-Model PyTorch Workloads for Closed-Loop Robotics Evaluation
- TorchTPU x Hugging Face: Native TPU Acceleration Across the HF Ecosystem
- Piecewise CUDA Graphs: Eager Breaks in CUDA Graph Capture for PyTorch LLM Inference
- Don't Just Trust the Model, Test the Physics: Evaluating PyTorch Models with PhysicsNeMo-CFD
- cuxray: Optimize Kernels Without a GPU
- Exploiting CUDA Locality Domains in PyTorch on Blackwell and Beyond
- Dynamic Geometric Computing: Stress-Gated Visible Feature Layers for Hidden-Regime Time Series
- LLMs & Natural Language Processing – Unlocking the Future of AI Communication
- Building Global Allies: How Male Mentors Can Accelerate Women in AI & Startups
- Your GPU Is Starving: The Hidden Cost of Loading LLMs at the Edge
- Attested Inference: a signed, independently-verified receipt for every model output
- AI-Driven Supply Chain Intelligence with PyTorch-Based Predictive Models
- Demystifying Nondeterministic PyTorch Model Edge Inference with the Google AI Edge Debugger
- Adaptive Consensus for Low-Latency Distributed Systems
- Continuous Causal Inference for Real-Time Advertising Measurement
- Quantize-Then-Refine: Two-Stage Scoring for Memory-Efficient GPU Retrieval
- From Scheduled to Real Overlap: All-Gather/Reduce-Scatter in PyTorch Distributed
- Moral Tensors and DecisionProofs: Compiling Language into an Auditable, Grounded Safety Layer
- Atlas + Erebus: A Personal AI Research Platform on Two Prosumer GPUs
- VLA Models on Intel Arc Pro: Enabling Real-Time Robot Control with PyTorch XPU and OpenVINO
- Stage-Specialized E-PD Disaggregation for SLO-Aware Multimodal Serving on Intel Arc Pro B50/B70
- AI in Analytics
- From PyTorch to a Phone: What Quantizes Cleanly On-Device, and What Doesn't
- Importance Aware Attention for Faster PyTorch Inference on Arm CPUs
- Training Embedding Models Resiliently for Multimodal Model Inference Routing
- Real-Time Ranking Systems for E-Commerce Search and Discovery
- Before the Loss Spike: Training Alignment for Precise PyTorch Debugging at Scale
- vLLM-plugin-FL -- Multi-Chip vLLM framework
- Scaling Large-scale Distributed Training with DeepSpeed using Muon Optimizer
- KernelOPT: An Open-Source Multi-Agent Harness for GPU Kernel Optimization
- A Multi-Layer Profiling Toolkit for Out-of-Tree PyTorch Accelerators
- REBAR: Routing Extensions for Burst-Aware RL
- Efficient MoE LLM Inference on Arm with vLLM and OpenVINO
- Maestro: Benchmarking Concurrent GPU Operations as They Run in Real Workloads
- A Practical Taxonomy of LLM Inference Bottlenecks: Prefill, Decode, Memory, and Scheduling
- Improving PyTorch Efficiency on ARM with SVE-Accelerated Memory Primitives
- weights_only Was Supposed to Save You: PyTorch Model RCE in 2026 and Safe Loading
- Escape Hatches for torch.compile
- CUDA Graph on Large Scale Recommender Systems
- Train Across Your Entire GPU Fleet: Heterogeneous Mixed-Generation Training in Pure PyTorch
- "Powering AI/ML with Python and Apache iceberg"
- Breaking torch.distributed on Purpose: Chaos Engineering for Large-Scale PyTorch Training
- AllReduce, AllGather, or AllToAll? How Parallelism Shapes PyTorch Network Traffic
- TPU Model Performance Auto-optimization
- Diagnosing LLM Streaming Bottlenecks: Profiling TTFT, Decode Throughput, and Inference Telemetry
- Building Self-Healing Infrastructure for AI Agents
- The 11-Minute Cold Start: A Visual Anatomy of Why GPU Pods Take Forever to Become Useful
- State Persistence in Containers
- The health benefits of a cheese based diet