vLLM Sessions at PyTorch Conference North America 2026
Advancements in deep learning and AI with vLLM technology
1vLLM Sessions at PyTorch Conference North America 2026
The PyTorch Conference North America 2026 will take place in San Jose, CA, on October 20-21. The conference features technical sessions, live demos, lightning talks, sponsored sessions, a keynote, and a Birds of a Feather discussion focusing on vLLM. Sessions will cover serving architecture, KV cache management, hardware portability, kernel optimization, and PyTorch integration.
2Attention Mechanism in vLLM
Lucas Wilkinson and Matthew Bonanni from Red Hat will explain how vLLM represents, serves, and optimizes the attention mechanism. The session will include approaches such as sliding windows, sparsity, compression, linear variants, and hybrid attention. Speakers will cover attention backends, KV-cache connectors, and the hybrid memory allocator.
3KV Transfer for Disaggregated LLM Serving
Nicolò Lucchesi from Mistral AI, Sunita Nadampalli from Amazon, and Zhanqiu Hu from Red Hat will discuss developments in vLLM's disaggregated serving stack for transferring KV cache between prefill and decode. Topics include hybrid-model transfer with heterogeneous tensor parallelism, bidirectional KV transfer, the KV Push connector, and KV cache leases for reliability.
4Tiered KV Cache Offloading in vLLM
Or Ozeri from IBM will present vLLM's tiered KV cache offloading framework, which routes transfers through CPU memory as a universal transport hub. The design minimizes GPU transfer overhead, consolidates I/O through a CPU buffer, avoids specialized transfer APIs, and remains independent of KV cache memory layout across hardware and model architectures.
5LMCache: LLM Prompt Caching Solution
Kuntai Du from Tensormesh, Inc. will discuss LMCache, which provides prompt caching across inference engines like vLLM, SGLang, and TensorRT-LLM, and storage systems like Mooncake, Redis, and AWS S3. The session will include a tutorial on deploying LMCache in Kubernetes.
6Model-Specific KV Cache Management in vLLM
Mengqing Cao from Huawei will address KV cache requirements that vary across model architectures such as MLA, SWA, Eagle, and DeepSeek-V4. The session will propose a model-customized KV Cache Planner built around a default planner plus model-specific planners.
7Elastic Expert Parallelism in vLLM
Itay Alroy from NVIDIA will explain how Elastic Expert Parallelism enables vLLM deployments to add or remove workers at runtime and redistribute experts across the updated worker set with minimal interruption to serving. The session will cover communication reconfiguration, CUDA Graph recapture, and expert rebalancing.
8Automatic Prefix Caching in Multi-Stage Pipelines
Ricardo Noriega and Alex Brooks from Red Hat will explore Automatic Prefix Caching for Stage Outputs in vLLM-Omni, an approach for extending vLLM's prefix caching to multi-stage models while minimizing GPU memory cost. The speakers will also describe how vLLM-Omni dynamically discovers cacheable tensors without requiring manual configuration.
9Hardware Portability and Accelerator Backends
Rob Mulla from Google will demonstrate TorchTPU's native, high-performance compilation path for PyTorch models on Cloud TPUs. The demo will highlight TorchTPU as a unified backend for inference engines like vLLM and SGLang, enabling model deployment with minimal code modifications.
10PyTorch Ecosystem on Trainium
Maen Suleiman from Amazon Web Services will cover PyTorch workflows on Trainium through TorchNeuron, including training with TorchTitan or Hugging Face Transformers v5, serving with vLLM-Neuron, profiling with Neuron Explorer, and adding NKI kernels directly to PyTorch code.
11Scaling vLLM Across Accelerators
The conference will explore how vLLM scales across different accelerators. This session will examine how vLLM is optimized on various hardware platforms and how its performance is enhanced.
12The Future and Impact of vLLM
The future of vLLM technology and its potential impact on deep learning and AI will be discussed in the final sessions of the conference. Participants will gain insights into vLLM's industrial applications and its role in research.
Original article: PyTorch Blog — PyTorch Konferansı 2026'da vLLM Oturumları