What we are looking for

Talks about vLLM itself — the engine, its internals, and what happens when you run it for real. Topics that fit:

• vLLM internals: V1 engine, scheduler, model runner, how a model actually loads
• KV cache and memory: prefix caching, KV connectors, offloading, LMCache, prefill/decode disaggregation
• Performance: continuous batching, quantization (LLM Compressor, FP8/INT4, QAT), speculative decoding, torch.compile and CUDA graphs, custom kernels
• Hardware backends: NVIDIA, AMD, Intel and Gaudi, TPU, and other accelerator plugins
• Distributed serving: tensor, pipeline and expert parallelism, prefill/decode split, llm-d, vLLM Production Stack
• Multimodal and vLLM-Omni: vision and audio serving, encoder cost, multimodal caching
• Agentic workloads: tool calling, structured output, prefix reuse across turns
• vLLM in RL and post-training: rollout engines, weight sync and hot reload
• Model enablement: adding or fixing model support, day-0 upstream work
• Benchmarking and observability: vllm bench, methodology, metrics, cost per token on live traffic
• War stories: what broke in production, what it cost, what you changed

Formats (per speaker)

• Full talk: 25 minutes plus 5 minutes Q&A
• Lightning talk: 10 minutes, no Q&A

First-time speakers

Encouraged. If you have the material but not the stage time, say so in your submission and we will help you shape it.

Code of conduct

All speakers and attendees follow the vLLM community code of conduct.