Welcome to vLLM office hours! These bi-weekly sessions are your chance to stay current with the vLLM ecosystem, ask questions, and hear directly from contributors and power users. This week's special topic: llm-d Project Update and Wide EP for Agentic Workloads vLLM project update: DSpark, DeepSeek's new speculative decoding method, now running in the vLLM nightly ahead of v0.25. Plus the v0.24 release itself, with Minimax M3 optimizations, continued sparse MLA performance work, Model Runner V2 expansion, dynamic speculative decoding, and Wide EP improvements. LLM Compressor update from maintainers Charles and Brian: faster turnaround publishing new models, DDP multi-GPU support for faster quantization, and a new converter abstraction for checkpoint format conversion. Then we dive into llm-d with Robert Shaw (vLLM Core Maintainer and co-lead of llm-d, Red Hat AI): why agentic workloads need distributed optimization beyond a single vLLM node, prefix cache-aware routing, KV cache tiering, prefill/decode disaggregation, and Wide EP for scaling large MoE models like GLM-5.2 across nodes. Includes a live walkthrough of serving GLM-5.2 on H200s, with real performance and cost numbers. Slides: https://docs.google.com/presentation/d/1ptF7tsojgPo7HQuR9SOnTV2VSsSZ6jfByZGwItzYz8M/ Want to join the discussion live on Google Meet? Get a calendar invite by filling out this form: https://red.ht/office-hours Timestamps: 00:00 vLLM intro 01:22 vLLM v0.24 project update 16:52 LLM Compressor update 23:40 llm-d deep dive: Wide EP for agentic workloads 49:07 Q&A

 320      19