State of vLLM 2025 | Ray Summit 2025 | Anyscale

CadVLM: Bridging Language and Vision in the Generation of Parametric CAD Sketches

Vivek Alamuri on Instagram: "Forget about bigger GPUs, there are a bunch of other things you can do first: 1️⃣ Use a Faster Serving Engine (Infrastructure change) 👉 Switch to vLLM or another optimized engine to leverage continous batching, PagedAttention mechanism to improve TTFT. 2️⃣ Add Streaming (Easy win) 👉 Stream tokens as they’re produced so users see output instantly. Significantly cuts perceived latency even if total generation time stays the same. 3️⃣ Quantize the Model (Model change)