Model Serving

vLLM

The open source engine inside the product IBM requires for a Spyre card. Worth knowing by name, because it is the reason the same card can answer many people at once instead of queuing them.

Vendor: Open source community Licence: Open source Operating environment: Linux Deployment: Open source project, supported commercially inside Red Hat AI Inference Server

vLLM is the open source project that does the actual work of serving a model. Red Hat wraps it, supports it, and sells it as Red Hat AI Inference Server, which is the product name IBM prints in the Spyre requirements. Knowing the engine's real name is useful when you are reading documentation or hiring.

The two ideas that make it matter

The first is PagedAttention. A model holds a running memory of the conversation so far, and that memory is expensive. Reserving the worst-case amount for every user wastes most of the card. PagedAttention hands out small blocks as they are needed, which is exactly what an operating system does with paged memory. Any AS/400 person will recognise the idea immediately.

The second is continuous batching. Rather than waiting for a full group of requests before starting work, new requests join the batch while it is already running. The card never sits idle waiting for a tidy queue to form.

Why an infrastructure buyer should care

These are not academic details. They decide how many concurrent users one 128 GB Spyre card supports, and therefore whether the project needs one card or four. Two shops can buy identical hardware and get very different capacity out of it purely from how the serving layer is configured.

You can run vLLM yourself without a subscription. For a supported production configuration on IBM Power with a Spyre card, IBM names Red Hat AI Inference Server or Red Hat OpenShift AI, not a self-assembled vLLM install.