Category archive

Serving a model is a scheduling problem with a budget attached.

Latency, throughput and cost in a language model system are set by the parts around the weights: the scheduler, the cache, the gateway, and the memory budget every request competes for. These articles measure those parts and treat the model itself as the one input that is already fixed.

7 articlesInference Systems

Articles in Inference Systems

Transformer Inference, From HTTP Request to the Next Token

A systems-level tour of prefill, decode, batching, memory movement, and the scheduling decisions that determine real-world LLM latency.

KV Cache Engineering: Pages, Prefix Reuse, and Memory Pressure

How to size, allocate, reuse, compress, and evict the state that makes autoregressive decoding practical.

Semantic Caching Without Serving the Wrong Answer Faster

A production design for similarity keys, freshness, authorization, invalidation, and the economics that determine when semantic reuse is safe.

Speculative Decoding Without Hand-Waving

Draft models, acceptance math, tree proposals, and the operational details that decide whether speculation makes serving faster.

The LLM Gateway Is a Policy Engine, Not a Proxy

Designing the routing, budgets, identity, resilience, and evidence layer between products and a changing model portfolio.

Fine-Tuning with LoRA: Data, Adapters, and Production Operations

When fine-tuning is justified, how low-rank adapters work, and what it takes to evaluate, serve, and update them safely.

LLM Cost Engineering: From Token Prices to Unit Economics

Modeling full request cost, eliminating waste, forecasting margins, and optimizing without disguising quality regressions.

Apply the category

Put this against a real system.

If a decision in Inference Systems is in front of you right now, the fastest version of this is a call: bring the architecture, the failure you are seeing, and the constraint you cannot move.

Discuss the system