Open Models at Scale: Fine-Tuning, MLOps and GPU Infrastructure (AI Engineering in Practice)
English | September 25, 2026 | ASIN: B0HL1VRQ3F | 137 pages | Epub | 239.81 KB
Everyone can find a LoRA tutorial. Almost nobody can find the one document that goes from "which base model" to "what does this cost per month, forever, with an on-call rotation attached."Teams default to closed AI APIs because the open-model path is scattered across a hundred papers and blog posts with no coherent throughline. This book is that throughline: the full operational path for a self-hosted open-weight model, from picking a base to running it in production, told through one composite deployment you can audit end to end.Cost and control are the stated motivation here, not open-source ideology. Every GPU price is labeled illustrative, every model claim is checked against a primary source, and every capability claim that rests on a leaderboard score gets the skepticism a contaminated benchmark deserves. What you will learnWhen self-hosting actually wins - a five-question decision framework for volume, latency, customization, data residency and operational capacity, instead of ideologyChoosing a base model without getting fooled - reading a license that actually binds you, dense versus mixture-of-experts parameter counts, and why a benchmark score is a hypothesis, not a verdictFine-tuning that fits your budget - full fine-tuning, LoRA and QLoRA compared on what they change in the weights and what they cost in GPU memory, with a working QLoRA runData curation that actually moves the needle - deduplication, decontamination against your eval set, PII handling, and why 85,000 clean examples beat 400,000 dirty onesEvaluation you can stake a rollout on - regression against the base model, confidence intervals, memorization probes and a promotion gate that outgrows eyeballing samplesServing fast - PagedAttention, continuous batching, chunked prefill and multi-adapter serving with vLLMScaling training across GPUs - data, tensor, pipeline, sequence and expert parallelism, and when each one earns its complexityMixture-of-experts, honestly priced - DeepSeek-V3's published training economics as a real, sourced reference point, and a framework for knowing when MoE is not yet worth adoptingMLOps that treats a bad model like a bad deploy - a training pipeline, a model registry, and a rollback path with an actual paper trailThe total cost of control - a full year-one reconciliation, and the break-even calculator to run with your own numbers instead of borrowing this book's Ten chapters. One running deployment. Python that runs.For ML engineers and platform teams deciding whether to self-host - and then doing it properly.
Links are Interchangeable - Single Extraction
