Free field guide
The LLM Inference Field Guide
The reference we wish existed when the pager went off: the numbers, mental models, and checklists for running LLM inference in production. Compact enough to actually read.
- The numbers that matter — the ~dozen constants every inference engineer should know cold: bytes per KV token, memory bandwidth ceilings, kernel launch overhead, realistic batch-scaling curves.
- KV cache sizing math — worked examples you can adapt in five minutes, not a research-paper derivation.
- How to read a GPU profile without lying to yourself — the traps that make slow kernels look fast and idle SMs look busy.
- A quantization decision tree — when INT8/FP8/AWQ actually pay for themselves, and when they quietly torch your quality.
- The benchmark checklist — the contamination sources (warm caches, shared GPUs, wrong percentiles) that invalidate most published numbers.
Get the field guide
Confirm your email and the guide lands in your inbox. You'll also get new lab posts — unsubscribe keeps the guide.
Double opt-in. Unsubscribe anytime. No spam, ever.