lora

Adapter math, the VRAM arithmetic that makes 4-bit a 4x unlock, and the honest decision rule for on-prem deployments. The deep-internals question Mistral and Cohere loops use to find real practitioners.

No data leaves, you cannot pre-inspect the corpus, and your toolchain arrives on a single USB. Design adaptation and evaluation when you will never observe a single input or output.

Each customer has their own LoRA on a shared 8B. The S-LoRA-style answer keeps one base in memory, pages adapters in and out, and tags every KV block with an adapter ID so reuse never crosses tenants.

Fine-tune a LoRA adapter on the user's typing on the phone's neural engine, add noise locally with DP-SGD, and aggregate only adapter deltas, so no individual's vocabulary is recoverable from the global update.