At a glance
- Did you know
- Multi-LoRA serving runs one base model for many fine-tuned adapters, batching requests for different adapters together.
- Rule of thumb
- Cap adapters per batch; swapping adapters in and out of GPU memory costs time.
- Watch out
- Adapter swaps and too many adapters per batch slow everyone.
Its settings in ScaleMySystems
What you can change on it in the lab, and what each number does.
- Concurrency (slots)
- Calls it can work on at once (per instance when autoscaling). Needed slots ≈ calls per second × seconds each takes. Extra calls wait in the queue, then are rejected.
- Queue capacity
- Calls allowed to wait for a free slot. A bigger queue absorbs bursts but adds waiting time; 0 rejects straight away.
- vCPU
- CPU cores available. Needed ≈ calls per second × CPU per request ÷ 1000. Above about 90% busy it slows down and may crash.
- CPU per request (ms)
- CPU time each call burns. Part of what the component is; make the code cheaper to lower it.
- Latency (p50) (ms)
- Typical time one call takes here, before any calls it makes itself. Slots are held for this long.
- Latency (p99) (ms)
- The slowest 1% of calls. Callers must set their timeout above this or they cut off healthy calls.
- Timeout to dependencies (ms)
- How long it waits for each call it makes before giving up. Set it above the dependency’s p99 but low enough that a dead dependency does not hold every slot.
- Error rate
- Share of calls that fail on their own (bugs, declines, flaky tools). An agent that calls it on every step multiplies this.
- Quota (req/min)
- Calls allowed per minute (an API key or provider limit). Calls over it get 429s. Needed = calls per second × 60, with headroom.
Problems that start with it
- Too many fine-tunes on one GPU One multi-LoRA server hosts 200 customer adapters on a shared base model. At 60 requests a second the shared batches overflow.
- Translation with per-language adapters A translation product serves 40 language adapters from one multi-LoRA server with 24 slots.
- Shadowing the new model doubles the load A fine-tuned candidate model is deployed as an adapter on the production GPU pool, and a traffic mirror copies 100% of live requests to it to compare answers. Production is 60 requests a second on a pool that serves about 80.
Full lesson coming soon
Lessons are being added one component at a time: how it works, its settings, a live exercise to break and fix it, interview questions and a short check. See the Circuit breaker lesson for what one looks like.