DAPT adapter router

Overview

A small service that answers a question with a base language model alone or with one domain-adaptive pretraining (DAPT) LoRA adapter attached. Every request names its target: the base model or an adapter id. The real setup is Qwen3.8-27B with trained adapters; until those exist, a 0.8B model of the same family runs the full path on a CPU.

How it works

Adapters are standard PEFT folders, and a config file lists each one with its id, its path and the base model it expects. The service runs its startup checks and loads the model before it accepts a request, then serves one request at a time. On the command line the answer goes to stdout, while the adapter report and a log line with the target and the latency go to stderr. Over HTTP, a POST to /ask returns the answer, the target and the latency. An unknown target stops with exit code 2 on the command line and HTTP 400 on the API, and both list the valid targets.

Testing

A script writes two development adapters with non-zero weights, so each one really changes the model’s output. The tests use them on a CPU, need no network and finish in about a minute.