Stacked on #578 (edit: now on main)

Summary

Adds a dedicated mode to UnslothService where training and inference run on separate GPUs. Training stays in-process on the trainer GPU(s); vLLM runs as a subprocess on the inference GPU. This eliminates the sleep/wake overhead of shared mode and enables true async overlap between training and inference.

Shared mode remains the default and is unchanged. Dedicated mode is opt-in via explicit GPU ID lists in _internal_config:

_internal_config = {
    "trainer_gpu_ids": [0],
    "inference_gpu_ids": [1],
}

Architecture

Shared Mode (unchanged):
  Single GPU timeshares training ↔ vLLM via sleep/wake

Dedicated Mode (new):
  Main process (GPU 0): Training (Unsloth + PEFT)
  Subprocess  (GPU 1): vLLM server (OpenAI-compatible API)
  Communication: HTTP only
  Adapter sync: disk checkpoint + /v1/load_lora_adapter (load_inplace=true)

What's included

• Config contracts (dev/model.py, dev/validate.py): trainer_gpu_ids and inference_gpu_ids with validation (non-overlapping, contiguous from 0, single inference GPU for now)
• vLLM subprocess entry point (vllm/dedicated_server.py): Applies ART patches, enables tool calling and runtime LoRA updates, then starts vLLM
• Dedicated mode in UnslothService (unsloth/service.py): Subprocess lifecycle (start, health check, close), adapter reload via HTTP after each train step, no sleep/wake
• LocalBackend routing (local/backend.py): Detects dedicated config, sets CUDA_VISIBLE_DEVICES before model init, routes to dedicated mode
• Unit tests: Config validation + dedicated server arg building

Testing

Benchmark: ART-E to convergence with both Shared and Dedicated mode: