Adapter Registry & Versioning

With hundreds of adapters in production, you need infrastructure to manage their lifecycle — training, evaluation, deployment, and retirement.

1. The Adapter Lifecycle

Train → Evaluate → Register → Deploy → Monitor → Update/Retire

Phase	Artifacts	Actions
Train	Adapter weights, config, training logs	LoRA/QLoRA training
Evaluate	Benchmark scores, human eval results	Automated + manual quality gates
Register	Metadata (base model, rank, training data hash)	Version and tag in registry
Deploy	Adapter loaded in serving system	A/B test with traffic splitting
Monitor	Quality metrics, latency, user feedback	Automated regression detection
Update	New adapter version	Retrain on updated data

2. Registry Design

# Example adapter registry entry
{
    "adapter_id": "medical-v3",
    "base_model": "llama-3.1-70b",
    "base_model_hash": "abc123...",     # ensure base model compatibility
    "lora_rank": 16,
    "lora_alpha": 32,
    "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj", 
                        "gate_proj", "up_proj", "down_proj"],
    "training_data_hash": "def456...",  # reproducibility
    "training_steps": 5000,
    "eval_scores": {
        "medqa": 0.89,
        "mmlu_medical": 0.91,
        "general_mmlu": 0.82,           # monitor for forgetting
    },
    "status": "canary",                  # canary → production → deprecated
    "created_at": "2024-10-20T12:00:00Z",
}

Version Compatibility

Critical: An adapter trained on Llama-3.1-70b cannot be used with Llama-3.1-70b-Instruct (different weights). The registry must track exact base model hashes.

3. A/B Testing Adapters

# Traffic splitting in inference router
def route_request(request):
    customer = request.customer_id
    adapter_config = get_adapter_config(customer)
    
    if adapter_config.ab_test_active:
        # 90% to current adapter, 10% to new candidate
        if hash(request.id) % 100 < 10:
            return adapter_config.candidate_adapter
        else:
            return adapter_config.current_adapter
    
    return adapter_config.current_adapter

Metrics to compare: quality scores, user satisfaction, latency, cost per request.

For the serving infrastructure that hosts these adapters, see Multi-Adapter Serving.

Self-Check

Design an adapter registry schema for a production system
Explain why base model hash tracking is critical
Set up A/B testing between adapter versions
Define quality gates for promoting an adapter from canary to production

1. The Adapter Lifecycle

2. Registry Design

Version Compatibility

3. A/B Testing Adapters

Self-Check

References