Skip to main content

Adapters

Adapters provide a stable interface for model I/O and enable hot-swapping specialized capabilities at runtime.

Why adapters​

  • Normalize inputs/outputs across providers
  • Swap domain-specialized behavior without retraining the base model
  • Reduce deployment friction by decoupling runtime from training artifacts

Hot-swappable adapters​

# Switch adapters in production without restart
llamafarm models swap-adapter --adapter medical_specialist --target production

Registry & versioning​

Track adapter metadata, versions, and performance.

# Register
llamafarm models register \
  --name "medical_qa_specialist" \
  --version "1.2.0" \
  --base-model "llama-2-7b" \
  --method "lora" \
  --dataset "medical_qa_v2" \
  --performance-metrics "./eval_results.json"

# List and pull
llamafarm models list --filter domain=medical
llamafarm models pull medical_qa_specialist:1.1.0

# Promote
llamafarm models promote medical_qa_specialist:1.2.0 --env production

Implementation notes​

  • Memory optimization: gradient checkpointing, mixed precision, quantization
  • Performance: batching, caching, compilation
  • Security: provenance, RBAC, audit logging