How do I package a LoRA fine-tune for local serving?
Merge the LoRA adapter into its base model, then quantize the merged model to GGUF. With the base already pulled, a2a.computer plans the merge→quantize path over governed capabilities.
from lora_adapterto ggufhave hf_dir
Available paths (live from the planner)
#11 hoploss 12%~90srisk high
lora_adapter→fuse→gguf
#21 hoploss 15%~180srisk high
lora_adapter→quantize_to_gguf→gguf
#32 hopsloss 14%~150srisk high
lora_adapter→fuse→safetensors→fuse→gguf
#42 hopsloss 14%~180srisk high
lora_adapter→merge_adapter→hf_dir→fuse→gguf
#52 hopsloss 17%~240srisk high
lora_adapter→fuse→safetensors→quantize_to_gguf→gguf
How a2a.computer knows this
The planner requires the base model (hf_dir) to be available and routes lora_adapter → merged model → GGUF. Each hop is a governed CHP capability; the merge step is huggingface.merge_adapter (merge_and_unload).
For agents
Capabilities:
chp.adapters.mlx.fusechp.adapters.huggingface.quantize_to_ggufchp.adapters.huggingface.merge_adapterMCP tools:
find_transmutation_pathsestimate_transmutationrealize_stateretrieve_evidenceLimitations & uncertainty
- –Requires the base model on hand — without it the planner returns NO_VALID_PATH, which is itself a demand signal.
- –Merging then quantizing compounds a small loss; the planner sums the per-edge loss estimates.