4.5. Merging Adapters

Merging bakes a trained adapter’s effect directly into a copy of the base model’s weights, so the result loads and runs like any other GGUF, with no .lora file needed alongside it at inference time.

# 1. Fine-tune
./juno lora --model-path /models/tinyllama.gguf
#   you > /train-qa What is your name? A: Juno
#   you > /save

# 2. Merge (produces /models/tinyllama-merged.gguf)
./juno merge --model-path /models/tinyllama.gguf

# 3. Run (no .lora file needed)
./juno local --model-path /models/tinyllama-merged.gguf
#   you > what is your name?
#   bot > Juno

The LoRA delta per weight element (~6x10^-4) is smaller than Q4_K quantization noise (~3x10^-3). Re-quantizing the merged weights back to Q4_K destroys the delta entirely. LoraMerge stores patched projection tensors as F32 and copies all other tensors verbatim in their original quantized form. All-linear merges expand file size substantially. The output is a valid GGUF v3 file.

Phi-3 fused-slice merge correctly patches attn_qkv at Q/K/V row ranges and ffn_up at gate/up row ranges without overwriting earlier slices.

Merge policies (--lora-merge):

Policy	Behaviour
`f32-preserve` (default)	Adapted tensors written as F32. Safe for all modes and architectures.
`source-type-projected`	Decode, add delta, re-encode with versioned `juno-kquant-v1` encoder (approximate requantization). Reports delta retention, RMSE, saturation per tensor. Use only when file size matters more than precision.
`sidecar-only`	Forbids bake-in merge; use overlay playback only.

Exact QA-LoRA zero-point merge into K-quants is not supported. Use f32-preserve or sidecar for production deployment.

Programmatic API:

LoraMerge.Result r = LoraMerge.merge(
    Path.of("TinyLlama.Q4_K_M.gguf"),
    Path.of("TinyLlama.Q4_K_M.lora"),
    Path.of("TinyLlama.Q4_K_M-merged.gguf"));

System.out.println("Patched " + r.adaptersApplied() + " tensors");
// Patched 44 tensors

See also¶