Minicpm5 Finetune Gguf Lora

Fine-tune MiniCPM5-1B → GGUF LoRA adapter

The framework skills (minicpm5-finetune-*) all emit a PEFT adapter (adapter_model.safetensors + adapter_config.json). But llama.cpp / llama-server — and the MiniCPM Desk Pet app's custom-LoRA upload — load a GGUF LoRA adapter (--lora some-adapter.gguf). This skill is the bridge: train a PEFT LoRA, then convert it to GGUF and run/upload it.

This is the path you want when the base model is served as GGUF (Desk Pet, Ollama, LM Studio, plain llama-server). If you serve the fp16 HF base with vLLM / transformers, you don't need GGUF — load the PEFT adapter directly.

Pipeline overview

  train (any minicpm5-finetune-* skill)        this skill
 ┌────────────────────────────────────┐   ┌──────────────────────────────────┐
 BASE (fp16 HF) ─► adapter_model.safetensors ─► convert_lora_to_gguf.py ─► adapter.gguf
                   adapter_config.json                                          │
                                                                  llama-server --lora / Desk Pet upload

Required input

Don't have an adapter yet? First run a training skill — start from the router minicpm5-finetune (or go straight to minicpm5-finetune-unsloth for single-GPU LoRA). Come back here with ADAPTER_DIR pointing at its output.

Steps

1. Get llama.cpp (has the converter)

git clone --depth=1 https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
pip install -r requirements.txt          # converter deps: torch, safetensors, gguf, transformers

convert_lora_to_gguf.py lives at the repo root. You only need the Python script for conversion; build the C++ binaries (step 4) only if you want to test locally.

2. Convert the PEFT adapter → GGUF

python convert_lora_to_gguf.py "$ADAPTER_DIR" \
    --base "$BASE_MODEL" \
    --outtype f16 \
    --outfile "$OUT_GGUF"

• --base only needs the base model's config (config.json, tokenizer.json) — not the weights. A local fp16 HF dir or an HF id both work.
• If the base is only on the Hub, use --base-model-id openbmb/MiniCPM5-1B instead of --base (it pulls just the config).
• Output is a single .gguf of roughly the same size as the input .safetensors (a 22 MB r=16 adapter → ~22 MB GGUF).

🔑 The #1 gotcha — base_model_name_or_path points at the training machine. PEFT writes the absolute path of the base used during training into adapter_config.json (e.g. /user/.../MiniCPM5-models-fixed/official). On any other machine the converter can't find it. Always pass --base (or --base-model-id) explicitly to override it — don't rely on whatever is baked into the config. The base you pass MUST match the one you trained on, or the adapter math is meaningless.

3. (Optional) Sanity-check the GGUF metadata

python -c "import gguf,sys; r=gguf.GGUFReader('$OUT_GGUF'); print('tensors:', len(r.tensors)); print('arch:', r.get_field('general.architecture').parts[-1].tobytes().decode() if r.get_field('general.architecture') else '?')"

Expect general.type = adapter and a nonzero tensor count. MiniCPM5-1B is a Llama-architecture model, so the converter treats it as llama — this is correct, not an error.

4. Test with llama-server before shipping

You need a GGUF base model too (the adapter is applied on top of it). Grab the released base:

huggingface-cli download openbmb/MiniCPM5-1B-GGUF MiniCPM5-1B-Q8_0.gguf --local-dir .

Then:

# CLI
llama-cli -m MiniCPM5-1B-Q8_0.gguf --lora "$OUT_GGUF" \
    -p "你好" -n 128 --temp 0.7 --top-p 0.95

# OpenAI-compatible server
llama-server -m MiniCPM5-1B-Q8_0.gguf --lora "$OUT_GGUF" --port 8080 --jinja

If the persona/behavior you trained shows up, the GGUF is good. If output is identical to the base, the adapter didn't load (check the --lora path and that --base matched in step 2).

Quant compatibility: a GGUF LoRA built from an fp16 adapter applies fine on top of a quantized base (Q8_0 / Q4_K_M). You do not need a separate adapter per base quant.

5. Use it in MiniCPM Desk Pet

The Desk Pet app loads exactly this kind of file. The fp16 GGUF you just built is upload-ready:

1. Open Settings → MiniCPM → 适配器 (LoRA).
2. Click 上传 (Upload), pick your .gguf — the app copies it into its <userData>/adapters/uploads/ directory and registers it.
3. Give it a display name + comma-separated aliases (used so you can switch personas by voice/chat, e.g. "换成 XX").
4. Select it in the adapter list to activate. The sidecar reloads llama-server with your --lora.

Constraints the app enforces (match them or the upload is rejected):

• Must be a single .gguf file (the GGUF LoRA from step 2 — not the .safetensors, not a merged full model).
• It is applied on top of whatever MiniCPM5-1B GGUF base the app already runs, so train against the MiniCPM5-1B base, not some other model.

Common pitfalls

• FileNotFoundError / can't load base model config during conversion → the adapter_config.json base path doesn't exist locally. Fix with --base <local-dir> or --base-model-id openbmb/MiniCPM5-1B (step 2 gotcha).
• Adapter loads but output is unchanged → base mismatch (trained on base A, applied on base B), or the persona needs its system prompt too — a LoRA biases style but the system prompt still matters. In Desk Pet the persona's system prompt is wired to the adapter; standalone llama-cli users must pass it themselves.
• KeyError on an unknown tensor / target module → the adapter targeted modules the converter doesn't map. Stick to the standard attention+MLP projections (q/k/v/o_proj, gate/up/down_proj) when training (all minicpm5-finetune-* skills already default to these).
• Uploaded .gguf but Desk Pet shows nothing → you uploaded the merged full model or a base GGUF by mistake. A LoRA adapter GGUF is small (tens of MB) and has general.type = adapter (step 3).

When NOT to use

• Serving the fp16 HF base with vLLM / transformers / SGLang → load the PEFT adapter directly, skip GGUF. See minicpm5-deploy-vllm / -transformers / -sglang.
• You want to bake the LoRA into the weights (no runtime --lora) → merge first (model.save_pretrained_merged(...) or peft ... merge_and_unload()), then convert the merged fp16 model with convert_hf_to_gguf.py (see the "Building your own GGUF" section of minicpm5-deploy-llama-cpp). That produces a standalone base GGUF, not an adapter.

Reference

• Training frameworks: minicpm5-finetune (router) and its sub-skills.
• Running GGUF: minicpm5-deploy-llama-cpp.
• Converter: convert_lora_to_gguf.py in ggerganov/llama.cpp.