Model Maker

CLI for local SFT fine-tuning, quantize, and run. One project = one model.
No GUI. npm package: modelmaker. This replaces the old modelstudio command.

Install → train (intended flow)

# 1. Install the CLI (once)
npm i -g modelmaker

# 2. Put your SFT JSONL in a folder, open that folder
cd path\to\your\jsonl_data

# 3. Train
modelmaker --train
# → numbered models → pick a number → Start training? [Y/n]

Local project install (no global):

npm i modelmaker
npx modelmaker --train

From this repo before publish:

cd ModelStudio-App
npm i -g .
# then anywhere:
modelmaker --train

After train

modelmaker quantize last   # HF/LoRA folder → GGUF (required for --run)
modelmaker --run last
modelmaker eval
modelmaker export last
modelmaker --train stop    # cancel training
modelmaker --run stop      # stop inference

--run needs a llama.cpp server. The npm package ships a Node fallback (modelstudio-server.js) that proxies to llama-server. Put llama-server on PATH, set MODELSTUDIO_LLAMA_SERVER, or from this repo run npm run build:server. Train and quantize do not need that.

Other commands

| Command | Purpose | |---------|---------| | modelmaker init . | Scaffold example JSONL in the current folder | | modelmaker doctor | Check Python / GPU / workers / inference backend | | modelmaker data check | Validate JSONL | | modelmaker --train --auto | Hardware-aware base model | | modelmaker --train --yes | Skip confirm (scripts) | | modelmaker quantize / quantize last | GGUF quantize (needed before --run) | | modelmaker jobs list | Job status |

Requires Node 18+. Training uses Python + workers/requirements.txt (see modelmaker doctor).

Training backends (CUDA / Intel Arc / AMD)

modelmaker --train --auto and modelmaker doctor detect the GPU vendor and pick a separate SFT flow:

| Flow | Hardware | Train stack | |------|----------|-------------| | cuda | NVIDIA | QDoRA/QLoRA, bitsandbytes NF4, flash-attn or SDPA, TF32, torch.compile | | xpu | Intel Arc / Core Ultra | bf16 LoRA on torch.xpu, IPEX, SDPA, no CUDA bitsandbytes | | rocm | AMD Radeon / Instinct | LoRA on ROCm/HIP, SDPA + TunableOp; bitsandbytes only if a HIP build works | | cpu | last resort | fp32 LoRA, slow |

Install one torch wheel, then the matching extras (never mix CUDA wheels onto Arc/AMD):

# NVIDIA
pip install torch --index-url https://download.pytorch.org/whl/cu124
pip install -r workers/requirements.txt -r workers/requirements-cuda.txt

# Intel Arc
pip install torch --index-url https://download.pytorch.org/whl/xpu
pip install -r workers/requirements.txt -r workers/requirements-xpu.txt

# AMD ROCm (Linux)
pip install torch --index-url https://download.pytorch.org/whl/rocm6.2
pip install -r workers/requirements.txt -r workers/requirements-rocm.txt

Verify:

modelmaker doctor
python workers/hw_probe.py
python workers/train_sft.py --dry-run --base-model Qwen/Qwen2.5-1.5B-Instruct --data-folder . --output-dir %TEMP%\mm-dry

HSA_OVERRIDE_GFX_VERSION is not set automatically. Only export it if AMD documents that override for your gfx arch.

Env vars

Legacy MODELSTUDIO_* names are still used (including default folders named ModelStudio):

| Variable | Purpose | |----------|---------| | MODELSTUDIO_EXPORT_ROOT | Export / project root (--export-root) | | MODELSTUDIO_PYTHON | Python executable | | MODELSTUDIO_WORKERS | Workers directory | | MODELSTUDIO_SERVER_BIN | Inference server binary or .js fallback | | MODELSTUDIO_LLAMA_SERVER | Path to llama-server | | LLAMA_SERVER | Same as MODELSTUDIO_LLAMA_SERVER |

Develop / test this repo

npm install
npm test
npm run smoke
npm run cli -- help

npm test and npm run smoke are dry-run (no GPU, no weight download).

GPU smoke (manual, once before a release)

On a machine with CUDA + Python torch CUDA + llama-server (or npm run build:server):

npm i -g .
cd path\to\jsonl
modelmaker doctor
modelmaker data check
modelmaker --train --auto --sample 50 --yes
modelmaker quantize last
modelmaker --run last
modelmaker eval
modelmaker export last
modelmaker --run stop

License

MIT