decant-dsl 0.3.0

Decant — A minimal DSL for AI model training & hosting with multi-GPU QLoRA support

Decant DSL

A minimal language for AI model training & hosting. Turn 300+ lines of boilerplate into ~20 lines of code.

Decant is a domain-specific language that lets you specify everything needed to train, fine-tune, or host an LLM in a single readable file. It compiles to VRAM-optimized Python using Unsloth, QLoRA, and Hugging Face Transformers — so anyone with a GPU can train their own AI.

# Minimal example: fine-tune Llama 3.2 3B with QLoRA + YaRN
purpose = "fine-tune"
model.base = "unsloth/Llama-3.2-3B-Instruct-bnb-4bit"
train.precision = "4bit"
train.compression = "qlora"
train.lora.rank = 16
dataset.path = "databricks/databricks-dolly-15k"
model.context = 2048
model.context_extend = 4096
model.context_method = "yarn"
train.steps = 100
train.batch.size = 2
train.learning.rate = 2e-4

Quick Start

# Install
pip install decant-dsl

# Validate your config
decant check train.nx

# Generate Python training code (don't run it)
decant build train.nx -o train.py

# Generate and run training
decant run train.nx

Language Reference

Purpose

purpose = "train"       # Train from scratch
purpose = "fine-tune"   # Fine-tune an existing model (recommended)
purpose = "host"        # Load and optimize model for inference only

Model

model.base = "org/model-name"
model.source = "huggingface"  # or "local"
model.local_path = "/path/to/model"  # if source is local
model.context = 2048
model.context_extend = 4096
model.context_method = "yarn"  # yarn, ntk, linear, none
model.rope.theta = 1000000

Training

# Precision & Compression
train.precision = "4bit"        # 4bit, 8bit, 16bit, 32bit
train.compression = "qlora"     # qlora, lora, full, none

# LoRA / QLoRA
train.lora.rank = 16            # 1-512
train.lora.alpha = 16
train.lora.dropout = 0.0
train.lora.targets = "all"      # all, attention, mlp, custom

# Hyper-parameters
train.steps = 100
train.epochs = 0
train.batch.size = 2
train.gradient.accumulation = 4
train.learning.rate = 2e-4
train.warmup.steps = 5
train.optimizer = "adamw_8bit"
train.scheduler = "cosine"
train.weight.decay = 0.0
train.max.grad.norm = 1.0
train.seed = 42

Dataset

dataset.path = "org/dataset-name"
dataset.source = "huggingface"  # or "local"
dataset.format = "alpaca"       # alpaca, chat, text, sharegpt, json
dataset.split = "train"
dataset.text.field = "text"
dataset.local_path = "/path/to/data.json"

Output

output.dir = "./outputs"
output.name = "my-finetuned-model"
output.push.to.hub = false
output.hub.id = "my-org/my-model"

Examples

See the examples/ directory:

• train-llama3-8b.nx — Fine-tune Llama 3.2 3B on Dolly-15k with QLoRA + YaRN
• host-mistral.nx — Host Mistral 7B with 4-bit quantization and 32K context
• train-local-dataset.nx — Train using a local dataset file

Error Messages

Decant translates cryptic Python/CUDA errors into plain English:

Error	What You'll See
CUDA OOM	"Your GPU ran out of VRAM — lower batch size or precision"
Module not found	"Missing Python package — run pip install"
Network error	"Could not download — check your internet connection"
File not found	"Resource not found — check the path or model name"
Permission error	"Access denied — you may need to log in to Hugging Face"

Architecture

.nx file → Lark Parser → Typed Config → Code Generator → Optimized Python → Execution
                  ↘                ↙
             Validation Engine  (with error translation)

Publishing to PyPI

pip install build twine
cd decant/
python -m build
python -m twine upload dist/*

Requirements

• Python 3.10+
• CUDA-capable GPU with 6+ GB VRAM (for 3B models with QLoRA)
• 12+ GB VRAM recommended for 7B models with QLoRA
• 24+ GB VRAM recommended for 13B+ models

License

MIT