amazingvmsloth 0.2.0

Blazing-fast LLM fine-tuning with minimal VRAM — multi-GPU, manual LoRA gradients, flash attention, 4-bit quant

amazingvmsloth

Blazing-fast LLM fine-tuning with minimal VRAM.

  \   / |    amazingvmsloth - Fast LLM Fine-Tuning
   O^O / \_/ \   Minimal VRAM. Maximum Speed.
  \        /
   "-____-"

Train 14B models on a 4GB GPU. Multi-GPU, 4-bit quantization, LoRA, CPU offloading, gradient checkpointing, and sequence packing — all built for speed on consumer hardware.

---

Install

pip install amazingvmsloth

Or from source:

git clone https://github.com/CollabVMgamez/amazingvmsloth.git
cd amazingvmsloth
pip install -e .

Requirements: Python 3.9+, PyTorch 2.1+, CUDA 11.8+ (optional, CPU training supported)

---

Quick Start

1. Wizard — let it pick settings for your hardware

amazingvmsloth wizard --model Qwen/Qwen2.5-0.5B

Analyzes your GPU/CPU and prints a ready-to-run command.

2. Train

amazingvmsloth train \
  --model Qwen/Qwen2.5-0.5B \
  --dataset tatsu-lab/alpaca \
  --epochs 3 \
  --batch-size 2 \
  --grad-accum 4 \
  --lora-r 16 \
  --output-dir ./output

Supports chat-format datasets too:

amazingvmsloth train \
  --model Qwen/Qwen2.5-0.5B \
  --dataset angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k \
  --dataset-format chat \
  --max-samples 1000 \
  --output-dir ./thinking_lora

3. Convert LoRA to merged model

amazingvmsloth merge \
  --model Qwen/Qwen2.5-0.5B \
  --lora ./output \
  --output ./merged_model

4. Run inference

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("./merged_model", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B")

messages = [{"role": "user", "content": "What is 2+2?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))

---

CLI Commands

Command	Description
wizard	Interactive config generator based on your hardware
train	Fine-tune a model with LoRA
merge	Merge LoRA weights into base model
convert	Merge + convert to GGUF (requires llama.cpp)
info	Show model info and VRAM estimates
bench	Benchmark vs unsloth

---

Hardware Tiers

GPU VRAM	Strategy
4-6 GB	4-bit quant, batch=1, grad accum=8, seq=512, tiny LoRA
6-12 GB	4-bit quant, batch=1-2, grad accum=4, seq=1024
12-24 GB	4-bit or full precision, batch=2-4, torch.compile
24+ GB	Full precision, no grad checkpointing, large batch
CPU only	fp32/bf16, torch.compile, physical-core threading

---

Key Features

• rsLoRA scaling for stable training at any rank
• 4-bit/8-bit quantization via bitsandbytes
• XFormers/SDPA attention patching (Flash Attention on Linux)
• Sequence packing for 2-3x throughput
• Gradient checkpointing with selective layer skipping
• Multi-GPU: DDP, FSDP, DeepSpeed, pipeline parallelism
• Layer offloading via accelerate.dispatch_model
• CPU training with IPEX, pre-packing, torch.compile
• PagedAdamW8bit optimizer for low-VRAM training
• Checkpoint resume with full RNG/optimizer state
• Tqdm progress bar with live loss + VRAM display

---

Example: 500 Steps on Dolly

amazingvmsloth train \
  --model Qwen/Qwen2.5-0.5B \
  --dataset databricks/databricks-dolly-15k \
  --dataset-format alpaca \
  --epochs 1 --batch-size 2 --grad-accum 2 \
  --max-samples 1000 --max-seq-length 512 \
  --lora-r 16 --output-dir ./dolly_lora --packing

This runs ~500 steps in ~10 minutes on a 4GB RTX 3050.

---

Project Structure

amazingvmsloth/
├── lora.py              # LoRA with rsLoRA, device-aware init
├── quantization.py      # 4-bit/8-bit quant, kbit training prep
├── attention.py         # SDPA/XFormers patching
├── trainer.py           # AmazingTrainer with tqdm, packing, offloading
├── cpu_trainer.py       # CpuTrainer for CPU-only training
├── packing.py           # Sequence packing collators
├── gradient.py          # GradientAccumulator
├── optimizer.py         # PagedAdamW8bit, CpuOffloadedAdamW
├── offload.py           # Layer offloading via accelerate
├── cli.py               # CLI entrypoint
├── wizard.py            # Hardware-aware config generator
├── bench.py             # Benchmark vs unsloth
└── utils/
    ├── banner.py        # Startup banner with GPU info
    ├── memory.py        # VRAM estimation
    ├── patching.py      # LoRA save/load helpers
    └── save_load.py     # Model save/merge

---

Benchmarks

On RTX 3050 4GB Laptop GPU:

Library	Time (1 epoch, 500 samples)	Peak VRAM
amazingvmsloth	5.3s	1.07 GB
unsloth	10.1s	0.96 GB

1.91x faster on small runs with pre-quantized models.

---

License

MIT