lmfast 0.3.3

Democratized Small Language Model Training - Train, fine-tune, distill, and deploy sub-500M parameter models on Colab T4 in 30-40 minutes

🚀 LMFast

Democratized Small Language Model Training - Train, fine-tune, distill, and deploy sub-500M parameter models on Colab T4 in 30-40 minutes with enterprise-grade features.

---

✨ Features

Feature	Description
🎯 T4 Optimized	Train on free Colab T4 (12GB) with QLoRA + gradient checkpointing
⚡ Fast Training	Unsloth integration for 2-5x faster fine-tuning
🧠 Distillation	Transfer knowledge from larger models to tiny ones
🤖 Agents	Tool-using agents and orchestration framework
📚 RAG	Lightweight document retrieval and indexing
🌐 Browser	Deploy to browser via ONNX/WebLLM (no server costs)
🛡️ Guardrails	PII detection, toxicity filtering, prompt injection protection
📊 Observability	Langfuse integration, metrics, attention visualization
🚀 Fast Inference	vLLM backend with OpenAI-compatible API
📦 Easy Export	GGUF, INT4, AWQ, GPTQ quantization
🧩 MCP	Native Model Context Protocol server support

---

🚀 Quick Start

Installation

# Basic installation
pip install lmfast

# With all features
pip install lmfast[all]

# Specific extras
pip install lmfast[fast]        # Unsloth for faster training
pip install lmfast[guardrails]  # Safety features
pip install lmfast[observability]  # Monitoring
pip install lmfast[inference]   # vLLM serving

Train in 5 Lines

from lmfast import SLMTrainer, SLMConfig, TrainingConfig
from datasets import load_dataset

# Load data
dataset = load_dataset("yahma/alpaca-cleaned", split="train[:1000]")

# Train
trainer = SLMTrainer(
    SLMConfig(model_name="HuggingFaceTB/SmolLM-135M"),
    TrainingConfig(max_steps=500)
)
trainer.train(dataset)
trainer.save("./my_slm")

CLI Usage

# Train a model
lmfast train --model HuggingFaceTB/SmolLM-135M --data yahma/alpaca-cleaned --output ./my_model

# Knowledge distillation
lmfast distill --teacher Qwen/Qwen2-1.5B --student HuggingFaceTB/SmolLM-135M --data my_data.json

# Start inference server
lmfast serve --model ./my_model --port 8000

# Export to GGUF
lmfast export --model ./my_model --output ./model.gguf --format gguf

# Interactive chat
lmfast generate --model ./my_model --interactive

---

📚 Documentation

Training

from lmfast import SLMTrainer, SLMConfig, TrainingConfig

# Configure for T4 GPU
model_config = SLMConfig(
    model_name="HuggingFaceTB/SmolLM-135M",
    max_seq_length=2048,
    load_in_4bit=True,  # QLoRA
)

training_config = TrainingConfig(
    max_steps=500,
    batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-4,
    lora_r=16,
    lora_alpha=32,
)

trainer = SLMTrainer(model_config, training_config)
trainer.train(dataset)

Knowledge Distillation

from lmfast.distillation import DistillationTrainer
from lmfast.core.config import DistillationConfig

config = DistillationConfig(
    teacher_model="Qwen/Qwen2-1.5B",
    temperature=2.0,
    alpha=0.5,
)

trainer = DistillationTrainer(
    student_model="HuggingFaceTB/SmolLM-135M",
    distillation_config=config,
)
trainer.distill(dataset)

Guardrails

from lmfast.guardrails import GuardrailsConfig, InputValidator, OutputFilter

config = GuardrailsConfig(
    enable_pii_detection=True,
    enable_toxicity_filter=True,
    enable_prompt_injection=True,
)

validator = InputValidator(config)
result = validator.validate(user_input)
if result.is_valid:
    # Process sanitized input
    output = model.generate(result.sanitized_input)

Observability

from lmfast.observability import SLMTracer, MetricsCollector

# Tracing (Langfuse integration)
tracer = SLMTracer(project_name="my_project")

with tracer.trace("inference") as span:
    span.set_attribute("model", "smollm-135m")
    response = model.generate(prompt)
    span.set_attribute("tokens", len(response))

# Metrics
collector = MetricsCollector()
collector.log("loss", 0.5, step=100)
collector.plot("loss")

Fast Inference

from lmfast.inference import SLMServer

# Create server
server = SLMServer("./my_model", use_vllm=True)

# Generate
response = server.generate("Hello, how are you?")

# Batch generation
responses = server.generate_batch(["Prompt 1", "Prompt 2"])

# Start OpenAI-compatible API
server.serve(port=8000)

---

🎯 Supported Models

Model	Parameters	T4 Compatible	Notes
SmolLM-135M	135M	✅	Fastest training
SmolLM-360M	360M	✅	Good balance
TinyLlama-1.1B	1.1B	✅ (with QLoRA)	More capable
Qwen2-0.5B	500M	✅	Multilingual
Phi-3-mini	3.8B	⚠️ (tight)	Most capable

---

📦 Package Structure

lmfast/
├── core/           # Config and model loading
├── training/       # Training and data processing
├── distillation/   # Knowledge distillation
├── guardrails/     # Safety and filtering
├── observability/  # Tracing and metrics
├── inference/      # Serving and quantization
└── cli/            # Command-line interface

---

🧪 Development

# Clone
git clone https://github.com/lmfast/lmfast
cd lmfast

# Create environment
conda env create -f environment.yml
conda activate lmfast

# Install in dev mode
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Format code
black lmfast/ tests/
ruff check lmfast/ tests/

---

📄 License

Apache 2.0 - See LICENSE for details.

---

🙏 Acknowledgments

• Unsloth for fast training
• HuggingFace for transformers ecosystem
• vLLM for fast inference
• Langfuse for observability