ModelForge v3 introduces a modular provider system that allows you to choose different model loading and optimization backends.

Providers are backends that handle:

• Loading models from model hubs or local storage
• Applying optimizations and patches
• Managing tokenizers
• Setting up device mappings

Different providers offer different trade-offs in terms of speed, memory usage, and compatibility.

✅ Running on native Windows
✅ Maximum compatibility needed
✅ Using models not supported by Unsloth
✅ Debugging or development
✅ First time using ModelForge

✅ Running on Linux or WSL
✅ Training large models (7B+ parameters)
✅ Need faster training times
✅ Have limited VRAM
✅ Using supported architectures (Llama, Mistral, etc.)

{
  "provider": "huggingface",
  "model_name": "meta-llama/Llama-3.2-3B",
  "task": "text-generation",
  ...
}

1. Go to Training tab
2. Select Provider dropdown
3. Choose huggingface or unsloth

curl -X POST http://localhost:8000/api/start_training \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "unsloth",
    "model_name": "meta-llama/Llama-3.2-3B",
    ...
  }'

Training Request
       ↓
Provider Factory
       ↓
  Load Provider ← (HuggingFace / Unsloth / Custom)
       ↓
  Load Model
       ↓
  Load Tokenizer
       ↓
  Validate Access
       ↓
Return (model, tokenizer)

All providers implement the same interface:

class ModelProvider(Protocol):
    def load_model(self, model_id: str, model_class: str, 
                   quantization_config, device_map, **kwargs) -> Model
    
    def load_tokenizer(self, model_id: str, **kwargs) -> Tokenizer
    
    def validate_model_access(self, model_id: str, model_class: str) -> bool
    
    def get_provider_name(self) -> str

This ensures all providers work seamlessly with ModelForge.

Based on Llama-3.2-3B with 1000 examples:

Based on Llama-3.2-7B with batch_size=4:

Results may vary based on hardware and configuration.

Supported Platforms:

• ✅ Windows 10/11
• ✅ Linux (all distributions)
• ✅ WSL 2
• ✅ Docker

Requirements:

• Python 3.11
• PyTorch 2.0+
• transformers 4.40+

Supported Platforms:

• ✅ Linux (Ubuntu 20.04+, Debian 11+)
• ✅ WSL 2 on Windows
• ✅ Docker with NVIDIA runtime
• ❌ Native Windows

Requirements:

• Python 3.11
• PyTorch 2.0+ with CUDA
• transformers 4.40+
• unsloth package
• NVIDIA GPU with CUDA 11.8+

Error: ProviderError: Unknown provider 'xyz'

Solution: Check spelling. Valid providers: huggingface, unsloth

Error: ProviderError: Unsloth is not installed

Solution: Use WSL or Docker. See Windows Installation.

Error: ModuleNotFoundError: No module named 'unsloth'

Solution: Install the provider:

pip install unsloth

Want to add a new provider? See Custom Providers Guide.

• HuggingFace Provider - Learn about the standard provider
• Unsloth Provider - Enable 2x faster training
• Custom Providers - Add your own provider
• Training Strategies - Choose a training strategy

Providers make ModelForge flexible and extensible! Choose the right one for your needs.