Initial Release

v0.1.1 — Initial release

What's included

• GGUF quantization (Q2_K through Q8_0) via llama.cpp
• GPTQ quantization (INT4/INT8) via gptqmodel — runs on Kaggle T4
• TFLite conversion via onnx2tf — runs on Google Colab
• Real benchmark: tok/s + perplexity via llama-cpp-python
• Simulated mobile benchmark: MAC count + latency across 7 SoC profiles
• Pareto frontier chart (interactive Plotly HTML)
• Gradio web UI with 4 tabs
• CLI: run and ui commands

Verified on

• Qwen2-0.5B, GTX 1060 6GB, Windows, CPU inference

Known limitations

• TFLite not supported on Windows (use Colab notebook)
• GPTQ requires 16GB+ VRAM locally (use Kaggle notebook)