MathFormer v2.0.0

MathFormer v2.0.0

Added

• Pure Python LLaMA Inference Engine: Introduced a brand-new llama.py module implementing the LLaMA architecture entirely from scratch using only Python standard library modules (math, struct, operator), with native safetensors weight loading and a dependency-free tokenizer.

Changed

• Zero External Dependencies: Removed all heavy third-party dependencies (torch, transformers, safetensors) from requirements.txt and pyproject.toml; the package now runs on pure Python with no external runtime dependencies.
• Internal Model Backend: Replaced the transformers.LlamaForCausalLM backend with the custom TinyLlama pure Python engine for all model inference.
• Project Description: Updated the package description to "A transformer-based math library (Pure Python)" to reflect the new zero-dependency architecture.
• Context Manager Usage: The MathFormerAPI context manager and __enter__/__exit__ have been removed; users should now use the MathFormer class directly as a context manager for single-model lifecycle management.

Removed

• MathFormerAPI.batch_predict(): Removed the batch prediction method and its underlying ThreadPoolExecutor-based concurrency from MathFormerAPI.
• MathFormerAPI.get_model_info(): Removed the model introspection method that previously exposed path, load state, and device information.
• MathFormerAPI.__enter__ / __exit__: Removed the context manager protocol from MathFormerAPI.

Improved

• Model Architecture: Compressed model configurations — reduced hidden_size from 16 to 8, intermediate_size from 64 to 32, num_attention_heads from 8 to 2, and num_hidden_layers from 2 to 1 — further shrinking the combined artifact size of all four models from 154,560 bytes to 28,288 bytes (approx. 150.9 KB to 27.6 KB).
• Documentation: Added comprehensive reStructuredText-style docstrings (:param, :type, :return, :rtype) to all public functions and classes across __init__.py, api.py, tokenizer.py, and llama.py.
• README: Corrected example outputs, fixed lazy_load parameter usage, and updated the context manager section to demonstrate the MathFormer class instead of MathFormerAPI.
• Attention Optimization: Fused QKV projections and streamlined attention score computation in the pure Python engine, using pre-bound operators and __slots__ on all model classes for reduced memory overhead.

Backward Compatibility

• Breaking: MathFormerAPI no longer supports batch_predict(), get_model_info(), or the context manager protocol (with MathFormerAPI() as api). Migrate batch workloads to explicit loops and use MathFormer for context-managed single-model usage.
• Public API functions (add, sub, mul, div, calculate, unload_models) remain fully compatible with the previous version.