salesforce/CodeTF

Provides unified training and inference for code LLMs with built-in support for quantized model serving (int8/int16), parameter-efficient fine-tuning via HuggingFace PEFT, and multi-GPU weight sharding. Includes tree-sitter-based AST parsers for 15+ languages to extract code attributes and preprocessed benchmarks (HumanEval, MBPP, CodeXGLUE) with evaluation metrics like pass@k. Bundles pre-trained (CodeT5, StarCoder, Incoder) and task-specific fine-tuned checkpoints across 8+ code intelligence tasks including summarization, translation, and defect prediction.

1,480 stars. No commits in the last 6 months.