CodeGeeX2 is a powerful, multilingual code generation model designed for developers seeking enhanced code completion, explanation, translation, and debugging capabilities. It offers significant performance improvements over its predecessor, making AI-assisted coding more accessible and efficient.

How It Works

CodeGeeX2 builds upon the ChatGLM2 architecture, incorporating extensive code pre-training on 600 billion tokens. This approach leverages ChatGLM2's efficient design and extends its capabilities specifically for code-related tasks. Key advantages include its 6B parameter size, which outperforms larger models like StarCoder-15B, and its optimized architecture supporting an 8192 token sequence length and faster inference.

Quick Start & Requirements

• Install/Run: Use transformers library:

  from transformers import AutoTokenizer, AutoModel
  tokenizer = AutoTokenizer.from_pretrained("THUDM/codegeex2-6b", trust_remote_code=True)
  model = AutoModel.from_pretrained("THUDM/codegeex2-6b", trust_remote_code=True, device='cuda')
  

• Prerequisites: Python, transformers, CUDA (for GPU acceleration). For bfloat16 support, a compatible GPU is needed; otherwise, use .half().cuda() for float16.
• Resources: Quantized versions require as little as 6GB VRAM.
• Docs: CodeGeeX Homepage

Highlighted Details

• Achieves 35.9% Pass@1 on HumanEval-X (Python), surpassing StarCoder-15B.
• Supports over 100 programming languages via IDE plugins (VS Code, Jetbrains).
• Offers features like context-aware completion and cross-file completion.
• Quantized to INT4 requires only 5.5 GB VRAM.

Maintenance & Community

• Developed by THUDM.
• Latest generation CodeGeeX4 is also available.
• Community discussion via WeChat group (link in README).

Licensing & Compatibility

• Code is licensed under Apache-2.0.
• Model weights require adherence to the Model License. Academic research is fully open; commercial use requires application.

Limitations & Caveats

• The base CodeGeeX2-6B model is for code generation, not chat. For conversational AI programming assistance, use the provided IDE plugins.
• Optimal performance requires specific prompt formatting, including language tags (e.g., # language: Python).