CodeGeeX

Generates executable code in Python, C++, Java, JavaScript, and Go using a 13B-parameter Transformer decoder with 40 layers trained on 850B+ tokens across 23 programming languages. The model uses a GPT-2 tokenizer extended with whitespace tokens (50,400 vocab) and processes up to 2,048 token sequences, enabling both zero-shot generation from natural language descriptions and continuation-based completion from partial code snippets. Inference supports single-GPU (27GB FP16), quantized (15GB 8-bit), and multi-GPU parallel deployment via checkpoint conversion and distributed inference scripts.

Translates code between Python, C++, Java, JavaScript, and Go by leveraging the multilingual Transformer decoder trained on parallel code examples across 23 languages. The model encodes source code as tokens and generates semantically equivalent target code by learning language-agnostic algorithmic patterns during training. Translation quality depends on the model's ability to abstract syntax and control flow across language boundaries; the 2,048 token limit constrains translation of large functions.

Provides end-to-end training infrastructure for fine-tuning CodeGeeX on custom datasets. The pipeline includes data processing scripts for tokenization and batching, training scripts supporting distributed training on Ascend 910 processors (or PyTorch equivalents), and checkpoint management for saving/resuming training. Training supports both full model fine-tuning and parameter-efficient approaches (e.g., LoRA, though not explicitly documented).

Provides a web-based UI for interactive code generation, allowing users to input natural language descriptions or code snippets and receive generated code without installing IDE extensions or managing inference servers. The web interface communicates with a backend CodeGeeX inference server via HTTP API, supporting the same four interaction modes as the IDE extension (completion, comment-to-code, explanation, summarization).

Integrates with VS Code (via aminer.codegeex extension) and JetBrains IDEs (IntelliJ IDEA, PyCharm, GoLand, CLion) to provide real-time code completion, code explanation, and code summarization. The extension communicates with a local or remote CodeGeeX inference server via HTTP/gRPC, sending cursor context (surrounding code, file type, position) and receiving token-level completions. Four interaction modes support different workflows: inline completion (Copilot-style), comment-to-code generation, code explanation, and function summarization.

Reduces the 13B-parameter model from 27GB (FP16) to 15GB through 8-bit quantization, enabling deployment on mid-range GPUs. The quantization process uses scripts/test_inference_quantized.sh to load checkpoints with reduced precision, trading inference speed and code quality for memory efficiency. Quantized models maintain functional correctness for most code generation tasks but show measurable degradation in complex reasoning and multi-step logic.

Distributes the 13B-parameter model across multiple GPUs using Megatron-LM style model parallelism, reducing per-GPU memory requirements to 6GB+ each. The deployment pipeline involves checkpoint conversion (scripts/convert_ckpt_parallel.sh) to shard model weights across GPUs, followed by parallel inference execution (scripts/test_inference_parallel.sh) that coordinates forward passes across devices. This approach enables inference on clusters of smaller GPUs or reduces latency through pipeline parallelism.

Provides a standardized evaluation platform (HumanEval-X benchmark) with 820 hand-crafted programming problems across Python, C++, Java, JavaScript, and Go. The benchmark includes functional correctness testing infrastructure that executes generated code against test cases, measuring pass@k metrics (percentage of problems solved with k attempts). Evaluation pipeline integrates with code generation utilities to automate the process of generating solutions, executing them, and computing metrics.

Generates natural language explanations of code snippets and function summaries by leveraging the Transformer decoder's ability to produce text from code tokens. The IDE extension exposes this capability through a 'explain code' interaction mode that sends selected code to the inference server and returns a human-readable explanation. Summarization works similarly, generating concise descriptions of function behavior, parameters, and return values.

Provides a Docker image (codegeex/codegeex:latest) with all dependencies pre-configured for GPU-accelerated inference. The container includes Python 3.7+, PyTorch, CUDA 11.0+, and CodeGeeX model checkpoint, enabling one-command deployment via docker run with NVIDIA Docker runtime. Container supports both single-GPU and multi-GPU inference through environment variable configuration.

Uses a GPT-2 tokenizer extended with whitespace tokens to create a 50,400-token vocabulary optimized for code across 23 programming languages. The tokenizer preserves whitespace significance (critical for Python indentation) and includes language-specific tokens for common keywords and operators. Tokenization is applied uniformly across all languages, enabling the same vocabulary for multilingual generation without language-specific tokenizers.

Provides utilities for loading, converting, and managing model checkpoints across different formats and deployment scenarios. The codegeex/torch/get_ckpt_qkv.py script extracts query-key-value projections for quantization, while convert_ckpt_parallel.sh converts checkpoints for distributed inference. Checkpoint management supports FP16 (27GB), 8-bit quantized (15GB), and parallel-distributed formats, with explicit conversion pipelines for each deployment mode.