code-completion

README

Code Completion

Generate code completions from a partial code snippet using a GGUF language model.

How it works

1. Reads a code prefix from the --code flag or piped stdin
2. Loads a GGUF model via inference.LoadFile
3. Constructs a completion prompt and generates the continuation
4. Prints the original code followed by the generated completion

Works best with code-capable models (CodeLlama, DeepSeek Coder, Qwen 2.5 Coder) but any instruction-tuned model can produce reasonable completions.

Usage

go build -o code-completion ./examples/code-completion/

# Pass code via flag
./code-completion --model path/to/model.gguf --code "func fibonacci(n int) int {"

# Pass code via stdin
echo "func add(a, b int) int {" | ./code-completion --model path/to/model.gguf

# Use GPU and lower temperature for more deterministic output
./code-completion --model path/to/model.gguf --device cuda --temperature 0.1 \
    --code "// quicksort sorts a slice of integers in place.
func quicksort(arr []int) {"

Flags

Flag	Default	Description
--model	(required)	Path to GGUF model file
--device	cpu	Compute device: "cpu", "cuda"
--code	(stdin)	Code prefix to complete
--max-tokens	256	Maximum tokens to generate
--temperature	0.2	Sampling temperature (lower = more deterministic)

Documentation

Overview

Command code-completion demonstrates using a language model for code completion.

It reads a code snippet (from a flag or stdin), appends a completion prompt, and generates the continuation. This pattern works with any code-capable GGUF model (e.g., CodeLlama, DeepSeek Coder, Qwen 2.5 Coder).

Usage:

go build -o code-completion ./examples/code-completion/
./code-completion --model path/to/model.gguf --code "func fibonacci(n int) int {"
echo "func add(a, b int) int {" | ./code-completion --model path/to/model.gguf