Skip to content
Use litecode

Editor autocomplete

litecode can serve fill-in-the-middle (FIM) autocomplete with the small fim model (default Qwen2.5-Coder-1.5B, about 20 tokens/s on a 4-core CPU).

1. Start the server#

litecode fim

It listens on http://127.0.0.1:8081/infill. Change the port with litecode config fim_port 8082.

RAM: the fim model uses about 1.5 GB. It can run next to the main model on a 16 GB PC.

2. VS Code#

The extension is in editors/vscode/. It has no dependencies.

  1. Copy the folder editors/vscode to ~/.vscode/extensions/litecode-complete (Windows: %USERPROFILE%\.vscode\extensions\litecode-complete).
  2. Restart VS Code.
  3. Type code. A grey suggestion appears after you stop typing. Press Tab to accept.

Settings (litecode.*): url, maxTokens (48), debounceMs (350), enabled.

3. Other editors#

Any editor plugin that supports the llama.cpp /infill API works. Example: llama.vim and llama.vscode by ggml-org. Point it at the URL above.

Test by hand:

curl -s http://127.0.0.1:8081/infill -d '{"input_prefix":"def add(a, b):\n    ","input_suffix":"\n","n_predict":16}'
PreviousMCP servers

Source: docs/EDITORS.md