Flatline LSP

Github Copilot-like LSP code completion server with local LLM powered by llama.cpp

About model

Flatline LSP uses CodeGen2.5 released under Apache v2 License by Salesforce.

https://huggingface.co/sokada/codegen25-7b-multi-gguf-with-dummy-tokenizer

Install

Please choose "openblas" or "cublas" and replace <backend_type> with it.

• openblas: for CPU, slow
• cublas: for NVIDIA GPU, fast

For Ubuntu20.04 & 22.04

Requirements

openblas

Install OpenBLAS library.

sudo apt-get install libopenblas0

cublas

Install CUDA library. See https://developer.nvidia.com/cuda-toolkit

Install flatline_lsp

curl -L https://raw.githubusercontent.com/okdshin/flatline_lsp/main/install.sh | bash -s -- -b <backend_type>

Uninstall flatline_lsp

# Uninstall: Just remove `$HOME/.flatline_lsp`
rm -R $HOME/.flatline_lsp

Editor setup

VSCode

See vscode-flatline-lsp-client

Neovim

Add this snippet to plugins.lua

local setup_flatline_lsp = function()
    vim.api.nvim_create_autocmd('FileType', {
        pattern = {'cpp', 'python'}, --other languages supported by the model can be added here
        callback = function()
            vim.lsp.start({
                name = 'flatline_lsp',
                capabilities = vim.lsp.protocol.make_client_capabilities(),
                cmd = { vim.fn.expand('$HOME/.flatline_lsp/flatline_lsp') },
                root_dir = vim.fs.dirname(vim.fs.find({ '.git' }, { upward = true })[1]),
            })
        end,
    })
end
setup_flatline_lsp()

Backend server

flatline_lsp depends on flatline-server (It is $HOME/.flatline_lsp/_internal/flatline/backend_server/flatline-server) and flatline_lsp automatically starts flatline-server when it is not started yet.

The process of flatline-server does not stop even when flatline_lsp is stopped. So if you want to stop flatline-server and release the resources (e.g. GPU memory), please kill the flatline-server process manually.

TODO prepare more sophisticated method