llama.vim

Local LLM-assisted text completion.

Fill-in-Middle (FIM) completions

Instruction-based editing

https://github.com/user-attachments/assets/641a6e72-f1a2-4fe5-b0fd-c2597c6f4cdc

https://github.com/user-attachments/assets/68bff15b-2d91-4800-985d-b7f110a0ccb7

Features

• Auto-suggest on cursor movement in Insert mode
• Accept a suggestion with Tab
• Accept the first line of a suggestion with Shift+Tab
• Instruction-based editing with lli
• Control max text generation time
• Configure scope of context around the cursor
• Ring context with chunks from open and edited files and yanked text
• Supports very large contexts even on low-end hardware via smart context reuse
• Display performance stats

Installation

Plugin setup

• vim-plug

  Plug 'ggml-org/llama.vim'
  

• Vundle

  cd ~/.vim/bundle
  git clone https://github.com/ggml-org/llama.vim
  

  Then add Plugin 'llama.vim' to your .vimrc in the vundle#begin() section.

• lazy.nvim

  {
      'ggml-org/llama.vim',
  }
  

Plugin configuration

You can customize llama.vim by setting the g:llama_config variable.

Examples:

1. Disable the inline info:

   " put before llama.vim loads
   let g:llama_config = { 'show_info': 0 }
   

2. Same thing but setting directly

   let g:llama_config.show_info = v:false
   

3. Disable auto FIM (Fill-In-the-Middle) completion with lazy.nvim

   {
       'ggml-org/llama.vim',
       init = function()
           vim.g.llama_config = {
               auto_fim = false,
           }
       end,
   }
   

4. Configure FIM keymaps:

   let g:llama_config.keymap_fim_trigger     = "llf"
   let g:llama_config.keymap_fim_accept_full = ""
   let g:llama_config.keymap_fim_accept_line = ""
   let g:llama_config.keymap_fim_accept_word = "ll]"
   

5. Configure instruction-based editing keymaps

   let g:llama_config.keymap_inst_trigger  = "lli"
   let g:llama_config.keymap_inst_rerun    = "llr"
   let g:llama_config.keymap_inst_continue = "llc"
   let g:llama_config.keymap_inst_accept   = ""
   let g:llama_config.keymap_inst_cancel   = ""
   

6. Configure named llama.cpp servers:

   let g:llama_config.profiles = {
       \ 'spark':   'http://192.168.0.66:8080',
       \ 'gmktec':  'http://192.168.0.65:8080',
       \ }
   let g:llama_config.profile = 'spark'
   

Use :LlamaProfile to show the available profiles and :LlamaProfile spark to switch both FIM and instruction requests to another server. Profile names support command-line completion. The selected profile is persisted and :LlamaProfileReset can be used to restore the profile (or custom endpoints) configured in .vimrc.

This only allows the host servers to be specified, not the models used which are configured using model_fim and model_inst. But using consistent naming accross servers, or using server aliases, one can make this independant of the actual selected profile. For example, using something like:

    \ 'model_fim':              'fim_model',
    \ 'model_inst':             'inst_model',

And then in the server configuration, something like the following can be used:

[qwen3.8-27B]
alias                = inst_model,pi

And likewise for the fim_model.

Please refer to :help llama_config or the source for the full list of options.

llama.cpp setup

The plugin requires a llama.cpp server instance to be running at g:llama_config.endpoint_fim and/or g:llama_config.endpoint_inst.

Mac OS

brew install llama.cpp

Windows

winget install llama.cpp

Any other OS

Either build from source or use the latest binaries: https://github.com/ggml-org/llama.cpp/releases

llama.cpp settings

Here are recommended settings, depending on the amount of VRAM that you have:

• More than 64GB VRAM:

  llama-server --fim-qwen-30b-default
  

• More than 16GB VRAM:

  llama-server --fim-qwen-7b-default
  

• Less than 16GB VRAM:

  llama-server --fim-qwen-3b-default
  

• Less than 8GB VRAM:

  llama-server --fim-qwen-1.5b-default
  

Use :help llama for more details.

Recommended LLMs

The plugin requires FIM-compatible models: HF collection

Examples

Using llama.vim on M1 Pro (2021) with Qwen2.5-Coder 1.5B Q8_0:

The orange text is the generated suggestion. The green text contains performance stats for the FIM request: the currently used context is 15186 tokens and the maximum is 32768. There are 30 chunks in the ring buffer with extra context (out of 64). So far, 1 chunk has been evicted in the current session and there are 0 chunks in queue. The newly computed prompt tokens for this request were 260 and the generated tokens were 24. It took 1245 ms to generate this suggestion after entering the letter c on the current line.

Using llama.vim on M2 Ultra with Qwen2.5-Coder 7B Q8_0:

https://github.com/user-attachments/assets/1f1eb408-8ac2-4bd2-b2cf-6ab7d6816754

Demonstrates that the global context is accumulated and maintained across different files and showcases the overall latency when working in a large codebase.

Another example on a small Swift code

Implementation details

The plugin aims to be very simple and lightweight and at the same time to provide high-quality and performant local FIM completions, even on consumer-grade hardware. Read more on how this is achieved in the following links:

• Initial implementation and technical description: https://github.com/ggml-org/llama.cpp/pull/9787
• Classic Vim support: https://github.com/ggml-org/llama.cpp/pull/9995

Other IDEs

• VS Code: https://github.com/ggml-org/llama.vscode