Debounce, LlamaCpp support, expose prompt as setup option, fix passing parameters to model (ollama)

README.md

@@ -128,7 +128,12 @@ cmp_ai:setup({
   provider = 'Ollama',
   provider_options = {
     model = 'codellama:7b-code',
+    prompt = function(lines_before, lines_after)
+      -- prompt depends on the model you use. Here is an example for deepseek coder
+      return prompt = '<PRE> ' .. lines_before .. ' <SUF>' .. lines_after .. ' <MID>', -- for codellama
+    end,
   },
+  debounce_delay = 600, -- ms llama may be GPU hungry, wait x ms after last key input, before sending request to it
   notify = true,
   notify_callback = function(msg)
     vim.notify(msg)
@@ -141,6 +146,54 @@ cmp_ai:setup({
   },
 })
 ```
+Models for Ollama are available at [here](https://ollama.ai/library). For code completions use model that supports it - e.g. [DeepSeek Base 6.7b](https://ollama.ai/library/deepseek-coder)
+
+To use with [LlamaCpp](https://github.com/ggerganov/llama.cpp):
+
+```lua
+local cmp_ai = require('cmp_ai.config')
+
+cmp_ai:setup {
+  max_lines = 30,
+  provider = "LlamaCpp",
+  provider_options = {
+    options = {
+      n_predict = 20,  -- number of generated predictions
+      min_p = 0.05, -- default 0.05,  Cut off predictions with probability below  Max_prob * min_p
+      -- repeat_last_n = 64, -- default 64
+      -- repeat_penalty = 1.100, -- default 1.1
+      -- see llama server link above - to see other options
+    },
+    prompt = function(lines_before, lines_after)
+      -- prompt depends on the model you use. Here is an example for deepseek coder
+      return "<s><｜fim▁begin｜>" .. lines_before .. "<｜fim▁hole｜>" .. lines_after .. "<｜fim▁end｜>" -- for deepseek coder
+    end,
+  },
+  debounce_delay = 600, -- ms llama may be GPU hungry, wait x ms after last key input, before sending request to it
+  notify = true,
+  notify_callback = function(msg)
+    vim.notify(msg)
+  end,
+  run_on_every_keystroke = false,
+  ignored_file_types = {
+    -- default is not to ignore
+    -- uncomment to ignore in lua:
+    -- lua = true
+  },
+}
+```
+
+
+[LlamaCpp Server](https://github.com/ggerganov/llama.cpp/blob/master/examples/server/README.md) has to be started manually with:
+
+```bash
+./server -m ./models/deepseek-coder-6.7b-base.Q4_K_M.gguf -ngl 50 -c 2048 --log-disable
+```
+
+LlamaCpp requires model in GGUP format. Here is the current model I use for coding:
+ - [DeepSeek Base 6.7b](https://huggingface.co/TheBloke/deepseek-coder-6.7B-base-GGUF/blob/main/deepseek-coder-6.7b-base.Q4_K_M.gguf)
+ It is good to have at least 12GB of VRAM to run it (works best with NVIDIA - due to CUDA acceleration). If you want you can grab smaller models too (faster to run, but lower quality of completions)
+
 
 ### `notify`
 

lua/cmp_ai/backends/llamacpp.lua

@@ -0,0 +1,51 @@
+local requests = require('cmp_ai.requests')
+
+LlamaCpp = requests:new(nil)
+
+function LlamaCpp:new(o, params)
+  o = o or {}
+  setmetatable(o, self)
+  self.__index = self
+  self.params = vim.tbl_deep_extend('keep', o or {}, {
+    base_url = 'http://localhost:8080/completion',
+    -- model = 'codellama:7b-code',
+    options = {
+      temperature = 0.2,
+    },
+  })
+  return o
+end
+
+function LlamaCpp:complete(lines_before, lines_after, cb)
+  local data = {
+    -- model = self.params.model,
+    -- prompt = '<PRE> ' .. lines_before .. ' <SUF>' .. lines_after .. ' <MID>', -- for codellama
+    prompt = "<s><｜fim▁begin｜>" .. lines_before .. "<｜fim▁hole｜>" .. lines_after .. "<｜fim▁end｜>", -- for deepseek coder
+    stream = false,
+  }
+  data = vim.tbl_extend('keep', data, self.params.options)
+  data.prompt = self.params.prompt(lines_before, lines_after)
+
+  self:Get(self.params.base_url, {}, data, function(answer)
+    local new_data = {}
+    -- vim.print('answer', answer)
+    if answer.error ~= nil then
+      vim.notify('Llamacp error: ' .. answer.error)
+      return
+    end
+    if answer.stop then
+      local result = answer.content:gsub('<EOT>', '')
+      -- vim.print('results', result)
+      table.insert(new_data, result)
+    end
+    cb(new_data)
+  end)
+end
+
+function LlamaCpp:test()
+  self:complete('def factorial(n)\n    if', '    return ans\n', function(data)
+    dump(data)
+  end)
+end
+
+return LlamaCpp

lua/cmp_ai/backends/ollama.lua

@@ -6,21 +6,21 @@ function Ollama:new(o, params)
   o = o or {}
   setmetatable(o, self)
   self.__index = self
-  self.params = vim.tbl_deep_extend('keep', params or {}, {
+  self.params = vim.tbl_deep_extend('keep', o or {}, {
     base_url = 'http://127.0.0.1:11434/api/generate',
     model = 'codellama:7b-code',
     options = {
       temperature = 0.2,
     },
   })
-
   return o
 end
 
 function Ollama:complete(lines_before, lines_after, cb)
   local data = {
     model = self.params.model,
-    prompt = '<PRE> ' .. lines_before .. ' <SUF>' .. lines_after .. ' <MID>',
+    -- prompt = '<PRE> ' .. lines_before .. ' <SUF>' .. lines_after .. ' <MID>', -- for codellama
+    prompt = "<s><｜fim▁begin｜>" .. lines_before .. "<｜fim▁hole｜>" .. lines_after .. "<｜fim▁end｜>", -- for deepseek coder
     stream = false,
     options = self.params.options,
   }

lua/cmp_ai/config.lua

@@ -5,6 +5,7 @@ local conf = {
   run_on_every_keystroke = true,
   provider = 'HF',
   provider_options = {},
+  debounce_delay = 200, -- ms
   notify = true,
   notify_callback = function(msg)
     vim.notify(msg)

lua/cmp_ai/source.lua

@@ -40,19 +40,70 @@ function Source:_do_complete(ctx, cb)
   local service = conf:get('provider')
   service:complete(before, after, function(data)
     self:end_complete(data, ctx, cb)
-    if conf:get('notify') then
-      conf:get('notify_callback')('Completion started')
-    end
+    -- why 2x ?
+    -- if conf:get('notify') then
+    --   conf:get('notify_callback')('Completion started')
+    -- end
   end)
 end
 
+function Source:trigger(ctx, callback)
+  if vim.fn.mode() == 'i' then
+    self:_do_complete(ctx, callback)
+  end
+end
+
+-- based on https://github.com/runiq/neovim-throttle-debounce/blob/main/lua/throttle-debounce/init.lua (MIT)
+local function debounce_trailing(fn, ms)
+	local timer = vim.loop.new_timer()
+	local wrapped_fn
+
+    function wrapped_fn(...)
+      local argv = {...}
+      local argc = select('#', ...)
+      -- timer:stop() -- seems not needed?
+      timer:start(ms, 0, function()
+        pcall(vim.schedule_wrap(fn), unpack(argv, 1, argc))
+      end)
+    end
+	return wrapped_fn, timer
+end
+
+local bounce_complete, ret_tim =  debounce_trailing(
+  Source.trigger,
+  conf:get('debounce_delay')
+)
+
+local self_cp, ctx_cp, call_cp -- variables to store last completion context
+
+local bounce_autogroup = vim.api.nvim_create_augroup("BounceCompletion", { clear = true })
+vim.api.nvim_create_autocmd({"TextChangedI","InsertEnter","TextChangedP"},{
+  pattern = "*",
+  callback = function()
+    if self_cp ~= nil then
+      bounce_complete(self_cp, ctx_cp, call_cp)
+    end
+  end,
+  group = bounce_autogroup
+})
+
+vim.api.nvim_create_autocmd({"InsertLeave"},{
+  pattern = "*",
+  callback = function()
+    ret_tim:stop()
+  end,
+ group = bounce_autogroup
+})
+
+
 --- complete
 function Source:complete(ctx, callback)
   if conf:get('ignored_file_types')[vim.bo.filetype] then
     callback()
     return
   end
-  self:_do_complete(ctx, callback)
+  self_cp, ctx_cp, call_cp = self, ctx, callback
+  bounce_complete(self_cp, ctx, callback)
 end
 
 function Source:end_complete(data, ctx, cb)