core : fix indent termination logic

Author: ggerganov

autoload/llama.vim

@@ -47,7 +47,7 @@ let s:default_config = {
     \ 'n_suffix':           64,
     \ 'n_predict':          128,
     \ 't_max_prompt_ms':    500,
-    \ 't_max_predict_ms':   1000,
+    \ 't_max_predict_ms':   2000,
     \ 'show_info':          2,
     \ 'auto_fim':           v:true,
     \ 'max_line_suffix':    8,
@@ -196,6 +196,15 @@ function! llama#init()
     if g:llama_config.ring_n_chunks > 0
         call s:ring_update()
     endif
+
+    " for debugging
+    "call timer_start(100, {-> s:update_status()})
+endfunction
+
+function! s:update_status()
+    let &statusline = 'indent = ' . s:indent_last
+
+    call timer_start(100, {-> s:update_status()})
 endfunction
 
 " compute how similar two chunks of text are
@@ -515,9 +524,9 @@ function! llama#fim(pos_x, pos_y, is_auto, prev, use_cache) abort
     endif
 
     let l:t_max_predict_ms = g:llama_config.t_max_predict_ms
-    if empty(a:prev)
-        " the first request is quick - we will launch a speculative request after this one is displayed
-        let l:t_max_predict_ms = 250
+    if !empty(a:prev)
+        " give more time for the speculative FIM
+        let l:t_max_predict_ms = min([3*g:llama_config.t_max_predict_ms, 3000])
     endif
 
     " compute multiple hashes that can be used to generate a completion for which the
@@ -547,8 +556,10 @@ function! llama#fim(pos_x, pos_y, is_auto, prev, use_cache) abort
         endfor
     endif
 
-    " TODO: this might be incorrect
-    let s:indent_last = l:indent
+    " update only for non-speculative fims
+    if empty(a:prev)
+        let s:indent_last = l:indent
+    endif
 
     " TODO: refactor in a function
     let l:text = getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')]))