Fix completion_tokens: count tokens not decoded characters

server.jl

@@ -270,12 +270,12 @@ function handle_request(request::HTTP.Request)
             for i in 1:n_completions
                 text = generate_streaming(CONFIG, PS, TOKENIZER, prompt_text;
                                           max_tokens, temperature, top_k=top_k_val, top_p=top_p_val)
-                finish_reason = length(text) >= max_tokens ? "length" : "stop"
+                finish_reason = "length"  # generate_streaming always produces exactly max_tokens tokens
                 push!(choices, Dict(
                     "index" => i - 1,
                     "message" => Dict("role" => "assistant", "content" => text),
                     "finish_reason" => finish_reason))
-                total_completion_tokens += length(text)
+                total_completion_tokens += max_tokens  # count tokens, not decoded chars
             end
 
             prompt_tokens = length(encode(TOKENIZER, prompt_text))