tools/main: llama-cli: prevent spurious assistant token (ggml-org#13402)

vinkal-chudgar · vinkal-chudgar · commit bcf14fd4c0fd · 2025-09-22T20:15:36.000Z
During prompt ingestion, prompt tokens are accepted into the sampler history (for repetition penalties). The conversation-mode path then appended `common_sampler_last(smpl)` to `assistant_ss` before any new token was sampled. At that point, "last" was a prompt-side token (e.g., an input prefix), so the assistant chat message began with an extra piece. Fix: append to `assistant_ss` only for a newly sampled (non-EOG) token. This affects only chat message assembly (`assistant_ss` / `chat_msgs` / `common_chat_format_single`); terminal stdout is unchanged. Sampling order/logits are unchanged. Fixes ggml-org#13402. Signed-off-by: Vinkal Chudgar <vinkal.chudgar@gmail.com>
diff --git a/tools/main/main.cpp b/tools/main/main.cpp
@@ -706,6 +706,10 @@ int main(int argc, char ** argv) {
             // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
 
             embd.push_back(id);
+            
+            if (params.conversation_mode && !waiting_for_first_input && !llama_vocab_is_eog(vocab, id)) {
+                assistant_ss << common_token_to_piece(ctx, id, false);
+            }
 
             // echo this to console
             input_echo = true;
@@ -826,9 +830,6 @@ int main(int argc, char ** argv) {
 
             // if current token is not EOG, we add it to current assistant message
             if (params.conversation_mode && !waiting_for_first_input) {
-                const auto id = common_sampler_last(smpl);
-                assistant_ss << common_token_to_piece(ctx, id, false);
-
                 if (!prompt.empty()) {
                     prompt.clear();
                     is_interacting = false;
