server : send partial stop string when <EOG> is reached (ggml-org#15007)

matteoserva · fm240223 · commit 5409c43b80bd · 2025-10-23T16:46:16.000+02:00
diff --git a/tools/server/server.cpp b/tools/server/server.cpp
@@ -2839,7 +2839,7 @@ struct server_context {
                     slot.generated_text.begin() + pos + stop_pos,
                     slot.generated_text.end());
                 pos = std::min(slot.n_sent_text, slot.generated_text.size());
-            } else if (slot.has_next_token) {
+            } else if (slot.has_next_token && !llama_vocab_is_eog(vocab, result.tok) ) {
                 stop_pos = slot.find_stopping_strings(str_test, token_str.size(), false);
                 send_text = stop_pos == std::string::npos;
             }

Original file line number	Diff line number	Diff line change
`@@ -2839,7 +2839,7 @@ struct server_context {`
`2839`	`2839`	`slot.generated_text.begin() + pos + stop_pos,`
`2840`	`2840`	`slot.generated_text.end());`
`2841`	`2841`	`pos = std::min(slot.n_sent_text, slot.generated_text.size());`
`2842`		`- } else if (slot.has_next_token) {`
	`2842`	`+ } else if (slot.has_next_token && !llama_vocab_is_eog(vocab, result.tok) ) {`
`2843`	`2843`	`stop_pos = slot.find_stopping_strings(str_test, token_str.size(), false);`
`2844`	`2844`	`send_text = stop_pos == std::string::npos;`
`2845`	`2845`	`}`