feat(serve): support ignore_eos on Chat Completions

docs/serving.md

@@ -108,6 +108,9 @@ The endpoint supports:
 - `temperature`, `top_p`, presence/frequency penalties, and signed integer `seed`;
 - the compatible `top_k` (`0..20`) and `min_p` (`0..1`) sampler extensions;
 - up to four non-empty stop strings, applied to both reasoning and answer output;
+- the boolean `ignore_eos` extension: `true` suppresses the checkpoint's own stop tokens so
+  generation runs to the requested token budget, while caller-supplied stop strings and stop token
+  ids still apply; omitted or `false` keeps them;
 - `n:1`, text-only `modalities`, and `response_format: {"type":"text"}`;
 - non-streaming responses and server-sent event streams;
 - `stream_options.include_usage`;

src/serve/openai_chat_request.cpp

@@ -753,6 +753,7 @@ void parse_parallel_tool_calls(const Json& body, const GenerationRequest& output
 }
 
 void parse_stop(const Json& body, GenerationRequest& output) {
+    output.ignore_eos = get_bool(body, "ignore_eos", false);
     if (!body.contains("stop") || body.at("stop").is_null()) { return; }
     output.stop_strings_apply_to_reasoning = true;
     const Json& stop                       = body.at("stop");

src/serve/request.h

@@ -178,6 +178,10 @@ struct GenerationRequest {
     ToolChoice tool_choice;
     std::vector<std::string> stop_strings;
     bool stop_strings_apply_to_reasoning = false;
+    // Benchmark/serving extension shared with vLLM, SGLang and llama.cpp: suppress the
+    // checkpoint's default stop tokens so generation runs to the requested token budget.
+    // Caller-supplied stop tokens and stop strings still apply.
+    bool ignore_eos = false;
     int max_tokens                       = 0; // resolved budget; zero means immediate output limit
     std::optional<bool> enable_thinking;      // unset => use the server default
     std::optional<std::uint32_t> thinking_budget;

src/serve/translate.cpp

@@ -306,6 +306,7 @@ ninfer::RequestOptions to_request_options(const GenerationRequest& request,
     options.output.raw                     = false;
     options.output.preserve_special_tokens = request.uses_tools() || request.has_tool_history();
     options.output.tool_name_max_length = static_cast<std::uint32_t>(request.tool_name_max_length);
+    options.stop.include_model_defaults = !request.ignore_eos;
     options.stop.strings.reserve(request.stop_strings.size() *
                                  (request.stop_strings_apply_to_reasoning ? 2U : 1U));
     for (const std::string& stop : request.stop_strings) {

tests/test_openai_schema.cpp

@@ -580,6 +580,25 @@ int test_stops_and_ranges() {
     failures +=
         check(api_error([&] { (void)parse(body); }).param == "stop", "empty stop string rejected");
 
+    body = base_request();
+    failures += check(options(parse(body).generation).stop.include_model_defaults,
+                      "an omitted ignore_eos keeps the checkpoint's own stop tokens");
+    body["ignore_eos"] = false;
+    failures += check(options(parse(body).generation).stop.include_model_defaults,
+                      "ignore_eos false keeps the checkpoint's own stop tokens");
+    body["ignore_eos"] = true;
+    failures += check(parse(body).generation.ignore_eos &&
+                          !options(parse(body).generation).stop.include_model_defaults,
+                      "ignore_eos suppresses the checkpoint's own stop tokens");
+    body["stop"] = Json::array({"A"});
+    failures += check(options(parse(body).generation).stop.strings.size() == 2 &&
+                          !options(parse(body).generation).stop.include_model_defaults,
+                      "ignore_eos leaves caller stop strings in place");
+    body.erase("stop");
+    body["ignore_eos"] = "true";
+    failures += check(api_error([&] { (void)parse(body); }).param == "ignore_eos",
+                      "a non-boolean ignore_eos is rejected");
+
     body                                  = base_request();
     body["top_k"]                         = 21;
     const GenerationRequest invalid_top_k = parse(body).generation;