Support finish reason in legacy pipelines

Author: Michal Kulakowski

src/llm/apis/openai_completions.cpp

@@ -409,17 +409,18 @@ std::string OpenAIChatCompletionsHandler::serializeUnaryResponse(ov::genai::Enco
 
     // choices: array of size N, where N is related to n request parameter
     jsonResponse.StartArray("choices");
-    int index = 0;
-    for (int i = 0; i < results.tokens.size(); i++) {
+    for (size_t i = 0; i < results.tokens.size(); ++i) {
         const std::vector<int64_t>& tokens = results.tokens[i];
         SPDLOG_LOGGER_TRACE(llm_calculator_logger, "Generated tokens: {}", tokens);
         ParsedOutput parsedOutput = parseOutputIfNeeded(tokens);
         jsonResponse.StartObject();
         // finish_reason: "stop" in regular scenario, "tool_calls" if output contains tool calls
-        auto finishReason = mapFinishReason(ov::genai::GenerationFinishReason::STOP, !parsedOutput.toolCalls.empty());
+        const ov::genai::GenerationFinishReason finishReasonRaw =
+            (i < results.finish_reasons.size()) ? results.finish_reasons[i] : ov::genai::GenerationFinishReason::STOP;
+        auto finishReason = mapFinishReason(finishReasonRaw, !parsedOutput.toolCalls.empty());
         jsonResponse.FinishReason(finishReason.value_or("unknown"));
         // index: integer; Choice index, only n=1 supported anyway
-        jsonResponse.Index(index++);
+        jsonResponse.Index(static_cast<int>(i));
 
         if (endpoint == Endpoint::CHAT_COMPLETIONS) {
             jsonResponse.MessageObject(parsedOutput);
@@ -462,7 +463,9 @@ std::string OpenAIChatCompletionsHandler::serializeUnaryResponse(ov::genai::VLMD
     OVMS_PROFILE_FUNCTION();
     usage.promptTokens = results.perf_metrics.get_num_input_tokens();
     usage.completionTokens = results.perf_metrics.get_num_generated_tokens();
-
+    if (results.finish_reasons.empty()) {
+        throw std::runtime_error("Missing finish reason in unary VLM responses generation result");
+    }
     OpenAiJsonResponse jsonResponse;
     jsonResponse.StartObject();
 
@@ -480,8 +483,14 @@ std::string OpenAIChatCompletionsHandler::serializeUnaryResponse(ov::genai::VLMD
         SPDLOG_LOGGER_TRACE(llm_calculator_logger, "Generated tokens: {}", generatedTokens);
         ParsedOutput parsedOutput = parseOutputIfNeeded(generatedTokens);
         jsonResponse.StartObject();
-        // finish_reason: "stop" in regular scenario, "tool_calls" if output contains tool calls
-        auto finishReason = mapFinishReason(ov::genai::GenerationFinishReason::STOP, !parsedOutput.toolCalls.empty());
+        ov::genai::GenerationFinishReason responsesFinishReason = ov::genai::GenerationFinishReason::STOP;
+        for (const auto& finishReason : results.finish_reasons) {
+            if (finishReason == ov::genai::GenerationFinishReason::LENGTH) {
+                responsesFinishReason = ov::genai::GenerationFinishReason::LENGTH;
+                break;
+            }
+        }
+        auto finishReason = mapFinishReason(responsesFinishReason, !parsedOutput.toolCalls.empty());
         jsonResponse.FinishReason(finishReason.value_or("unknown"));
         // index: integer; Choice index, only n=1 supported anyway
         jsonResponse.Index(index++);

src/llm/apis/openai_responses.cpp

@@ -653,10 +653,17 @@ std::string OpenAIResponsesHandler::serializeUnaryResponse(ov::genai::EncodedRes
     usage.promptTokens = results.perf_metrics.get_num_input_tokens();
     usage.completionTokens = results.perf_metrics.get_num_generated_tokens();
     std::vector<ParsedOutput> parsedOutputs;
+    ov::genai::GenerationFinishReason responsesFinishReason = ov::genai::GenerationFinishReason::STOP;
     for (const auto& tokens : results.tokens) {
         parsedOutputs.push_back(parseOutputIfNeeded(tokens));
     }
-    return serializeUnaryResponseImpl(parsedOutputs);
+    for (const auto& finishReason : results.finish_reasons) {
+        if (finishReason == ov::genai::GenerationFinishReason::LENGTH) {
+            responsesFinishReason = ov::genai::GenerationFinishReason::LENGTH;
+            break;
+        }
+    }
+    return serializeUnaryResponseImpl(parsedOutputs, responsesFinishReason);
 }
 
 std::string OpenAIResponsesHandler::serializeUnaryResponse(ov::genai::VLMDecodedResults& results, const std::string& textResponse) {
@@ -677,7 +684,14 @@ std::string OpenAIResponsesHandler::serializeUnaryResponse(ov::genai::VLMDecoded
             parsedOutputs.push_back(std::move(output));
         }
     }
-    return serializeUnaryResponseImpl(parsedOutputs);
+    ov::genai::GenerationFinishReason responsesFinishReason = ov::genai::GenerationFinishReason::STOP;
+    for (const auto& finishReason : results.finish_reasons) {
+        if (finishReason == ov::genai::GenerationFinishReason::LENGTH) {
+            responsesFinishReason = ov::genai::GenerationFinishReason::LENGTH;
+            break;
+        }
+    }
+    return serializeUnaryResponseImpl(parsedOutputs, responsesFinishReason);
 }
 
 // --- Streaming event building blocks ---

src/llm/language_model/legacy/servable.cpp

@@ -229,7 +229,11 @@ absl::Status LegacyServable::preparePartialResponse(std::shared_ptr<GenAiServabl
         if (!executionContext->lastStreamerCallbackOutput.empty()) {
             lastTextChunk = lastTextChunk + executionContext->lastStreamerCallbackOutput;
         }
-        std::string serializedChunk = executionContext->apiHandler->serializeStreamingChunk(lastTextChunk, ov::genai::GenerationFinishReason::STOP);
+        ov::genai::GenerationFinishReason finishReason = ov::genai::GenerationFinishReason::STOP;
+        if (!legacyExecutionContext->results.finish_reasons.empty()) {
+            finishReason = legacyExecutionContext->results.finish_reasons[0];
+        }
+        std::string serializedChunk = executionContext->apiHandler->serializeStreamingChunk(lastTextChunk, finishReason);
         if (!serializedChunk.empty()) {
             executionContext->response = wrapTextInServerSideEventMessage(serializedChunk);
         }

src/test/llm/llmnode_test.cpp

@@ -2685,7 +2685,7 @@ INSTANTIATE_TEST_SUITE_P(
     ::testing::Values(
         // params:     model name, generate expected output, check logprobs, check finish reason, test speculative decoding, supports empty handshake msg
         TestParameters{"lm_cb_regular", true, true, true, false, true},
-        TestParameters{"lm_legacy_regular", false, false, false, false, false},
+        TestParameters{"lm_legacy_regular", false, false, true, false, false},
         TestParameters{"vlm_cb_regular", false, true, true, false, true},
         TestParameters{"vlm_legacy_regular", false, false, false, false, false}));