From 0863f35bae15b23f0ee935725316358ade3845df Mon Sep 17 00:00:00 2001
From: phantomlei3 <phantomlei3@gmail.com>
Date: Thu, 16 Jul 2026 11:19:01 +0800
Subject: [PATCH] bugfix: mark prefill complete after empty token decoding.

---
 .../scheduler/continuous_scheduler_test.cpp   | 30 +++++++++++++++++++
 .../scheduler/async_response_processor.cpp    | 12 ++++----
 2 files changed, 36 insertions(+), 6 deletions(-)

diff --git a/tests/core/scheduler/continuous_scheduler_test.cpp b/tests/core/scheduler/continuous_scheduler_test.cpp
index 6535de92ab..297444724e 100644
--- a/tests/core/scheduler/continuous_scheduler_test.cpp
+++ b/tests/core/scheduler/continuous_scheduler_test.cpp
@@ -1011,6 +1011,36 @@ TEST(ContinuousSchedulerTest, BatchRejectedStreamsCancelAtSchedulingBoundary) {
   EXPECT_TRUE(requests[1]->cancelled());
   EXPECT_TRUE(scheduler->get_running_requests().empty());
 }
+
+TEST(ContinuousSchedulerTest,
+     BatchStreamMarksPrefillFinishedWhenFirstTokenDecodesEmpty) {
+  ContinuousScheduler::Options opt =
+      create_scheduler_options(1024, 16, 0, 1024, 1);
+  auto engine = std::make_unique<FakeEngine>(32, 4);
+  auto scheduler = std::make_unique<TestContinuousScheduler>(engine.get(), opt);
+
+  auto request = generate_request_with_prompt_tokens({1, 2, 3, 4}, 4, 30000);
+  request->state().stream = true;
+  make_request_decode_ready(request);
+
+  bool callback_called = false;
+  bool outputs_empty = false;
+  bool finished_on_prefill_instance = false;
+  request->state().outputs_func =
+      [&](const std::vector<RequestOutput>& outputs) {
+        callback_called = true;
+        outputs_empty = outputs.size() == 1 && outputs[0].outputs.empty();
+        finished_on_prefill_instance =
+            outputs.size() == 1 && outputs[0].finished_on_prefill_instance;
+        return std::vector<bool>{true};
+      };
+
+  scheduler->reject_streams({request});
+
+  EXPECT_TRUE(callback_called);
+  EXPECT_TRUE(outputs_empty);
+  EXPECT_TRUE(finished_on_prefill_instance);
+}
 // ============== Async RL training: Pause/Resume tests ==============
 
 // TEST: pause()/resume() state transitions are correct and idempotent.
diff --git a/xllm/core/scheduler/async_response_processor.cpp b/xllm/core/scheduler/async_response_processor.cpp
index ab1aa157aa..635006fc2c 100644
--- a/xllm/core/scheduler/async_response_processor.cpp
+++ b/xllm/core/scheduler/async_response_processor.cpp
@@ -302,12 +302,12 @@ void AsyncResponseProcessor::batch_process_stream_requests(
         auto seq_output = seq->generate_streaming_output(size, *tokenizer_);
         if (seq_output.has_value()) {
           req_output->outputs.push_back(std::move(seq_output.value()));
-          if (seq->num_generated_tokens() == 1) {
-            // currently only support one sequence when enable_service_routing
-            // IMPROVE LATER: support enable_schedule_overlap in Default mode
-            // for stream request
-            req_output->finished_on_prefill_instance = true;
-          }
+        }
+        if (seq->num_generated_tokens() == 1) {
+          // currently only support one sequence when enable_service_routing
+          // IMPROVE LATER: support enable_schedule_overlap in Default mode
+          // for stream request
+          req_output->finished_on_prefill_instance = true;
         }
       }
       counter->decrement_count();