Add a test case to test retract (#1662)

python/sglang/srt/managers/schedule_batch.py

@@ -590,9 +590,11 @@ class ScheduleBatch:
 
         retracted_reqs = []
         seq_lens_cpu = self.seq_lens.cpu().numpy()
+        first_iter = True
         while (
             self.token_to_kv_pool.available_size()
             < len(sorted_indices) * global_config.retract_decode_steps
+            or first_iter
         ):
             if len(sorted_indices) == 1:
                 # Corner case: only one request left
@@ -601,6 +603,7 @@ class ScheduleBatch:
                 ), "No space left for only one request"
                 break
 
+            first_iter = False
             idx = sorted_indices.pop()
             req = self.reqs[idx]
             retracted_reqs.append(req)

python/sglang/srt/managers/scheduler.py

@@ -77,6 +77,9 @@ logger = logging.getLogger(__name__)
 # Crash on warning if we are running CI tests
 crash_on_warning = os.getenv("SGLANG_IS_IN_CI", "false") == "true"
 
+# Test retract decode
+test_retract = os.getenv("SGLANG_TEST_RETRACT", "false") == "true"
+
 
 class Scheduler:
     """A scheduler that manages a tensor parallel GPU worker."""
@@ -611,10 +614,11 @@ class Scheduler:
         return new_batch
 
     def update_running_batch(self):
+        global test_retract
         batch = self.running_batch
 
         # Check if decode out of memory
-        if not batch.check_decode_mem():
+        if not batch.check_decode_mem() or (test_retract and batch.batch_size() > 10):
             old_ratio = self.new_token_ratio
 
             retracted_reqs, new_token_ratio = batch.retract_decode()

test/srt/run_suite.py

@@ -17,6 +17,7 @@ suites = {
         "test_large_max_new_tokens.py",
         "test_openai_server.py",
         "test_pytorch_sampling_backend.py",
+        "test_retract_decode.py",
         "test_server_args.py",
         "test_skip_tokenizer_init.py",
         "test_srt_engine.py",

test/srt/test_retract_decode.py

+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_child_process
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
+
+
+class TestRetractDecode(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = DEFAULT_MODEL_NAME_FOR_TEST
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.process = popen_launch_server(
+            cls.model, cls.base_url, timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH
+        )
+
+    @classmethod
+    def tearDownClass(cls):
+        kill_child_process(cls.process.pid)
+
+    def test_mmlu(self):
+        args = SimpleNamespace(
+            base_url=self.base_url,
+            model=self.model,
+            eval_name="mmlu",
+            num_examples=64,
+            num_threads=32,
+        )
+
+        metrics = run_eval(args)
+        assert metrics["score"] >= 0.65
+
+
+if __name__ == "__main__":
+    unittest.main()