Fix chunked prefill condition (#1594)

python/sglang/srt/managers/schedule_policy.py

@@ -249,7 +249,10 @@ class PrefillAdder:
                 return AddReqResult.NO_TOKEN
             tokens_freed += tokens_occupied
 
-        if req.extend_input_len <= self.rem_chunk_tokens:
+        if (
+            self.rem_chunk_tokens is None
+            or req.extend_input_len <= self.rem_chunk_tokens
+        ):
             self.can_run_list.append(req)
             self._prefill_one_req(
                 0,

test/srt/test_chunked_prefill.py

@@ -8,6 +8,7 @@ from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     popen_launch_server,
+    run_bench_serving,
 )
 
 
@@ -62,6 +63,16 @@ class TestChunkedPrefill(unittest.TestCase):
             disable_radix_cache=False, enable_mixed_chunk=False, chunked_prefill_size=-1
         )
 
+    def test_no_chunked_prefill_without_radix_cache(self):
+        res = run_bench_serving(
+            model=DEFAULT_MODEL_NAME_FOR_TEST,
+            num_prompts=10,
+            request_rate=float("inf"),
+            other_server_args=["--disable-radix-cache", "--chunked-prefill-size", "-1"],
+        )
+
+        assert res["completed"] == 10
+
 
 if __name__ == "__main__":
     unittest.main()