The problem is that the optimization no longer checks for pending signals in TARGET(CALL_NO_KW_BUILTIN_FAST). The patch below fix my issue. I guess that other opcode needs an additional CHECK_EVAL_BREAKER().

diff --git a/Python/ceval.c b/Python/ceval.c
index 9aaddd99eda..7cc0f805366 100644
--- a/Python/ceval.c
+++ b/Python/ceval.c
@@ -4822,6 +4822,7 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, InterpreterFrame *frame, int thr
                 */
                 goto error;
             }
+            CHECK_EVAL_BREAKER();
             DISPATCH();
         }