[bugfix] avoid attention padding tokens computation in pcg (#17706)

test/registered/scheduler/test_priority_scheduling.py

@@ -41,8 +41,6 @@ class TestPriorityScheduling(CustomTestCase):
                 "--max-queued-requests",  # Enforce max queued request number is 3
                 "3",
                 "--enable-priority-scheduling",  # Enable priority scheduling
-                # Disable PCG to avoid padding in flashinfer backend. Ref: https://github.com/sgl-project/sglang/pull/21452
-                "--disable-piecewise-cuda-graph",
             ),
             return_stdout_stderr=(cls.stdout, cls.stderr),
         )
@@ -249,7 +247,6 @@ class TestPrioritySchedulingMultipleRunningRequests(CustomTestCase):
                 "--max-queued-requests",  # Enforce max queued request number is 3
                 "3",
                 "--enable-priority-scheduling",  # Enable priority scheduling
-                "--disable-piecewise-cuda-graph",
             ),
             return_stdout_stderr=(cls.stdout, cls.stderr),
         )

test/registered/scheduler/test_scheduler_control.py

@@ -318,8 +318,6 @@ class TestAbortWithWaitingTimeout(WaitingTimeoutMixin, CustomTestCase):
                 timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
                 other_args=[
                     "--max-running-requests=1",
-                    # Disable PCG to avoid padding in flashinfer backend. Ref: https://github.com/sgl-project/sglang/pull/21452
-                    "--disable-piecewise-cuda-graph",
                 ],
             )