Add execute_begin method and reset is_tensor_computed to be false during each step

Author: leideng

ucm/sparse/kvcomp/kvcomp_hbm.py

@@ -11,6 +11,7 @@
 from vllm.forward_context import ForwardContext
 from vllm.v1.attention.backends.mla.common import MLACommonMetadata
 from vllm.v1.request import Request, RequestStatus
+from vllm.v1.core.sched.output import SchedulerOutput
 
 from ucm.logger import init_logger
 from ucm.sparse.base import (
@@ -459,6 +460,9 @@ def request_finished_in_scheduler(self, request_id: Union[int, str]):
         """
         pass
 
+    def execute_begin(self, scheduler_output: SchedulerOutput):
+        self.is_tensor_computed = False
+
     def estimate_num_slots_sparsed(self, request: Request) -> int:
         return INVALID_SLOT