chore: remove unused variables in pyexecutor (NVIDIA#6280)

Signed-off-by: junq <[email protected]>

Author: QiJune

tensorrt_llm/_torch/pyexecutor/executor_request_queue.py

@@ -10,7 +10,6 @@
 import torch
 
 from tensorrt_llm._utils import nvtx_range
-from tensorrt_llm.bindings.executor import RequestType
 
 from ..distributed import Distributed
 from .llm_request import ExecutorRequest, executor_request_to_llm_request
@@ -61,7 +60,6 @@ def __init__(self, dist: Distributed, enable_attention_dp: bool,
         self.num_fetch_requests_cur_rank = 0
         self.expected_num_active_requests = 0
         self.new_active_requests_queue_latency_ms = 0
-        self.has_context_request = False
         self.is_shutdown = False
         self.should_exclude_last_generation_logits = False
 
@@ -318,7 +316,6 @@ def _balance_requests_across_ranks(
             self, new_requests: List[RequestQueueItem],
             all_ranks_num_active_requests: List[int]) -> List[RequestQueueItem]:
         """Balance requests across ranks for attention DP."""
-        self.has_context_request = False
         new_requests_cur_rank = []
 
         if new_requests and self.expected_num_active_requests > all_ranks_num_active_requests[
@@ -364,15 +361,6 @@ def _balance_requests_across_ranks(
                 elif val.rank == self.dist.tp_rank:
                     break
 
-            # Check for context requests
-            if self.is_disaggregated:
-                for req_item in new_requests_cur_rank:
-                    if req_item.request.request_type == RequestType.REQUEST_TYPE_CONTEXT_ONLY:
-                        self.has_context_request = True
-                        break
-            else:
-                self.has_context_request = len(new_requests_cur_rank) > 0
-
         return new_requests_cur_rank
 
     def _collect_py_objects_from_requests(

tensorrt_llm/_torch/pyexecutor/py_executor.py

@@ -169,7 +169,6 @@ def __init__(self,
         self.draft_model_engine = draft_model_engine
 
         # enqueue and _fetch_new_requests used data
-        self.active = True
         self.next_req_id = max_batch_size  # The first max_batch_size request IDs are reserved for dummy requests
         self.max_beam_width = max_beam_width
         self.max_draft_len = max_draft_len
@@ -196,7 +195,6 @@ def __init__(self,
         self.max_num_active_requests = model_engine.get_max_num_sequences()
         self.active_requests: List[LlmRequest] = []
         self.expected_num_active_requests = 0
-        self.has_context_request = False
         self.ctx_in_transmission_requests = []
         self.previous_batch: Optional[BatchState] = None
         self.num_scheduled_requests: int = 0
@@ -1148,7 +1146,7 @@ def _check_disagg_gen_transfer_status(self):
     @nvtx_range("_pad_attention_dp_dummy_request")
     def _pad_attention_dp_dummy_request(self):
         """
-        Pad with a dummy request, if required, to ensure every attention_dp rank has at least one active request.
+        Pad with a generation dummy request, if required, to ensure every attention_dp rank has at least one active request.
         """
         if not self.enable_attention_dp:
             return
@@ -1166,8 +1164,8 @@ def _pad_attention_dp_dummy_request(self):
         if self.expected_num_active_requests - num_active_request > 0 and num_active_request == 0:
             llm_request = self.kv_cache_manager.add_dummy_requests(
                 request_ids=[0],
-                is_gen=not self.has_context_request,
-                prepare_resource=not self.has_context_request,
+                is_gen=True,
+                prepare_resource=True,
                 max_num_draft_tokens=self.max_draft_len,
             )[0]
             llm_request.is_attention_dp_dummy = True