feat: pp balance

vllm/envs.py

@@ -291,6 +291,7 @@ if TYPE_CHECKING:
     VLLM_USE_FUSED_RMS_ROPE: bool = False
     VLLM_USE_FUSED_FILL_RMS_CAT: bool = False
     VLLM_W8A8_BACKEND: int = 3
+    VLLM_USE_PP_BALANCE = True
     VLLM_REJECT_SAMPLE_OPT: bool = False
 
 
@@ -1831,6 +1832,9 @@ environment_variables: dict[str, Callable[[], Any]] = {
     "VLLM_USE_FUSED_FILL_RMS_CAT":
         lambda: (os.environ.get("VLLM_USE_FUSED_FILL_RMS_CAT", "False").lower() in
                  ("true", "1")),
+    "VLLM_USE_PP_BALANCE":
+        lambda: (os.environ.get("VLLM_USE_PP_BALANCE", "True").lower() in
+                 ("true", "1")),
     # W8A8 GEMM backend selection for vLLM quantized models.
     # lightop/triton: 1
     # cutlass: 2 (will remove in the future)

vllm/v1/core/sched/scheduler.py

@@ -343,7 +343,10 @@ class Scheduler(SchedulerInterface):
 
         # For logging.
         scheduled_timestamp = time.monotonic()
-
+        if self.use_pp and envs.VLLM_USE_PP_BALANCE:
+            pipeline_size = self.parallel_config.pipeline_parallel_size
+            max_batch_running = (len(self.waiting) + len(self.running)
+                                 + pipeline_size - 1 ) // pipeline_size
         # First, schedule the RUNNING requests.
         req_index = 0
         while req_index < len(self.running) and token_budget > 0:
@@ -352,7 +355,12 @@ class Scheduler(SchedulerInterface):
             # do not schedule another step for the same request while it still has
             # output placeholders for PP.
             # TODO: support PP + async scheduling without this limit
-            if self.use_pp and request.num_output_placeholders > 0:
+            if self.use_pp:
+                if (envs.VLLM_USE_PP_BALANCE and
+                        len(scheduled_new_reqs) + len(scheduled_resumed_reqs)
+                        + len(scheduled_running_reqs) >= max_batch_running):
+                    break
+                if request.num_output_placeholders > 0:
                     req_index += 1
                     continue
 
@@ -543,7 +551,10 @@ class Scheduler(SchedulerInterface):
             while self.waiting and token_budget > 0:
                 if len(self.running) == self.max_num_running_reqs:
                     break
-
+                if (self.use_pp and envs.VLLM_USE_PP_BALANCE and
+                    len(scheduled_new_reqs) + len(scheduled_resumed_reqs)
+                    + len(scheduled_running_reqs) >= max_batch_running):
+                        break
                 request = self.waiting.peek_request()
                 # KVTransfer: skip request if still waiting for remote kvs.
                 if request.status == RequestStatus.WAITING_FOR_REMOTE_KVS: