[BugFix] Fix PP/async scheduling with pooling models (#28899)

Signed-off-by: Nick Hill <nhill@redhat.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

vllm/v1/engine/core.py

@@ -184,6 +184,7 @@ class EngineCore:
             vllm_config.ec_transfer_config is not None
             and vllm_config.ec_transfer_config.is_ec_producer
         )
+        self.is_pooling_model = vllm_config.model_config.runner_type == "pooling"
 
         self.request_block_hasher: Callable[[Request], list[BlockHash]] | None = None
         if vllm_config.cache_config.enable_prefix_caching or kv_connector is not None:
@@ -392,7 +393,7 @@ class EngineCore:
             if not self.ec_producer:
                 model_executed = scheduler_output.total_num_scheduled_tokens > 0
 
-            if not model_executed:
+            if self.is_pooling_model or not model_executed:
                 # No sampling required (no requests scheduled).
                 future = cast(Future[ModelRunnerOutput], exec_future)
             else:

vllm/v1/executor/ray_executor.py

@@ -99,9 +99,9 @@ class RayDistributedExecutor(Executor):
         # KV connector setup
         self.has_connector = self.vllm_config.kv_transfer_config is not None
 
-        self.ec_producer = (
-            self.vllm_config.ec_transfer_config is not None
-            and self.vllm_config.ec_transfer_config.is_ec_producer
+        self.uses_sampler = self.vllm_config.model_config.runner_type != "pooling" and (
+            self.vllm_config.ec_transfer_config is None
+            or not self.vllm_config.ec_transfer_config.is_ec_producer
         )
 
         self.scheduler_output: SchedulerOutput | None = None
@@ -401,7 +401,7 @@ class RayDistributedExecutor(Executor):
                 "after execute_model() returns None."
             )
 
-        if self.ec_producer or not scheduler_output.total_num_scheduled_tokens:
+        if not self.uses_sampler or not scheduler_output.total_num_scheduled_tokens:
             # Model will not execute, call model runner immediately.
             return self._execute_dag(scheduler_output, None, non_block)