[Bugfix] Strengthen the check of X-data-parallel-rank in Hybrid LB mode (#32314)

Signed-off-by: Tianchen Ding <dtcccc@linux.alibaba.com>

vllm/v1/engine/coordinator.py

@@ -62,12 +62,10 @@ class DPCoordinator:
         assert dp_size > 1, "Coordinator only used for data parallel"
 
         host = parallel_config.data_parallel_master_ip
-        external_lb = parallel_config.data_parallel_external_lb
-        hybrid_lb = parallel_config.data_parallel_hybrid_lb
 
         # Assume coordinator is colocated with front-end procs when not in
         # either external or hybrid DP LB mode.
-        local_only = not (external_lb or hybrid_lb)
+        local_only = not parallel_config.local_engines_only
         front_publish_address = get_engine_client_zmq_addr(
             local_only=local_only, host=host
         )

vllm/v1/engine/core_client.py

@@ -507,12 +507,7 @@ class MPClient(EngineCoreClient):
             offline_mode = parallel_config.data_parallel_rank_local is not None
             # Client manages local+remote EngineCores in pure internal LB case.
             # Client manages local EngineCores in hybrid and external LB case.
-            local_engines_only = (
-                parallel_config.data_parallel_hybrid_lb
-                or parallel_config.data_parallel_external_lb
-            )
-
-            num_ranks = dp_local_size if local_engines_only else dp_size
+            num_ranks = dp_local_size if parallel_config.local_engines_only else dp_size
             self.engine_ranks_managed = (
                 [dp_rank] if offline_mode else list(range(dp_rank, dp_rank + num_ranks))
             )

vllm/v1/engine/input_processor.py

@@ -458,13 +458,14 @@ class InputProcessor:
         self._validate_lora(lora_request)
         self._validate_params(params)
 
-        data_parallel_size = self.vllm_config.parallel_config.data_parallel_size
-        if data_parallel_rank is not None and not (
-            0 <= data_parallel_rank < data_parallel_size
-        ):
+        parallel_config = self.vllm_config.parallel_config
+        dp_size = parallel_config.data_parallel_size
+        dp_local_size = parallel_config.data_parallel_size_local
+        num_ranks = dp_local_size if parallel_config.local_engines_only else dp_size
+        if data_parallel_rank is not None and not (0 <= data_parallel_rank < num_ranks):
             raise ValueError(
                 f"data_parallel_rank {data_parallel_rank} "
-                f"is out of range [0, {data_parallel_size})."
+                f"is out of range [0, {num_ranks})."
             )
 
         if arrival_time is None:

vllm/v1/engine/utils.py

@@ -787,10 +787,7 @@ def launch_core_engines(
     local_start_index = parallel_config.data_parallel_rank_local
     dp_rank = parallel_config.data_parallel_rank
     host = parallel_config.data_parallel_master_ip
-    local_engines_only = (
-        parallel_config.data_parallel_hybrid_lb
-        or parallel_config.data_parallel_external_lb
-    )
+    local_engines_only = parallel_config.local_engines_only
 
     # In offline mode there is an LLM instance per DP rank and
     # one core engine per LLM, see