[Core][TPU] Support TPU Data Parallalism (vllm-project#27365)

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

Author: wenxindongwork

vllm/entrypoints/llm.py

@@ -67,6 +67,7 @@
     RequestOutput,
     ScoringRequestOutput,
 )
+from vllm.platforms import current_platform
 from vllm.pooling_params import PoolingParams
 from vllm.sampling_params import BeamSearchParams, RequestOutputKind, SamplingParams
 from vllm.tasks import PoolingTask
@@ -289,7 +290,11 @@ def __init__(
         # warn about single-process data parallel usage.
         _dp_size = int(kwargs.get("data_parallel_size", 1))
         _distributed_executor_backend = kwargs.get("distributed_executor_backend")
-        if _dp_size > 1 and not _distributed_executor_backend == "external_launcher":
+        if (
+            _dp_size > 1
+            and not _distributed_executor_backend == "external_launcher"
+            and not current_platform.is_tpu()
+        ):
             raise ValueError(
                 f"LLM(data_parallel_size={_dp_size}) is not supported for single-"
                 "process usage and may hang. Please use "