[Model] enable data parallel for Llama4 vision encoder (vllm-project#18368)

Signed-off-by: yzhen <[email protected]>
Co-authored-by: yZhen <[email protected]>
Co-authored-by: yzhen <[email protected]>

Author: 3 people

vllm/config.py

@@ -1790,6 +1790,10 @@ class is dynamically inherited by the worker class. This is used to inject
     rank: int = 0
     """Global rank in distributed setup."""
 
+    enable_multimodal_encoder_data_parallel: bool = False
+    """ Use data parallelism instead of tensor parallelism for vision encoder. 
+    Only support LLama4 for now"""
+
     @property
     def world_size_across_dp(self) -> int:
         """world_size_across_dp is TPxPPxDP, it is the size of the world

vllm/engine/arg_utils.py

@@ -423,6 +423,9 @@ class EngineArgs:
     use_tqdm_on_load: bool = LoadConfig.use_tqdm_on_load
     pt_load_map_location: str = LoadConfig.pt_load_map_location
 
+    enable_multimodal_encoder_data_parallel: bool = \
+        ParallelConfig.enable_multimodal_encoder_data_parallel
+
     def __post_init__(self):
         # support `EngineArgs(compilation_config={...})`
         # without having to manually construct a
@@ -637,6 +640,9 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
                                     **parallel_kwargs["worker_cls"])
         parallel_group.add_argument("--worker-extension-cls",
                                     **parallel_kwargs["worker_extension_cls"])
+        parallel_group.add_argument(
+            "--enable-multimodal-encoder-data-parallel",
+            **parallel_kwargs["enable_multimodal_encoder_data_parallel"])
 
         # KV cache arguments
         cache_kwargs = get_kwargs(CacheConfig)
@@ -1078,6 +1084,8 @@ def create_engine_config(
             distributed_executor_backend=self.distributed_executor_backend,
             worker_cls=self.worker_cls,
             worker_extension_cls=self.worker_extension_cls,
+            enable_multimodal_encoder_data_parallel=self.
+            enable_multimodal_encoder_data_parallel,
         )
 
         speculative_config = self.create_speculative_config(