[Misc] Move DP for ViT code inside model executor dir (#25459)

DarkLight1337 · web-flow · commit babad6e5dde3 · 2025-09-23T09:20:52.000Z
Signed-off-by: DarkLight1337 &lt;tlleungac@connect.ust.hk&gt;
diff --git a/tests/models/test_vision.py b/tests/models/test_vision.py
diff --git a/tests/multimodal/test_utils.py b/tests/multimodal/test_utils.py
diff --git a/vllm/model_executor/models/glm4_1v.py b/vllm/model_executor/models/glm4_1v.py
@@ -69,7 +69,6 @@
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate, PromptUpdateDetails)
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
-from vllm.multimodal.utils import run_dp_sharded_mrope_vision_model
 from vllm.platforms import _Backend
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.config import uses_mrope
@@ -83,7 +82,7 @@
 from .utils import (AutoWeightsLoader, WeightsMapper,
                     init_vllm_registered_model, maybe_prefix,
                     merge_multimodal_embeddings)
-from .vision import get_vit_attn_backend
+from .vision import get_vit_attn_backend, run_dp_sharded_mrope_vision_model
 
 logger = init_logger(__name__)
 
diff --git a/vllm/model_executor/models/idefics2_vision_model.py b/vllm/model_executor/models/idefics2_vision_model.py
@@ -34,7 +34,8 @@
                                                RowParallelLinear)
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
-from vllm.multimodal.utils import run_dp_sharded_vision_model
+
+from .vision import run_dp_sharded_vision_model
 
 
 class Idefics2VisionEmbeddings(nn.Module):
diff --git a/vllm/model_executor/models/intern_vit.py b/vllm/model_executor/models/intern_vit.py
@@ -28,7 +28,8 @@
                                                RowParallelLinear)
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
-from vllm.multimodal.utils import run_dp_sharded_vision_model
+
+from .vision import run_dp_sharded_vision_model
 
 NORM2FN = {
     'rms_norm': RMSNorm,
diff --git a/vllm/model_executor/models/kimi_vl.py b/vllm/model_executor/models/kimi_vl.py
@@ -76,13 +76,13 @@
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate)
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
-from vllm.multimodal.utils import run_dp_sharded_mrope_vision_model
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs import KimiVLConfig, MoonViTConfig
 from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekV2Config
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
 
 from .utils import PPMissingLayer, is_pp_missing_parameter, maybe_prefix
+from .vision import run_dp_sharded_mrope_vision_model
 
 
 # For dummy input only
diff --git a/vllm/model_executor/models/mllama4.py b/vllm/model_executor/models/mllama4.py
@@ -50,14 +50,14 @@
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate, PromptUpdateDetails)
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
-from vllm.multimodal.utils import run_dp_sharded_vision_model
 from vllm.sequence import IntermediateTensors
 from vllm.utils.tensor_schema import TensorSchema, TensorShape
 
 from .interfaces import MultiModalEmbeddings, SupportsMultiModal, SupportsPP
 from .llama4 import Llama4ForCausalLM
 from .utils import (AutoWeightsLoader, flatten_bn, maybe_prefix,
                     merge_multimodal_embeddings)
+from .vision import run_dp_sharded_vision_model
 
 
 class Llama4ImagePatchInputs(TensorSchema):
diff --git a/vllm/model_executor/models/qwen2_5_vl.py b/vllm/model_executor/models/qwen2_5_vl.py
@@ -59,7 +59,6 @@
 from vllm.model_executor.models.module_mapping import MultiModelKeys
 from vllm.multimodal import MULTIMODAL_REGISTRY
 from vllm.multimodal.inputs import MultiModalFieldConfig
-from vllm.multimodal.utils import run_dp_sharded_mrope_vision_model
 from vllm.platforms import _Backend
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.config import uses_mrope
@@ -74,7 +73,7 @@
 from .utils import (AutoWeightsLoader, WeightsMapper, cast_overflow_tensors,
                     init_vllm_registered_model, maybe_prefix,
                     merge_multimodal_embeddings)
-from .vision import get_vit_attn_backend
+from .vision import get_vit_attn_backend, run_dp_sharded_mrope_vision_model
 
 logger = init_logger(__name__)
 
diff --git a/vllm/model_executor/models/qwen2_vl.py b/vllm/model_executor/models/qwen2_vl.py
@@ -66,7 +66,6 @@
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate)
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
-from vllm.multimodal.utils import run_dp_sharded_mrope_vision_model
 from vllm.platforms import _Backend, current_platform
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.config import uses_mrope
@@ -78,7 +77,7 @@
 from .utils import (AutoWeightsLoader, WeightsMapper,
                     init_vllm_registered_model, maybe_prefix,
                     merge_multimodal_embeddings)
-from .vision import get_vit_attn_backend
+from .vision import get_vit_attn_backend, run_dp_sharded_mrope_vision_model
 
 logger = init_logger(__name__)
 
diff --git a/vllm/model_executor/models/qwen3_vl.py b/vllm/model_executor/models/qwen3_vl.py
@@ -83,7 +83,7 @@
 from .qwen3 import Qwen3ForCausalLM, Qwen3Model
 from .utils import (AutoWeightsLoader, PPMissingLayer, WeightsMapper,
                     maybe_prefix, merge_multimodal_embeddings)
-from .vision import get_vit_attn_backend
+from .vision import get_vit_attn_backend, run_dp_sharded_mrope_vision_model
 
 logger = init_logger(__name__)
 
@@ -1214,8 +1214,6 @@ def _process_image_input(
         else:
             pixel_values = image_input["pixel_values"].type(self.visual.dtype)
             if self.use_data_parallel:
-                from vllm.multimodal.utils import (
-                    run_dp_sharded_mrope_vision_model)
                 return run_dp_sharded_mrope_vision_model(self.visual,
                                                          pixel_values,
                                                          grid_thw_list,
@@ -1245,8 +1243,6 @@ def _process_video_input(
             pixel_values_videos = video_input["pixel_values_videos"].type(
                 self.visual.dtype)
             if self.use_data_parallel:
-                from vllm.multimodal.utils import (
-                    run_dp_sharded_mrope_vision_model)
                 return run_dp_sharded_mrope_vision_model(self.visual,
                                                          pixel_values_videos,
                                                          grid_thw_list,
diff --git a/vllm/model_executor/models/step3_vl.py b/vllm/model_executor/models/step3_vl.py
@@ -31,7 +31,6 @@
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate, PromptUpdateDetails)
 from vllm.multimodal.profiling import BaseDummyInputsBuilder
-from vllm.multimodal.utils import run_dp_sharded_vision_model
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs import Step3VisionEncoderConfig
 from vllm.transformers_utils.tokenizer import AnyTokenizer
@@ -40,6 +39,7 @@
 from .utils import (AutoWeightsLoader, WeightsMapper, flatten_bn,
                     init_vllm_registered_model, maybe_prefix,
                     merge_multimodal_embeddings)
+from .vision import run_dp_sharded_vision_model
 
 
 class Step3VLImagePixelInputs(TypedDict):
diff --git a/vllm/model_executor/models/vision.py b/vllm/model_executor/models/vision.py
diff --git a/vllm/multimodal/utils.py b/vllm/multimodal/utils.py
