[Bugfix] Fix Intern / Radio Pooling Tests

vllm/model_executor/models/vision.py

@@ -10,7 +10,12 @@
 import torch
 from transformers import PretrainedConfig
 
-from vllm.config import MultiModalConfig, VllmConfig, get_current_vllm_config
+from vllm.config import (
+    ModelConfig,
+    MultiModalConfig,
+    VllmConfig,
+    get_current_vllm_config,
+)
 from vllm.distributed import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,
@@ -95,20 +100,32 @@ def _get_vit_attn_backend(
     )
 
 
+def _get_current_mm_config() -> MultiModalConfig | None:
+    """
+    Get the current MultiModalConfig if one exists.
+    """
+    try:
+        vllm_config: VllmConfig = get_current_vllm_config()
+    except AssertionError:
+        multimodal_config = None
+    else:
+        # We have a vLLM config; if we don't have a model config for
+        # any reason, set the MM config to None by default as well.
+        model_config: ModelConfig | None = vllm_config.model_config
+        multimodal_config: MultiModalConfig | None = (
+            model_config.multimodal_config if model_config is not None else None
+        )
+    return multimodal_config
+
+
 def get_vit_attn_backend(
     head_size: int,
     dtype: torch.dtype,
 ) -> AttentionBackendEnum:
     """
     Get the attention backend for Vision Transformer.
     """
-    try:
-        vllm_config: VllmConfig = get_current_vllm_config()
-        multimodal_config: MultiModalConfig | None = (
-            vllm_config.model_config.multimodal_config
-        )
-    except AssertionError:
-        multimodal_config = None
+    multimodal_config: MultiModalConfig | None = _get_current_mm_config()
 
     attn_backend_override = (
         multimodal_config.mm_encoder_attn_backend
@@ -127,13 +144,7 @@ def is_vit_use_data_parallel():
     """
     Get the tensor parallel type for Vision Transformer.
     """
-    try:
-        vllm_config: VllmConfig = get_current_vllm_config()
-        multimodal_config: MultiModalConfig | None = (
-            vllm_config.model_config.multimodal_config
-        )
-    except AssertionError:
-        multimodal_config = None
+    multimodal_config: MultiModalConfig | None = _get_current_mm_config()
 
     mm_encoder_tp_mode = (
         multimodal_config.mm_encoder_tp_mode if multimodal_config is not None else None