[BugFix] Fix mm_encoder_attn_backend arg type checking (#28599)

Signed-off-by: Nick Hill <nhill@redhat.com>

.buildkite/test-pipeline.yaml

@@ -57,7 +57,7 @@ steps:
   - pytest -v -s -m 'not cpu_test' multimodal
   - pytest -v -s utils_
 
-- label: Async Engine, Inputs, Utils, Worker Test (CPU) # 4 mins
+- label: Async Engine, Inputs, Utils, Worker, Config Test (CPU) # 4 mins
   timeout_in_minutes: 10
   source_file_dependencies:
   - vllm/
@@ -66,6 +66,7 @@ steps:
   - tests/multimodal
   - tests/standalone_tests/lazy_imports.py
   - tests/transformers_utils
+  - tests/config
   no_gpu: true
   commands:
   - python3 standalone_tests/lazy_imports.py
@@ -73,6 +74,7 @@ steps:
   - pytest -v -s test_outputs.py
   - pytest -v -s -m 'cpu_test' multimodal
   - pytest -v -s transformers_utils
+  - pytest -v -s config
 
 - label: Python-only Installation Test # 10min
   timeout_in_minutes: 20

vllm/config/multimodal.py

@@ -170,6 +170,9 @@ class MultiModalConfig:
     def _validate_mm_encoder_attn_backend(
         cls, value: str | AttentionBackendEnum | None
     ) -> AttentionBackendEnum | None:
+        # We need to import the real type here (deferred to avoid circular import).
+        from vllm.attention.backends.registry import AttentionBackendEnum
+
         if value is None or isinstance(value, AttentionBackendEnum):
             return value