diff --git a/components/src/dynamo/vllm/handlers.py b/components/src/dynamo/vllm/handlers.py
index 36305c283051..c8a9f86725e2 100644
--- a/components/src/dynamo/vllm/handlers.py
+++ b/components/src/dynamo/vllm/handlers.py
@@ -4035,16 +4035,7 @@ async def generate(self, request, context):
         # Use context ID for request tracking and correlation with decode phase
         request_id = context.id()
         logger.debug("Prefill Request ID: %s", request_id)
-        try:
-            self._multimodal_request_processor.validate_multimodal_request(request)
-        except ValueError as exc:
-            logger.error("Request %s: %s", request_id, exc)
-            yield {
-                "status": "error",
-                "message": str(exc),
-                "disaggregated_params": None,
-            }
-            return
+        self._multimodal_request_processor.validate_multimodal_request(request)
 
         # Token-in-token-out mode: internal protocol format
         with time_and_log_code_section(f"[PREFILL] request: {request_id} generate"):
diff --git a/components/src/dynamo/vllm/multimodal_utils/request_processor.py b/components/src/dynamo/vllm/multimodal_utils/request_processor.py
index a12513cb2d53..2b9da813b39a 100644
--- a/components/src/dynamo/vllm/multimodal_utils/request_processor.py
+++ b/components/src/dynamo/vllm/multimodal_utils/request_processor.py
@@ -38,6 +38,7 @@
 )
 from dynamo.common.multimodal.video_loader import VideoLoader
 from dynamo.common.utils import nvtx_utils as _nvtx
+from dynamo.llm.exceptions import InvalidArgument
 
 from .hash_utils import compute_mm_uuids_from_images
 from .model import ModelFamily, construct_qwen_decode_mm_data, resolve_model_family
@@ -462,8 +463,8 @@ def _expand_kimi_k3_pads(
         return expanded
 
     @staticmethod
-    def _multimodal_disabled_error() -> ValueError:
-        return ValueError(
+    def _multimodal_disabled_error() -> InvalidArgument:
+        return InvalidArgument(
             "Received multimodal data but multimodal processing is not enabled. "
             "Use --enable-multimodal flag to enable multimodal processing."
         )
diff --git a/components/src/dynamo/vllm/tests/multimodal_utils/test_vllm_request_processor.py b/components/src/dynamo/vllm/tests/multimodal_utils/test_vllm_request_processor.py
index 65003097c401..d8b16d214995 100644
--- a/components/src/dynamo/vllm/tests/multimodal_utils/test_vllm_request_processor.py
+++ b/components/src/dynamo/vllm/tests/multimodal_utils/test_vllm_request_processor.py
@@ -412,7 +412,7 @@ async def test_rejects_malformed_encoder_image_item_before_dispatch():
 async def test_rejects_media_when_multimodal_is_disabled():
     processor = _processor(enabled=False)
 
-    with pytest.raises(ValueError, match="--enable-multimodal"):
+    with pytest.raises(mod.InvalidArgument, match="--enable-multimodal"):
         await _prepare_prompt(
             processor,
             {
@@ -424,7 +424,7 @@ async def test_rejects_media_when_multimodal_is_disabled():
             DisaggregationMode.AGGREGATED,
         )
 
-    with pytest.raises(ValueError, match="--enable-multimodal"):
+    with pytest.raises(mod.InvalidArgument, match="--enable-multimodal"):
         await _prepare_prompt(
             processor,
             {
@@ -436,7 +436,7 @@ async def test_rejects_media_when_multimodal_is_disabled():
             DisaggregationMode.AGGREGATED,
         )
 
-    with pytest.raises(ValueError, match="--enable-multimodal"):
+    with pytest.raises(mod.InvalidArgument, match="--enable-multimodal"):
         await _prepare_prompt(
             processor,
             {
@@ -456,7 +456,7 @@ async def test_decode_cannot_hide_disabled_media_with_expanded_tokens():
         enabled=False,
     )
 
-    with pytest.raises(ValueError, match="--enable-multimodal"):
+    with pytest.raises(mod.InvalidArgument, match="--enable-multimodal"):
         await _prepare_prompt(
             processor,
             {
diff --git a/components/src/dynamo/vllm/tests/test_vllm_worker_handler.py b/components/src/dynamo/vllm/tests/test_vllm_worker_handler.py
index fad0a4e58683..c679c2f3a5f1 100644
--- a/components/src/dynamo/vllm/tests/test_vllm_worker_handler.py
+++ b/components/src/dynamo/vllm/tests/test_vllm_worker_handler.py
@@ -675,7 +675,7 @@ async def test_text_mode_rejects_multimodal_input_when_disabled(
         handler.use_vllm_tokenizer = True
         handler._multimodal_request_processor.enable_multimodal = False
 
-        with pytest.raises(ValueError, match="--enable-multimodal"):
+        with pytest.raises(mod.InvalidArgument, match="--enable-multimodal"):
             async for _ in handler.generate(request_payload, MagicMock()):
                 pass
 
@@ -984,26 +984,21 @@ async def test_prefill_delegates_mode_policy_to_shared_processor():
 
 
 @pytest.mark.asyncio
-async def test_prefill_returns_structured_error_when_multimodal_is_disabled():
+async def test_prefill_raises_typed_error_when_multimodal_is_disabled():
     handler = mod.PrefillWorkerHandler.__new__(mod.PrefillWorkerHandler)
     processor = SimpleNamespace(
         validate_multimodal_request=MagicMock(
-            side_effect=ValueError("use --enable-multimodal")
+            side_effect=mod.InvalidArgument("use --enable-multimodal")
         )
     )
     handler._multimodal_request_processor = processor
     context = MagicMock()
     context.id.return_value = "request-prefill-disabled"
 
-    chunks = [chunk async for chunk in handler.generate({}, context)]
+    with pytest.raises(mod.InvalidArgument, match="use --enable-multimodal"):
+        [chunk async for chunk in handler.generate({}, context)]
 
-    assert chunks == [
-        {
-            "status": "error",
-            "message": "use --enable-multimodal",
-            "disaggregated_params": None,
-        }
-    ]
+    processor.validate_multimodal_request.assert_called_once_with({})
 
 
 # ── Deferred abort (disagg decode KV-transfer safety) tests ────────