diff --git a/grpc_servicer/pyproject.toml b/grpc_servicer/pyproject.toml
index 28b0c3120..0f37d04ec 100644
--- a/grpc_servicer/pyproject.toml
+++ b/grpc_servicer/pyproject.toml
@@ -30,7 +30,7 @@ classifiers = [
 ]
 
 [project.optional-dependencies]
-vllm = ["vllm>=0.17.0"]
+vllm = ["vllm>=0.19.0"]
 sglang = ["sglang>=0.5.10rc0"]
 
 [project.urls]
diff --git a/grpc_servicer/smg_grpc_servicer/vllm/servicer.py b/grpc_servicer/smg_grpc_servicer/vllm/servicer.py
index 2a4acfe01..31d1be63e 100755
--- a/grpc_servicer/smg_grpc_servicer/vllm/servicer.py
+++ b/grpc_servicer/smg_grpc_servicer/vllm/servicer.py
@@ -15,17 +15,14 @@
 from transformers import BatchFeature
 from vllm import PoolingParams, SamplingParams, TokensPrompt
 from vllm.engine.protocol import EngineClient
-from vllm.inputs import token_inputs
+from vllm.inputs.engine import MultiModalInput as VllmMultiModalInput
+from vllm.inputs.engine import mm_input, tokens_input
 from vllm.logger import init_logger
 from vllm.logprobs import PromptLogprobs, SampleLogprobs
 from vllm.multimodal.inputs import (
     MultiModalFieldConfig,
     MultiModalKwargsItems,
     PlaceholderRange,
-    mm_inputs,
-)
-from vllm.multimodal.inputs import (
-    MultiModalInputs as VllmMultiModalInputs,
 )
 from vllm.outputs import CompletionOutput, RequestOutput
 from vllm.sampling_params import RequestOutputKind, StructuredOutputsParams
@@ -210,7 +207,7 @@ async def Embed(
             if not request.HasField("tokenized"):
                 raise ValueError("EmbedRequest requires tokenized input")
 
-            prompt = token_inputs(
+            prompt = tokens_input(
                 prompt_token_ids=list(request.tokenized.input_ids),
                 prompt=request.tokenized.original_text or None,
             )
@@ -367,8 +364,8 @@ def _build_preprocessed_mm_inputs(
         self,
         tokenized: vllm_engine_pb2.TokenizedInput,
         mm_proto: vllm_engine_pb2.MultimodalInputs,
-    ) -> VllmMultiModalInputs:
-        """Build vLLM MultiModalInputs from preprocessed proto data.
+    ) -> VllmMultiModalInput:
+        """Build vLLM MultiModalInput from preprocessed proto data.
 
         Bypasses HF processor entirely — pixel values and model-specific
         tensors were already computed by the Rust router.  Field layouts
@@ -450,7 +447,7 @@ def _build_preprocessed_mm_inputs(
                 )
             mm_placeholders["image"] = placeholders
 
-        return mm_inputs(
+        return mm_input(
             prompt_token_ids=prompt_token_ids,
             mm_kwargs=mm_kwargs,
             mm_hashes=mm_hashes,