feat: Adding nixl read() multimodal support for vLLM backend

components/src/dynamo/vllm/handlers.py

@@ -98,6 +98,8 @@ def __init__(
         self.engine_monitor = VllmEngineMonitor(runtime, engine)
         self.image_loader = ImageLoader()
         self._connector = None  # Lazy-initialized on first Decoded variant
+        self.temp_dirs: list[tempfile.TemporaryDirectory] = []
+        self.model_max_len = model_max_len
 
     @abstractmethod
     async def generate(self, request, context) -> AsyncGenerator[dict, None]:

lib/llm/Cargo.toml

@@ -152,6 +152,8 @@ image = { version = "0.25", features = ["serde"] }
 tokio-rayon = {version = "2" }
 ndarray = { version = "0.16" }
 flate2 = { version = "1.0" }
+ndarray-npy = { version = "0.9" }
+ndarray-interp = { version = "0.5" }
 
 # Publishers
 zeromq = "0.4.1"