update variable name + add --frontend-decoding flag

Signed-off-by: Krishnan Prashanth <[email protected]>

components/src/dynamo/vllm/args.py

@@ -65,6 +65,7 @@ class Config:
     multimodal_decode_worker: bool = False
     multimodal_encode_prefill_worker: bool = False
     mm_prompt_template: str = "USER: <image>\n<prompt> ASSISTANT:"
+    frontend_decoding: bool = False
     # dump config to file
     dump_config_to: Optional[str] = None
 
@@ -175,6 +176,16 @@ def parse_args() -> Config:
             "'USER: <image> please describe the image ASSISTANT:'."
         ),
     )
+    parser.add_argument(
+        "--frontend-decoding",
+        action="store_true",
+        help=(
+            "EXPERIMENTAL: Enable frontend decoding of multimodal images. "
+            "When enabled, images are decoded in the Rust frontend and transferred to the backend via NIXL RDMA. "
+            "Requires building Dynamo's Rust components with '--features media-nixl'. "
+            "Without this flag, images are decoded in the Python backend (default behavior)."
+        ),
+    )
     parser.add_argument(
         "--store-kv",
         type=str,

components/src/dynamo/vllm/main.py

@@ -16,8 +16,6 @@
 
 from dynamo.common.config_dump import dump_config
 from dynamo.llm import (
-    MediaDecoder,
-    MediaFetcher,
     ModelInput,
     ModelRuntimeConfig,
     ModelType,
@@ -311,6 +309,26 @@ async def register_vllm_model(
     data_parallel_size = getattr(vllm_config.parallel_config, "data_parallel_size", 1)
     runtime_config.data_parallel_size = data_parallel_size
 
+    # Conditionally enable frontend decoding if --frontend-decoding flag is set
+    media_decoder = None
+    media_fetcher = None
+    if config.frontend_decoding:
+        try:
+            from dynamo.llm import MediaDecoder, MediaFetcher
+
+            media_decoder = MediaDecoder()
+            media_fetcher = MediaFetcher()
+            logger.info(
+                "Frontend decoding enabled: images will be decoded in Rust frontend "
+                "and transferred via NIXL RDMA"
+            )
+        except ImportError as e:
+            raise RuntimeError(
+                "Frontend decoding (--frontend-decoding) requires building Dynamo's "
+                "Rust components with '--features media-nixl'. "
+                f"Import failed: {e}"
+            ) from e
+
     await register_llm(
         model_input,
         model_type,
@@ -321,8 +339,8 @@ async def register_vllm_model(
         migration_limit=migration_limit,
         runtime_config=runtime_config,
         custom_template_path=config.custom_jinja_template,
-        media_decoder=MediaDecoder(),
-        media_fetcher=MediaFetcher(),
+        media_decoder=media_decoder,
+        media_fetcher=media_fetcher,
     )
 
 

lib/llm/src/preprocessor/media/rdma.rs

@@ -124,9 +124,10 @@ pub fn get_nixl_metadata(agent: &NixlAgent, _storage: &SystemStorage) -> Result<
 
     // Compress metadata before base64 encoding (matches Python nixl_connect behavior)
     // Backend expects: b64:<base64_of_compressed_bytes>
-    let mut encoder = ZlibEncoder::new(Vec::new(), Compression::new(6));
-    encoder.write_all(&nixl_md)?;
-    let compressed = encoder.finish()?;
+    // Note: Python nixl_connect automatically decompresses when seeing "b64:" prefix
+    let mut zlib_encoder = ZlibEncoder::new(Vec::new(), Compression::new(6));
+    zlib_encoder.write_all(&nixl_md)?;
+    let compressed = zlib_encoder.finish()?;
 
     let b64_encoded = general_purpose::STANDARD.encode(&compressed);
     Ok(format!("b64:{}", b64_encoded))