build: Enable frontend decoding (media-nixl) in regular dynamo builds (#4618)

Signed-off-by: Krishnan Prashanth <kprashanth@nvidia.com>

.github/workflows/container-validation-dynamo.yml

@@ -48,15 +48,15 @@ jobs:
         working-directory: ./deploy
         run: |
           docker compose up -d nats-server etcd-server
-      - name: Run Rust checks (block-manager + integration tests)
+      - name: Run Rust checks (block-manager + media-nixl + integration tests)
         run: |
           docker run --rm -w /workspace/lib/llm \
             --name ${{ env.CONTAINER_ID }}_rust_checks \
             ${{ steps.define_image_tag.outputs.image_tag }} \
             bash -ec 'rustup component add rustfmt clippy && \
                       cargo fmt -- --check && \
-                      cargo clippy --features block-manager --no-deps --all-targets -- -D warnings && \
-                      cargo test --locked --all-targets --features=block-manager && \
+                      cargo clippy --features block-manager,media-nixl --no-deps --all-targets -- -D warnings && \
+                      cargo test --locked --all-targets --features=block-manager,media-nixl && \
                       cargo test --locked --features integration -- --nocapture'
       - name: Cleanup services
         if: always()

container/Dockerfile

@@ -14,6 +14,7 @@ ARG BASE_IMAGE_TAG
 
 ARG PYTHON_VERSION
 ARG ENABLE_KVBM
+ARG ENABLE_MEDIA_NIXL
 ARG CARGO_BUILD_JOBS
 
 # Define general architecture ARGs for supporting both x86 and aarch64 builds.
@@ -278,7 +279,11 @@ RUN --mount=type=secret,id=aws-key-id,env=AWS_ACCESS_KEY_ID \
     cd /opt/dynamo && \
     uv build --wheel --out-dir /opt/dynamo/dist && \
     cd /opt/dynamo/lib/bindings/python && \
-    maturin build --release --out /opt/dynamo/dist && \
+    if [ "$ENABLE_MEDIA_NIXL" = "true" ]; then \
+        maturin build --release --features dynamo-llm/media-nixl --out /opt/dynamo/dist; \
+    else \
+        maturin build --release --out /opt/dynamo/dist; \
+    fi && \
     if [ "$ENABLE_KVBM" = "true" ]; then \
         cd /opt/dynamo/lib/bindings/kvbm && \
         maturin build --release --out target/wheels && \
@@ -354,7 +359,7 @@ USER dynamo
 ENV HOME=/home/dynamo \
     DYNAMO_HOME=/opt/dynamo \
     CARGO_TARGET_DIR=/opt/dynamo/target
-ENV LD_LIBRARY_PATH=${NIXL_LIB_DIR}:${NIXL_PLUGIN_DIR}:/usr/local/ucx/lib:/usr/local/ucx/lib/ucx:${LD_LIBRARY_PATH}
+ENV LD_LIBRARY_PATH=${NIXL_LIB_DIR}:${NIXL_PLUGIN_DIR}:/usr/local/ucx/lib:/usr/local/ucx/lib/ucx:/usr/local/cuda/compat/lib.real:${LD_LIBRARY_PATH}
 
 # Create and activate virtual environment
 ARG PYTHON_VERSION

container/README.md

@@ -230,6 +230,7 @@ Follow the instructions in [`deploy/inference-gateway/README.md`](../deploy/infe
 The base image contains the core Dynamo runtime components, NATS server, etcd, and Python dependencies:
 ```bash
 # Build the base dev image (framework=none for frontend-only deployment)
+# Note: --framework none defaults ENABLE_MEDIA_NIXL=false
 ./build.sh --framework none --target dev
 ```
 

container/build.sh

@@ -297,6 +297,9 @@ get_options() {
         --enable-kvbm)
             ENABLE_KVBM=true
             ;;
+        --enable-media-nixl)
+            ENABLE_MEDIA_NIXL=true
+            ;;
         --make-efa)
             NIXL_UCX_REF=$NIXL_UCX_EFA_REF
             ;;
@@ -469,6 +472,7 @@ show_help() {
     echo "  [--release-build perform a release build]"
     echo "  [--make-efa Enables EFA support for NIXL]"
     echo "  [--enable-kvbm Enables KVBM support in Python 3.12]"
+    echo "  [--enable-media-nixl Enable media processing with NIXL support (default: true for frameworks, false for none)]"
     echo "  [--use-sccache enable sccache for Rust/C/C++ compilation caching]"
     echo "  [--sccache-bucket S3 bucket name for sccache (required with --use-sccache)]"
     echo "  [--sccache-region S3 region for sccache (required with --use-sccache)]"
@@ -810,6 +814,18 @@ if [  ! -z ${ENABLE_KVBM} ]; then
     BUILD_ARGS+=" --build-arg ENABLE_KVBM=${ENABLE_KVBM} "
 fi
 
+# ENABLE_MEDIA_NIXL: Enable media processing with NIXL support
+# Used in base Dockerfile for maturin build feature flag.
+# Can be explicitly overridden with --enable-media-nixl flag
+if [ -z "${ENABLE_MEDIA_NIXL}" ]; then
+    if [[ $FRAMEWORK == "VLLM" ]] || [[ $FRAMEWORK == "TRTLLM" ]] || [[ $FRAMEWORK == "SGLANG" ]]; then
+        ENABLE_MEDIA_NIXL=true
+    else
+        ENABLE_MEDIA_NIXL=false
+    fi
+fi
+BUILD_ARGS+=" --build-arg ENABLE_MEDIA_NIXL=${ENABLE_MEDIA_NIXL} "
+
 # NIXL_UCX_REF: Used in base Dockerfile only.
 #               Passed to framework Dockerfile.{vllm,sglang,...} where it's NOT used.
 if [ -n "${NIXL_UCX_REF}" ]; then

lib/llm/src/preprocessor.rs

@@ -28,7 +28,7 @@ use tracing;
 
 use crate::model_card::{ModelDeploymentCard, ModelInfo};
 #[cfg(feature = "media-nixl")]
-use crate::preprocessor::media::{MediaDecoder, MediaFetcher, MediaLoader};
+use crate::preprocessor::media::MediaLoader;
 use crate::preprocessor::prompt::OAIChatLikeRequest;
 use crate::protocols::common::preprocessor::{
     MultimodalData, MultimodalDataMap, PreprocessedRequestBuilder,

lib/llm/src/preprocessor/media/README.md

@@ -38,6 +38,14 @@ register_llm(
 ```
 
 
+## Known Limitations
+
+> [!WARNING]
+> **Incompatible with `Dockerfile.frontend`**: Frontend media decoding (enabled with `--features media-nixl`) is not supported when using `Dockerfile.frontend`. The frontend image built from `Dockerfile.frontend` does not enable the feature + include the required NIXL/UCX dependencies.
+
+> [!WARNING]
+> **Requires GPU node**: The frontend must run on a node with GPU access. During media processing, decoded tensors are written to GPU memory via NIXL, which requires `libcuda.so.1` to be available. Running the frontend on a CPU-only node will fail with something like: `Failed to initialize required backends: [UCX: No UCX plugin found]`.
+
 ## TODOs
 
 ### Modalities

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

@@ -166,7 +166,7 @@ mod tests {
             ..Default::default()
         };
 
-        let loader: MediaLoader = MediaLoader::new(media_decoder, fetcher).unwrap();
+        let loader: MediaLoader = MediaLoader::new(media_decoder, Some(fetcher)).unwrap();
 
         let image_url = ImageUrl::from(format!("{}/llm-optimize-deploy-graphic.png", server.url()));
         let content_part = ChatCompletionRequestUserMessageContentPart::ImageUrl(