ci(mm-rdma): dedicated native wheel for the RDMA e2e

The RDMA multimodal e2e needs a gateway built with --features mm-rdma, but the
shared wheel is cross-compiled with maturin --zig (LLVM libc++), which can't
link nixl-sys's GNU-libstdc++ C++ stub. Build a dedicated mm-rdma wheel natively
on the bare Ubuntu CPU runner (GNU libstdc++, same runner image as the GPU e2e
lanes), upload it as smg-wheel-mm-rdma, and run test_multimodal_rdma.py in a
dedicated e2e-1gpu-mm-rdma job via e2e-gpu-job's new wheel_artifact input. The
shared wheel and other e2e lanes are untouched; the RDMA test is excluded from
the shared vLLM lane.

Signed-off-by: Simo Lin <linsimo.mark@gmail.com>

Author: slin1237

.github/workflows/e2e-gpu-job.yml

@@ -54,6 +54,11 @@ on:
         type: string
         default: "nixl"
         description: "KV transfer backend for vLLM PD workers: nixl or mooncake"
+      wheel_artifact:
+        required: false
+        type: string
+        default: "smg-wheel"
+        description: "Name of the wheel artifact to install (e.g. smg-wheel-mm-rdma)"
 
 jobs:
   run:
@@ -92,7 +97,7 @@ jobs:
       - name: Download wheel artifact
         uses: actions/download-artifact@v8
         with:
-          name: smg-wheel
+          name: ${{ inputs.wheel_artifact }}
           path: wheel/
 
       - name: Download WASM test fixtures

.github/workflows/pr-test-rust.yml

@@ -212,6 +212,37 @@ jobs:
           python3 -c "from smg.smg_rs import Router; print('Rust extension: OK')"
           python3 -m smg.launch_router --help > /dev/null && echo "Entry point: OK"
 
+  build-wheel-mm-rdma:
+    # Dedicated mm-rdma wheel for the RDMA e2e, built natively on the bare Ubuntu
+    # runner (GNU libstdc++). The shared wheel is cross-compiled with
+    # `maturin --zig`, which ships LLVM libc++ and can't link nixl-sys's
+    # GNU-libstdc++ C++ stub; building it separately here keeps that off the
+    # shared wheel so it can't break the other e2e lanes. Runs on the same runner
+    # image as the GPU e2e lanes, so the native wheel loads there unchanged.
+    needs: detect-changes
+    if: >-
+      always()
+      && !cancelled()
+      && (github.event_name != 'pull_request'
+          || (needs.detect-changes.result == 'success'
+              && (needs.detect-changes.outputs.common == 'true'
+                  || needs.detect-changes.outputs.chat-completions == 'true')))
+    runs-on: k8s-runner-cpu
+    permissions:
+      contents: read
+    steps:
+      - uses: actions/checkout@v7
+      - name: Setup Rust
+        uses: ./.github/actions/setup-rust
+      - name: Build mm-rdma wheel
+        run: bash scripts/ci_build_wheel_mm_rdma.sh
+      - name: Upload mm-rdma wheel artifact
+        uses: actions/upload-artifact@v7
+        with:
+          name: smg-wheel-mm-rdma
+          path: bindings/python/dist/*.whl
+          retention-days: 1
+
   python-unit-tests:
     needs: build-wheel
     runs-on: k8s-runner-cpu
@@ -508,6 +539,9 @@ jobs:
           - engine: vllm
             timeout: 24
             test_timeout: 18
+            # The RDMA multimodal test needs the dedicated mm-rdma wheel + a NIXL
+            # runner; it runs in e2e-1gpu-mm-rdma, not this shared-wheel lane.
+            test_filter: "--ignore=e2e_test/chat_completions/test_multimodal_rdma.py"
           - engine: trtllm
             timeout: 32
             test_timeout: 18
@@ -529,6 +563,31 @@ jobs:
       timeout: ${{ matrix.timeout }}
       test_timeout: ${{ matrix.test_timeout }}
       test_dirs: ${{ matrix.test_dirs || 'e2e_test/chat_completions' }}
+      test_filter: ${{ matrix.test_filter || '' }}
+    secrets: inherit
+
+  e2e-1gpu-mm-rdma:
+    # Isolated RDMA multimodal e2e: installs the dedicated mm-rdma wheel and runs
+    # only test_multimodal_rdma.py against a vLLM worker on a NIXL-capable runner.
+    needs: [build-wheel, build-wheel-mm-rdma, detect-changes]
+    if: >-
+      always()
+      && !cancelled()
+      && needs.build-wheel.result == 'success'
+      && needs.build-wheel-mm-rdma.result == 'success'
+      && (github.event_name != 'pull_request'
+          || (needs.detect-changes.result == 'success'
+              && (needs.detect-changes.outputs.common == 'true'
+                  || needs.detect-changes.outputs.chat-completions == 'true')))
+    uses: ./.github/workflows/e2e-gpu-job.yml
+    with:
+      engine: vllm
+      gpu_tier: "1"
+      runner: 1-gpu-h100
+      timeout: 24
+      test_timeout: 18
+      test_dirs: e2e_test/chat_completions/test_multimodal_rdma.py
+      wheel_artifact: smg-wheel-mm-rdma
     secrets: inherit
 
   e2e-1gpu-completions:
@@ -1037,7 +1096,7 @@ jobs:
           path: benchmark_go_bindings/
 
   finish:
-    needs: [pre-commit, python-lint, grpc-proto-build-check, build-wheel, python-unit-tests, unit-tests, benchmarks, e2e-1gpu-chat, e2e-1gpu-completions, e2e-1gpu-embeddings, e2e-1gpu-gateway, e2e-1gpu-responses, e2e-2gpu-pd, e2e-4gpu-chat, e2e-4gpu-gateway, e2e-4gpu-epd, e2e-vendor, go-unit-tests, go-bindings-e2e]
+    needs: [pre-commit, python-lint, grpc-proto-build-check, build-wheel, build-wheel-mm-rdma, python-unit-tests, unit-tests, benchmarks, e2e-1gpu-chat, e2e-1gpu-mm-rdma, e2e-1gpu-completions, e2e-1gpu-embeddings, e2e-1gpu-gateway, e2e-1gpu-responses, e2e-2gpu-pd, e2e-4gpu-chat, e2e-4gpu-gateway, e2e-4gpu-epd, e2e-vendor, go-unit-tests, go-bindings-e2e]
     if: always()
     runs-on: k8s-runner-cpu
     permissions: {}
@@ -1048,10 +1107,12 @@ jobs:
                 "${{ needs.python-lint.result }}" == "failure" || \
                 "${{ needs.grpc-proto-build-check.result }}" == "failure" || \
                 "${{ needs.build-wheel.result }}" == "failure" || \
+                "${{ needs.build-wheel-mm-rdma.result }}" == "failure" || \
                 "${{ needs.python-unit-tests.result }}" == "failure" || \
                 "${{ needs.unit-tests.result }}" == "failure" || \
                 "${{ needs.benchmarks.result }}" == "failure" || \
                 "${{ needs.e2e-1gpu-chat.result }}" == "failure" || \
+                "${{ needs.e2e-1gpu-mm-rdma.result }}" == "failure" || \
                 "${{ needs.e2e-1gpu-completions.result }}" == "failure" || \
                 "${{ needs.e2e-1gpu-embeddings.result }}" == "failure" || \
                 "${{ needs.e2e-1gpu-gateway.result }}" == "failure" || \

bindings/python/Cargo.toml

@@ -36,6 +36,7 @@ workspace = true
 default = ["pyo3/extension-module"]
 opencv-video = ["smg/opencv-video"]
 vendored-openssl = ["smg/vendored-openssl"]
+mm-rdma = ["smg/mm-rdma"]
 
 [profile.ci]
 inherits = "release"

scripts/ci_build_wheel_mm_rdma.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+# Build the mm-rdma smg wheel for the RDMA e2e, natively on the bare Ubuntu
+# runner (GNU libstdc++).
+#
+# Why a separate build (not the shared ci_build_wheel.sh): the shared wheel is
+# cross-compiled with `maturin --zig`, which ships LLVM libc++. nixl-sys (pulled
+# by the mm-rdma feature) compiles a C++ stub that references GNU libstdc++
+# (std::cerr / _ZSt4cerr), so it must be linked against GNU libstdc++ or the
+# wheel fails to import ("undefined symbol: _ZSt4cerr"). Building natively with
+# the runner's GNU toolchain links the C++ stub cleanly, and the wheel runs on
+# the GPU e2e runner unchanged (same runner image, same libstdc++.so.6).
+#
+# Assumes ./.github/actions/setup-rust already ran (Rust + build-essential/g++ +
+# protoc). This wheel is CI-only (installed by the RDMA e2e job), so a plain
+# linux tag is fine.
+set -euxo pipefail
+
+# libclang for nixl-sys bindgen; setup-rust installs build-essential (g++ ->
+# libstdc++) but not clang.
+export DEBIAN_FRONTEND=noninteractive
+if command -v sudo >/dev/null 2>&1; then
+    sudo apt-get update
+    sudo apt-get install -y --no-install-recommends libclang-dev
+else
+    apt-get update
+    apt-get install -y --no-install-recommends libclang-dev
+fi
+
+# setup-rust adds cargo to GITHUB_PATH; source for good measure (and local runs).
+if [ -f "$HOME/.cargo/env" ]; then
+    source "$HOME/.cargo/env"
+fi
+export RUSTC_WRAPPER="${RUSTC_WRAPPER:-sccache}"
+
+python3 -m pip install --upgrade pip maturin
+
+echo "Building mm-rdma wheel (native, GNU libstdc++)..."
+cd bindings/python
+# abi3 wheel (pyo3 abi3-py38) -> one build works for all 3.8+.
+# --manylinux off: CI-only wheel; skip auditwheel's manylinux policy and
+# dynamically link libstdc++.so.6 (present on the GPU e2e runner).
+maturin build \
+    --profile ci \
+    --features vendored-openssl,mm-rdma \
+    --manylinux off \
+    --out dist
+echo "mm-rdma wheel: OK"
+ls -lh dist/