fix(container): align sglang nixl_ref with PyPI nixl-cu12 1.0.x (#8205)

Signed-off-by: Keiven Chang <keivenc@nvidia.com>

container/context.yaml

@@ -85,7 +85,7 @@ sglang:
     runtime_image: lmsysorg/sglang
     base_image_tag: 25.11-cuda13.0-devel-ubuntu24.04
     runtime_image_tag: v0.5.10.post1-cu130-runtime
-  nixl_ref: 0.10.0
+  nixl_ref: v1.0.1
   enable_media_ffmpeg: "true"
   enable_gpu_memory_service: "true"
   enable_kvbm: "false"

container/templates/wheel_builder.Dockerfile

@@ -152,13 +152,25 @@ RUN --mount=type=cache,target=/var/cache/dnf,sharing=locked \
         librdmacm-devel \
         numactl-devel \
         # Libfabric support
-        hwloc \
-        hwloc-devel \
         libcurl-devel \
         openssl-devel \
         libuuid-devel \
         zlib-devel
 
+# Build hwloc >= 2.3 from source (RHEL8 ships 2.2 which lacks hwloc_location API
+# required by nixl v1.0.x libfabric topology code)
+ARG HWLOC_VERSION=2.12.0
+RUN HWLOC_SERIES="$(echo "${HWLOC_VERSION}" | cut -d. -f1-2)" && \
+    cd /tmp && \
+    curl --retry 3 -LO "https://download.open-mpi.org/release/hwloc/v${HWLOC_SERIES}/hwloc-${HWLOC_VERSION}.tar.gz" && \
+    tar xf hwloc-${HWLOC_VERSION}.tar.gz && \
+    cd hwloc-${HWLOC_VERSION} && \
+    ./configure --prefix=/usr/local && \
+    make -j$(nproc) && \
+    make install && \
+    ldconfig && \
+    rm -rf /tmp/hwloc-*
+
 # Set GCC toolset 14 as the default compiler (CUDA requires GCC <= 14)
 ENV PATH="/opt/rh/gcc-toolset-14/root/usr/bin:${PATH}" \
     LD_LIBRARY_PATH="/opt/rh/gcc-toolset-14/root/usr/lib64:${LD_LIBRARY_PATH}" \

pyproject.toml

@@ -63,7 +63,7 @@ vllm = [
 sglang = [
     "uvloop",
     "sglang[diffusion]==0.5.10.post1",
-    "nixl[cu12]<=0.10.1",
+    "nixl[cu12]>=1.0.0,<1.1.0",
     "cupy-cuda12x>=13.0.0",
 ]