diff --git a/.ci/docs/setup_nvidia_gpu_with_rdma_support_on_ubuntu.md b/.ci/docs/setup_nvidia_gpu_with_rdma_support_on_ubuntu.md
index f3b59dc8a8..11d58da8ac 100644
--- a/.ci/docs/setup_nvidia_gpu_with_rdma_support_on_ubuntu.md
+++ b/.ci/docs/setup_nvidia_gpu_with_rdma_support_on_ubuntu.md
@@ -34,8 +34,11 @@ sudo reboot
 If running on nvlink hosts like DGX we should also install fabric manager
 ```bash
 sudo apt install nvidia-fabricmanager-<version> # should be same as kernel version nvidia-fabricmanager-575
+sudo systemctl enable --now nvidia-fabricmanager
 ```
 
+**Important**: On NVSwitch-enabled systems (DGX A100, H100), `nvidia-fabricmanager` must be running before GPU initialization. Without it, CUDA will fail with "system not yet initialized" errors. Ensure the service is enabled at boot.
+
 Verify with `nvidia-smi`. Driver compatibility is critical for RDMA support[^1_1][^1_3].
 
 
diff --git a/.ci/jenkins/lib/test-matrix.yaml b/.ci/jenkins/lib/test-matrix.yaml
index 2284b1724b..a0c671af09 100644
--- a/.ci/jenkins/lib/test-matrix.yaml
+++ b/.ci/jenkins/lib/test-matrix.yaml
@@ -25,7 +25,7 @@ timeout_minutes: 240
 # label is defined at jenkins slave configuration, we want to run the job on a gpu agent and be able to esaly replace it without having to change this file
 runs_on_agents:
   - {nodeLabel: 'H100'}
-  # - {nodeLabel: 'DGX'}
+  - {nodeLabel: 'DGX'}
 
 matrix:
   axes:
@@ -35,7 +35,8 @@ matrix:
       - x86_64
     ucx_version:
       - master
-      - v1.20.x
+
+#     - v1.20.x
 
 taskName: "${name}/${arch}/ucx-${ucx_version}/${axis_index}"
 
@@ -43,7 +44,7 @@ env:
   CONTAINER_WORKSPACE: /workspace
   INSTALL_DIR: ${CONTAINER_WORKSPACE}/nixl_install
   # Manual timeout - ci-demo doesn't handle docker exec
-  TEST_TIMEOUT: 30
+  TEST_TIMEOUT: 60
   # NPROC for bare-metal: containers see all host CPUs, need to limit parallelism
   NPROC: 16
 
@@ -92,7 +93,9 @@ steps:
           done
         fi
       fi
-
+      # print environment
+      set
+      env
 
   - name: Build GPU Test Environment
     parallel: false
diff --git a/.gitlab/test_cpp.sh b/.gitlab/test_cpp.sh
index 8c911134ea..8ec97a04f2 100755
--- a/.gitlab/test_cpp.sh
+++ b/.gitlab/test_cpp.sh
@@ -92,7 +92,8 @@ if $TEST_LIBFABRIC ; then
     ./bin/nixl_example LIBFABRIC
 fi
 ./bin/nixl_etcd_example
-./bin/ucx_backend_test
+# TODO: Remove UCX_GDR_COPY_SHARED_MD=n once UCX is fixed.
+UCX_GDR_COPY_SHARED_MD=n ./bin/ucx_backend_test
 mkdir -p /tmp/telemetry_test
 NIXL_TELEMETRY_ENABLE=y NIXL_TELEMETRY_DIR=/tmp/telemetry_test ./bin/agent_example &
 sleep 5
diff --git a/.gitlab/test_nixlbench.sh b/.gitlab/test_nixlbench.sh
index 6585b8346a..8d84fed519 100755
--- a/.gitlab/test_nixlbench.sh
+++ b/.gitlab/test_nixlbench.sh
@@ -60,6 +60,9 @@ wait_for_etcd
 echo "==== Running Nixlbench tests ===="
 cd ${INSTALL_DIR}
 
+# TODO: Remove UCX_GDR_COPY_SHARED_MD=n once UCX is fixed.
+export UCX_GDR_COPY_SHARED_MD=n
+
 DEFAULT_NB_PARAMS="--filepath /tmp --total_buffer_size 80000000 --start_block_size 4096 --max_block_size 16384 --start_batch_size 1 --max_batch_size 4"
 
 run_nixlbench() {