diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index 6aeb2684b..3f51b15ea 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -155,6 +155,8 @@ build-infra-runner:
     - kapsule
     - rust
   variables:
+    # GPU: compile jobs need CUDA for ml crate linking
+    KUBERNETES_RUNTIME_CLASS_NAME: nvidia
     # Give compile pods all available CPU for fast Rust compilation
     KUBERNETES_CPU_REQUEST: "7000m"
     KUBERNETES_CPU_LIMIT: "7800m"
@@ -423,6 +425,7 @@ build-training:
   # Picked up by gitlab-runner-rl which mounts training-data-l4-pvc + sccache-l4-pvc
   # Runner default node_selector = ci-compile (L4-1-24G)
   variables:
+    KUBERNETES_RUNTIME_CLASS_NAME: nvidia
     # RL is single-threaded (~1 core actual), 2000m allows 2 concurrent on L4
     KUBERNETES_CPU_REQUEST: "2000m"
     KUBERNETES_CPU_LIMIT: "3800m"
@@ -452,6 +455,7 @@ build-training:
     - kapsule
     - gpu
   variables:
+    KUBERNETES_RUNTIME_CLASS_NAME: nvidia
     # Route to L40S pool (48GB VRAM for large supervised models)
     KUBERNETES_NODE_SELECTOR_POOL: "k8s.scaleway.com/pool-name=ci-training"
     # Supervised is GPU-bound — minimal CPU. Low request allows 3-4 concurrent jobs.