Merge pull request #44 from NVIDIA/fix/gke-remove-deprecated-cfs-sysctls

nvidia-tuning-gke: drop deprecated CFS sysctls + fix base image registry + bump TUNING_VERSION to 1.1.6 + permissive CI guard

Author: ayuskauskas

.github/workflows/pr_build.yaml

@@ -106,8 +106,9 @@ jobs:
           if [ ! -f "$CONFIG_FILE" ]; then
             # Check if Dockerfile exists and has a FROM that references a skyhook-packages image
             if [ -f "$DOCKERFILE" ]; then
-              # Check if FROM line contains "skyhook-packages" (indicating it inherits from another package)
-              if grep -q "^FROM.*skyhook-packages" "$DOCKERFILE"; then
+              # Check if FROM line contains "skyhook-packages" or "nodewright-packages"
+              # (indicating it inherits from another package in this repo's published image set).
+              if grep -Eq "^FROM.*(skyhook|nodewright)-packages" "$DOCKERFILE"; then
                 echo "config_exists=false" >> $GITHUB_OUTPUT
                 echo "config_changed=false" >> $GITHUB_OUTPUT
                 echo "No config.json found, but Dockerfile inherits from skyhook-packages image - validation skipped"
@@ -152,7 +153,7 @@ jobs:
         run: |
           DOCKERFILE="${{ matrix.package }}/Dockerfile"
           
-          if [ -f "$DOCKERFILE" ] && grep -q "^FROM.*skyhook-packages" "$DOCKERFILE"; then
+          if [ -f "$DOCKERFILE" ] && grep -Eq "^FROM.*(skyhook|nodewright)-packages" "$DOCKERFILE"; then
             echo "inherits_from_skyhook_packages=true" >> $GITHUB_OUTPUT
             echo "Package inherits from skyhook-packages image - will use validate-inherited target"
           else

makefile

@@ -108,8 +108,8 @@ validate-inherited: ## Validate an inherited package (inherits from skyhook-pack
 		echo "ERROR: Dockerfile not found for package $(PACKAGE)"; \
 		exit 1; \
 	fi
-	@if ! grep -q "^FROM.*skyhook-packages" "$(PACKAGE)/Dockerfile"; then \
-		echo "ERROR: Package $(PACKAGE) does not inherit from skyhook-packages. Use 'make validate-standalone' instead."; \
+	@if ! grep -Eq "^FROM.*(skyhook|nodewright)-packages" "$(PACKAGE)/Dockerfile"; then \
+		echo "ERROR: Package $(PACKAGE) does not inherit from skyhook-packages or nodewright-packages. Use 'make validate-standalone' instead."; \
 		exit 1; \
 	fi
 	@echo "Building container for validation: $(PACKAGE)"

nvidia-tuning-gke/Dockerfile

@@ -21,8 +21,8 @@
 # Config step: prepare_nvidia_configs.sh (populate configmaps from profile)
 #              then update_settings.sh (base tuning apply).
 
-ARG TUNING_VERSION=1.1.4
-FROM ghcr.io/nvidia/skyhook-packages/tuning:${TUNING_VERSION}
+ARG TUNING_VERSION=1.1.6
+FROM ghcr.io/nvidia/nodewright-packages/tuning:${TUNING_VERSION}
 
 COPY profiles/ /skyhook-package/profiles/
 COPY skyhook_dir/prepare_nvidia_configs.sh /skyhook-package/skyhook_dir/

nvidia-tuning-gke/profiles/gb200/inference/sysctl.conf

@@ -10,5 +10,3 @@ vm.max_map_count=262144
 vm.min_free_kbytes=65536
 vm.overcommit_memory=1
 vm.swappiness=1
-kernel.sched_latency_ns=1000000
-kernel.sched_min_granularity_ns=100000

nvidia-tuning-gke/profiles/h100/inference/sysctl.conf

@@ -4,5 +4,3 @@ net.ipv4.conf.default.arp_announce = 2
 net.ipv4.conf.all.arp_ignore = 1
 net.ipv4.conf.default.arp_ignore = 1
 vm.swappiness=1
-kernel.sched_latency_ns=1000000
-kernel.sched_min_granularity_ns=100000

tests/integration/nvidia_tuning_gke/test_prepare_nvidia_configs.py

@@ -41,7 +41,7 @@
 @pytest.mark.parametrize(
     "accelerator,intent,expected_sysctl_line,expect_containerd",
     [
-        ("h100", "inference", "kernel.sched_latency_ns=1000000", False),
+        ("h100", "inference", "vm.swappiness=1", False),
         ("h100", "multiNodeTraining", "net.core.default_qdisc=fq", False),
         ("gb200", "inference", "vm.swappiness=1", True),
         ("gb200", "multiNodeTraining", "net.core.default_qdisc=fq", True),