Merge pull request #87 from nebius/main

Backporting from main

soperator/Makefile

@@ -4,7 +4,6 @@ SHELL = /usr/bin/env bash -o pipefail
 SOPERATOR_VERSION		= $(shell cat VERSION)
 SUBVERSION				= $(shell cat SUBVERSION)
 VERSION					= $(SOPERATOR_VERSION)-$(SUBVERSION)
-DOCKER_REGISTRY_NAME	= soperator
 
 ifeq ($(shell uname), Darwin)
     SED_COMMAND = sed -i ''
@@ -16,10 +15,6 @@ endif
 sync-version: ## Sync Soperator version from file
 	@echo 'Soperator version is - $(SOPERATOR_VERSION)'
 
-	@# region modules/slurm/locals.tf
-	@$(SED_COMMAND) "s|\(oci://cr.eu-north1.nebius.cloud/\)[^\"]*|\1$(DOCKER_REGISTRY_NAME)|" modules/slurm/locals.tf
-	@# endregion modules/slurm/locals.tf
-
 	@# region installations/example/terraform.tfvars
 	@echo 'Syncing installations/example/terraform.tfvars'
 	@$(SED_COMMAND) -E 's/slurm_operator_version *= *"[0-9]+.[0-9]+.[0-9]+[^ ]*"/slurm_operator_version = "$(SOPERATOR_VERSION)"/' installations/example/terraform.tfvars

soperator/SUBVERSION

@@ -1 +1 @@
-0
+1

soperator/installations/example/.envrc

@@ -36,9 +36,8 @@ NEBIUS_GROUP_EDITORS_ID=$(nebius iam group get-by-name \
 # Separate declaration and assignment for service account
 NEBIUS_SA_TERRAFORM_ID=$(nebius iam service-account list \
   --parent-id "${NEBIUS_PROJECT_ID}" \
-  --filter "name=slurm-terraform-sa" \
   --format json \
-  | jq -r '.items[0].metadata.id')
+  | jq -r '.items[] | select(.metadata.name == "slurm-terraform-sa").metadata.id')
 
 if [ -z "$NEBIUS_SA_TERRAFORM_ID" ]; then
   NEBIUS_SA_TERRAFORM_ID=$(nebius iam service-account create \

soperator/installations/example/main.tf

@@ -4,7 +4,7 @@ locals {
     controller = module.resources.this[var.slurm_nodeset_controller.resource.platform][var.slurm_nodeset_controller.resource.preset]
     workers    = [for worker in var.slurm_nodeset_workers : module.resources.this[worker.resource.platform][worker.resource.preset]]
     login      = module.resources.this[var.slurm_nodeset_login.resource.platform][var.slurm_nodeset_login.resource.preset]
-    accounting = module.resources.this[var.slurm_nodeset_accounting.resource.platform][var.slurm_nodeset_accounting.resource.preset]
+    accounting = var.slurm_nodeset_accounting != null ? module.resources.this[var.slurm_nodeset_accounting.resource.platform][var.slurm_nodeset_accounting.resource.preset] : null
   }
 
   use_node_port = var.slurm_login_service_type == "NodePort"
@@ -210,11 +210,11 @@ module "slurm" {
       memory_gibibytes            = local.resources.login.memory_gibibytes
       ephemeral_storage_gibibytes = ceil(var.slurm_nodeset_login.boot_disk.size_gibibytes / 2)
     }
-    accounting = {
+    accounting = var.accounting_enabled ? {
       cpu_cores                   = local.resources.accounting.cpu_cores
       memory_gibibytes            = local.resources.accounting.memory_gibibytes
       ephemeral_storage_gibibytes = ceil(var.slurm_nodeset_accounting.boot_disk.size_gibibytes / 2)
-    }
+    } : null
   }
 
   login_service_type         = var.slurm_login_service_type

soperator/installations/example/variables.tf

@@ -331,7 +331,7 @@ variable "slurm_nodeset_workers" {
   validation {
     condition = length([for worker in var.slurm_nodeset_workers :
       1 if worker.size % worker.split_factor != 0
-    ]) == 1
+    ]) == 0
     error_message = "Worker count must be divisible by split_factor."
   }
 }

soperator/mlperf/gpt3-impl-4.0-nvidia/run.sub

@@ -234,6 +234,7 @@ cleanup_preload_shared() {
 #########################################################################
 if [ -n "${CONTAINER_PRELOAD_SHARED_PATH}" ]; then
     CONT_FILE="${CONTAINER_PRELOAD_SHARED_PATH}/containers/${SLURM_JOBID}_$(basename ${CONT}).squashfs"
+    mkdir -p "${CONTAINER_PRELOAD_SHARED_PATH}/containers"
     # Prepull container image to the shared filesystem
     srun --ntasks=1 enroot import --output ${CONT_FILE} docker://${CONT}
 else

soperator/mlperf/gpt3-impl-4.0-nvidia/start.sh

@@ -137,4 +137,3 @@ sbatch \
 	run.sub
 
 squeue
-