Add support for GPU sharing on GKE

mmv1/third_party/terraform/tests/resource_container_node_pool_test.go.erb

@@ -2274,6 +2274,10 @@ resource "google_container_node_pool" "np_with_gpu" {
       type  = "nvidia-tesla-a100"
       gpu_partition_size = "1g.5gb"
       count = 1
+	  gpu_sharing_config {
+		gpu_sharing_strategy = "TIME_SHARING"
+		max_shared_clients_per_gpu = 2
+	  }
     }
   }
 }

mmv1/third_party/terraform/utils/node_config.go.erb

@@ -102,6 +102,30 @@ func schemaNodeConfig() *schema.Schema {
 								ForceNew:         true,
 								Description: `Size of partitions to create on the GPU. Valid values are described in the NVIDIA mig user guide (https://docs.nvidia.com/datacenter/tesla/mig-user-guide/#partitioning)`,
 							},
+							"gpu_sharing_config": &schema.Schema{
+								Type:         schema.TypeList,
+								MaxItems:     1,
+								Optional:     true,
+								ForceNew:     true,
+								ConfigMode: schema.SchemaConfigModeAttr,
+								Description:  `The configuration for GPU sharing options.`,
+								Elem: &schema.Resource{
+									Schema: map[string]*schema.Schema{
+										"gpu_sharing_strategy": &schema.Schema{
+											Type:        schema.TypeString,
+											Required:    true,
+											ForceNew:    true,
+											Description: `The type of GPU sharing strategy to enable on the GPU node. Possible value is TIME_SHARING`,
+										},
+										"max_shared_clients_per_gpu": &schema.Schema{
+											Type:        schema.TypeInt,
+											Required:    true,
+											ForceNew:    true,
+											Description: `The max number of containers that can share a GPU.`,
+										},
+									},
+								},
+							},
 						},
 					},
 				},
@@ -491,11 +515,23 @@ func expandNodeConfig(v interface{}) *container.NodeConfig {
 			if data["count"].(int) == 0 {
 				continue
 			}
-			guestAccelerators = append(guestAccelerators, &container.AcceleratorConfig{
+			guestAcceleratorConfig := &container.AcceleratorConfig{
 				AcceleratorCount: int64(data["count"].(int)),
 				AcceleratorType:  data["type"].(string),
 				GpuPartitionSize: data["gpu_partition_size"].(string),
-			})
+			}
+
+			if v, ok := data["gpu_sharing_config"]; ok {
+				if len(v.([]interface{})) > 0 {
+					gpuSharingConfig := data["gpu_sharing_config"].([]interface{})[0].(map[string]interface{})
+					guestAcceleratorConfig.GpuSharingConfig = &container.GPUSharingConfig{
+						GpuSharingStrategy: gpuSharingConfig["gpu_sharing_strategy"].(string),
+						MaxSharedClientsPerGpu: int64(gpuSharingConfig["max_shared_clients_per_gpu"].(int)),
+					}
+				}
+			}
+
+			guestAccelerators = append(guestAccelerators, guestAcceleratorConfig)
 		}
 		nc.Accelerators = guestAccelerators
 	}
@@ -795,11 +831,20 @@ func flattenNodeConfig(c *container.NodeConfig) []map[string]interface{} {
 func flattenContainerGuestAccelerators(c []*container.AcceleratorConfig) []map[string]interface{} {
 	result := []map[string]interface{}{}
 	for _, accel := range c {
-		result = append(result, map[string]interface{}{
+		accelerator := map[string]interface{}{
 			"count": accel.AcceleratorCount,
 			"type":  accel.AcceleratorType,
 			"gpu_partition_size": accel.GpuPartitionSize,
-		})
+		}
+		if accel.GpuSharingConfig != nil {
+			accelerator["gpu_sharing_config"] = []map[string]interface{}{
+				{
+					"gpu_sharing_strategy": accel.GpuSharingConfig.GpuSharingStrategy,
+					"max_shared_clients_per_gpu": accel.GpuSharingConfig.MaxSharedClientsPerGpu,
+				},
+			}
+		}
+		result = append(result, accelerator)
 	}
 	return result
 }

mmv1/third_party/terraform/website/docs/r/container_cluster.html.markdown

@@ -843,6 +843,16 @@ linux_node_config {
 
 * `gpu_partition_size` (Optional) - Size of partitions to create on the GPU. Valid values are described in the NVIDIA mig [user guide](https://docs.nvidia.com/datacenter/tesla/mig-user-guide/#partitioning).
 
+* `gpu_sharing_config` (Optional) - The configuration for GPU sharing options. Structure is [documented below](#nested_gpu_sharing_config).
+
+<a name="nested_gpu_sharing_config"></a>The `gpu_sharing_config` block supports:
+
+* `gpu_sharing_strategy` (Required) - The type of GPU sharing strategy to enable on the GPU node. 
+    Accepted values are:
+    * `"TIME_SHARING"`: Allow multiple containers to have [time-shared](https://cloud.google.com/kubernetes-engine/docs/concepts/timesharing-gpus) access to a single GPU device. 
+
+* `max_shared_clients_per_gpu` (Required) - The max number of containers that can share a GPU. 
+
 <a name="nested_workload_identity_config"></a> The `workload_identity_config` block supports:
 
 * `workload_pool` (Optional) - The workload pool to attach all Kubernetes service accounts to.