Jetstream Maxtext Module

modules/jetstream-maxtext-deployment/README.md

@@ -0,0 +1,65 @@
+## Bash equivalent of this module
+
+Assure the following are set before running:
+   - BUCKET_NAME: Bucket name to be used in your checkpoint path
+   - METRICS_PORT: Port to emit custom metrics on
+   - (optional) MAXENGINE_SERVER_IMAGE: Maxengine server container image
+   - (optional) JETSTREAM_HTTP_SERVER_IMAGE: Jetstream HTTP server container image
+
+```
+if [ -z "$MAXENGINE_SERVER_IMAGE" ]; then
+    MAXENGINE_SERVER_IMAGE="us-docker.pkg.dev\/cloud-tpu-images\/inference\/maxengine-server:v0.2.2"
+fi
+
+if [ -z "$JETSTREAM_HTTP_SERVER_IMAGE" ]; then
+    JETSTREAM_HTTP_SERVER_IMAGE="us-docker.pkg.dev\/cloud-tpu-images\/inferenc\/jetstream-http:v0.2.2"
+fi
+
+if [ -z "$BUCKET_NAME" ]; then
+    echo "Must provide BUCKET_NAME in environment" 1>&2
+    exit 2;
+fi
+
+JETSTREAM_MANIFEST=$(mktemp)
+cat ./templates/deployment.yaml.tftpl >> "$JETSTREAM_MANIFEST"
+
+PODMONITORING_MANIFEST=$(mktemp)
+cat ./templates/podmonitoring.yaml.tftpl >> "$PODMONITORING_MANIFEST"
+
+if [ "$METRICS_PORT" != "" ]; then
+    cat $PODMONITORING_MANIFEST | sed "s/\${metrics_port}/$METRICS_PORT/g" >> "$PODMONITORING_MANIFEST"
+    cat $JETSTREAM_MANIFEST | sed "s/\${metrics_port_arg}/prometheus_port=$METRICS_PORT/g" >> "$JETSTREAM_MANIFEST"
+
+    cat $PODMONITORING_MANIFEST | kubectl apply -f -
+else
+    cat $JETSTREAM_MANIFEST | sed "s/\${metrics_port_arg}//g" >> "$JETSTREAM_MANIFEST"
+fi
+
+cat $JETSTREAM_MANIFEST \
+| sed "s/\${maxengine_server_image}/$MAXENGINE_SERVER_IMAGE/g" \
+| sed "s/\${jetstream_http_server_image}/$JETSTREAM_HTTP_SERVER_IMAGE/g" \
+| sed "s/\${load_parameters_path_arg}/load_parameters=gs:\/\/$BUCKET_NAME\/final\/unscanned\/gemma_7b-it\/0\/checkpoints\/0\/items/g" >> "$JETSTREAM_MANIFEST"
+
+cat $JETSTREAM_MANIFEST | kubectl apply -f -
+```
+
+### Metrics Adapter
+
+#### Custom Metrics Stackdriver Adapter
+
+Follow the [Custom-metrics-stackdriver-adapter README](LINK HERE)
+
+#### Prometheus Adapter
+
+Follow the [Prometheus-adapter README](AWAITING OTHER MERGE), a few notes:
+
+This module requires the cluster name to be passed in manually via the CLUSTER_NAME variable to filter incoming metrics. This is a consequence of differing cluster name schemas between GKE and standard k8s clusters. Instructions for each are as follows for if the cluster name isnt already known. For GKE clusters, Remove any charachters prior to and including the last underscore with `kubectl config current-context | awk -F'_' ' { print $NF }'` to get the cluster name. For other clusters, The cluster name is simply: `kubectl config current-context`.
+
+Instructions to set the PROMETHEUS_HELM_VALUES_FILE env var as follows:
+
+```
+PROMETHEUS_HELM_VALUES_FILE=$(mktemp)
+sed "s/\${cluster_name}/$CLUSTER_NAME/g" ../templates/values.yaml.tftpl >> "$PROMETHEUS_HELM_VALUES_FILE"
+```
+
+

modules/jetstream-maxtext-deployment/apply.sh

@@ -0,0 +1,34 @@
+if [ -z "$MAXENGINE_SERVER_IMAGE" ]; then
+    MAXENGINE_SERVER_IMAGE="us-docker.pkg.dev\/cloud-tpu-images\/inference\/maxengine-server:v0.2.2"
+fi
+
+if [ -z "$JETSTREAM_HTTP_SERVER_IMAGE" ]; then
+    JETSTREAM_HTTP_SERVER_IMAGE="us-docker.pkg.dev\/cloud-tpu-images\/inferenc\/jetstream-http:v0.2.2"
+fi
+
+if [ -z "$BUCKET_NAME" ]; then
+    echo "Must provide BUCKET_NAME in environment" 1>&2
+    exit 2;
+fi
+
+JETSTREAM_MANIFEST=$(mktemp)
+cat ./templates/deployment.yaml.tftpl >> "$JETSTREAM_MANIFEST"
+
+PODMONITORING_MANIFEST=$(mktemp)
+cat ./templates/podmonitoring.yaml.tftpl >> "$PODMONITORING_MANIFEST"
+
+if [ "$METRICS_PORT" != "" ]; then
+    cat $PODMONITORING_MANIFEST | sed "s/\${metrics_port}/$METRICS_PORT/g" >> "$PODMONITORING_MANIFEST"
+    cat $JETSTREAM_MANIFEST | sed "s/\${metrics_port_arg}/prometheus_port=$METRICS_PORT/g" >> "$JETSTREAM_MANIFEST"
+
+    cat $PODMONITORING_MANIFEST | kubectl apply -f -
+else
+    cat $JETSTREAM_MANIFEST | sed "s/\${metrics_port_arg}//g" >> "$JETSTREAM_MANIFEST"
+fi
+
+cat $JETSTREAM_MANIFEST \
+| sed "s/\${maxengine_server_image}/$MAXENGINE_SERVER_IMAGE/g" \
+| sed "s/\${jetstream_http_server_image}/$JETSTREAM_HTTP_SERVER_IMAGE/g" \
+| sed "s/\${load_parameters_path_arg}/load_parameters=gs:\/\/$BUCKET_NAME\/final\/unscanned\/gemma_7b-it\/0\/checkpoints\/0\/items/g" >> "$JETSTREAM_MANIFEST"
+
+cat $JETSTREAM_MANIFEST | kubectl apply -f -

modules/jetstream-maxtext-deployment/main.tf

@@ -0,0 +1,81 @@
+/**
+ * Copyright 2024 Google LLC
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+locals {
+  deployment_template         = "${path.module}/templates/deployment.yaml.tftpl"
+  service_template            = "${path.module}/templates/service.yaml.tftpl"
+  podmonitoring_template      = "${path.module}/templates/podmonitoring.yaml.tftpl"
+  cmsa_jetstream_hpa_template = "${path.module}/templates/custom-metrics-stackdriver-adapter/hpa.jetstream.yaml.tftpl"
+}
+
+resource "kubernetes_manifest" "jetstream-deployment" {
+  count = 1
+  manifest = yamldecode(templatefile(local.deployment_template, {
+    maxengine_server_image      = var.maxengine_server_image
+    jetstream_http_server_image = var.jetstream_http_server_image
+    load_parameters_path_arg    = format("load_parameters_path=gs://%s/final/unscanned/gemma_7b-it/0/checkpoints/0/items", var.bucket_name)
+    metrics_port_arg            = var.metrics_port != null ? format("prometheus_port=%d", var.metrics_port) : "",
+  }))
+}
+
+resource "kubernetes_manifest" "jetstream-service" {
+  count    = 1
+  manifest = yamldecode(file(local.service_template))
+}
+
+resource "kubernetes_manifest" "jetstream-podmonitoring" {
+  count = var.metrics_port != null ? 1 : 0
+  manifest = yamldecode(templatefile(local.podmonitoring_template, {
+    metrics_port = var.metrics_port != null ? var.metrics_port : "",
+  }))
+}
+
+
+## CMSA module pending https://github.com/GoogleCloudPlatform/ai-on-gke/pull/718/files merge
+module "custom_metrics_stackdriver_adapter" {
+  count  = var.metrics_adapter == "custom-metrics-stackdriver-adapter" ? 1 : 0
+  source = "../custom-metrics-stackdriver-adapter"
+  workload_identity = {
+    enabled    = true
+    project_id = var.project_id
+  }
+}
+
+## Prometheus adapter module pending https://github.com/GoogleCloudPlatform/ai-on-gke/pull/716/files merge
+module "prometheus_adapter" {
+  count = var.metrics_adapter == "prometheus-adapter" ? 1 : 0
+  source = "../prometheus-adapter"
+  credentials_config = {
+    kubeconfig = {
+      path : "~/.kube/config"
+    }
+  }
+  project_id = var.project_id
+  cluster_name = var.cluster_name
+  config_file = templatefile("/templates/prometheus-adapter/values.yaml.tftpl", {
+    cluster_name = var.cluster_name
+  })
+}
+
+resource "kubernetes_manifest" "hpa_custom_metric" {
+  count = (var.custom_metrics_enabled && var.hpa_type != null || var.hpa_type != "memory_used") && var.hpa_averagevalue_target != null ? 1 : 0
+  manifest = yamldecode(templatefile(local.cmsa_jetstream_hpa_template, {
+    hpa_type                = try(var.hpa_type, "")
+    hpa_averagevalue_target = try(var.hpa_averagevalue_target, 1)
+    hpa_min_replicas        = var.hpa_min_replicas
+    hpa_max_replicas        = var.hpa_max_replicas
+  }))
+}

modules/jetstream-maxtext-deployment/templates/custom-metrics-stackdriver-adapter/hpa.jetstream.yaml.tftpl

@@ -0,0 +1,30 @@
+apiVersion: autoscaling/v2
+kind: HorizontalPodAutoscaler
+metadata:
+  name: jetstream-hpa
+  namespace: default
+spec:
+  scaleTargetRef:
+    apiVersion: apps/v1
+    kind: Deployment
+    name: maxengine-server
+  minReplicas: ${hpa_min_replicas}
+  maxReplicas: ${hpa_max_replicas}
+  metrics:
+%{ if length(regexall("jetstream_.*", hpa_type)) > 0 }
+  - type: Pods
+    pods:
+      metric:
+        name: prometheus.googleapis.com|${hpa_type}|gauge
+      target:
+        type: AverageValue
+        averageValue: ${hpa_averagevalue_target}
+%{ else }
+  - type: External
+    external:
+      metric:
+        name: kubernetes.io|node|accelerator|memory_used
+      target:
+        type: AverageValue
+        averageValue: ${hpa_averagevalue_target}
+%{ endif }

modules/jetstream-maxtext-deployment/templates/deployment.yaml.tftpl

@@ -0,0 +1,51 @@
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: maxengine-server
+  namespace: default
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: maxengine-server
+  template:
+    metadata:
+      labels:
+        app: maxengine-server
+    spec:
+      nodeSelector:
+        cloud.google.com/gke-tpu-topology: 2x2
+        cloud.google.com/gke-tpu-accelerator: tpu-v5-lite-podslice
+      containers:
+      - name: maxengine-server
+        image: ${maxengine_server_image}
+        imagePullPolicy: Always
+        securityContext:
+          privileged: true
+        args:
+        - model_name=gemma-7b
+        - tokenizer_path=assets/tokenizer.gemma
+        - per_device_batch_size=4
+        - max_prefill_predict_length=1024
+        - max_target_length=2048
+        - async_checkpointing=false
+        - ici_fsdp_parallelism=1
+        - ici_autoregressive_parallelism=-1
+        - ici_tensor_parallelism=1
+        - scan_layers=false
+        - weight_dtype=bfloat16
+        - attention=dot_product
+        - ${load_parameters_path_arg}
+        - ${metrics_port_arg}
+        ports:
+        - containerPort: 9000
+        resources:
+          requests:
+            google.com/tpu: 4
+          limits:
+            google.com/tpu: 4
+      - name: jetstream-http
+        image: ${jetstream_http_server_image}
+        imagePullPolicy: Always
+        ports:
+        - containerPort: 8000

modules/jetstream-maxtext-deployment/templates/podmonitoring.yaml.tftpl

@@ -0,0 +1,15 @@
+apiVersion: monitoring.googleapis.com/v1
+kind: PodMonitoring
+metadata:
+  name: jetstream-podmonitoring
+  namespace: default
+spec:
+  endpoints:
+  - interval: 1s
+    path: "/"
+    port: ${metrics_port}
+  targetLabels:
+    metadata:
+    - pod
+    - container
+    - node

modules/jetstream-maxtext-deployment/templates/prometheus-adapter/hpa.jetstream.yaml.tftpl

@@ -0,0 +1,30 @@
+apiVersion: autoscaling/v2
+kind: HorizontalPodAutoscaler
+metadata:
+  name: jetstream-hpa
+  namespace: ${namespace}
+spec:
+  scaleTargetRef:
+    apiVersion: apps/v1
+    kind: Deployment
+    name: maxengine-server
+  minReplicas: ${hpa_min_replicas}
+  maxReplicas: ${hpa_max_replicas}
+  metrics:
+%{ if length(regexall("jetstream_.*", hpa_type)) > 0 }
+  - type: Pods
+    pods:
+      metric:
+        name: ${hpa_type}
+      target:
+        type: AverageValue
+        averageValue: ${hpa_averagevalue_target}
+%{ else }
+  - type: External
+    external:
+      metric:
+        name: ${hpa_type}
+      target:
+        type: AverageValue
+        averageValue: ${hpa_averagevalue_target}
+%{ endif }

modules/jetstream-maxtext-deployment/templates/prometheus-adapter/values.yaml.tftpl

@@ -0,0 +1,38 @@
+rules:
+  default: false
+  external:
+  - seriesQuery: 'jetstream_prefill_backlog_size'
+    resources:
+      template: <<.Resource>>
+    name:
+      matches: ""
+      as: "jetstream_prefill_backlog_size"
+    metricsQuery: sum(<<.Series>>{<<.LabelMatchers>>,cluster="${cluster_name}"})
+  - seriesQuery: 'jetstream_transfer_backlog_size'
+    resources:
+      template: <<.Resource>>
+    name:
+      matches: ""
+      as: "jetstream_transfer_backlog_size"
+    metricsQuery: sum(<<.Series>>{<<.LabelMatchers>>,cluster="${cluster_name}"})
+  - seriesQuery: 'jetstream_generate_backlog_size'
+    resources:
+      template: <<.Resource>>
+    name:
+      matches: ""
+      as: "jetstream_generate_backlog_size"
+    metricsQuery: sum(<<.Series>>{<<.LabelMatchers>>,cluster="${cluster_name}"})
+  - seriesQuery: 'jetstream_slots_used_percentage'
+    resources:
+      template: <<.Resource>>
+    name:
+      matches: ""
+      as: "jetstream_slots_used_percentage"
+    metricsQuery: sum(<<.Series>>{<<.LabelMatchers>>,cluster="${cluster_name}"})
+  - seriesQuery: 'kubernetes_io:node_accelerator_memory_used'
+    resources:
+      template: <<.Resource>>
+    name:
+      matches: ""
+      as: "accelerator_memory_used_percentage"
+    metricsQuery: avg(kubernetes_io:node_accelerator_memory_used{cluster_name="${cluster_name}"}) / avg(kubernetes_io:node_accelerator_memory_total{cluster_name="${cluster_name}"})

modules/jetstream-maxtext-deployment/templates/service.yaml.tftpl

@@ -0,0 +1,17 @@
+apiVersion: v1
+kind: Service
+metadata:
+  name: jetstream-svc
+  namespace: default
+spec:
+  selector:
+    app: maxengine-server
+  ports:
+  - protocol: TCP
+    name: jetstream-http
+    port: 8000
+    targetPort: 8000
+  - protocol: TCP
+    name: jetstream-grpc
+    port: 9000
+    targetPort: 9000