Deprecating TimesFM 1.0, 2.0 and 2.5 dockers on MG.

MG_DOCKER_CODES_PIPER_ORIGIN_REV_ID: 914520347

.cloud-build/cleanup/cleanup.py

@@ -1,5 +1,14 @@
 from typing import List
 from ratemate import RateLimit
+import argparse
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--dry_run",
+                    type=bool,
+                    default=False)
+args = parser.parse_args()
+
+
 from resource_cleanup_manager import (
     DatasetResourceCleanupManager,
     ModelResourceCleanupManager,
@@ -7,6 +16,15 @@ from resource_cleanup_manager import (
     ResourceCleanupManager,
     MatchingEngineIndexEndpointResourceCleanupManager,
     MatchingEngineIndexResourceCleanupManager,
+    FeatureStoreLegacyCleanupManager,
+    FeatureStoreCleanupManager,
+    PipelineJobCleanupManager,
+    TrainingJobCleanupManager,
+    HyperparameterTuningCleanupManager,
+    BatchPredictionJobCleanupManager,
+    ExperimentCleanupManager,
+    BucketCleanupManager,
+    ArtifactRegistryCleanupManager
 )
 
 rate_limit = RateLimit(max_count=25, per=60, greedy=False)
@@ -18,12 +36,14 @@ def run_cleanup_managers(managers: List[ResourceCleanupManager], is_dry_run: boo
 
         print(f"Fetching {type_name}'s...")
         resources = manager.list()
-        print(f"Found {len(resources)} {type_name}'s")
+        try:
+            print(f"Found {len(resources)} {type_name}'s")
+        except Exception as e:
+            print(f"{type_name} {e}")
         for resource in resources:
             try:
                 if not manager.is_deletable(resource):
                     continue
-
                 if is_dry_run:
                     resource_name = manager.resource_name(resource)
                     print(f"Will delete '{type_name}': {resource_name}")
@@ -36,9 +56,7 @@ def run_cleanup_managers(managers: List[ResourceCleanupManager], is_dry_run: boo
         print("")
 
 
-is_dry_run = False
-
-if is_dry_run:
+if args.dry_run:
     print("Starting cleanup in dry run mode...")
 
 # List of all cleanup managers
@@ -48,6 +66,15 @@ managers: List[ResourceCleanupManager] = [
     ModelResourceCleanupManager(),  # ModelResourceCleanupManager must follow EndpointResourceCleanupManager due to deployed models blocking model deletion.
     MatchingEngineIndexEndpointResourceCleanupManager(),
     MatchingEngineIndexResourceCleanupManager(),
+    FeatureStoreLegacyCleanupManager(),
+    FeatureStoreCleanupManager(),
+    PipelineJobCleanupManager(),
+    TrainingJobCleanupManager(),
+    HyperparameterTuningCleanupManager(),
+    BatchPredictionJobCleanupManager(),
+    ExperimentCleanupManager(), # Experiment missing _resource_noun
+    BucketCleanupManager(),
+    ArtifactRegistryCleanupManager()
 ]
 
-run_cleanup_managers(managers=managers, is_dry_run=is_dry_run)
+run_cleanup_managers(managers=managers, is_dry_run=args.dry_run)

.cloud-build/cleanup/resource_cleanup_manager.py

@@ -1,10 +1,26 @@
+'''
+READ FIRST BEFORE MAKING CHANGES
+-  Create a convention for resources created from vertex-ai-samples GH. We already have one IIRC
+- Only delete those objects as part of our clean-up script.
+- Don't run any tests on python-docs-samples-tests project, especially ones that affect resources created outside of our purview
+- Add --dry-run option to the clean-up script. This option will just output the list of resources the script will delete instead of actually deleting the resources.
+- Have a larger conversation in DEE before touching any resources that were not created as part of vertex-ai-samples
+'''
+import os
 import abc
 from typing import Any, Type
 
 from google.cloud import aiplatform
 from google.cloud.aiplatform import base
+from google.cloud.aiplatform_v1beta1 import (FeatureOnlineStoreAdminServiceClient,
+                                             FeatureOnlineStore)
+from google.cloud import storage
 from proto.datetime_helpers import DatetimeWithNanoseconds
 
+PROJECT_ID = "python-docs-samples-tests"
+REGION = "us-central1"
+API_ENDPOINT = f"{REGION}-aiplatform.googleapis.com"
+
 # If a resource was updated within this number of seconds, do not delete.
 RESOURCE_UPDATE_BUFFER_IN_SECONDS = 60 * 60 * 8
 
@@ -69,7 +85,7 @@ class VertexAIResourceCleanupManager(ResourceCleanupManager):
     def delete(self, resource):
         resource.delete()
 
-    def get_seconds_since_modification(self, resource: Any) -> bool:
+    def get_seconds_since_modification(self, resource: Any) -> float:
         update_time = resource.update_time
         current_time = DatetimeWithNanoseconds.now(tz=update_time.tzinfo)
         return (current_time - update_time).total_seconds()
@@ -97,8 +113,6 @@ class EndpointResourceCleanupManager(VertexAIResourceCleanupManager):
     vertex_ai_resource = aiplatform.Endpoint
 
     def delete(self, resource):
-        # TODO: Remove this once https://github.com/googleapis/python-aiplatform/issues/1441 is fixed
-        resource._sync_gca_resource()
         for deployed_model_id in [
             models.id for models in resource._gca_resource.deployed_models
         ]:
@@ -119,4 +133,173 @@ class MatchingEngineIndexEndpointResourceCleanupManager(VertexAIResourceCleanupM
 
     def delete(self, resource):
         resource.undeploy_all()
-        resource.delete(force=True)
+        resource.delete(force=True)
+
+class FeatureStoreLegacyCleanupManager(VertexAIResourceCleanupManager):
+    # TODO: only deleting legacy
+    #    not deleting ingestions jobs
+    #       ingest_from_xxx methods do not return a job ID, there is no list command, aka no python way to delete
+    #    not deleting batch serving jobs
+    #       batch_serve_to_xxx methods do not return a job ID, there is no list command, aka no python way to delete
+    vertex_ai_resource = aiplatform.Featurestore
+
+    def resource_name(self, resource: Any) -> str:
+        return resource.name
+
+    def delete(self, resource):
+        resource.delete(force=True)
+
+
+class FeatureStoreCleanupManager(VertexAIResourceCleanupManager):
+    # for FS 2.0
+    # TODO: use _v1beta1, and gapic clients
+    #    delete features, feature groups, feature views, feature online stores
+    vertex_ai_resource = FeatureOnlineStore
+
+    admin_client = FeatureOnlineStoreAdminServiceClient(
+        client_options={"api_endpoint": API_ENDPOINT}
+    )
+
+    def resource_name(self, resource: Any) -> str:
+        return resource.name
+
+    def type_name(self) -> str:
+        return "FeatureOnlineStore"
+
+    def list(self) -> Any:
+        try:
+            return self.admin_client.list_feature_online_stores(parent=f"projects/{PROJECT_ID}/locations/{REGION}")
+        except Exception as e:
+            print(e)
+            return []
+
+    def delete(self, resource):
+        try:
+            self.admin_client.delete_feature_online_store(name=resource.name, force=True)
+        except Exception as e:
+            print(e)
+
+
+class PipelineJobCleanupManager(VertexAIResourceCleanupManager):
+    vertex_ai_resource = aiplatform.PipelineJob
+
+class TrainingJobCleanupManager(VertexAIResourceCleanupManager):
+    vertex_ai_resource = aiplatform.training_jobs._CustomTrainingJob
+
+    job_types = [
+            aiplatform.AutoMLImageTrainingJob,
+            aiplatform.AutoMLTextTrainingJob,
+            aiplatform.AutoMLTabularTrainingJob,
+            aiplatform.AutoMLVideoTrainingJob,
+            aiplatform.AutoMLForecastingTrainingJob,
+            aiplatform.CustomJob,
+            aiplatform.CustomTrainingJob,
+            aiplatform.CustomContainerTrainingJob,
+            aiplatform.CustomPythonPackageTrainingJob
+    ]
+
+    def list(self) -> Any:
+        return [
+            job
+            for job_type in self.job_types
+            for job in job_type.list()
+        ]
+
+class HyperparameterTuningCleanupManager(VertexAIResourceCleanupManager):
+    vertex_ai_resource = aiplatform.HyperparameterTuningJob
+
+
+class BatchPredictionJobCleanupManager(VertexAIResourceCleanupManager):
+    vertex_ai_resource = aiplatform.BatchPredictionJob
+
+class ExperimentCleanupManager(VertexAIResourceCleanupManager):
+    vertex_ai_resource = aiplatform.Experiment
+
+    @property
+    def type_name(self) -> str:
+        return "Experiment"
+
+    def resource_name(self, resource: Any) -> str:
+        return resource.name
+
+    def get_seconds_since_modification(self, resource: Any) -> float:
+        update_time = resource._metadata_context.update_time
+        current_time = DatetimeWithNanoseconds.now()
+        return float(current_time.timestamp() - update_time.timestamp())
+
+class BucketCleanupManager(ResourceCleanupManager):
+    vertex_ai_resource = storage.bucket.Bucket
+
+    def list(self) -> Any:
+        storage_client = storage.Client()
+        return list(storage_client.list_buckets())
+
+    def delete(self, resource):
+        try:
+            resource.delete(force=True)
+        except Exception as e:
+            print(e)
+
+    @property
+    def type_name(self) -> str:
+        return "Bucket"
+
+    def get_seconds_since_modification(self, resource: Any) -> float:
+        # Bucket has no last_update property, only time created
+        created_time = resource.time_created
+        current_time = DatetimeWithNanoseconds.now()
+        return float(current_time.timestamp() - created_time.timestamp())
+
+    def resource_name(self, resource: Any) -> str:
+        return resource.name
+
+    def is_deletable(self, resource: Any) -> bool:
+        time_difference = self.get_seconds_since_modification(resource)
+
+        if not self.resource_name(resource).startswith('your-bucket-name'):
+            print(f"Skipping '{resource}' not a Vertex AI notebook bucket")
+            return False
+
+        # Check that it wasn't created too recently, to prevent race conditions
+        if time_difference <= RESOURCE_UPDATE_BUFFER_IN_SECONDS:
+            print(
+                f"Skipping '{resource}' due to update_time being '{time_difference}', which is less than '{RESOURCE_UPDATE_BUFFER_IN_SECONDS}'."
+            )
+            return False
+        return True
+
+class ArtifactRegistryCleanupManager(ResourceCleanupManager):
+    vertex_ai_resource = "Artifact Registry"
+
+    def list(self) -> Any:
+        import subprocess
+
+        result = subprocess.run(["gcloud artifacts repositories list --location=us-central1"], 
+                                shell=True, capture_output=True, text=True)
+
+        ret = []
+        lines = result.stdout.split('\n')[2:]
+        for line in lines:
+            repo = line.split(' ')[0]
+            if repo.startswith("my-docker-repo"):
+                ret.append(repo)
+
+        return ret
+
+    def delete(self, resource):
+        os.system(f"! gcloud artifacts repositories delete {resource} --location=us-central1")
+
+    @property
+    def type_name(self) -> str:
+        return "ArtifactRepository"
+
+    def resource_name(self, resource: Any) -> str:
+        return resource
+
+    # delete repository regardless of age
+    def get_seconds_since_modification(self, resource: Any) -> float:
+        return RESOURCE_UPDATE_BUFFER_IN_SECONDS + 1
+    
+    def is_deleteable(self, resource: Any) -> bool:
+        return True
+

.cloud-build/execute_changed_notebooks_cli.py

@@ -17,6 +17,8 @@
 
 import argparse
 import pathlib
+import os
+import csv
 
 import execute_changed_notebooks_helper
 
@@ -36,9 +38,22 @@ parser = argparse.ArgumentParser(description="Run changed notebooks.")
 parser.add_argument(
     "--test_paths_file",
     type=pathlib.Path,
-    help="The path to the file that has newline-limited folders of notebooks that should be tested.",
+    help="The path to the file that has newline-delimited folders of notebooks that should be tested.",
     required=True,
 )
+parser.add_argument(
+    "--test_percent",
+    type=int,
+    help="The percent of notebooks to be tested (between 1 and 100).",
+    required=False,
+    default=100,
+)
+parser.add_argument(
+    "--build_id",
+    type=str,
+    help="The build id (which may be a Cloud Build job specific or user explicit.",
+    required=True
+)
 parser.add_argument(
     "--base_branch",
     help="The base git branch to diff against to find changed files.",
@@ -107,24 +122,98 @@ parser.add_argument(
     default=True,
     help="Should run notebooks in parallel.",
 )
+parser.add_argument(
+    "--concurrent_notebooks",
+    type=int,
+    help="Maximum number of parallel notebook executions per minute",
+    default=10,
+    required=False,
+)
+parser.add_argument(
+    "--run_first_file",
+    type=pathlib.Path,
+    help="The path to the file that has newline-delimited of notebooks to run in the first batch",
+    default=None,
+    required=False,
+)
+parser.add_argument(
+    "--aiplatform_whl",
+    type=str,
+    help="The GCS path to a whl version google-cloud-aiplatform",
+    default=None,
+    required=False,
+)
+parser.add_argument(
+    "--dry_run",
+    type=str2bool,
+    default=False,
+    help="Dry run for testing - no execution",
+)
 
 args = parser.parse_args()
 
-notebooks = execute_changed_notebooks_helper.get_changed_notebooks(
+changed_notebooks = execute_changed_notebooks_helper.get_changed_notebooks(
     test_paths_file=args.test_paths_file,
     base_branch=args.base_branch,
 )
 
-execute_changed_notebooks_helper.process_and_execute_notebooks(
-    notebooks=notebooks,
-    container_uri=args.container_uri,
-    staging_bucket=args.staging_bucket,
-    artifacts_bucket=args.artifacts_bucket,
-    should_parallelize=args.should_parallelize,
-    timeout=args.timeout,
-    variable_project_id=args.variable_project_id,
-    variable_region=args.variable_region,
-    variable_service_account=args.variable_service_account,
-    variable_vpc_network=args.variable_vpc_network,
-    private_pool_id=args.private_pool_id,
+
+results_bucket = f"{args.artifacts_bucket}"
+# artifacts_bucket may get set by trigger to a full gs:// folder path
+if results_bucket.startswith("gs://"):
+    results_bucket = results_bucket[5:]
+results_bucket = results_bucket.split('/')[0]
+results_file = f"build_results/{args.build_id}.json"
+
+if args.test_percent == 100:
+    notebooks = changed_notebooks
+    accumulative_results = {}
+else:
+    accumulative_results = execute_changed_notebooks_helper.load_results(results_bucket, results_file)
+
+    notebooks = [changed_notebook for changed_notebook in changed_notebooks if execute_changed_notebooks_helper.select_notebook(changed_notebook, accumulative_results, args.test_percent)]
+    # cap the number of notebooks to the specified percentage
+    max_notebooks = int((len(changed_notebooks) * (args.test_percent/100)))
+    if (len(notebooks) > max_notebooks):
+        notebooks = notebooks[:max_notebooks]
+
+run_first = []
+if args.run_first_file:
+    if not os.path.isfile(args.run_first_file):
+        print("Error: file does not exist", args.run_first_file)
+    else:
+        with open(args.run_first_file, 'r') as csvfile:
+            reader = csv.reader(csvfile)
+            for row in reader:
+                notebook = row[0]
+                run_first.append(notebook)
+
+    for notebook in run_first:
+        if notebook in notebooks:
+            # remove from existing list
+            notebooks.remove(notebook)
+            # add back to the front of the list
+            notebooks.insert(0, notebook)
+            print(f"Run first: {notebook}")
+
+if args.dry_run:
+    print("Dry run ...\n")
+    for notebook in notebooks:
+        print(f"Would execute: {notebook}")
+else:
+    execute_changed_notebooks_helper.process_and_execute_notebooks(
+        notebooks=notebooks,
+        container_uri=args.container_uri,
+        staging_bucket=args.staging_bucket,
+        artifacts_bucket=args.artifacts_bucket,
+        results_file=results_file,
+        should_parallelize=args.should_parallelize,
+        timeout=args.timeout,
+        variable_project_id=args.variable_project_id,
+        variable_region=args.variable_region,
+        variable_service_account=args.variable_service_account,
+        variable_vpc_network=args.variable_vpc_network,
+        private_pool_id=args.private_pool_id,
+        concurrent_notebooks=args.concurrent_notebooks,
+        aiplatform_whl=args.aiplatform_whl
 )

.cloud-build/execute_changed_notebooks_helper.py

@@ -21,25 +21,34 @@ import json
 import git
 import operator
 import os
+import io
+import json
 import pathlib
 import re
 import subprocess
+import random
+from google.cloud import storage
 import utils
-from typing import List, Optional
+from typing import List, Optional, Dict, Any
 from utils import util
 
 import execute_notebook_helper
 import execute_notebook_remote
 import nbformat
 from google.cloud.devtools.cloudbuild_v1.types import BuildOperationMetadata
-from ratemate import RateLimit
 from tabulate import tabulate
 from utils import NotebookProcessors, util
 
 # A buffer so that workers finish before the orchestrating job
 WORKER_TIMEOUT_BUFFER_IN_SECONDS: int = 60 * 60
+
 PYTHON_VERSION = "3.9"  # Set default python version
 
+# rolling time window for accumulating build results for selecting notebooks
+MAX_RESULTS_AGE_SECONDS: int = (60 * 60) * 24 * 60  # 60 days
+# maximum time since last run to force a run on the current build
+MAX_AGE_BEFORE_FORCE_RUN: int = (60 * 60) * 24 * 30
+
 
 def format_timedelta(delta: datetime.timedelta) -> str:
     """Formats a timedelta duration to [N days] %H:%M:%S format"""
@@ -65,7 +74,9 @@ def format_timedelta(delta: datetime.timedelta) -> str:
 @dataclasses.dataclass
 class NotebookExecutionResult:
     name: str
+    path: str
     duration: datetime.timedelta
+    start_time: datetime.datetime
     is_pass: bool
     log_url: str
     output_uri: str
@@ -81,6 +92,97 @@ class NotebookExecutionResult:
             return None
 
 
+def load_results(results_bucket: str,
+                 results_file: str) -> Dict[str, Any]:
+    '''
+    Load accumulated notebook test results
+    '''
+
+    print("Loading existing accumulative results ...")
+    accumulative_results = {}
+    try:
+        client = storage.Client()
+        bucket = client.bucket(results_bucket)
+
+        build_results_dir = os.path.dirname(results_file)
+        blobs = client.list_blobs(results_bucket, prefix=build_results_dir)
+        for blob in blobs:
+            time_created = blob.time_created.replace(tzinfo=None)
+            if (datetime.datetime.now().replace(tzinfo=None) - time_created).total_seconds() > MAX_RESULTS_AGE_SECONDS:
+                continue
+
+            content = util.download_blob_into_memory(results_bucket, blob.name, download_as_text=True)
+
+            try:
+                build_results = json.loads(content)
+            except:
+                continue  # skip corrupted build results files
+            for notebook in build_results:
+                if notebook in accumulative_results:
+                    accumulative_results[notebook]['passed'] += build_results[notebook]['passed']
+                    accumulative_results[notebook]['failed'] += build_results[notebook]['failed']
+                    if accumulative_results[notebook]['last_time_ran'] < time_created:
+                        accumulative_results[notebook]['last_time_ran'] = time_created
+                else:
+                    accumulative_results[notebook] = build_results[notebook]
+                    accumulative_results[notebook]['failed_on_latest_run'] = build_results[notebook]['failed']
+                    accumulative_results[notebook]['last_time_ran'] = time_created
+
+        print(accumulative_results)
+    except Exception as e:
+        print(e)
+
+    # If there are no accumulative results, an empty dict is returned
+    return accumulative_results
+
+def select_notebook(changed_notebook: str,
+                    accumulative_results: Dict[str, Any],
+                    test_percent: int) -> bool:
+    '''
+    Algorithm to randomly select a notebook, but weight the propbability of selected based on past failures
+    '''
+
+    if changed_notebook in accumulative_results:
+        pass_count = accumulative_results[changed_notebook]['passed']
+        fail_count = accumulative_results[changed_notebook]['failed']
+        failed_on_latest_run = accumulative_results[changed_notebook]['failed_on_latest_run']
+        last_time_ran = accumulative_results[changed_notebook]['last_time_ran']
+    else:
+        pass_count = 1
+        fail_count = 0
+        failed_on_latest_run = 0
+        last_time_ran = datetime.datetime.now().replace(tzinfo=None)
+
+    # If notebook has not been ran in a long time, force running it
+    if (datetime.datetime.now().replace(tzinfo=None) - last_time_ran).total_seconds() > MAX_AGE_BEFORE_FORCE_RUN:
+        should_test_do_to_age = True
+    else:
+        should_test_do_to_age  = False
+
+
+    # if failed on the last time it was ran, select the notebook
+    if failed_on_latest_run:
+        inferred_failure_rate = 1
+    # otherwise, calculate the frequency of failure
+    else:
+        inferred_failure_rate = fail_count / (pass_count + fail_count)
+
+    # If failure rate is high, the chance of testing should be higher
+    should_test_due_to_failure = random.uniform(0, 1) <= inferred_failure_rate
+
+    #if accumulative_resultsi[changed_notebook]['latest_date_ran']
+
+    # Additionally, only test a percentage of these
+    should_test_due_to_random_subset = random.uniform(0, 1) <= (test_percent / 100)
+
+    if should_test_due_to_failure or should_test_due_to_random_subset or should_test_do_to_age:
+        print(f"Selected: {changed_notebook}, {should_test_due_to_failure}, {should_test_due_to_random_subset}")
+        return True
+    else:
+        print(f"Not Selected: {changed_notebook}, pass {pass_count}, fail {fail_count}")
+        return False
+
+
 def _process_notebook(
     notebook_path: str,
     variable_project_id: str,
@@ -136,7 +238,7 @@ def _get_notebook_python_version(notebook_path: str) -> str:
 
             # Look for the python version specification pattern
             re_match = re.search(
-                "python version = (\d\.\d)", markdown, flags=re.IGNORECASE
+                r"python version = (\d+\.\d+)", markdown, flags=re.IGNORECASE
             )
             if re_match:
                 # get the version number
@@ -156,7 +258,6 @@ def _create_tag(filepath: str) -> str:
     return tag
 
 
-rate_limit = RateLimit(max_count=50, per=60, greedy=True)
 
 
 def process_and_execute_notebook(
@@ -170,9 +271,8 @@ def process_and_execute_notebook(
     private_pool_id: Optional[str],
     deadline: datetime.datetime,
     notebook: str,
-    should_get_tail_logs: bool = False,
+    should_get_tail_logs: bool = True,
 ) -> NotebookExecutionResult:
-    rate_limit.wait()  # wait before creating the task
 
     print(f"Running notebook: {notebook}")
 
@@ -191,7 +291,9 @@ def process_and_execute_notebook(
 
     result = NotebookExecutionResult(
         name=tag,
+        path=notebook,
         duration=datetime.timedelta(seconds=0),
+        start_time=datetime.datetime.now(),
         is_pass=False,
         output_uri=notebook_output_uri,
         log_url="",
@@ -201,7 +303,6 @@ def process_and_execute_notebook(
     )
 
     # TODO: Handle cases where multiple notebooks have the same name
-    time_start = datetime.datetime.now()
     operation = None
     try:
         # Get the python version for running the notebook if specified
@@ -247,9 +348,10 @@ def process_and_execute_notebook(
         # Block and wait for the result
         operation_result = operation.result(timeout=timeout_in_seconds)
 
-        result.duration = datetime.datetime.now() - time_start
+        result.duration = datetime.datetime.now() - result.start_time
         result.is_pass = True
         print(f"{notebook} PASSED in {format_timedelta(result.duration)}.")
+
     except Exception as error:
         result.error_message = str(error)
 
@@ -263,12 +365,12 @@ def process_and_execute_notebook(
             # Use gcloud to get tail
             try:
                 result.error_message = subprocess.check_output(
-                    ["gsutil", "cat", "-r", "-1000", log_file_uri], encoding="UTF-8"
+                    ["gcloud", "storage", "cat", "--range", "-1000", log_file_uri], encoding="UTF-8"
                 )
             except Exception as error:
                 result.error_message = str(error)
 
-        result.duration = datetime.datetime.now() - time_start
+        result.duration = datetime.datetime.now() - result.start_time
         result.is_pass = False
 
         print(
@@ -336,12 +438,68 @@ def get_changed_notebooks(
 
     return notebooks
 
+def _save_results(results: List[NotebookExecutionResult],
+                  artifacts_bucket: str,
+                  results_file: str):
+
+    artifacts_bucket = artifacts_bucket.replace("gs://", "").split('/')[0]
+
+    print("Updating build results ...")
+    build_results = {}
+    for result in results:
+        if result.is_pass:
+            pass_count = 1
+            fail_count = 0
+        else:
+            pass_count = 0
+            fail_count = 1
+        if result.error_message is None:
+            error_type = ''
+        elif '500 Internal' in result.error_message or 'INTERNAL' in result.error_message or 'internal error' in result.error_message:
+            error_type = 'INTERNAL'
+        elif 'context deadline exceeded' in result.error_message or 'TIMEOUT' in result.error_message:
+            error_type = 'TIMEOUT'
+        elif 'Quota' in result.error_message or 'quotas are exceeded' in result.error_message:
+            error_type = 'QUOTA'
+        elif 'ServiceUnavailable' in result.error_message:
+            error_type = 'SERVICEUNAVAILABLE'
+        elif 'ModuleNotFoundError' in result.error_message:
+            error_type = 'IMPORT'
+        elif result.is_pass:
+            error_type = ''
+        else:
+            error_type = 'undetermined'
+
+        if error_type != '':
+            log_url = result.log_url
+        else:
+            log_url = ''
+
+        build_results[result.path] = {
+                'duration': result.duration.total_seconds(),
+                'start_time': str(result.start_time),
+                'passed': pass_count,
+                'failed': fail_count,
+                'error_type': error_type,
+                'log_url': log_url
+        }
+        print(f"adding {result.path}")
+
+    print(f"Saving accumulative results to {results_file}, nentries {len(build_results)}")
+    content = json.dumps(build_results)
+
+    client = storage.Client()
+    bucket = client.get_bucket(artifacts_bucket)
+    bucket.blob(str(results_file)).upload_from_string(content, 'text/json')
+
+
 
 def process_and_execute_notebooks(
     notebooks: List[str],
     container_uri: str,
     staging_bucket: str,
     artifacts_bucket: str,
+    results_file: str,
     should_parallelize: bool,
     timeout: int,
     variable_project_id: str,
@@ -349,6 +507,8 @@ def process_and_execute_notebooks(
     variable_service_account: str,
     variable_vpc_network: Optional[str] = None,
     private_pool_id: Optional[str] = None,
+    concurrent_notebooks: Optional[int] = 10,
+    aiplatform_whl: Optional[str] = None,
 ):
     """
     Run the notebooks that exist under the folders defined in the test_paths_file.
@@ -369,6 +529,8 @@ def process_and_execute_notebooks(
             Required. The GCS staging bucket to write source code to.
         artifacts_bucket (str):
             Required. The GCS staging bucket to write executed notebooks to.
+        results_file (str):
+            Required: The path to the artifacts bucket to save results 
         variable_project_id (str):
             Required. The value for PROJECT_ID to inject into notebooks.
         variable_region (str):
@@ -377,6 +539,8 @@ def process_and_execute_notebooks(
             Required. Should run notebooks in parallel using a thread pool as opposed to in sequence.
         timeout (str):
             Required. Timeout string according to https://cloud.google.com/build/docs/build-config-file-schema#timeout.
+        concurrent_notebooks (int): Max number of notebooks per minute to run in parallel.
+        aiplatform_whl: alternate whl version of Vertex AI SDK to install
     """
 
     # Calculate deadline
@@ -393,7 +557,9 @@ def process_and_execute_notebooks(
             print(
                 "Running notebooks in parallel, so no logs will be displayed. Please wait..."
             )
-            with concurrent.futures.ThreadPoolExecutor(max_workers=100) as executor:
+
+
+            with concurrent.futures.ThreadPoolExecutor(max_workers=concurrent_notebooks) as executor:
                 print(f"Max workers: {executor._max_workers}")
 
                 notebook_execution_results = list(
@@ -471,7 +637,7 @@ def process_and_execute_notebooks(
             print("=" * 100)
 
             build_id = results_sorted[0].build_id
-            logs_bucket_name = (results_sorted[0].logs_bucket).removeprefix("gs://")
+            logs_bucket_name = (results_sorted[0].logs_bucket).replace("gs://", "")
             log_file_name = f"log-{build_id}.txt"
 
             log_contents = util.download_blob_into_memory(
@@ -489,6 +655,10 @@ def process_and_execute_notebooks(
             else:
                 print(log_contents)
 
+        _save_results(results_sorted, 
+                      artifacts_bucket, 
+                      results_file)
+
         print("\n=== END RESULTS===\n")
 
         total_notebook_duration = functools.reduce(

.cloud-build/execute_notebook_helper.py

@@ -56,8 +56,8 @@ def execute_notebook(
     print("\n=== DOWNLOAD EXECUTED NOTEBOOK ===\n")
     print(f"Please debug the executed notebook by downloading the executed notebook:")
 
-    print("Option 1. Using gsutil. Run the following command in your terminal.")
-    print(f'\tgsutil cp "{output_file_or_uri}" .')
+    print("Option 1. Using gcloud storage. Run the following command in your terminal.")
+    print(f'\tgcloud storage cp "{output_file_or_uri}" .')
 
     print("Option 2. Using this link.")
     print(f"\thttps://storage.googleapis.com/{output_file_or_uri[5:]}")

.cloud-build/notebook-execution-test-cloudbuild.yaml

@@ -36,7 +36,7 @@ steps:
     - -c
     - |
       . workspace/env/bin/activate &&
-      python3 .cloud-build/execute_changed_notebooks_cli.py --test_paths_file "${_TEST_PATHS_FILE}" --base_branch "${_FORCED_BASE_BRANCH}" --container_uri ${_PYTHON_IMAGE} --staging_bucket ${_GCS_STAGING_BUCKET}  --artifacts_bucket ${_GCS_STAGING_BUCKET}/executed_notebooks/PR_${_PR_NUMBER}/BUILD_${BUILD_ID} --variable_project_id ${PROJECT_ID} --variable_region ${_GCP_REGION} --variable_service_account ${_GCP_SERVICE_ACCOUNT} --variable_vpc_network "${_GPC_VPC_NETWORK_NAME}" `if [ ! -z "${_PRIVATE_POOL_NAME}" ]; then echo "--private_pool_id ${_PRIVATE_POOL_NAME}"; fi`
+      python3 .cloud-build/execute_changed_notebooks_cli.py --test_paths_file "${_TEST_PATHS_FILE}" --base_branch "${_FORCED_BASE_BRANCH}" --container_uri ${_PYTHON_IMAGE} --staging_bucket ${_GCS_STAGING_BUCKET}  --artifacts_bucket ${_GCS_STAGING_BUCKET}/executed_notebooks/PR_${_PR_NUMBER}/BUILD_${BUILD_ID} --variable_project_id ${PROJECT_ID} --variable_region ${_GCP_REGION} --variable_service_account ${_GCP_SERVICE_ACCOUNT} --variable_vpc_network "${_GCP_VPC_NETWORK_NAME}" `if [ ! -z "${_PRIVATE_POOL_NAME}" ]; then echo "--private_pool_id ${_PRIVATE_POOL_NAME}"; fi` --build_id ${BUILD_ID} --test_percent=${_TEST_PERCENT} --concurrent_notebooks=${_CONCURRENT_NOTEBOOKS} --run_first_file=${_RUN_FIRST_FILE}
     env:
     - 'IS_TESTING=1'
 timeout: 86400s

.cloud-build/requirements.txt

@@ -3,11 +3,15 @@ numpy
 jupyter
 nbconvert
 papermill
-pandas
 matplotlib
 tabulate
 google-cloud-aiplatform
 google-cloud-storage
 google-cloud-build
+google-cloud-storage
 ratemate
 GitPython
+tqdm
+fsspec
+pandas
+

.cloud-build/run_first.csv

@@ -0,0 +1,8 @@
+notebooks/official/pipelines/google_cloud_pipeline_components_automl_text.ipynb
+notebooks/official/generative_ai/rlhf_tune_llm.ipynb
+notebooks/official/generative_ai/tune_peft.ipynb
+notebooks/official/prediction/llm_streaming_prediction.ipynb
+notebooks/official/migration/sdk-automl-text-classification-batch-prediction.ipynb
+notebooks/official/vizier/get_started_vertex_vizier.ipynb
+notebooks/official/workbench/sentiment_analysis/Sentiment_Analysis.ipynb
+notebooks/official/model_monitoring/get_started_with_model_monitoring_automl.ipynb

.cloud-build/test_notebook_vm.txt

@@ -1,5 +1,3 @@
-notebooks/official/vizier/gapic-vizier-multi-objective-optimization.ipynb
-notebooks/official/pipelines/lightweight_functions_component_io_kfp.ipynb
 notebooks/official/ml_metadata/sdk-metric-parameter-tracking-for-locally-trained-models.ipynb
 notebooks/official/custom/custom-tabular-bq-managed-dataset.ipynb
 .cloud-build/tests/python_version_test.ipynb

.cloud-build/tests/flaky.py

@@ -0,0 +1,46 @@
+# grep PASSED tests.txt | cut -c 10-100 >passed.txt
+
+import os
+
+repo_dir = '/home/jupyter/vertex-ai-samples/'
+repo_dir_len = len(repo_dir)
+official_dir = repo_dir + 'notebooks/official'
+
+entries = os.scandir(official_dir)
+folders = []
+for entry in entries:
+    if entry.is_dir():
+        folders.append(entry.path)
+
+# Passing
+with open('passed.txt', 'r') as pass_file:
+    notebook_names = pass_file.readlines()
+        
+notebooks = []
+for folder in folders:
+    entries = os.scandir(folder)
+    for entry in entries:
+        for notebook in notebook_names:
+            if entry.name == notebook.rstrip():
+                notebooks.append(entry.path[repo_dir_len:])
+                
+with open('passing_tests.txt', 'w') as f:
+    for notebook in notebooks:
+        f.write(notebook + '\n')
+        
+        
+# Failing
+with open('failed.txt', 'r') as fail_file:
+    notebook_names = fail_file.readlines()
+        
+notebooks = []
+for folder in folders:
+    entries = os.scandir(folder)
+    for entry in entries:
+        for notebook in notebook_names:
+            if entry.name == notebook.rstrip():
+                notebooks.append(entry.path[repo_dir_len:])
+                
+with open('failing_tests.txt', 'w') as f:
+    for notebook in notebooks:
+        f.write(notebook + '\n')

.cloud-build/tests/select_notebooks_test.py

@@ -0,0 +1,33 @@
+import sys
+
+from execute_changed_notebooks_helper import (load_results, select_notebook)
+
+
+def test_load_results():
+    bucket: str = "cloud-build-notebooks-presubmit"
+    bucket_file: str = "build_results"
+
+    accum = load_results(bucket, bucket_file)
+
+    print(accum)
+
+    assert len(accum) > 0
+
+def test_select_notebook():
+    bucket: str = "cloud-build-notebooks-presubmit"
+    bucket_file: str = "build_results"
+
+    accum = load_results(bucket, bucket_file)
+
+    n_select = 0
+    n_notselect = 0
+    for notebook in accum:
+        if select_notebook(notebook, accum, 50):
+            n_select += 1
+        else:
+            n_notselect += 1
+
+    print(f"SELECTED {n_select}, NOT SELECTED {n_notselect}")
+
+    assert n_select > 0
+    assert n_notselect > 0

.cloud-build/utils/NotebookProcessors.py

@@ -35,7 +35,7 @@ class RemoveNoExecuteCells(Preprocessor):
 
 
 class UpdateVariablesPreprocessor(Preprocessor):
-    def __init__(self, replacement_map: Dict):
+    def __init__(self, replacement_map: Dict[str, str]):
         self._replacement_map = replacement_map
 
     @staticmethod
@@ -98,3 +98,28 @@ class UniqueStringsPreprocessor(Preprocessor):
             executable_cells.append(cell)
         notebook.cells = executable_cells
         return notebook, resources
+
+class VertexAIInstallProprocessor(Preprocessor):
+    def __init__(self, vertex_ai_wheel):
+        self.vertex_ai_wheel = vertex_ai_wheel
+
+    @staticmethod
+    def update_vertex_ai_install(content: str):
+        if "google-cloud-aiplatform" not in content:
+            return content
+        return (
+            f"gcloud storage cp {self.vertex_ai_wheel} google-cloud-aiplatform.whl\n" + 
+            content.replace("google-cloud-aiplatform\n", "google-cloud-aiplatform.whl\n")
+            .replace("google-cloud-aiplatform ", "google-cloud-aiplatform.whl ")
+        )
+
+    def preprocess(self, notebook, resources=None):
+        executable_cells = []
+        for cell in notebook.cells:
+            if cell.cell_type == "code":
+                cell.source = self.update_vertex_ai_install(
+                    content=cell.source,
+                )
+
+            executable_cells.append(cell)
+        notebook.cells = executable_cells

.cloud-build/utils/build_results_viewer.py

@@ -0,0 +1,73 @@
+'''
+Viewer for the weekly regression testing of the official notebooks
+
+Cloud Storage location: gs://cloud-build-notebooks-presubmit/build_results/
+'''
+import argparse
+import json
+from  util import download_file
+import csv
+import datetime
+from google.cloud import storage
+
+BUILD_BUCKET = "cloud-build-notebooks-presubmit"
+BUILD_FOLDER = "build_results"
+
+
+parser = argparse.ArgumentParser()
+parser.add_argument('--file', dest='file',
+                    default=None, type=str, help='build results filei (local or GCS)')
+args = parser.parse_args()
+
+investigate = {}
+with open('investigate.csv', 'r') as csvfile:
+    reader = csv.reader(csvfile)
+    for row in reader:
+        investigate[row[0][:-6]] = row[1]
+
+if not args.file:
+    client = storage.Client()
+    blobs = client.list_blobs(BUILD_BUCKET, prefix=BUILD_FOLDER)
+    newest_time = datetime.datetime(2000, 1, 1)
+    for blob in blobs:
+        # individual PR
+        if blob.size < 2000:
+            continue
+        time_created = blob.time_created.replace(tzinfo=None)
+        if time_created > newest_time:
+            newest_time = time_created
+            args.file = f"gs://{BUILD_BUCKET}/{blob.name}"
+
+if args.file.startswith("gs://"):
+    path = args.file[5:]
+    bucket = path.split('/')[0]
+    file   = path[len(bucket)+1:]
+    download_file(bucket, file, "build.json")
+    args.file = "build.json"
+
+with open(args.file, 'r') as f:
+    results = json.load(f)
+
+for item in results.items():
+    notebook = item[0][len("/notebooks/official/")-1:-6]
+    if item[1]['passed']:
+        passed = "PASS"
+    else:
+        if notebook in investigate:
+            passed = "INVG"
+        else:
+            passed = "FAIL"
+      
+    error = item[1]['error_type']
+  
+    if passed == "FAIL":
+        if error == '':
+            error = "undetermined"
+        if 'log_url' in item[1]:
+            log_url = item[1]['log_url']
+        else:
+            log_url = ''
+    else:
+        log_url = ''
+
+    print(f"{notebook:75} {passed} {error:10} {log_url}")

.cloud-build/utils/investigate.csv

@@ -0,0 +1,19 @@
+notebook,status
+prediction/llm_streaming_prediction.ipynb,wait_for_fix
+custom/get_started_with_vertex_endpoint_and_shared_vm.ipynb,issue 2527
+feature_store/online_feature_serving_and_fetching_bigquery_data_with_feature_store.ipynb,wait_for_reaper
+feature_store/online_feature_serving_and_vector_retrieval_bigquery_data_with_feature_store.ipynb,wait_for_reaper
+pipelines/google_cloud_pipeline_components_TPU_model_train_upload_deploy.ipynb,wait_for_fix
+explainable_ai/sdk_custom_image_classification_batch_explain.ipynb,issue 2528
+explainable_ai/sdk_custom_image_classification_online_explain.ipynb,issue 2528
+explainable_ai/sdk_custom_tabular_regression_online_explain.ipynb,issue 2528
+explainable_ai/sdk_automl_tabular_binary_classification_batch_explain.ipynb,issue 2528
+explainable_ai/xai_image_classification_feature_attributions.ipynb,issue 2528
+matching_engine,sdk_matching_engine_create_stack_overflow_embeddings.ipynb,issue 2530
+automl/automl_forecasting_bqml_arima_plus_comparison.ipynb,flaky
+model_evaluation/custom_tabular_regression_model_evaluation.ipynb,regr
+experiments/get_started_with_vertex_experiments.ipynb,regr
+experiments/comparing_local_trained_models.ipynb,regr
+generative_ai/tune_peft.ipynb,internal
+pipelines/custom_model_training_and_batch_prediction.ipynb,regr
+feature_store/online_feature_serving_and_fetching_bigquery_data_with_feature_store_optimized.ipynb,wait_for_reaper

.cloud-build/utils/util.py

@@ -15,7 +15,7 @@ def download_file(bucket_name: str, blob_name: str, destination_file: str) -> st
     remote_file_path = "".join(["gs://", "/".join([bucket_name, blob_name])])
 
     subprocess.check_output(
-        ["gsutil", "cp", remote_file_path, destination_file], encoding="UTF-8"
+        ["gcloud", "storage", "cp", remote_file_path, destination_file], encoding="UTF-8"
     )
 
     return destination_file
@@ -27,7 +27,7 @@ def upload_file(
 ) -> str:
     """Copies a local file to a GCS path"""
     subprocess.check_output(
-        ["gsutil", "cp", local_file_path, remote_file_path], encoding="UTF-8"
+        ["gcloud", "storage", "cp", local_file_path, remote_file_path], encoding="UTF-8"
     )
 
     return remote_file_path

.cloud-build/web-index/do_not_index.txt

@@ -0,0 +1,11 @@
+sdk2_remote_tabnet_training.ipynb
+remote_hyperparameter_tuning.ipynb
+remote_prediction.ipynb
+remote_training_bigframes_pytorch.ipynb
+remote_training_bigframes_sklearn.ipynb
+remote_training_bigframes_tensorflow.ipynb
+remote_training_lightning.ipynb
+remote_training_pytorch.ipynb
+remote_training_sklearn.ipynb
+remote_training_tensorflow_with_autologging.ipynb
+

.cloud-build/web-index/notebook-index-cloudbuild.yaml

@@ -0,0 +1,23 @@
+steps:
+  # Fetch full repo for diff purposes
+  - name: gcr.io/cloud-builders/git
+    args: [fetch, --unshallow, --quiet]
+  # Create a virtual environment
+  - name: ${_PYTHON_IMAGE}
+    entrypoint: /bin/sh
+    args:
+    - -c
+    - python3 -m venv workspace/env
+  # Install Python dependencies and run testing script
+  - name: ${_PYTHON_IMAGE}
+    entrypoint: /bin/sh
+    args:
+    - -c
+    - |
+      . workspace/env/bin/activate &&
+      python3 notebooks/notebook_template_review.py --web --title --steps --desc --linkback --notebook-dir=notebooks/official --skip-file=${_DO_NOT_INDEX_FILE} >web.html
+artifacts:
+  objects:
+          location: gs://${_GCS_ARTIFACTS_BUCKET}/webdoc
+          paths: ['web.html']
+timeout: 86400s

.github/dependabot.yml

@@ -0,0 +1,10 @@
+version: 2
+updates:
+
+  # Ignore model garden dockerfiles:
+  - package-ecosystem: "npm"
+    directory: "/community-content/vertex_model_garden"
+    schedule:
+      interval: "monthly"
+    ignore:
+      - dependency-name: "*"

.github/workflows/ci.yaml

@@ -7,11 +7,11 @@ jobs:
     runs-on: ubuntu-latest
     steps:
     - name: Set up Python
-      uses: actions/setup-python@v4
+      uses: actions/setup-python@v6
       with:
-        python-version: '3.x'      
+        python-version: '3.12'      
     - name: Fetch pull request branch
-      uses: actions/checkout@v3
+      uses: actions/checkout@v6
       with:
         fetch-depth: 0      
     - name: Fetch base main branch

.github/workflows/linter/Dockerfile

@@ -4,7 +4,7 @@
 # 2. To lint specific notebooks:
 # docker run -v ${PWD}:/setup/app gcr.io/python-docs-samples-tests/notebook_linter:latest notebooks/1.ipynb notebooks/2.ipynb
 
-FROM python:3.10
+FROM python:3.14
 
 WORKDIR setup
 

.github/workflows/linter/requirements.txt

@@ -2,9 +2,9 @@ git+https://github.com/tensorflow/docs
 ipython
 jupyter
 nbconvert
-black==22.10.0
-pyupgrade==2.38.4
-isort==5.10.1
-flake8==6.0.0
-nbqa==1.5.3
+black==26.3.1
+pyupgrade==3.21.2
+isort==7.0.0
+flake8==7.3.0
+nbqa==1.9.1
 

.github/workflows/linter/run_linter.sh

@@ -58,7 +58,7 @@ done
 # Only check notebooks in test folders modified in this pull request.
 # Note: Use process substitution to persist the data in the array
 if [ ${#notebooks[@]} -eq 0 ]; then
-    echo "Checking for changed notebooked using git"
+    echo "Checking for changed notebooks using git"
     while read -r file || [ -n "$line" ]; do
         notebooks+=("$file")
     done < <(git diff --name-only main... | grep '\.ipynb$')
@@ -84,7 +84,7 @@ if [ ${#notebooks[@]} -gt 0 ]; then
                 # python3 -m nbqa black "$notebook" --check
                 # BLACK_RTN=$?
                 echo "Running pyupgrade..."
-                python3 -m nbqa pyupgrade "$notebook"
+                python3 -m nbqa pyupgrade --exit-zero-even-if-changed "$notebook"
                 PYUPGRADE_RTN=$?
                 echo "Running isort..."
                 python3 -m nbqa isort "$notebook" --check
@@ -97,7 +97,7 @@ if [ ${#notebooks[@]} -gt 0 ]; then
                 python3 -m nbqa black "$notebook"
                 BLACK_RTN=$?
                 echo "Running pyupgrade..."
-                python3 -m nbqa pyupgrade "$notebook"
+                python3 -m nbqa pyupgrade --exit-zero-even-if-changed "$notebook"
                 PYUPGRADE_RTN=$?
                 echo "Running isort..."
                 python3 -m nbqa isort "$notebook"

CONTRIBUTING.md

@@ -44,12 +44,10 @@ Finally, run this code block to check for errors. Each step will attempt to
 automatically fix any issues. If the fixes can't be performed automatically,
 then you will need to manually address them before submitting your PR.
 
+Note: For official, only submit one notebook per PR.
+
 ```shell
-nbqa black "$notebook"
-nbqa pyupgrade "$notebook"
-nbqa isort "$notebook"
-nbqa flake8 "$notebook" --extend-ignore=W391,E501,F821,E402,F404,W503,E203,E722,W293,W291
-python3 -m tensorflow_docs.tools.nbfmt --remove_outputs "$notebook"
+docker run -v ${PWD}:/setup/app gcr.io/cloud-devrel-public-resources/notebook_linter:latest your_notebook
 ```
 
 ## Code Reviews

README.md

@@ -1,37 +1,192 @@
-# Google Cloud Vertex AI Samples
+# ![Google Cloud](https://avatars.githubusercontent.com/u/2810941?s=60&v=4) Google Cloud Vertex AI Samples
 
-[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](LICENSE)
-
-Welcome to the Google Cloud [Vertex AI](https://cloud.google.com/vertex-ai/docs/) sample repository.
+This repository contains notebooks, code samples, sample apps, skills, and other resources that demonstrate how to use, develop and manage machine learning and generative AI workflows using Google Cloud Vertex AI.
 
 ## Overview
 
-The repository contains [notebooks](https://github.com/GoogleCloudPlatform/vertex-ai-samples/tree/master/notebooks) and [community content](https://github.com/GoogleCloudPlatform/vertex-ai-samples/tree/master/community-content) that demonstrate how to develop and manage ML workflows using Google Cloud Vertex AI.
+[Vertex AI](https://cloud.google.com/vertex-ai) is a fully-managed, unified AI development platform for building and using generative AI. This repository is designed to help you get started with Vertex AI. Whether you're new to Vertex AI or an experienced ML practitioner, you'll find valuable resources here.
+
+⚠️ For more Vertex AI Generative AI notebook samples, please visit the Vertex AI [Generative AI](https://github.com/GoogleCloudPlatform/generative-ai) GitHub repository.
+
+## Explore, learn and contribute
+
+You can explore, learn, and contribute to this repository to unleash the full potential of machine learning on Vertex AI! 
+
+### Explore and learn
+
+Explore this repository, follow the links in the header section of each of the notebooks to -
+
+- Open and run the notebook in [Colab](https://colab.google/)
+- Open and run the notebook in [Colab Enterprise](https://cloud.google.com/colab/docs/introduction)
+- Open and run the notebook in [Vertex AI Workbench](https://cloud.google.com/vertex-ai/docs/workbench/introduction)
+- View the notebook on Github
+  
+### Contribute
+
+See the [Contributing Guide](https://github.com/GoogleCloudPlatform/vertex-ai-samples/blob/master/CONTRIBUTING.md).
+
+## Get started
+
+To get started using Vertex AI, you must have a Google Cloud project. 
+
+- If you don't have a Google Cloud project, you can learn and build on GCP for free using [Free Trail](https://cloud.google.com/free).
+- Once you have a Google Cloud project, you can learn more about [setting up a project and a development environment](https://cloud.google.com/vertex-ai/docs/start/cloud-environment).
+
 
 ## Repository structure
 
-```bash
-├── community-content - Sample code and tutorials contributed by the community
+```text
 ├── notebooks
-│   ├── community - Notebooks contributed by the community
 │   ├── official - Notebooks demonstrating use of each Vertex AI service
 │   │   ├── automl
 │   │   ├── custom
 │   │   ├── ...
+│   ├── community - Notebooks contributed by the community
+│   │   ├── model_garden
+│   │   ├── ...
+├── community-content - Sample code and tutorials contributed by the community
+├── docs - Deep-dive documentation and advanced setup guides
+└── skills - Suite of AI Agent "Skills" for Vertex AI
+    ├── README.md               # Developer guide for Vertex AI skills
+    ├── vertex-ai/              # Primary router for Vertex AI tasks
+    │   └── SKILL.md            # Entry point that routes across capabilities
+    ├── genai-sdk/              # Gemini API usage with Gen AI SDK
+    │   └── SKILL.md            # Guides for Python, JS/TS, Go, Java, C#
+    ├── vertex-deploy/          # Deploying models to Endpoints
+    │   └── SKILL.md            # Commands for open models & custom weights
+    ├── vertex-inference/       # Inferencing with GenAI models
+    │   └── SKILL.md            # Code samples for Gemini and OpenMaaS
+    └── vertex-tuning/          # Secondary router for model fine-tuning
+        ├── SKILL.md            # Router for tuning tasks
+        ├── gemini/             # Fine-tuning first-party Gemini models
+        │   └── SKILL.md
+        └── open-model/         # Fine-tuning third-party open models
+            └── SKILL.md
 ```
+## Examples
 
-## Contributing
+<!-- markdownlint-disable MD033 -->
+<table>
 
-Contributions welcome! See the [Contributing Guide](https://github.com/GoogleCloudPlatform/vertex-ai-samples/blob/master/CONTRIBUTING.md).
+  <tr>
+    <th style="text-align: center;">Category</th>
+    <th style="text-align: center;">Product</th>
+    <th style="text-align: center;">Description</th>
+  </tr>
+  <tr>
+    <td>Model</td>
+    <td>
+      <a href="notebooks/community/model_garden"><code>Model Garden/</code></a>
+    </td>
+    <td>
+      Curated collection of first-party, open-source, and third-party models available on Vertex AI including Gemini, Gemma, Llama 3, Claude 3 and many more.
+    </td>
+  </tr>
+  <tr>
+    <td>Data</td>
+    <td>
+      <a href="notebooks/official/feature_store"><code>Feature Store/</code></a>
+    </td>
+    <td>
+      Set up and manage online serving using Vertex AI Feature Store.
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/datasets"><code>datasets/</code></a>
+    </td>
+    <td>
+      Use BigQuery and Data Labeling service with Vertex AI.
+    </td>
+  </tr>
+  <tr>
+    <td>Model development</td>
+    <td>
+      <a href="notebooks/official/automl"><code>automl/</code></a>
+    </td>
+    <td>
+      Train and make predictions on AutoML models
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/custom"><code>custom/</code></a>
+    </td>
+    <td>
+      Create, deploy and serve custom  models on Vertex AI
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/ray_on_vertex_ai"><code>ray_on_vertex_ai/</code></a>
+    </td>
+    <td>
+      Use Colab Enterprise and Vertex AI SDK for Python to connect to the Ray Cluster.
+    </td>
+  </tr>
+  <tr>
+    <td>Deploy and use</td>
+    <td>
+      <a href="notebooks/official/prediction"><code>prediction/</code></a>
+    </td>
+    <td>
+      Build, train and deploy models using prebuilt containers for custom training and prediction.
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/model_registry"><code>model_registry/</code></a>
+    </td>
+    <td>
+      Use Model Registry to create and register a model.
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/explainable_ai"><code>Explainable AI/</code></a>
+    </td>
+    <td>
+      Use Vertex Explainable AI's feature-based and example-based explanations to explain how or why a model produced a specific prediction.
+    </td>
+  </tr>
+  <tr>
+    <td></td>
+    <td>
+      <a href="notebooks/official/ml_metadata"><code>ml_metadata/</code></a>
+    </td>
+    <td>
+      Record the metadata and artifacts and query that metadata to help analyze, debug, and audit the performance of your ML system.
+    </td>
+  </tr>
+  <tr>
+    <td>Tools</td>
+    <td>
+      <a href="notebooks/official/pipelines"><code>Pipelines/</code></a>
+    </td>
+    <td>
+      Use `Vertex AI Pipelines` and `Google Cloud Pipeline Components` to build, tune, or deploy a custom model.
+    </td>
+  </tr>
+</table>
+<!-- markdownlint-enable MD033 -->
 
-## Getting help
 
-Please use the [issues page](https://github.com/GoogleCloudPlatform/vertex-ai-samples/issues) to provide feedback or submit a bug report.
+## Get help
+
+Please use the [Issues page](https://github.com/GoogleCloudPlatform/vertex-ai-samples/issues) to provide feedback or submit a bug report.
 
 ## Disclaimer
 
 This is not an officially supported Google product. The code in this repository is for demonstrative purposes only.
 
-## Feedback
 
-Please feel free to fill out our [survey](https://bit.ly/vertex-ai-samples-survey) to give us feedback on the repo and its content.
+## References
+- [Vertex AI Jupyter Notebook tutorials](https://cloud.google.com/vertex-ai/docs/tutorials/jupyter-notebooks)
+- Vertex AI [Generative AI](https://github.com/GoogleCloudPlatform/generative-ai) GitHub repository
+- [Vertex AI documentaton](https://cloud.google.com/vertex-ai/docs)
+  

community-content/CODEOWNERS

@@ -9,3 +9,25 @@
 /Train_tabular_models_with_many_frameworks_and_import_to_Vertex_AI_using_Pipelines @Ark-kun
 /pipeline_components @Ark-kun
 /pipeline_components/image_ml_model_training @lakeyk
+/prediction_featurestore_integration @googleapis/vertex-prediction-team
+/vertex_model_garden/model_oss/notebook_util @minwoo33park
+/vertex_model_garden/model_oss/util @weigary
+/vertex_model_garden/model_oss/diffusers @weigary
+/vertex_model_garden/model_oss/keras @dstnluong-google
+/vertex_model_garden/model_oss/transformers @dstnluong-google
+/vertex_model_garden/model_oss/pic2word @jismailyan-google
+/vertex_model_garden/model_oss/open_clip @lydhr
+/vertex_model_garden/model_oss/movinet @KCFindstr
+/vertex_model_garden/model_oss/data_converter @KCFindstr
+/vertex_model_garden/model_oss/peft @weigary
+/vertex_model_garden/model_oss/peft/templates @rayandasoriya
+/vertex_model_garden/model_oss/lm-evaluation-harness @kathyyu-google
+/vertex_model_garden/model_oss/tfvision @dstnluong-google
+/vertex_model_garden/model_oss/fvlm @minwoo33park
+/vertex_model_garden/model_oss/imagebind @kathyyu-google
+/vertex_model_garden/model_oss/llava @py4
+/vertex_model_garden/model_oss/vllm @kathyyu-google
+/vertex_model_garden/benchmarking_reports @lavraicse
+/vertex_model_garden/model_oss/autogluon @lavraicse
+/vertex_distributed_training/a3mega/llama-3-8b-nemo-pretraining @mstyer-google @erwinh85 @mchrestkha
+

community-content/Train_tabular_models_with_many_frameworks_and_import_to_Vertex_AI_using_Pipelines/Train_tabular_classification_model_using_XGBoost_and_import_to_Vertex_AI/pipeline.py

@@ -9,7 +9,7 @@ binarize_column_using_Pandas_on_CSV_data_op = components.load_component_from_url
 split_rows_into_subsets_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/dataset_manipulation/Split_rows_into_subsets/in_CSV/component.yaml")
 train_XGBoost_model_on_CSV_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/XGBoost/Train/component.yaml")
 xgboost_predict_on_CSV_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/XGBoost/Predict/component.yaml")
-upload_XGBoost_model_to_Google_Cloud_Vertex_AI_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/google-cloud/Vertex_AI/Models/Upload_XGBoost_model/component.yaml")
+upload_XGBoost_model_to_Google_Cloud_Vertex_AI_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/d5c9918850a6cc70004c4269dae066cfe2e664eb/community-content/pipeline_components/google-cloud/Vertex_AI/Models/Upload_XGBoost_model/component.yaml")
 deploy_model_to_endpoint_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/google-cloud/Vertex_AI/Models/Deploy_to_endpoint/component.yaml")
 
 # %% Pipeline definition

community-content/Train_tabular_models_with_many_frameworks_and_import_to_Vertex_AI_using_Pipelines/Train_tabular_classification_model_using_all_frameworks_and_import_to_Vertex_AI/pipeline.py

@@ -23,7 +23,7 @@ upload_PyTorch_model_archive_to_Google_Cloud_Vertex_AI_op = components.load_comp
 # XGBoost
 train_XGBoost_model_on_CSV_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/XGBoost/Train/component.yaml")
 xgboost_predict_on_CSV_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/XGBoost/Predict/component.yaml")
-upload_XGBoost_model_to_Google_Cloud_Vertex_AI_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/google-cloud/Vertex_AI/Models/Upload_XGBoost_model/component.yaml")
+upload_XGBoost_model_to_Google_Cloud_Vertex_AI_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/d5c9918850a6cc70004c4269dae066cfe2e664eb/community-content/pipeline_components/google-cloud/Vertex_AI/Models/Upload_XGBoost_model/component.yaml")
 
 # Scikit-learn
 #train_linear_regression_model_using_scikit_learn_from_CSV_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/1f5cf6e06409b704064b2086c0a705e4e6b4fcde/community-content/pipeline_components/ML_frameworks/Scikit_learn/Train_linear_regression_model/from_CSV/component.yaml")

community-content/Train_tabular_models_with_many_frameworks_and_import_to_Vertex_AI_using_Pipelines/Train_tabular_regression_model_using_XGBoost_and_import_to_Vertex_AI/pipeline.py

@@ -8,7 +8,7 @@ fill_all_missing_values_using_Pandas_on_CSV_data_op = components.load_component_
 split_rows_into_subsets_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/dataset_manipulation/Split_rows_into_subsets/in_CSV/component.yaml")
 train_XGBoost_model_on_CSV_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/XGBoost/Train/component.yaml")
 xgboost_predict_on_CSV_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/XGBoost/Predict/component.yaml")
-upload_XGBoost_model_to_Google_Cloud_Vertex_AI_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/google-cloud/Vertex_AI/Models/Upload_XGBoost_model/component.yaml")
+upload_XGBoost_model_to_Google_Cloud_Vertex_AI_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/d5c9918850a6cc70004c4269dae066cfe2e664eb/community-content/pipeline_components/google-cloud/Vertex_AI/Models/Upload_XGBoost_model/component.yaml")
 deploy_model_to_endpoint_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/google-cloud/Vertex_AI/Models/Deploy_to_endpoint/component.yaml")
 
 # %% Pipeline definition

community-content/Train_tabular_models_with_many_frameworks_and_import_to_Vertex_AI_using_Pipelines/Train_tabular_regression_model_using_all_frameworks_and_import_to_Vertex_AI/pipeline.py

@@ -22,7 +22,7 @@ upload_PyTorch_model_archive_to_Google_Cloud_Vertex_AI_op = components.load_comp
 # XGBoost
 train_XGBoost_model_on_CSV_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/XGBoost/Train/component.yaml")
 xgboost_predict_on_CSV_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/XGBoost/Predict/component.yaml")
-upload_XGBoost_model_to_Google_Cloud_Vertex_AI_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/399405402d95f4a011e2d2e967c96f8508ba5688/community-content/pipeline_components/google-cloud/Vertex_AI/Models/Upload_XGBoost_model/component.yaml")
+upload_XGBoost_model_to_Google_Cloud_Vertex_AI_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/d5c9918850a6cc70004c4269dae066cfe2e664eb/community-content/pipeline_components/google-cloud/Vertex_AI/Models/Upload_XGBoost_model/component.yaml")
 
 # Scikit-learn
 train_linear_regression_model_using_scikit_learn_from_CSV_op = components.load_component_from_url("https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/1f5cf6e06409b704064b2086c0a705e4e6b4fcde/community-content/pipeline_components/ML_frameworks/Scikit_learn/Train_linear_regression_model/from_CSV/component.yaml")

community-content/cpr-examples/timm_serving/requirements.txt

@@ -1,5 +1,5 @@
 absl-py==1.1.0
-fastapi==0.75.2
+fastapi==0.109.1
 uvicorn==0.18.2
 timm==0.5.4
 smart_open==6.0.0

community-content/pipeline_components/google-cloud/Vertex_AI/Models/Upload_PyTorch_model_archive/component.yaml

@@ -148,7 +148,7 @@ implementation:
 
       # Downloading the model archive from GCS
       # TODO: Fix gsutil bugs (requires project ID, has auth issues) and use gsutil instead.
-      # gsutil cp "$model_archive_uri" "$model_archive_local_path"
+      # gcloud storage cp "$model_archive_uri" "$model_archive_local_path"
       pip install google-cloud-storage
       python -c '
       import sys

community-content/pipeline_components/google-cloud/Vertex_AI/Models/Upload_XGBoost_model/component.yaml

@@ -64,8 +64,8 @@ implementation:
           labels["component-source"] = "github-com-ark-kun-pipeline-components"
 
           # The serving container decides the model type based on the model file extension.
-          # So we need to rename the mode file (e.g. /tmp/inputs/model/data) to *.pkl
-          _, renamed_model_path = tempfile.mkstemp(suffix=".pkl")
+          # So we need to rename the mode file (e.g. /tmp/inputs/model/data) to *.bst
+          _, renamed_model_path = tempfile.mkstemp(suffix=".bst")
           shutil.copyfile(src=model_path, dst=renamed_model_path)
 
           model = aiplatform.Model.upload_xgboost_model_file(

community-content/pipeline_components/google-cloud/storage/download/component.yaml

@@ -24,12 +24,12 @@ implementation:
 
             # Checking whether the URI points to a single blob, a directory or a URI pattern
             # URI points to a blob when that URI does not end with slash and listing that URI only yields the same URI
-            if [[ "$uri" != */ ]] && (gsutil ls "$uri" | grep --fixed-strings --line-regexp "$uri"); then
+            if [[ "$uri" != */ ]] && (gcloud storage ls "$uri" | grep --fixed-strings --line-regexp "$uri"); then
                 mkdir -p "$(dirname "$output_path")"
-                gsutil -m cp -r "$uri" "$output_path"
+                gcloud storage cp --recursive "$uri" "$output_path"
             else
                 mkdir -p "$output_path" # When source path is a directory, gsutil requires the destination to also be a directory
-                gsutil -m rsync -r "$uri" "$output_path" # gsutil cp has different path handling than Linux cp. It always puts the source directory (name) inside the destination directory. gsutil rsync does not have that problem.
+                gcloud storage rsync --recursive "$uri" "$output_path" # gsutil cp has different path handling than Linux cp. It always puts the source directory (name) inside the destination directory. gsutil rsync does not have that problem.
             fi
         - inputValue: GCS path
         - outputPath: Data

community-content/pipeline_components/image_ml_model_training/load_image_classification_model/component.yaml

@@ -87,7 +87,7 @@ outputs:
 - {name: image_size_path, type: HeightWidth}
 implementation:
   container:
-    image: us-docker.pkg.dev/vertex-ai/ready-to-go-image-classification/image-components:v0.1
+    image: us-docker.pkg.dev/vertex-ai/ready-to-go-image-classification/image-components:v0.2
     # command is a list of strings (command-line arguments).
     # The YAML language has two syntaxes for lists and you can use either of them.
     # Here we use the "flow syntax" - comma-separated strings inside square brackets.
@@ -109,4 +109,4 @@ implementation:
       {inputValue: l2_regularization_penalty},
       --image-size-path,
       {outputPath: image_size_path},
-    ]
+    ]

community-content/pipeline_components/image_ml_model_training/preprocess_image_data/component.yaml

@@ -34,7 +34,7 @@ outputs:
     path for the validation data,'}
 implementation:
   container:
-    image: us-docker.pkg.dev/vertex-ai/ready-to-go-image-classification/image-components:v0.1
+    image: us-docker.pkg.dev/vertex-ai/ready-to-go-image-classification/image-components:v0.2
     # command is a list of strings (command-line arguments).
     # The YAML language has two syntaxes for lists and you can use either of them.
     # Here we use the "flow syntax" - comma-separated strings inside square brackets.

community-content/pipeline_components/image_ml_model_training/train_image_classification_model/component.yaml

@@ -55,7 +55,7 @@ outputs:
     for the saved model,'}
 implementation:
   container:
-    image: us-docker.pkg.dev/vertex-ai/ready-to-go-image-classification/image-components:v0.1
+    image: us-docker.pkg.dev/vertex-ai/ready-to-go-image-classification/image-components:v0.2
     # command is a list of strings (command-line arguments).
     # The YAML language has two syntaxes for lists and you can use either of them.
     # Here we use the "flow syntax" - comma-separated strings inside square brackets.

community-content/pipeline_components/image_ml_model_training/transcode_tfrecord_image_dataset_from_csv/component.yaml

@@ -20,7 +20,7 @@ outputs:
     path for the TFRecord image data}
 implementation:
   container:
-    image: us-docker.pkg.dev/vertex-ai/ready-to-go-image-classification/image-components:v0.1
+    image: us-docker.pkg.dev/vertex-ai/ready-to-go-image-classification/image-components:v0.2
     # command is a list of strings (command-line arguments).
     # The YAML language has two syntaxes for lists and you can use either of them.
     # Here we use the "flow syntax" - comma-separated strings inside square brackets.

community-content/pipeline_components/image_ml_model_training/transcode_tfrecord_image_dataset_from_jsonl/component.yaml

@@ -22,7 +22,7 @@ outputs:
     path for the TFRecord image data}
 implementation:
   container:
-    image: us-docker.pkg.dev/vertex-ai/ready-to-go-image-classification/image-components:v0.1
+    image: us-docker.pkg.dev/vertex-ai/ready-to-go-image-classification/image-components:v0.2
     # command is a list of strings (command-line arguments).
     # The YAML language has two syntaxes for lists and you can use either of them.
     # Here we use the "flow syntax" - comma-separated strings inside square brackets.

community-content/pytorch_image_classification_distributed_data_parallel_training_with_vertex_sdk/trainer/Dockerfile

@@ -1,16 +1,40 @@
-FROM pytorch/pytorch:1.8.1-cuda11.1-cudnn8-runtime
+# Stage 1: Build Environment
+FROM pytorch/pytorch:1.8.1-cuda11.1-cudnn8-runtime AS builder
 
+# Install necessary tools and dependencies
 RUN apt-get update && \
     apt-get install -y curl gnupg && \
     echo "deb [signed-by=/usr/share/keyrings/cloud.google.gpg] http://packages.cloud.google.com/apt cloud-sdk main" | tee -a /etc/apt/sources.list.d/google-cloud-sdk.list && \
-    curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key --keyring /usr/share/keyrings/cloud.google.gpg  add - && \
+    curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key --keyring /usr/share/keyrings/cloud.google.gpg add - && \
     apt-get update -y && \
-    apt-get install google-cloud-sdk -y
+    apt-get install -y google-cloud-sdk
 
+# Copy application code
 COPY . /trainer
 
+# Set working directory
 WORKDIR /trainer
 
-RUN pip install -r requirements.txt
+# Install Python dependencies
+RUN pip install --no-cache-dir -r requirements.txt
 
-ENTRYPOINT ["python", "-m", "task"]
+# Stage 2: Runtime Environment
+FROM pytorch/pytorch:1.8.1-cuda11.1-cudnn8-runtime
+
+# Install Google Cloud SDK
+RUN apt-get update && \
+    apt-get install -y curl gnupg && \
+    echo "deb [signed-by=/usr/share/keyrings/cloud.google.gpg] http://packages.cloud.google.com/apt cloud-sdk main" | tee -a /etc/apt/sources.list.d/google-cloud-sdk.list && \
+    curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | apt-key --keyring /usr/share/keyrings/cloud.google.gpg add - && \
+    apt-get update -y && \
+    apt-get install -y google-cloud-sdk && \
+    apt-get clean && rm -rf /var/lib/apt/lists/*
+
+# Copy from the builder stage
+COPY --from=builder /trainer /trainer
+
+# Set working directory
+WORKDIR /trainer
+
+# Set the entry point
+ENTRYPOINT ["python", "-m", "task"]

community-content/pytorch_image_classification_distributed_data_parallel_training_with_vertex_sdk/trainer/requirements.txt

@@ -1,3 +1,3 @@
-torch==1.8.1
+torch==2.8.0
 torchvision==0.9.1
 tensorboard==2.5.0

community-content/pytorch_image_classification_single_gpu_with_vertex_sdk_and_torchserve/trainer/requirements.txt

@@ -1,3 +1,3 @@
-torch==1.8.1
+torch==2.7.0
 torchvision==0.9.1
 tensorboard==2.5.0

community-content/pytorch_image_classification_single_gpu_with_vertex_sdk_and_torchserve/vertex_prediction_with_custom_torchserve_container.ipynb

@@ -110,7 +110,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil ls $gcs_output_uri_prefix"
+        "! gcloud storage ls $gcs_output_uri_prefix"
       ]
     },
     {
@@ -192,7 +192,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil cp -r $gcs_output_uri_prefix/model ./model_server/"
+        "! gcloud storage cp --recursive $gcs_output_uri_prefix/model ./model_server/"
       ]
     },
     {
@@ -556,7 +556,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil rm -rf $gcs_output_uri_prefix"
+        "! gcloud storage rm --recursive --continue-on-error $gcs_output_uri_prefix"
       ]
     },
     {

community-content/pytorch_image_classification_single_gpu_with_vertex_sdk_and_torchserve/vertex_training_with_custom_container.ipynb

@@ -412,7 +412,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil ls $gcs_output_uri_prefix"
+        "! gcloud storage ls $gcs_output_uri_prefix"
       ]
     }
   ],

community-content/pytorch_text_classification_using_vertex_sdk_and_gcloud/custom_container/scripts/train-cloud.sh

@@ -77,4 +77,4 @@ echo "After the job is completed successfully, model files will be saved at $JOB
 
 # # Verify the model was exported
 # echo "Verify the model was exported:"
-# gsutil ls ${JOB_DIR}/
+# gcloud storage ls ${JOB_DIR}/

community-content/pytorch_text_classification_using_vertex_sdk_and_gcloud/predictor/Dockerfile.serve

@@ -34,4 +34,4 @@ RUN echo "service_envelope=json\n"     "inference_address=http://0.0.0.0:${AIP_H
 USER model-server
 
 # run Torchserve HTTP serve to respond to prediction requests
-CMD ["echo", "AIP_STORAGE_URI=${AIP_STORAGE_URI}", ";",     "gsutil", "cp", "-r", "${AIP_STORAGE_URI}/${MODEL_NAME}.mar", "/home/model-server/model-store/", ";",     "ls", "-ltr", "/home/model-server/model-store/", ";",     "torchserve", "--start", "--ts-config=/home/model-server/config.properties",     "--models", "${MODEL_NAME}=${MODEL_NAME}.mar",     "--model-store", "/home/model-server/model-store"]
+CMD ["echo", "AIP_STORAGE_URI=${AIP_STORAGE_URI}", ";",     "gcloud", "storage", "cp", "--recursive", "${AIP_STORAGE_URI}/${MODEL_NAME}.mar", "/home/model-server/model-store/", ";",     "ls", "-ltr", "/home/model-server/model-store/", ";",     "torchserve", "--start", "--ts-config=/home/model-server/config.properties",     "--models", "${MODEL_NAME}=${MODEL_NAME}.mar",     "--model-store", "/home/model-server/model-store"]

community-content/pytorch_text_classification_using_vertex_sdk_and_gcloud/python_package/scripts/train-cloud.sh

@@ -67,4 +67,4 @@ echo "After the job is completed successfully, model files will be saved at $JOB
 
 # # Verify the model was exported
 # echo "Verify the model was exported:"
-# gsutil ls ${JOB_DIR}/
+# gcloud storage ls ${JOB_DIR}/

community-content/tf_agents_bandits_movie_recommendation_with_kfp_and_vertex_sdk/mlops_pipeline_tf_agents_bandits_movie_recommendation/mlops_pipeline_tf_agents_bandits_movie_recommendation.ipynb

@@ -478,8 +478,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil mb -l $REGION $BUCKET_NAME"
-      ]
+        "! gcloud storage buckets create --location $REGION $BUCKET_NAME"      ]
     },
     {
       "cell_type": "markdown",
@@ -498,8 +497,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil ls -al $BUCKET_NAME"
-      ]
+        "! gcloud storage ls --all-versions --long $BUCKET_NAME"      ]
     },
     {
       "cell_type": "markdown",
@@ -582,8 +580,7 @@
       "outputs": [],
       "source": [
         "# Download the sample data into your RAW_DATA_PATH\n",
-        "! gsutil cp \"gs://cloud-samples-data/vertex-ai/community-content/tf_agents_bandits_movie_recommendation_with_kfp_and_vertex_sdk/u.data\" $RAW_DATA_PATH"
-      ]
+        "! gcloud storage cp \"gs://cloud-samples-data/vertex-ai/community-content/tf_agents_bandits_movie_recommendation_with_kfp_and_vertex_sdk/u.data\" $RAW_DATA_PATH"      ]
     },
     {
       "cell_type": "code",
@@ -1621,9 +1618,7 @@
         "! gcloud scheduler jobs delete $SIMULATOR_SCHEDULER_JOB --quiet\n",
         "\n",
         "# Delete Cloud Storage objects that were created.\n",
-        "! gsutil -m rm -r $PIPELINE_ROOT\n",
-        "! gsutil -m rm -r $TRAINING_ARTIFACTS_DIR"
-      ]
+        "! gcloud storage rm --recursive $PIPELINE_ROOT\n",        "! gcloud storage rm --recursive $TRAINING_ARTIFACTS_DIR"      ]
     }
   ],
   "metadata": {

community-content/tf_agents_bandits_movie_recommendation_with_kfp_and_vertex_sdk/mlops_pipeline_tf_agents_bandits_movie_recommendation/src/logger/requirements.txt

@@ -1,4 +1,4 @@
 google-cloud-bigquery==2.20.0
-tensorflow==2.7.2
-pillow==9.0.1
+tensorflow==2.12.1
+pillow==12.2.0
 tf-agents==0.8.0

community-content/tf_agents_bandits_movie_recommendation_with_kfp_and_vertex_sdk/mlops_pipeline_tf_agents_bandits_movie_recommendation/src/prediction_container/requirements.txt

@@ -1,4 +1,4 @@
 google-cloud-pubsub==2.5.0
-pillow==9.0.1
+pillow==12.2.0
 tf-agents==0.8.0
-tensorflow==2.7.2
+tensorflow==2.12.1

community-content/tf_agents_bandits_movie_recommendation_with_kfp_and_vertex_sdk/mlops_pipeline_tf_agents_bandits_movie_recommendation/src/simulator/requirements.txt

@@ -1,5 +1,5 @@
 dataclasses==0.6
 google-cloud-aiplatform==1.8.1
-tensorflow==2.7.2
-pillow==9.0.1
+tensorflow==2.12.1
+pillow==12.1.1
 tf-agents==0.8.0

community-content/tf_agents_bandits_movie_recommendation_with_kfp_and_vertex_sdk/step_by_step_sdk_tf_agents_bandits_movie_recommendation/step_by_step_sdk_tf_agents_bandits_movie_recommendation.ipynb

@@ -398,6 +398,7 @@
         "if not IS_GOOGLE_CLOUD_NOTEBOOK:\n",
         "    if \"google.colab\" in sys.modules:\n",
         "        from google.colab import auth as google_auth\n",
+        "\n",
         "        google_auth.authenticate_user()\n",
         "\n",
         "    # If you are running this notebook locally, replace the string below with the\n",
@@ -472,7 +473,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil mb -l $REGION $BUCKET_NAME"
+        "! gcloud storage buckets create --location $REGION $BUCKET_NAME"
       ]
     },
     {
@@ -492,7 +493,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil ls -al $BUCKET_NAME"
+        "! gcloud storage ls --all-versions --long $BUCKET_NAME"
       ]
     },
     {
@@ -565,7 +566,7 @@
       "outputs": [],
       "source": [
         "# Copy the sample data into your DATA_PATH\n",
-        "! gsutil cp \"gs://cloud-samples-data/vertex-ai/community-content/tf_agents_bandits_movie_recommendation_with_kfp_and_vertex_sdk/u.data\"  $DATA_PATH"
+        "! gcloud storage cp \"gs://cloud-samples-data/vertex-ai/community-content/tf_agents_bandits_movie_recommendation_with_kfp_and_vertex_sdk/u.data\"  $DATA_PATH"
       ]
     },
     {
@@ -579,11 +580,15 @@
         "# Set hyperparameters.\n",
         "BATCH_SIZE = 8  # @param {type:\"integer\"} Training and prediction batch size.\n",
         "TRAINING_LOOPS = 5  # @param {type:\"integer\"} Number of training iterations.\n",
-        "STEPS_PER_LOOP = 2  # @param {type:\"integer\"} Number of driver steps per training iteration.\n",
+        "STEPS_PER_LOOP = (\n",
+        "    2  # @param {type:\"integer\"} Number of driver steps per training iteration.\n",
+        ")\n",
         "\n",
         "# Set MovieLens simulation environment parameters.\n",
         "RANK_K = 20  # @param {type:\"integer\"} Rank for matrix factorization in the MovieLens environment; also the observation dimension.\n",
-        "NUM_ACTIONS = 20  # @param {type:\"integer\"} Number of actions (movie items) to choose from.\n",
+        "NUM_ACTIONS = (\n",
+        "    20  # @param {type:\"integer\"} Number of actions (movie items) to choose from.\n",
+        ")\n",
         "PER_ARM = False  # Use the non-per-arm version of the MovieLens environment.\n",
         "\n",
         "# Set agent parameters.\n",
@@ -621,7 +626,8 @@
       "source": [
         "# Define RL environment.\n",
         "env = movielens_py_environment.MovieLensPyEnvironment(\n",
-        "    DATA_PATH, RANK_K, BATCH_SIZE, num_movies=NUM_ACTIONS, csv_delimiter=\"\\t\")\n",
+        "    DATA_PATH, RANK_K, BATCH_SIZE, num_movies=NUM_ACTIONS, csv_delimiter=\"\\t\"\n",
+        ")\n",
         "environment = tf_py_environment.TFPyEnvironment(env)\n",
         "\n",
         "# Define RL agent/algorithm.\n",
@@ -631,7 +637,8 @@
         "    tikhonov_weight=TIKHONOV_WEIGHT,\n",
         "    alpha=AGENT_ALPHA,\n",
         "    dtype=tf.float32,\n",
-        "    accepts_per_arm_features=PER_ARM)\n",
+        "    accepts_per_arm_features=PER_ARM,\n",
+        ")\n",
         "print(\"TimeStep Spec (for each batch):\\n\", agent.time_step_spec, \"\\n\")\n",
         "print(\"Action Spec (for each batch):\\n\", agent.action_spec, \"\\n\")\n",
         "print(\"Reward Spec (for each batch):\\n\", environment.reward_spec(), \"\\n\")\n",
@@ -639,7 +646,8 @@
         "# Define RL metric.\n",
         "optimal_reward_fn = functools.partial(\n",
         "    environment_utilities.compute_optimal_reward_with_movielens_environment,\n",
-        "    environment=environment)\n",
+        "    environment=environment,\n",
+        ")\n",
         "regret_metric = tf_bandit_metrics.RegretMetric(optimal_reward_fn)\n",
         "metrics = [regret_metric]"
       ]
@@ -704,35 +712,38 @@
         "    if training_data_spec_transformation_fn is None:\n",
         "        data_spec = agent.policy.trajectory_spec\n",
         "    else:\n",
-        "        data_spec = training_data_spec_transformation_fn(\n",
-        "            agent.policy.trajectory_spec)\n",
-        "    replay_buffer = trainer.get_replay_buffer(data_spec, environment.batch_size,\n",
-        "                                              steps_per_loop)\n",
+        "        data_spec = training_data_spec_transformation_fn(agent.policy.trajectory_spec)\n",
+        "    replay_buffer = trainer.get_replay_buffer(\n",
+        "        data_spec, environment.batch_size, steps_per_loop\n",
+        "    )\n",
         "\n",
         "    # `step_metric` records the number of individual rounds of bandit interaction;\n",
         "    # that is, (number of trajectories) * batch_size.\n",
         "    step_metric = tf_metrics.EnvironmentSteps()\n",
         "    metrics = [\n",
         "        tf_metrics.NumberOfEpisodes(),\n",
-        "        tf_metrics.AverageEpisodeLengthMetric(batch_size=environment.batch_size)\n",
+        "        tf_metrics.AverageEpisodeLengthMetric(batch_size=environment.batch_size),\n",
         "    ]\n",
         "    if additional_metrics:\n",
         "        metrics += additional_metrics\n",
         "\n",
         "    if isinstance(environment.reward_spec(), dict):\n",
-        "        metrics += [tf_metrics.AverageReturnMultiMetric(\n",
-        "            reward_spec=environment.reward_spec(),\n",
-        "            batch_size=environment.batch_size)]\n",
-        "    else:\n",
         "        metrics += [\n",
-        "            tf_metrics.AverageReturnMetric(batch_size=environment.batch_size)]\n",
+        "            tf_metrics.AverageReturnMultiMetric(\n",
+        "                reward_spec=environment.reward_spec(), batch_size=environment.batch_size\n",
+        "            )\n",
+        "        ]\n",
+        "    else:\n",
+        "        metrics += [tf_metrics.AverageReturnMetric(batch_size=environment.batch_size)]\n",
         "\n",
         "    # Store intermediate metric results, indexed by metric names.\n",
         "    metric_results = defaultdict(list)\n",
         "\n",
         "    if training_data_spec_transformation_fn is not None:\n",
-        "        def add_batch_fn(data): return replay_buffer.add_batch(training_data_spec_transformation_fn(data)) \n",
-        "        \n",
+        "\n",
+        "        def add_batch_fn(data):\n",
+        "            return replay_buffer.add_batch(training_data_spec_transformation_fn(data))\n",
+        "\n",
         "    else:\n",
         "        add_batch_fn = replay_buffer.add_batch\n",
         "\n",
@@ -742,10 +753,12 @@
         "        env=environment,\n",
         "        policy=agent.collect_policy,\n",
         "        num_steps=steps_per_loop * environment.batch_size,\n",
-        "        observers=observers)\n",
+        "        observers=observers,\n",
+        "    )\n",
         "\n",
         "    training_loop = trainer.get_training_loop_fn(\n",
-        "        driver, replay_buffer, agent, steps_per_loop)\n",
+        "        driver, replay_buffer, agent, steps_per_loop\n",
+        "    )\n",
         "    saver = policy_saver.PolicySaver(agent.policy)\n",
         "\n",
         "    for _ in range(training_loops):\n",
@@ -783,7 +796,8 @@
         "    environment=environment,\n",
         "    training_loops=TRAINING_LOOPS,\n",
         "    steps_per_loop=STEPS_PER_LOOP,\n",
-        "    additional_metrics=metrics)\n",
+        "    additional_metrics=metrics,\n",
+        ")\n",
         "\n",
         "tf.profiler.experimental.stop()"
       ]
@@ -1092,11 +1106,15 @@
       },
       "outputs": [],
       "source": [
-        "RUN_HYPERPARAMETER_TUNING = True  # Execute hyperparameter tuning instead of regular training.\n",
+        "RUN_HYPERPARAMETER_TUNING = (\n",
+        "    True  # Execute hyperparameter tuning instead of regular training.\n",
+        ")\n",
         "TRAIN_WITH_BEST_HYPERPARAMETERS = False  # Do not train.\n",
         "\n",
         "HPTUNING_RESULT_DIR = \"hptuning/\"  # @param {type: \"string\"} Directory to store the best hyperparameter(s) in `BUCKET_NAME` and locally (temporarily).\n",
-        "HPTUNING_RESULT_PATH = os.path.join(HPTUNING_RESULT_DIR, \"result.json\")  # @param {type: \"string\"} Path to the file containing the best hyperparameter(s)."
+        "HPTUNING_RESULT_PATH = os.path.join(\n",
+        "    HPTUNING_RESULT_DIR, \"result.json\"\n",
+        ")  # @param {type: \"string\"} Path to the file containing the best hyperparameter(s)."
       ]
     },
     {
@@ -1124,7 +1142,7 @@
         "    image_uri: str,\n",
         "    args: List[str],\n",
         "    location: str = \"us-central1\",\n",
-        "    api_endpoint: str = \"us-central1-aiplatform.googleapis.com\"\n",
+        "    api_endpoint: str = \"us-central1-aiplatform.googleapis.com\",\n",
         ") -> None:\n",
         "    \"\"\"Creates a hyperparameter tuning job using a custom container.\n",
         "\n",
@@ -1197,8 +1215,8 @@
         "\n",
         "    # Create job\n",
         "    response = client.create_hyperparameter_tuning_job(\n",
-        "        parent=parent,\n",
-        "        hyperparameter_tuning_job=hyperparameter_tuning_job)\n",
+        "        parent=parent, hyperparameter_tuning_job=hyperparameter_tuning_job\n",
+        "    )\n",
         "    job_id = response.name.split(\"/\")[-1]\n",
         "    print(\"Job ID:\", job_id)\n",
         "    print(\"Job config:\", response)\n",
@@ -1242,7 +1260,8 @@
         "    image_uri=f\"gcr.io/{PROJECT_ID}/{HPTUNING_TRAINING_CONTAINER}:latest\",\n",
         "    args=args,\n",
         "    location=REGION,\n",
-        "    api_endpoint=f\"{REGION}-aiplatform.googleapis.com\")"
+        "    api_endpoint=f\"{REGION}-aiplatform.googleapis.com\",\n",
+        ")"
       ]
     },
     {
@@ -1292,7 +1311,8 @@
         "    name = client.hyperparameter_tuning_job_path(\n",
         "        project=project,\n",
         "        location=location,\n",
-        "        hyperparameter_tuning_job=hyperparameter_tuning_job_id)\n",
+        "        hyperparameter_tuning_job=hyperparameter_tuning_job_id,\n",
+        "    )\n",
         "    response = client.get_hyperparameter_tuning_job(name=name)\n",
         "    return response"
       ]
@@ -1313,7 +1333,8 @@
         "        location=REGION,\n",
         "        api_endpoint=f\"{REGION}-aiplatform.googleapis.com\")\n",
         "    if response.state.name == 'JOB_STATE_SUCCEEDED':\n",
-        "        print(\"Job succeeded.\\nJob Time:\", response.update_time - response.create_time)\n",
+        "        print(\"Job succeeded.\n",
+        "Job Time:\", response.update_time - response.create_time)\n",
         "        trials = response.trials\n",
         "        print(\"Trials:\", trials)\n",
         "        break\n",
@@ -1348,8 +1369,8 @@
         "if trials:\n",
         "    # Dict mapping from metric names to the best metric values seen so far\n",
         "    best_objective_values = dict.fromkeys(\n",
-        "        [metric.metric_id for metric in trials[0].final_measurement.metrics],\n",
-        "        -np.inf)\n",
+        "        [metric.metric_id for metric in trials[0].final_measurement.metrics], -np.inf\n",
+        "    )\n",
         "    # Dict mapping from metric names to a list of the best combination(s) of\n",
         "    # hyperparameter(s). Each combination is a dict mapping from hyperparameter\n",
         "    # names to their values.\n",
@@ -1358,12 +1379,13 @@
         "        # `final_measurement` and `parameters` are `RepeatedComposite` objects.\n",
         "        # Reference the structure above to extract the value of your interest.\n",
         "        for metric in trial.final_measurement.metrics:\n",
-        "            params = {\n",
-        "                param.parameter_id: param.value for param in trial.parameters}\n",
+        "            params = {param.parameter_id: param.value for param in trial.parameters}\n",
         "            if metric.value > best_objective_values[metric.metric_id]:\n",
         "                best_params[metric.metric_id] = [params]\n",
         "            elif metric.value == best_objective_values[metric.metric_id]:\n",
-        "                best_params[param.parameter_id].append(params)  # Handle cases where multiple hyperparameter values lead to the same performance.\n",
+        "                best_params[param.parameter_id].append(\n",
+        "                    params\n",
+        "                )  # Handle cases where multiple hyperparameter values lead to the same performance.\n",
         "    print(\"Best hyperparameter value(s):\")\n",
         "    for metric, params in best_params.items():\n",
         "        print(f\"Metric={metric}: {sorted(params)}\")\n",
@@ -1443,7 +1465,9 @@
       },
       "outputs": [],
       "source": [
-        "PREDICTION_CONTAINER = \"prediction-custom-container\"  # @param {type:\"string\"} Name of the container image."
+        "PREDICTION_CONTAINER = (\n",
+        "    \"prediction-custom-container\"  # @param {type:\"string\"} Name of the container image.\n",
+        ")"
       ]
     },
     {
@@ -1475,7 +1499,7 @@
         "  machineType: 'E2_HIGHCPU_8'\"\"\".format(\n",
         "    PROJECT_ID=PROJECT_ID,\n",
         "    PREDICTION_CONTAINER=PREDICTION_CONTAINER,\n",
-        "    ARTIFACTS_DIR=ARTIFACTS_DIR\n",
+        "    ARTIFACTS_DIR=ARTIFACTS_DIR,\n",
         ")\n",
         "\n",
         "with open(\"cloudbuild.yaml\", \"w\") as fp:\n",
@@ -1592,8 +1616,12 @@
       },
       "outputs": [],
       "source": [
-        "RUN_HYPERPARAMETER_TUNING = False  # Execute regular training instead of hyperparameter tuning.\n",
-        "TRAIN_WITH_BEST_HYPERPARAMETERS = True  # @param {type:\"bool\"} Whether to use learned hyperparameters in training."
+        "RUN_HYPERPARAMETER_TUNING = (\n",
+        "    False  # Execute regular training instead of hyperparameter tuning.\n",
+        ")\n",
+        "TRAIN_WITH_BEST_HYPERPARAMETERS = (\n",
+        "    True  # @param {type:\"bool\"} Whether to use learned hyperparameters in training.\n",
+        ")"
       ]
     },
     {
@@ -1633,10 +1661,12 @@
         "job = aiplatform.CustomContainerTrainingJob(\n",
         "    display_name=\"train-movielens\",\n",
         "    container_uri=f\"gcr.io/{PROJECT_ID}/{HPTUNING_TRAINING_CONTAINER}:latest\",\n",
-        "    command=[\"python3\", \"-m\", \"src.training.task\"] + args,  # Pass in training arguments, including hyperparameters.\n",
+        "    command=[\"python3\", \"-m\", \"src.training.task\"]\n",
+        "    + args,  # Pass in training arguments, including hyperparameters.\n",
         "    model_serving_container_image_uri=f\"gcr.io/{PROJECT_ID}/{PREDICTION_CONTAINER}:latest\",\n",
         "    model_serving_container_predict_route=\"/predict\",\n",
-        "    model_serving_container_health_route=\"/health\")\n",
+        "    model_serving_container_health_route=\"/health\",\n",
+        ")\n",
         "\n",
         "print(\"Training Spec:\", job._managed_model)\n",
         "\n",
@@ -1645,7 +1675,8 @@
         "    replica_count=1,\n",
         "    machine_type=\"n1-standard-4\",\n",
         "    accelerator_type=\"ACCELERATOR_TYPE_UNSPECIFIED\",\n",
-        "    accelerator_count=0)"
+        "    accelerator_count=0,\n",
+        ")"
       ]
     },
     {
@@ -1784,7 +1815,7 @@
         "! gcloud ai models delete $model.name --quiet\n",
         "\n",
         "# Delete Cloud Storage objects that were created\n",
-        "! gsutil -m rm -r $ARTIFACTS_DIR"
+        "! gcloud storage rm --recursive $ARTIFACTS_DIR"
       ]
     }
   ],

community-content/tf_keras_image_classification_distributed_multi_worker_with_vertex_sdk/multi_worker_vertex_training_on_cpu_with_custom_container.ipynb

@@ -324,7 +324,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil ls $gcs_output_uri_prefix"
+        "! gcloud storage ls $gcs_output_uri_prefix"
       ]
     },
     {
@@ -344,7 +344,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil rm -rf $gcs_output_uri_prefix"
+        "! gcloud storage rm --recursive --continue-on-error $gcs_output_uri_prefix"
       ]
     }
   ],

community-content/tf_keras_image_classification_distributed_multi_worker_with_vertex_sdk/multi_worker_vertex_training_on_gpu_with_custom_container.ipynb

@@ -328,7 +328,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil ls $gcs_output_uri_prefix"
+        "! gcloud storage ls $gcs_output_uri_prefix"
       ]
     },
     {
@@ -348,7 +348,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil rm -rf $gcs_output_uri_prefix"
+        "! gcloud storage rm --recursive --continue-on-error $gcs_output_uri_prefix"
       ]
     }
   ],

community-content/tf_keras_image_classification_distributed_multi_worker_with_vertex_sdk/trainer/requirements.txt

@@ -1 +1 @@
-tensorflow==2.7.2
+tensorflow==2.12.1

community-content/tf_keras_text_classification_distributed_single_worker_gpus_with_gcloud_local_run_and_vertex_sdk/trainer/requirements.txt

@@ -1 +1 @@
-tensorflow==2.7.2
+tensorflow==2.12.1

community-content/tf_keras_text_classification_distributed_single_worker_gpus_with_gcloud_local_run_and_vertex_sdk/vertex_training_with_local_mode_container.ipynb

@@ -341,7 +341,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil ls $gcs_output_uri_prefix"
+        "! gcloud storage ls $gcs_output_uri_prefix"
       ]
     },
     {
@@ -361,7 +361,7 @@
       },
       "outputs": [],
       "source": [
-        "! gsutil rm -rf $gcs_output_uri_prefix"
+        "! gcloud storage rm --recursive --continue-on-error $gcs_output_uri_prefix"
       ]
     }
   ],

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/README.md

@@ -0,0 +1,126 @@
+# Vertex AI Training: Llama 3.1 8B pre-training using Nvidia A3 Mega VMs (H100)
+This document provides a step-by-step guide for pre-training a Llama 3.1 8B model on the `en-wiki` dataset using multiple [Vertex AI Custom Training](https://cloud.google.com/vertex-ai/docs/training/overview) `a3-megagpu-8g` nodes.
+
+We will use a custom container based on NVIDIA's [NeMo Framework](https://docs.nvidia.com/nemo-framework/user-guide/24.07/overview.html) to demonstrate a scalable, multi-node training workflow. All required artifacts and commands are included.
+
+## 1. Prerequisites
+
+### 1.1. Google Cloud Project setup
+- **Enable APIs:** Ensure the Vertex AI API is [enabled for your project](http://console.cloud.google.com/flows/enableapi?apiid=aiplatform.googleapis.com).
+- **H100 Mega Quota:** A3 Mega VMs are powered by H100 GPUs. Request quota for `custom_model_training_nvidia_h100_mega_gpus` in one of the [supported regions](https://cloud.google.com/vertex-ai/docs/general/locations#accelerator_support). If using Spot VMs, request `custom_model_training_preemptible_nvidia_h100_mega_gpus` quota instead.
+- **Reservations (Optional but recommended):** For guaranteed capacity, [create a reservation](https://cloud.google.com/compute/docs/instances/reservations-shared) and ensure the reservation is shared with the Vertex AI service account. This guide requires a minimum of **16 H100 GPUs** (2 full A3 Mega nodes).
+
+### 1.2. GCS bucket
+Create a [Cloud Storage bucket](https://cloud.google.com/storage/docs/creating-buckets) in the same region where you have quota. If you're using Hierarchical Namespace for your bucket, you may need to update permissions of the Vertex AI Custom Code Service Agent .
+
+This bucket is used for:
+- Staging the training application.
+- Storing model checkpoints and logs.
+- Storing data if you use your own data.
+
+
+## 2. Setup & configuration
+
+### 2.1. Clone the repo
+First clone the repo into your development environment.
+
+```bash
+git clone https://github.com/GoogleCloudPlatform/vertex-ai-samples.git
+```
+
+Navigate to the root folder for this sample.
+
+### 2.2. Environment Setup
+First, configure your local environment. These variables are used in subsequent commands.
+
+```bash
+# Required: Update with your values
+export PROJECT_ID="<your-project-id>"
+export REPOSITORY="<your-artifact-registry-repo-name>" # e.g., "my-containers"
+export BUCKET="<your-gcs-bucket-name>"
+
+# Optional: Change if needed
+export REGION="us-central1"
+
+# --- Do not change the lines below ---
+export ARTIFACT_REGISTRY="${REGION}-docker.pkg.dev/${PROJECT_ID}/${REPOSITORY}"
+export REPO_ROOT=$(git rev-parse --show-toplevel)
+```
+
+## 3. Build and push a docker container image to Artifact Registry
+Normally, you can use any custom training container on Vertex AI Training. In this example you build a NeMo Docker image that is based on the [Nvidia’s NeMo 24.09](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/nemo/tags) image. Use Cloud Build to build and push the container image.
+
+This document picked NeMo as the demonstrating container since it’s a widely adopted GPU LLM training framework providing high performance and versatile training functionalities.
+
+In addition to the base image, some customizations are included to form the final prebuilt image:
+- Some dependencies are installed to integrate with Vertex AI Training.
+- An entrypoint script that sets up required environments and calls the training job.
+- Some patches are applied to the NeMo code to let it  load the dataset from a GCS bucket.
+
+Run this command to build the container and push the container into the Google Artifact Registry.
+
+```bash
+cd "${REPO_ROOT}/community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining"
+export IMAGE_NAME="vertex-nemo-llama"
+gcloud builds submit . \
+    --project="${PROJECT_ID}" \
+    --region="${REGION}" \
+    --config=docker/cloudbuild.yml \
+    --substitutions="_ARTIFACT_REGISTRY=${ARTIFACT_REGISTRY},_IMAGE_NAME=${IMAGE_NAME}" \
+    --timeout="2h" \
+    --machine-type="e2-highcpu-32"
+```
+
+## 4. Launch the Training Job
+
+
+###  4.1. Job Configuration File
+Once the container is built, update the job_config.json to set up the training job.
+File: job_config.json
+```json
+{
+  "project_id": "<project-id>",
+  "region": "<region>",
+  "zone": "<zone if using reservation>",
+  "bucket": "<bucket>",
+  "dataset_bucket": "github-repo/data/third-party/enwiki-latest-pages-articles",
+  "image_uri": "<docker image uri from artifact registry>",
+  "strategy": "spot",
+  "nodes": "2",
+  "machine_type": "a3-megagpu-8g",
+  "gpu_type": "NVIDIA_H100_MEGA_80GB",
+  "gpus_per_node": "8",
+  "recipe_name": "llama3_1_8b_pretrain_a3mega",
+  "job_prefix": "vertex-spot-",
+  "reservation_name": ""
+}
+```
+
+### 4.2 Launch the Training Job
+
+First, create a Python virtual environment using your tool of choice, then install
+the requirements specified in `requirements.txt`. Using `pip`, the command would be:
+```bash
+pip install -r requirements.txt
+```
+
+Now launch the Vertex AI training job using the provided Python script.
+
+```bash
+python3 scripts/launch.py --config_file=job_config.json
+```
+
+This script reads job_config.json, defines the cluster specification (2 nodes, 8 GPUs each), and submits the custom training job to Vertex AI.
+
+## 5. Monitor and Clean Up
+
+### 5.1. Monitoring
+Vertex AI Console: Track the job's status in the Google Cloud Console under Vertex AI > Training > Custom Jobs.
+Logs: View detailed logs in Cloud Logging by filtering for your job name.
+Checkpoints: Model checkpoints are saved to your GCS bucket at the path specified in your training script's configuration.
+
+### 5.2. Cleaning Up
+To avoid ongoing charges, delete the resources you created:
+- The Artifact Registry image.
+- The contents of the GCS bucket (checkpoints, logs).
+- The Vertex AI Custom Job will eventually complete or fail, incurring no further cost.

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/configs/llama3_1_8b_pretrain_a3mega.yaml

@@ -0,0 +1,265 @@
+# Reference:
+# https://github.com/NVIDIA/NeMo-Framework-Launcher/blob/24.07/launcher_scripts/conf/training/llama/llama3_1_8b.yaml
+name: llama3_1_8b_pretrain_a3mega
+restore_from_path: null  # used when starting from a .nemo file
+
+trainer:
+  devices: 8
+  num_nodes: 1
+  accelerator: gpu
+  precision: bf16
+  logger: false  # logger provided by exp_manager
+  enable_checkpointing: false
+  use_distributed_sampler: false
+  max_epochs: -1  # PTL default. In practice, max_steps will be reached first.
+  max_steps: 30  # consumed_samples = global_step * micro_batch_size * data_parallel_size * accumulate_grad_batches
+  log_every_n_steps: 1
+  val_check_interval: null
+  limit_val_batches: 1
+  limit_test_batches: 1
+  accumulate_grad_batches: 1  # do not modify, grad acc is automatic for training megatron models
+  gradient_clip_val: 1.0
+  benchmark: false
+  enable_model_summary: false  # default PTL callback for this does not support model parallelism, instead we log manually
+
+exp_manager:
+  explicit_log_dir: null
+  exp_dir: /data
+  name: ${name}
+  create_dllogger_logger: true
+  dllogger_logger_kwargs:
+    verbose: true
+    stdout: true
+    json_file: "/data/dllogger.json"
+  create_wandb_logger: false
+  wandb_logger_kwargs:
+    project: null
+    name: null
+  resume_if_exists: true
+  resume_ignore_no_checkpoint: true
+  create_checkpoint_callback: false
+  checkpoint_callback_params:
+    monitor: val_loss
+    save_top_k: 3
+    mode: min
+    always_save_nemo: false  # saves nemo file during validation, not implemented for model parallel
+    save_nemo_on_train_end: false  # not recommended when training large models on clusters with short time limits
+    filename: 'megatron_gpt--{val_loss:.2f}-{step}-{consumed_samples}'
+    model_parallel_size: ${multiply:${model.tensor_model_parallel_size}, ${model.pipeline_model_parallel_size}}
+  seconds_to_sleep: 5  # Allows node_rank!=0 to sleep and let node0 to init, like preparing data
+
+model:
+  mcore_gpt: true
+  # specify micro_batch_size, global_batch_size, and model parallelism
+  # gradient accumulation will be done automatically based on data_parallel_size
+  micro_batch_size: 1  # limited by GPU memory
+  global_batch_size: 1024  # will use more micro batches to reach global batch size
+  tensor_model_parallel_size: 1  # intra-layer model parallelism
+  pipeline_model_parallel_size: 2  # inter-layer model parallelism
+  context_parallel_size: 1
+  virtual_pipeline_model_parallel_size: null  # interleaved pipeline
+  ## Sequence Parallelism
+  # Makes tensor parallelism more memory efficient for LLMs (20B+) by parallelizing layer norms and dropout sequentially
+  # See Reducing Activation Recomputation in Large Transformer Models: https://arxiv.org/abs/2205.05198 for more details.
+  sequence_parallel: false
+
+  fsdp: false
+  fsdp_cpu_offload: true
+  fsdp_sharding_strategy: "full"  # Method to shard model states. Available options are 'full', 'hybrid', and 'grad'.
+  fsdp_grad_reduce_dtype: "16"  # Gradient reduction data type.
+  fsdp_sharded_checkpoint: false  # Store and load FSDP shared checkpoint.
+  fsdp_use_orig_params: false  # Set to True to use FSDP for specific peft scheme.
+
+  # Distributed checkpoint setup
+  dist_ckpt_format: "torch_dist"  # Set to 'torch_dist' to use PyTorch distributed checkpoint format.
+  dist_ckpt_load_on_device: true  # whether to load checkpoint weights directly on GPU or to CPU
+  dist_ckpt_parallel_save: true  # if true, each worker will write its own part of the dist checkpoint
+  dist_ckpt_parallel_save_within_dp: false  # if true, save will be parallelized only within a DP group (whole world otherwise), which might slightly reduce the save overhead
+  dist_ckpt_parallel_load: false  # if true, each worker will load part of the dist checkpoint and exchange with NCCL. Might use some extra GPU memory
+  dist_ckpt_torch_dist_multiproc: 2  # number of extra processes per rank used during ckpt save with PyTorch distributed format
+  dist_ckpt_assume_constant_structure: false  # set to True only if the state dict structure doesn't change within a single job. Allows caching some computation across checkpoint saves.
+  dist_ckpt_parallel_dist_opt: true  # parallel save/load of a DistributedOptimizer. 'True' allows performant save and reshardable checkpoints. Set to 'False' only in order to minimize the number of checkpoint files.
+  dist_ckpt_load_strictness: null  # defines checkpoint keys mismatch behavior (only during dist-ckpt load). Choices: assume_ok_unexpected (default - try loading without any check), log_all (log mismatches), raise_all (raise mismatches)
+
+  # model architecture
+  encoder_seq_length: 8192
+  max_position_embeddings: ${.encoder_seq_length}
+  num_layers: 32  # 8b: 32 | 70b: 80 | 405b: 126
+  hidden_size: 4096  # 8b: 4096 | 70b: 8192 | 405b: 16384
+  ffn_hidden_size: 14336  # 8b: 14336 | 70b: 28672 | 405b: 53248
+  num_attention_heads: 32  # 8b: 32 | 70b: 64 | 405b: 128
+  num_query_groups: 8  # Number of query groups for group query attention. If None, normal attention is used. 8b: 8 | 70b: 8 | 405b: 16
+  init_method_std: 0.01  # Standard deviation of the zero mean normal distribution used for weight initialization.  8b: 0.01 | 70b: 0.008944 | 405b: 0.02
+  use_scaled_init_method: true  # use scaled residuals initialization
+  hidden_dropout: 0.0  # Dropout probability for hidden state transformer.
+  attention_dropout: 0.0  # Dropout probability for attention
+  ffn_dropout: 0.0  # Dropout probability in the feed-forward layer.
+  kv_channels: null  # Projection weights dimension in multi-head attention. Set to hidden_size // num_attention_heads if null
+  apply_query_key_layer_scaling: true  # scale Q * K^T by 1 / layer-number.
+  normalization: 'rmsnorm'  # Normalization layer to use. Options are 'layernorm', 'rmsnorm'
+  layernorm_epsilon: 1e-5
+  do_layer_norm_weight_decay: false  # True means weight decay on all params
+  make_vocab_size_divisible_by: 128  # Pad the vocab size to be divisible by this value for computation efficiency.
+  pre_process: true  # add embedding
+  post_process: true  # add pooler
+  persist_layer_norm: true  # Use of persistent fused layer norm kernel.
+  bias: false  # Whether to use bias terms in all weight matrices.
+  activation: 'fast-swiglu'  # Options ['gelu', 'geglu', 'swiglu', 'reglu', 'squared-relu', 'fast-geglu', 'fast-swiglu', 'fast-reglu']
+  headscale: false  # Whether to learn extra parameters that scale the output of the each self-attention head.
+  transformer_block_type: 'pre_ln'  # Options ['pre_ln', 'post_ln', 'normformer']
+  openai_gelu: false  # Use OpenAI's GELU instead of the default GeLU
+  normalize_attention_scores: true  # Whether to scale the output Q * K^T by 1 / sqrt(hidden_size_per_head). This arg is provided as a configuration option mostly for compatibility with models that have been weight-converted from HF. You almost always want to se this to True.
+  position_embedding_type: 'rope'  # Position embedding type. Options ['learned_absolute', 'rope']
+  rotary_percentage: 1.0  # If using position_embedding_type=rope, then the per head dim is multiplied by this.
+  attention_type: 'multihead'  # Attention type. Options ['multihead']
+  share_embeddings_and_output_weights: false  # Share embedding and output layer weights.
+  scale_positional_embedding: true  # This is false for llama3 models. Only used for >= llama3.1.
+
+  # Use GPT2BPETokenizer for test, because the testing dataset is tokenized by this tokenizer.
+  # https://docs.nvidia.com/nemo-framework/user-guide/24.07/playbooks/singlenodepretrain.html#data-download-and-pre-processing
+  tokenizer:
+    library: megatron
+    type: GPT2BPETokenizer
+    model: null  # /path/to/tokenizer.model
+    vocab_file: null
+    merge_file: null
+    delimiter: null  # only used for tabular tokenizer
+    sentencepiece_legacy: false  # Legacy=True allows you to add special tokens to sentencepiece tokenizers.
+
+  # Mixed precision
+  native_amp_init_scale: 4294967296  # 2 ** 32
+  native_amp_growth_interval: 1000
+  hysteresis: 2  # Gradient scale hysteresis
+  fp32_residual_connection: false  # Move residual connections to fp32
+  fp16_lm_cross_entropy: false  # Move the cross entropy unreduced loss calculation for lm head to fp16
+
+  # Megatron O2-style half-precision
+  megatron_amp_O2: true  # Enable O2-level automatic mixed precision using main parameters
+  grad_allreduce_chunk_size_mb: 125
+
+  # Fusion
+  grad_div_ar_fusion: true  # Fuse grad division into torch.distributed.all_reduce. Only used with O2 and no pipeline parallelism..
+  gradient_accumulation_fusion: true  # Fuse weight gradient accumulation to GEMMs. Only used with pipeline parallelism and O2.
+  bias_activation_fusion: true  # Use a kernel that fuses the bias addition from weight matrices with the subsequent activation function.
+  bias_dropout_add_fusion: true  # Use a kernel that fuses the bias addition, dropout and residual connection addition.
+  masked_softmax_fusion: true  # Use a kernel that fuses the attention softmax with it's mask.
+  apply_rope_fusion: true  # Use a kernel to add rotary positional embeddings. Only used if position_embedding_type=rope
+  cross_entropy_loss_fusion: true
+
+  # Miscellaneous
+  seed: 1234
+  resume_from_checkpoint: null  # manually set the checkpoint file to load from
+  use_cpu_initialization: false  # Init weights on the CPU (slow for large models)
+  onnx_safe: false  # Use work-arounds for known problems with Torch ONNX exporter.
+  apex_transformer_log_level: 30  # Python logging level displays logs with severity greater than or equal to this
+  gradient_as_bucket_view: true  # PyTorch DDP argument. Allocate gradients in a contiguous bucket to save memory (less fragmentation and buffer memory)
+  sync_batch_comm: false  # Enable stream synchronization after each p2p communication between pipeline stages
+
+  ## Activation Checkpointing
+  # NeMo Megatron supports 'selective' activation checkpointing where only the memory intensive part of attention is checkpointed.
+  # These memory intensive activations are also less compute intensive which makes activation checkpointing more efficient for LLMs (20B+).
+  # See Reducing Activation Recomputation in Large Transformer Models: https://arxiv.org/abs/2205.05198 for more details.
+  # 'full' will checkpoint the entire transformer layer.
+  activations_checkpoint_granularity: null  # 'selective' or 'full'
+  activations_checkpoint_method: null  # 'uniform', 'block'
+  # 'uniform' divides the total number of transformer layers and checkpoints the input activation
+  # of each chunk at the specified granularity. When used with 'selective', 'uniform' checkpoints all attention blocks in the model.
+  # 'block' checkpoints the specified number of layers per pipeline stage at the specified granularity
+  activations_checkpoint_num_layers: null
+  # when using 'uniform' this creates groups of transformer layers to checkpoint. Usually set to 1. Increase to save more memory.
+  # when using 'block' this this will checkpoint the first activations_checkpoint_num_layers per pipeline stage.
+  num_micro_batches_with_partial_activation_checkpoints: null
+  # This feature is valid only when used with pipeline-model-parallelism.
+  # When an integer value is provided, it sets the number of micro-batches where only a partial number of Transformer layers get checkpointed
+  # and recomputed within a window of micro-batches. The rest of micro-batches in the window checkpoint all Transformer layers. The size of window is
+  # set by the maximum outstanding micro-batch backpropagations, which varies at different pipeline stages. The number of partial layers to checkpoint
+  # per micro-batch is set by 'activations_checkpoint_num_layers' with 'activations_checkpoint_method' of 'block'.
+  # This feature enables using activation checkpoint at a fraction of micro-batches up to the point of full GPU memory usage.
+  activations_checkpoint_layers_per_pipeline: null
+  # This feature is valid only when used with pipeline-model-parallelism.
+  # When an integer value (rounded down when float is given) is provided, it sets the number of Transformer layers to skip checkpointing at later
+  # pipeline stages. For example, 'activations_checkpoint_layers_per_pipeline' of 3 makes pipeline stage 1 to checkpoint 3 layers less than
+  # stage 0 and stage 2 to checkpoint 6 layers less stage 0, and so on. This is possible because later pipeline stage
+  # uses less GPU memory with fewer outstanding micro-batch backpropagations. Used with 'num_micro_batches_with_partial_activation_checkpoints',
+  # this feature removes most of activation checkpoints at the last pipeline stage, which is the critical execution path.
+
+  ## Transformer Engine
+  transformer_engine: true
+  fp8: false  # enables fp8 in TransformerLayer forward
+  fp8_e4m3: false  # sets fp8_format = recipe.Format.E4M3
+  fp8_hybrid: false  # sets fp8_format = recipe.Format.HYBRID
+  fp8_margin: 0  # scaling margin
+  fp8_interval: 1  # scaling update interval
+  fp8_amax_history_len: 1024  # Number of steps for which amax history is recorded per tensor
+  fp8_amax_compute_algo: 'max'  # 'most_recent' or 'max'. Algorithm for computing amax from history
+  ub_tp_comm_overlap: false  # do not turn on because of b/397797926
+  use_flash_attention: true
+  gc_interval: 100
+
+  ## Offloading Activations/Weights to CPU
+  cpu_offloading: false
+  cpu_offloading_num_layers: ${sum:${.num_layers},-1}  # This value should be between [1,num_layers-1] as we don't want to offload the final layer's activations and expose any offloading duration for the final layer
+  cpu_offloading_activations: true
+  cpu_offloading_weights: true
+
+  data:
+    # Path to data must be specified by the user.
+    # Supports List, String and Dictionary
+    # List : can override from the CLI: "model.data.data_prefix=[.5,/raid/data/pile/my-gpt3_00_text_document,.5,/raid/data/pile/my-gpt3_01_text_document]",
+    # Or see example below:
+    # data_prefix:
+    # - .5
+    # - /raid/data/pile/my-gpt3_00_text_document
+    # - .5
+    # - /raid/data/pile/my-gpt3_01_text_document
+    # Dictionary: can override from CLI "model.data.data_prefix"={"train":[1.0, /path/to/data], "validation":/path/to/data, "test":/path/to/test}
+    # Or see example below:
+    # "model.data.data_prefix: {train:[1.0,/path/to/data], validation:[/path/to/data], test:[/path/to/test]}"
+    data_prefix: [1.0, /data/hfbpe_gpt_training_data_text_document]
+    index_mapping_dir: null  # path to save index mapping .npy files, by default will save in the same location as data_prefix
+    data_impl: mmap
+    splits_string: 900,50,50
+    seq_length: ${model.encoder_seq_length}
+    skip_warmup: true
+    num_workers: 2
+    dataloader_type: single  # cyclic
+    reset_position_ids: false  # Reset position ids after end-of-document token
+    reset_attention_mask: false  # Reset attention mask after end-of-document token
+    eod_mask_loss: false  # Mask loss for the end of document tokens
+    validation_drop_last: true  # Set to false if the last partial validation samples is to be consumed
+    no_seqlen_plus_one_input_tokens: false  # Set to True to disable fetching (sequence length + 1) input tokens, instead get (sequence length) input tokens and mask the last token
+    pad_samples_to_global_batch_size: false  # Set to True if you want to pad the last partial batch with -1's to equal global batch size
+    shuffle_documents: true  # Set to False to disable documents shuffling. Sample index will still be shuffled
+
+  # Nsys profiling options
+  nsys_profile:
+    enabled: false
+    start_step: 0  # Global batch to start profiling
+    end_step: 1  # Global batch to end profiling
+    ranks: [0]  # Global rank IDs to profile
+    gen_shape: false  # Generate model and kernel details including input shapes
+
+  memory_profile:
+    enabled: false
+    start_step: 0
+    end_step: 1
+    ranks: [0]
+    output_path: /data  # Must be a dir
+
+  optim:
+    name: distributed_fused_adam  # E.g., fused_adam or set _target_: torch.optim.AdamW field
+    lr: 2e-5
+    weight_decay: 0.01
+    betas:
+    - 0.9
+    - 0.98
+    bucket_cap_mb: 125
+    overlap_grad_sync: true
+    overlap_param_sync: true
+    contiguous_grad_buffer: true
+    contiguous_param_buffer: true
+    sched:
+      name: CosineAnnealing
+      warmup_steps: 400
+      constant_steps: 0
+      min_lr: 2e-6

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/docker/cloudbuild.yml

@@ -0,0 +1,26 @@
+# 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.
+
+steps:
+- name: 'gcr.io/cloud-builders/docker'
+  args:
+  - 'build'
+  - '--tag=${_ARTIFACT_REGISTRY}/${_IMAGE_NAME}'
+  - '--file=docker/vertex-dist-recipes.Dockerfile'
+  - '.'
+  automapSubstitutions: true
+  env:
+  - 'DOCKER_BUILDKIT=1'
+images:
+- '${_ARTIFACT_REGISTRY}/${_IMAGE_NAME}'

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/docker/patches/24.09/gpu_stats.patch

@@ -0,0 +1,41 @@
+diff --git a/nemo/collections/nlp/parts/megatron_trainer_builder.py b/nemo/collections/nlp/parts/megatron_trainer_builder.py
+index b2c85cde4..a3a9670c3 100644
+--- a/nemo/collections/nlp/parts/megatron_trainer_builder.py
++++ b/nemo/collections/nlp/parts/megatron_trainer_builder.py
+@@ -19,6 +19,7 @@ from lightning_fabric.utilities.exceptions import MisconfigurationException
+ from omegaconf import DictConfig
+ from pytorch_lightning import Trainer
+ from pytorch_lightning.callbacks import ModelSummary
++from pytorch_lightning.callbacks import Callback
+ from pytorch_lightning.plugins.environments import TorchElasticEnvironment
+
+ from nemo.collections.common.metrics.perf_metrics import FLOPsMeasurementCallback
+@@ -38,6 +39,23 @@ from nemo.utils.callbacks.dist_ckpt_io import (
+     AsyncFinalizerCallback,
+     DistributedCheckpointIO,
+ )
++from vmg.util.device_stats import gpu_stats_str
++
++class GpuStatsMon(Callback):
++    def on_train_start(self, trainer, pl_module) -> None:
++        rank=pl_module.global_rank
++        print(f'train_start: {rank=} {gpu_stats_str()}', flush=True)
++
++    def on_train_batch_start(self, trainer, pl_module, batch, batch_idx) -> None:
++        rank=pl_module.global_rank
++        print(f'batch_start: {rank=} {gpu_stats_str()}', flush=True)
++
++    def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx) -> None:
++        rank=pl_module.global_rank
++        print(f'batch_end: {rank=} {gpu_stats_str()}', flush=True)
+
+
+ class MegatronTrainerBuilder:
+@@ -178,6 +196,7 @@ class MegatronTrainerBuilder:
+         if self.cfg.get('exp_manager', {}).get('log_tflops_per_sec_per_gpu', True):
+             callbacks.append(FLOPsMeasurementCallback(self.cfg))
+
++        callbacks.append(GpuStatsMon())
+         return callbacks
+
+     def create_trainer(self, callbacks=None) -> Trainer:

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/docker/patches/24.09/local_rank.patch

@@ -0,0 +1,41 @@
+diff -ruN old-datasets/blended_megatron_dataset_builder.py datasets/blended_megatron_dataset_builder.py
+--- old-datasets/blended_megatron_dataset_builder.py	2025-05-02 04:08:45.369199665 +0000
++++ datasets/blended_megatron_dataset_builder.py	2025-05-02 04:10:47.369119891 +0000
+@@ -2,6 +2,7 @@
+
+ import logging
+ import math
++import os
+ from concurrent.futures import ThreadPoolExecutor
+ from typing import Any, Callable, Iterable, List, Optional, Type, Union
+
+@@ -353,7 +354,7 @@
+         num_dataset_builder_threads = self.config.num_dataset_builder_threads
+
+         if torch.distributed.is_initialized():
+-            rank = torch.distributed.get_rank()
++            rank = int(os.getenv("LOCAL_RANK", "0"))
+             # First, build on rank 0
+             if rank == 0:
+                 num_workers = num_dataset_builder_threads
+@@ -475,7 +476,7 @@
+             Optional[Union[DistributedDataset, Iterable]]: The DistributedDataset instantion, the Iterable instantiation, or None
+         """
+         if torch.distributed.is_initialized():
+-            rank = torch.distributed.get_rank()
++            rank = int(os.getenv("LOCAL_RANK", "0"))
+
+             dataset = None
+
+diff -ruN old-datasets/gpt_dataset.py datasets/gpt_dataset.py
+--- old-datasets/gpt_dataset.py	2025-05-02 04:08:45.369199665 +0000
++++ datasets/gpt_dataset.py	2025-05-02 04:09:30.309170278 +0000
+@@ -351,7 +351,7 @@
+
+         if not path_to_cache or (
+             not cache_hit
+-            and (not torch.distributed.is_initialized() or torch.distributed.get_rank() == 0)
++            and (not torch.distributed.is_initialized() or int(os.getenv("LOCAL_RANK", "0")) == 0)
+         ):
+
+             log_single_rank(

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/docker/patches/24.09/nemo2hf.patch

@@ -0,0 +1,13 @@
+diff --git a/scripts/checkpoint_converters/convert_llama_nemo_to_hf.py b/scripts/checkpoint_converters/convert_llama_nemo_to_hf.py
+index 8da15148d..005cae6c9 100644
+--- a/scripts/checkpoint_converters/convert_llama_nemo_to_hf.py
++++ b/scripts/checkpoint_converters/convert_llama_nemo_to_hf.py
+@@ -104,6 +104,8 @@ def convert(input_nemo_file, output_hf_file, precision=None, cpu_only=False) ->
+     dummy_trainer = Trainer(devices=1, accelerator='cpu', strategy=NLPDDPStrategy())
+     model_config = MegatronGPTModel.restore_from(input_nemo_file, trainer=dummy_trainer, return_config=True)
+     model_config.tensor_model_parallel_size = 1
++    model_config.virtual_pipeline_model_parallel_size = None
++    model_config.sequence_parallel = False
+     model_config.pipeline_model_parallel_size = 1
+     if cpu_only:
+         map_location = torch.device('cpu')

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/docker/patches/24.09/sigabort.patch

@@ -0,0 +1,24 @@
+diff --git a/examples/nlp/language_modeling/tuning/megatron_gpt_finetuning.py b/examples/nlp/language_modeling/tuning/megatron_gpt_finetuning.py
+index bfe8ea359..dfeaf93b5 100644
+--- a/examples/nlp/language_modeling/tuning/megatron_gpt_finetuning.py
++++ b/examples/nlp/language_modeling/tuning/megatron_gpt_finetuning.py
+@@ -13,6 +13,8 @@
+ # limitations under the License.
+
+ import torch.multiprocessing as mp
++import torch.distributed as dist
++
+ from omegaconf.omegaconf import OmegaConf
+
+ from nemo.collections.nlp.models.language_modeling.megatron_gpt_sft_model import MegatronGPTSFTModel
+@@ -76,6 +78,10 @@ def main(cfg) -> None:
+
+     trainer.fit(model)
+
++    if dist.is_available() and dist.is_initialized():
++        dist.barrier()
++        dist.destroy_process_group()
++
+
+ if __name__ == '__main__':
+     main()

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/docker/patches/24.09/throughput_calc.patch

@@ -0,0 +1,13 @@
+diff --git a/src/utils/training_metrics/process_training_results.py b/src/utils/training_metrics/process_training_results.py
+index 3e82a66..e61e1d8 100644
+--- a/src/utils/training_metrics/process_training_results.py
++++ b/src/utils/training_metrics/process_training_results.py
+@@ -134,7 +134,7 @@ def get_average_step_time(file: str, start_step: int, end_step: int) -> float:
+     for line in datajson:
+         if line.get("step") != "PARAMETER":
+             step = line.get("step")
+-            if step >= start_step and step <= end_step:
++            if step >= start_step and step <= end_step and "train_step_timing in s" in line["data"]:
+                 time_step_accumulator += line["data"].get("train_step_timing in s")
+                 num_steps += 1
+     if num_steps == 0:

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/docker/requirements.txt

@@ -0,0 +1,10 @@
+dllogger@git+https://github.com/NVIDIA/dllogger@v1.0.0
+
+# Fixing these libraries versions to avoid conflicting or broken packages.
+immutabledict==4.2.1
+protobuf==5.29.6
+opencv-python-headless==4.11.0.86
+docutils==0.16
+urllib3==2.6.3
+google-cloud-storage==3.0.0
+retrying

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/docker/uninstall.txt

@@ -0,0 +1,18 @@
+# cuml-cu12==24.8.0 was installed in nemo:24.09
+# Removing cuml=24.4.0 to avoid conflicting packages.
+cudf==24.4.0
+cugraph==24.4.0
+cugraph-service-server==24.4.0
+cuml==24.4.0
+dask-cudf==24.4.0
+raft-dask==24.4.0
+cugraph-dgl==24.4.0
+cugraph-pyg==24.4.0
+# The following packages are removed temporarily to avoid conflicting packages
+# and can be brought back if needed.
+tensorrt-llm==0.12.0
+img2dataset==1.45.0
+Sphinx==8.1.3
+sphinxcontrib-bibtex==2.6.3
+torchx==0.7.0
+nemo-run

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/docker/vertex-dist-recipes.Dockerfile

@@ -0,0 +1,66 @@
+# Dockerfile wrapping NeMo.
+#
+# To workaround base nemo docker image using too many layers, we use Multi-stage
+# build to first collect the additional files we'll need.
+FROM alpine:latest AS prep_files
+WORKDIR /workspace
+RUN mkdir -p configs vdt vdt/util
+COPY scripts/*.py vdt/
+COPY scripts/util/*.py vdt/util/
+COPY configs/* configs/
+COPY docker/patches/24.09/* vdt/patches/
+RUN chmod a+rwX -R vdt
+# Copy license.
+RUN wget https://raw.githubusercontent.com/GoogleCloudPlatform/vertex-ai-samples/main/LICENSE
+
+# Available tags
+# https://catalog.ngc.nvidia.com/orgs/nvidia/containers/nemo/tags
+# It installs NeMo source code in /opt/NeMo folder, with tag=r2.0.0
+FROM nvcr.io/nvidia/nemo:24.09
+
+RUN apt-get update && apt-get install -y sudo zsh tmux && \
+    rm -rf /var/lib/apt/lists*
+
+RUN echo "deb [signed-by=/usr/share/keyrings/cloud.google.gpg] http://packages.cloud.google.com/apt cloud-sdk main" | \
+    tee -a /etc/apt/sources.list.d/google-cloud-sdk.list && \
+    curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | \
+    apt-key --keyring /usr/share/keyrings/cloud.google.gpg add - && \
+    apt-get update -y && apt-get install google-cloud-sdk -y && \
+    rm -rf /var/lib/apt/lists*
+
+# Install libraries with pip
+ENV PIP_ROOT_USER_ACTION=ignore
+
+# We expect this will be run in the root directory of the vertex-dist-recipes repo
+ARG HOST_SRC_DIR="."
+
+# The pre-installed NeMo introduces a lot of deps conflicts.
+# We uninstall the confilicting libs and reinstall some of them as needed.
+COPY ${HOST_SRC_DIR}/docker/uninstall.txt /tmp/uninstall.txt
+RUN cat /tmp/uninstall.txt | grep -v '#' | xargs pip uninstall -y
+COPY ${HOST_SRC_DIR}/docker/requirements.txt /tmp/requirements.txt
+RUN pip install -r /tmp/requirements.txt
+
+# Make sure there's no inconsistent pip libraries.
+RUN pip check
+
+WORKDIR /workspace
+
+# Copy configs
+COPY ${HOST_SRC_DIR}/configs/* /opt/NeMo/examples/nlp/language_modeling/conf/
+
+# Copy all additional files we need from `prep_files` image.
+COPY --from=prep_files /workspace/ .
+
+# Install for `src/utils/training_metrics/process_training_results.py` to report
+# throughput and MFU numbers.
+RUN git clone https://github.com/AI-Hypercomputer/gpu-recipes.git
+
+# This hack is needed for multi-node training while not using a sharing file system.
+RUN patch --verbose -l -d /opt/megatron-lm/megatron/core/datasets -p1 -i /workspace/vdt/patches/local_rank.patch; \
+    git -C /workspace/gpu-recipes apply /workspace/vdt/patches/throughput_calc.patch; \
+    git -C /opt/NeMo apply /workspace/vdt/patches/nemo2hf.patch; \
+    git -C /opt/NeMo apply /workspace/vdt/patches/sigabort.patch;
+# git -C /opt/NeMo apply /workspace/vdt/patches/gpu_stats.patch;
+
+# Do not put an entrypoint here. Specify the entrypoint in the docker run script.

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/job_config.json

@@ -0,0 +1,16 @@
+{
+  "project_id": "<your_project_id>",
+  "region": "us-central1",
+  "zone": "us-central1-c",
+  "bucket": "<your_bucket",
+  "dataset_bucket": "github-repo/data/third-party/enwiki-latest-pages-articles",
+  "image_uri": "<your_image_uri>",
+  "strategy": "spot",
+  "nodes": "2",
+  "machine_type": "a3-megagpu-8g",
+  "gpu_type": "NVIDIA_H100_MEGA_80GB",
+  "gpus_per_node": "8",
+  "recipe_name": "llama3_1_8b_pretrain_a3mega",
+  "job_prefix": "vertex-ai",
+  "reservation_name": ""
+}

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/requirements.txt

@@ -0,0 +1,49 @@
+absl-py==2.2.2
+annotated-types==0.7.0
+anyio==4.9.0
+black==26.3.1
+cachetools==5.5.2
+certifi==2025.4.26
+charset-normalizer==3.4.2
+click==8.1.8
+docstring_parser==0.16
+google-api-core==2.24.2
+google-auth==2.40.1
+google-cloud-aiplatform==1.92.0
+google-cloud-bigquery==3.31.0
+google-cloud-core==2.4.3
+google-cloud-resource-manager==1.14.2
+google-cloud-storage==2.19.0
+google-crc32c==1.7.1
+google-genai==1.14.0
+google-resumable-media==2.7.2
+googleapis-common-protos==1.70.0
+grpc-google-iam-v1==0.14.2
+grpcio==1.71.0
+grpcio-status==1.71.0
+h11==0.16.0
+httpcore==1.0.9
+httpx==0.28.1
+idna==3.10
+mypy_extensions==1.1.0
+numpy==2.2.5
+packaging==25.0
+pathspec==0.12.1
+platformdirs==4.3.8
+proto-plus==1.26.1
+protobuf==5.29.6
+pyasn1==0.6.3
+pyasn1_modules==0.4.2
+pydantic==2.11.4
+pydantic_core==2.33.2
+python-dateutil==2.9.0.post0
+pytz==2025.2
+requests==2.33.0
+rsa==4.9.1
+shapely==2.1.0
+six==1.17.0
+sniffio==1.3.1
+typing-inspection==0.4.0
+typing_extensions==4.13.2
+urllib3==2.6.3
+websockets==15.0.1

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/scripts/launch.py

@@ -0,0 +1,173 @@
+"""Launch script for Vertex distributed training"""
+
+# Copy the sample_job_config.json file to job_config.json
+# to define the job parameters.
+#
+# Run like this:
+#
+#   python3 vertex_dist_train/launch.py --config_file=job_config.json
+#
+
+import datetime
+import json
+import os
+import pprint
+from collections.abc import Sequence
+from typing import Any, List
+
+from absl import app, flags
+from google.cloud import aiplatform
+from google.cloud.aiplatform_v1.types.custom_job import Scheduling
+from pytz import timezone
+
+FLAGS = flags.FLAGS
+flags.DEFINE_string("config_file", None, "Path to JSON config file")
+flags.DEFINE_boolean(
+    "debug", False, "Debug mode: just print the command, don't run it."
+)
+
+
+def launch_job(
+    job_name: str,
+    project: str,
+    region: str,
+    gcs_bucket: str,
+    image_uri: str,
+    entrypoint_cmd: List[str],
+    trainer_args: List[Any],
+    num_nodes: int,
+    machine_type: str,
+    num_gpus_per_node: int,
+    gpu_type: str,
+    strategy: str,
+    reservation_name: str = "",
+):
+    assert strategy in ("dws", "spot", "reservation")
+    aiplatform.init(
+        project=project, location=region, staging_bucket=gcs_bucket
+    )
+
+    train_job = aiplatform.CustomContainerTrainingJob(
+        display_name=job_name,
+        container_uri=image_uri,
+        command=entrypoint_cmd,
+    )
+
+    job_args = dict(
+        args=trainer_args,
+        enable_web_access=True,
+        replica_count=num_nodes,
+        machine_type=machine_type,
+        accelerator_type=gpu_type,
+        accelerator_count=num_gpus_per_node,
+        boot_disk_size_gb=1000,
+        restart_job_on_worker_restart=True,
+        #restart_job_on_worker_restart=False,
+    )
+
+    if strategy == "spot":
+        job_args.update({"scheduling_strategy": Scheduling.Strategy.SPOT.name})
+    elif strategy == "dws":
+        job_args.update(
+            {"scheduling_strategy": Scheduling.Strategy.FLEX_START.name}
+        )
+    elif strategy == "reservation":
+        assert reservation_name != "", (
+            "If using a reservation, provide the reservation_name in the "
+            "format `projects/{project_id_or_number}/zones/{zone}/"
+            "reservations/{reservation_name}`"
+        )
+        job_args.update(
+            {
+                "reservation_affinity_type": "SPECIFIC_RESERVATION",
+                "reservation_affinity_key": "compute.googleapis.com/reservation-name",
+                "reservation_affinity_values": [reservation_name],
+            }
+        )
+
+    pprint.pprint(job_args)
+    if not FLAGS.debug:
+        train_job.submit(**job_args)
+
+
+def main(argv: Sequence[str]) -> None:
+    config_file_path = FLAGS.config_file
+    print(f"Reading job config from {config_file_path}")
+    with open(config_file_path, encoding="utf-8") as config_file:
+        config = json.load(config_file)
+
+    project_id = config["project_id"]
+    region = config["region"]
+    zone = config["zone"]
+    bucket = config["bucket"]
+    dataset_bucket = config["dataset_bucket"]
+    n_nodes = int(config["nodes"])
+    machine_type = config["machine_type"]
+    num_gpus_per_node = int(config["gpus_per_node"])
+    gpu_type = config["gpu_type"]
+    reservation_name = config.get("reservation_name")
+    reservation_full_name = (
+        f"projects/{project_id}/zones/{zone}/reservations/{reservation_name}"
+        if "reservation_name" in config
+        else ""
+    )
+
+    strategy = config["strategy"]
+    recipe_name = config["recipe_name"]
+    job_prefix = config["job_prefix"]
+    image_uri = config["image_uri"]
+
+    # Job name
+    timestamp = (
+        datetime.datetime.now()
+        .astimezone(timezone("US/Pacific"))
+        .strftime("%Y%m%d_%H%M%S")
+    )
+    job_name = f"{recipe_name}-{timestamp}"
+    if job_prefix:
+        job_name = f"{job_prefix}-{job_name}"
+
+    base_output_dir = os.path.join("/gcs", bucket, job_name)
+
+    # Training command and args
+    entrypoint_cmd = ["python3", "vdt/run.py"]
+
+    dataset_bucket = f"gs://{config['dataset_bucket']}"
+
+    trainer_args = [
+        f"--train_data_gcs={dataset_bucket}",
+        "/opt/NeMo/examples/nlp/language_modeling/megatron_gpt_pretraining.py",
+        "--config-path=conf/",
+        f"--config-name={recipe_name}.yaml",
+        f"exp_manager.explicit_log_dir={base_output_dir}",
+        f"exp_manager.dllogger_logger_kwargs.json_file={base_output_dir}/dllogger.json",
+        "+exp_manager.create_tensorboard_logger=true",
+        "exp_manager.create_checkpoint_callback=false",
+        f"trainer.num_nodes={n_nodes}",
+        f"trainer.devices={num_gpus_per_node}",
+        "trainer.max_steps=10",
+        "trainer.log_every_n_steps=1",
+        "model.tokenizer.vocab_file=/data/gpt2-vocab.json",
+        "model.tokenizer.merge_file=/data/gpt2-merges.txt",
+        "model.data.data_prefix=[1.0,/data/hfbpe_gpt_training_data_text_document]",
+    ]
+
+    launch_job(
+        job_name=job_name,
+        project=project_id,
+        region=region,
+        gcs_bucket=bucket,
+        image_uri=image_uri,
+        entrypoint_cmd=entrypoint_cmd,
+        trainer_args=trainer_args,
+        num_nodes=n_nodes,
+        machine_type=machine_type,
+        num_gpus_per_node=num_gpus_per_node,
+        gpu_type=gpu_type,
+        strategy=strategy,
+        reservation_name=reservation_full_name,
+    )
+
+
+if __name__ == "__main__":
+    app.run(main)

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/scripts/run.py

@@ -0,0 +1,85 @@
+"""Entrypoint for Vertex Distributed Training container."""
+
+import argparse
+import os
+import sys
+from collections.abc import Sequence
+from subprocess import STDOUT, check_output, run
+
+from absl import app, flags, logging
+from util import cluster_spec
+
+from retrying import retry
+
+# PyTorch barrier call which synchronizes all of the nodes before launching the training process.
+# This makes sure that processes will block until all processes are ready.
+# Improves the reliability of spot VM usage for multi-node training jobs
+
+@retry(stop_max_attempt_number=100, wait_exponential_multiplier=1000)
+def barrier_with_retry() -> None:
+    import torch
+    logging.info("Starting barrier on RANK {}".format(os.environ["RANK"]))
+    torch.distributed.init_process_group()
+    torch.distributed.barrier()
+    torch.distributed.destroy_process_group()
+    logging.info("Finished barrier on RANK {}".format(os.environ["RANK"]))
+
+def main(unused_argv: Sequence[str]) -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--train_data_gcs",
+        type=str,
+        help="Download training data from gcs path",
+    )
+    args, unknown = parser.parse_known_args()
+
+    for key, val in os.environ.items():
+        logging.info("ENV %s=%s", key, val)
+
+    if args.train_data_gcs:
+        local_dir = "/data"
+        if not os.path.exists(local_dir):
+            os.mkdir(local_dir)
+        logging.info("downloading %s to %s...", args.train_data_gcs, local_dir)
+        check_output(
+            [
+                "gcloud",
+                "storage",
+                "cp",
+                "-r",
+                f"{args.train_data_gcs}/*",
+                local_dir,
+            ],
+            stderr=STDOUT,
+        )
+        logging.info("%s downloaded.", args.train_data_gcs)
+
+    primary_node_addr, primary_node_port, node_rank, num_nodes = (
+        cluster_spec.get_cluster_spec()
+    )
+
+    cmd = [
+        "torchrun",
+        "--nproc-per-node=8",
+        f"--nnodes={num_nodes}",
+        f"--node_rank={node_rank}",
+    ]
+    if num_nodes > 1:
+        cmd += [
+            "--max-restarts=3",
+            "--rdzv-backend=static",
+            f'--rdzv_id={os.getenv("CLOUD_ML_JOB_ID", primary_node_port)}',
+            f"--rdzv-endpoint={primary_node_addr}:{primary_node_port}",
+        ]
+    cmd += unknown
+
+    logging.info("launching with cmd: \n%s", " \\\n".join(cmd))
+    barrier_with_retry()
+    run(cmd, stdout=sys.stdout, stderr=sys.stdout, check=True)
+
+
+if __name__ == "__main__":
+    logging.get_absl_handler().python_handler.stream = sys.stdout
+    app.run(
+        main, flags_parser=lambda _args: flags.FLAGS(_args, known_only=True)
+    )

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/scripts/util/cluster_spec.py

@@ -0,0 +1,81 @@
+"""Get cluster info from environment variables."""
+
+import dataclasses
+import json
+import os
+
+from absl import logging
+
+
+@dataclasses.dataclass
+class ClusterInfo:
+    """Contains information about the cluster.
+
+    Attributes:
+      primary_node_addr: The address of the primary node.
+      primary_node_port: The port of the primary node.
+      node_rank: The rank of the node.
+      num_nodes: The number of nodes in the cluster.
+    """
+
+    primary_node_addr: str | None = None
+    primary_node_port: str | None = None
+    node_rank: int = 0
+    num_nodes: int = 1
+
+    # Allows unpacking operation like
+    # primary_node_addr, primary_node_port, _, _ = ClusterInfo()
+    # See https://stackoverflow.com/a/70753113
+    def __iter__(self):
+        return iter(dataclasses.astuple(self))
+
+
+def get_cluster_spec() -> ClusterInfo:
+    """Parses CLUSTER_SPEC environment variable and returns the cluster info.
+
+    Returns:
+      A ClusterInfo object.
+    """
+    cluster_spec = os.getenv("CLUSTER_SPEC", None)
+
+    # If CLUSTER_SPEC is not set, use individual vars to construct cluster info.
+    if not cluster_spec:
+        cluster_info = ClusterInfo(
+            primary_node_addr=os.getenv("MASTER_ADDR", None),
+            primary_node_port=os.getenv("MASTER_PORT", None),
+            node_rank=int(os.getenv("RANK", "0")),
+            num_nodes=int(os.getenv("NNODES", "1")),
+        )
+        return cluster_info
+
+    cluster_data = json.loads(cluster_spec)
+    # Get primary node info
+    primary_node = cluster_data["cluster"]["workerpool0"][0]
+    logging.info("primary node: %s", primary_node)
+    primary_node_addr, primary_node_port = primary_node.split(":")
+    logging.info("primary node address: %s", primary_node_addr)
+    logging.info("primary node port: %s", primary_node_port)
+
+    # Determine node rank of this machine
+    workerpool = cluster_data["task"]["type"]
+    if workerpool == "workerpool0":
+        node_rank = 0
+    elif workerpool == "workerpool1":
+        # Add 1 for the primary node, since `index` is the index of workerpool1.
+        node_rank = cluster_data["task"]["index"] + 1
+    else:
+        raise ValueError(
+            "Only workerpool0 and workerpool1 are supported. Unknown workerpool:"
+            f" {workerpool}"
+        )
+    logging.info("node rank: %s", node_rank)
+
+    # Calculate total nodes.
+    num_nodes = 1  # For the primary node.
+    if "workerpool1" in cluster_data["cluster"]:
+        num_nodes += len(cluster_data["cluster"]["workerpool1"])
+    logging.info("num nodes: %s", num_nodes)
+
+    return ClusterInfo(
+        primary_node_addr, primary_node_port, node_rank, num_nodes
+    )

community-content/vertex-distributed-training/a3mega/llama-3-8b-nemo-pretraining/scripts/util/cluster_spec_test.py

@@ -0,0 +1,59 @@
+"""Add tests for cluster_spec.py."""
+
+import os
+
+from . import cluster_spec
+
+
+# TODO(styer): Use pytest instead
+class ClusterSpecTest(googletest.TestCase):
+
+    def setUp(self):
+        super().setUp()
+        self.curr_env_var = os.environ.copy()
+
+    def tearDown(self):
+        super().tearDown()
+        os.environ = self.curr_env_var
+
+    def test_get_cluster_spec_from_env_vars(self):
+        os.environ["CLUSTER_SPEC"] = ""
+        os.environ["MASTER_ADDR"] = "127.0.0.1"
+        os.environ["MASTER_PORT"] = "8080"
+        os.environ["RANK"] = "0"
+        os.environ["NNODES"] = "2"
+        cluster_info = cluster_spec.get_cluster_spec()
+        self.assertEqual(cluster_info.primary_node_addr, "127.0.0.1")
+        self.assertEqual(cluster_info.primary_node_port, "8080")
+        self.assertEqual(cluster_info.node_rank, 0)
+        self.assertEqual(cluster_info.num_nodes, 2)
+
+    def test_get_cluster_spec_from_cluster_spec(self):
+        os.environ[
+            "CLUSTER_SPEC"
+        ] = """
+    {
+      "cluster": {
+        "workerpool0": [
+          "127.0.0.1:8080"
+        ],
+        "workerpool1": [
+          "127.0.0.2:8080",
+          "127.0.0.3:8080"
+        ]
+      },
+      "task": {
+        "type": "workerpool1",
+        "index": 0
+      }
+    }
+    """
+        cluster_info = cluster_spec.get_cluster_spec()
+        self.assertEqual(cluster_info.primary_node_addr, "127.0.0.1")
+        self.assertEqual(cluster_info.primary_node_port, "8080")
+        self.assertEqual(cluster_info.node_rank, 1)
+        self.assertEqual(cluster_info.num_nodes, 3)
+
+
+if __name__ == "__main__":
+    googletest.main()

community-content/vertex_cpr_samples/README.md

@@ -0,0 +1,15 @@
+# Vertex AI custom prediction routines samples
+
+## Overview
+Vertex Custom Prediction Routines(CPR) simplify the process of building custom containers 
+and make local model testing easy. Here are the sameple codes for different libraries.
+
+
+### Objectives
+The objective is to provide various samples for Vertex Custom Prediction Routine(CPR).
+
+
+### Supporting libraries
+* torch
+* sklearn
+* xgboost

community-content/vertex_cpr_samples/sklearn/predictor_Classifier.py

@@ -0,0 +1,33 @@
+import numpy as np
+import os
+import pickle
+
+from google.cloud.aiplatform.constants import prediction
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from sklearn.datasets import load_breast_cancer
+from sklearn.linear_model import RidgeClassifier
+
+class LinearRegressionPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists(prediction.MODEL_FILENAME_PKL):
+            self._model = pickle.load(open(prediction.MODEL_FILENAME_PKL, "rb"))
+        else:
+            self._model = RidgeClassifier()
+        X, y = load_breast_cancer(return_X_y=True)
+        self._model.fit(X, y)
+
+    def preprocess(self, prediction_input: dict) -> np.ndarray:
+        instances = prediction_input["instances"]
+        return np.asarray(instances)
+
+    def predict(self, instances: np.ndarray) -> np.ndarray:
+        return self._model.predict(instances)
+
+    def postprocess(self, prediction_results: np.ndarray) -> dict:
+        return {"predictions": prediction_results.tolist()}

community-content/vertex_cpr_samples/sklearn/predictor_LinearRegression.py

@@ -0,0 +1,33 @@
+import numpy as np
+import os
+import pickle
+
+from google.cloud.aiplatform.constants import prediction
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from sklearn.datasets import make_blobs
+from sklearn.linear_model import LinearRegression 
+
+class LinearRegressionPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists(prediction.MODEL_FILENAME_PKL):
+            self._model = pickle.load(open(prediction.MODEL_FILENAME_PKL, "rb"))
+        else:
+            self._model = LogisticRegression()
+        X, y = make_blobs(n_samples=100, centers=2, n_features=2, random_state=1)
+        self._model.fit(X, y)
+
+    def preprocess(self, prediction_input: dict) -> np.ndarray:
+        instances = prediction_input["instances"]
+        return np.asarray(instances)
+
+    def predict(self, instances: np.ndarray) -> np.ndarray:
+        return self._model.predict_proba(instances)
+
+    def postprocess(self, prediction_results: np.ndarray) -> dict:
+        return {"predictions": prediction_results.tolist()}

community-content/vertex_cpr_samples/sklearn/predictor_SGDClassifier.py

@@ -0,0 +1,33 @@
+import numpy as np
+import os
+import pickle
+
+from google.cloud.aiplatform.constants import prediction
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from sklearn.linear_model import SGDClassifier
+
+class SGDClassifierPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists(prediction.MODEL_FILENAME_PKL):
+            self._model = pickle.load(open(prediction.MODEL_FILENAME_PKL, "rb"))
+        else:
+            self._model = SGDClassifier(max_iter=5)
+        X = [[0., 0.], [1., 1.]]
+        y = [0, 1]
+        self._model.fit(X, y)
+
+    def preprocess(self, prediction_input: dict) -> np.ndarray:
+        instances = prediction_input["instances"]
+        return np.asarray(instances)
+
+    def predict(self, instances: np.ndarray) -> np.ndarray:
+        return self._model.predict(instances)
+
+    def postprocess(self, prediction_results: np.ndarray) -> dict:
+        return {"predictions": prediction_results.tolist()}

community-content/vertex_cpr_samples/torch/predictor_resnet.py

@@ -0,0 +1,34 @@
+import os
+import torch
+
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from torchvision.models import detection, resnet50, ResNet50_Weights
+from typing import Dict, List
+
+class ResNetPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists("model.pth.tar"):
+            self.model = detection.fasterrcnn_resnet50_fpn(pretrained=True)
+            stat_dic = torch.load("model.pth.tar")
+            self.model.load_state_dict(stat_dic['state_dict'])
+        else:
+            weights = ResNet50_Weights.DEFAULT
+            self.model = resnet50(weights=weights)
+        self.model.eval()
+
+    def preprocess(self, prediction_input: dict) -> torch.Tensor:
+        instances = prediction_input["instances"]
+        return torch.Tensor(instances)
+
+    @torch.inference_mode()
+    def predict(self, instances: torch.Tensor) -> List[str]:
+        return self._model(instances)
+
+    def postprocess(self, prediction_results: List[str]) -> Dict:
+        return {"predictions": prediction_results}

community-content/vertex_cpr_samples/torch/predictor_transformers.py

@@ -0,0 +1,73 @@
+import ast
+import json
+import os
+import pickle
+import torch
+
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from transformers import AutoModelForQuestionAnswering
+from typing import Dict, List
+
+class TorchTransformersPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        
+        if os.path.isfile("setup_config.json"):
+            with open("setup_config.json") as setup_config_file:
+                self.setup_config = json.load(setup_config_file)
+
+        if os.path.exists("model.pt"):
+            self.model = AutoModelForQuestionAnswering.from_pretrained("model.pt")
+            self.model.eval()
+        else:
+            raise ValueError("One of the following model files must be provided: model.pt.")
+
+    def preprocess(self, prediction_input: dict) -> torch.Tensor:
+        max_length = self.setup_config["max_length"]
+        instances = prediction_input["instances"]
+        question_context = ast.literal_eval(instances)
+        question = question_context["question"]
+        context = question_context["context"]
+        inputs = self.tokenizer.encode_plus(
+            question,
+            context,
+            max_length=int(max_length),
+            pad_to_max_length=True,
+            add_special_tokens=True,
+            return_tensors="pt",
+        )
+        input_ids = inputs["input_ids"]
+        attention_mask = inputs["attention_mask"]
+        return torch.Tensor(input_ids, attention_mask)
+    
+    @torch.inference_mode()
+    def predict(self, instances: torch.Tensor) -> List[str]:
+        input_ids, attention_mask = instances
+        outputs = self._model(input_ids, attention_mask)
+        answer_start_scores = outputs.start_logits
+        answer_end_scores = outputs.end_logits
+        
+        num_rows, num_cols = answer_start_scores.shape
+        inferences = []
+        for i in range(num_rows):
+            answer_start_scores_one_seq = answer_start_scores[i].unsqueeze(0)
+            answer_start = torch.argmax(answer_start_scores_one_seq)
+            answer_end_scores_one_seq = answer_end_scores[i].unsqueeze(0)
+            answer_end = torch.argmax(answer_end_scores_one_seq) + 1
+            prediction = self.tokenizer.convert_tokens_to_string(
+                self.tokenizer.convert_ids_to_tokens(
+                    input_ids[i].tolist()[answer_start:answer_end]
+                )
+            )
+            inferences.append(prediction)
+        return inferences
+
+    def postprocess(self, prediction_results: List[str]) -> Dict:
+        return {"predictions": prediction_results}
+     
+

community-content/vertex_cpr_samples/xgboost/predictor_Classifier.py

@@ -0,0 +1,37 @@
+import os
+import numpy as np
+import pickle
+import xgboost as xgb
+
+from google.cloud.aiplatform.constants import prediction
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+from sklearn.datasets import make_blobs
+from xgboost import XGBClassifier
+
+
+class ClassifierPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists(prediction.MODEL_FILENAME_PKL):
+            booster = pickle.load(open(prediction.MODEL_FILENAME_PKL, "rb"))
+        else:
+            X, y = make_blobs(n_samples=100, centers=2, n_features=2, random_state=1)
+            model = XGBClassifier()
+            model.fit(X, y)
+            booster = model.get_booster()
+        self._booster = booster
+
+    def preprocess(self, prediction_input: dict) -> xgb.DMatrix:
+        instances = prediction_input["instances"]
+        return xgb.DMatrix(instances)
+
+    def predict(self, instances: xgb.DMatrix) -> np.ndarray:
+        return self._booster.predict(instances)
+
+    def postprocess(self, prediction_results: np.ndarray) -> dict:
+        return {"predictions": prediction_results.tolist()}

community-content/vertex_cpr_samples/xgboost/predictor_xgbranker.py

@@ -0,0 +1,41 @@
+import os
+import numpy as np
+import pandas as pd
+import pickle
+import xgboost as xgb
+
+from google.cloud.aiplatform.constants import prediction
+from google.cloud.aiplatform.utils import prediction_utils
+from google.cloud.aiplatform.prediction.predictor import Predictor
+
+class XGBRankerPredictor(Predictor):
+
+    def __init__(self):
+        return
+
+    def load(self, artifacts_uri: str) -> None:
+        prediction_utils.download_model_artifacts(artifacts_uri)
+        if os.path.exists(prediction.MODEL_FILENAME_PKL):
+            booster = pickle.load(open(prediction.MODEL_FILENAME_PKL, "rb"))
+            self._booster = booster
+        else:
+            N = 500
+            dates = pd.date_range(start='2023-01-01', end='2023-01-12', periods=N)
+            X = pd.DataFrame(np.random.randn(N, 5), columns=list('ABCDE'), index=dates)
+            y = pd.Series(np.random.randint(0, 10, size=N), index=dates, name='label')
+            group = X.groupby(dates + pd.offsets.MonthEnd(0)).size()
+            sample_weight = pd.Series(np.arange(len(group)), index=group.index)
+            model = xgb.XGBRanker(objective='rank:pairwise', max_depth=3, learning_rate=0.1, booster='gbtree', tree_method='hist', n_jobs=4, n_estimators=50, enable_categorical=False, random_state=42)
+            model.fit(X=X, y=y, group=group, sample_weight=sample_weight, verbose=True)
+            booster = model.get_booster()
+            self._booster = booster
+
+    def preprocess(self, prediction_input: dict) -> xgb.DMatrix:
+        instances = prediction_input["instances"]
+        return xgb.DMatrix(instances)
+
+    def predict(self, instances: xgb.DMatrix) -> np.ndarray:
+        return self._booster.predict(instances, output_margin=False, ntree_limit=0)
+
+    def postprocess(self, prediction_results: np.ndarray) -> dict:
+        return {"predictions": prediction_results.tolist()}

community-content/vertex_model_garden/.pylintrc

@@ -0,0 +1,4 @@
+[MASTER]
+
+generated-members=get_concrete_function,cv2.*
+ignored-modules=tensorflow,google.cloud