Add elastic pause/resume functionality to MaxText.

PiperOrigin-RevId: 890594825

Author: abhinavclemson

src/maxtext/configs/base.yml

@@ -1193,3 +1193,9 @@ distill_temperature: 1.0
 # 0.0 value disables this feature.
 distill_beta: 0.0
 distill_layer_indices: None
+
+##### Elastic training parameters
+# Elastic training is Pathways-specific and does not work on McJAX.
+elastic_enabled: false
+elastic_timeout_seconds: 300
+elastic_max_retries: 10

src/maxtext/configs/types.py

@@ -1550,6 +1550,26 @@ class Goodput(BaseModel):
   enable_gcp_step_deviation_metrics: bool = Field(True, description="Enable GCP step deviation metrics.")
 
 
+class ElasticTraining(BaseModel):
+  """Configuration for elastic training and fault tolerance.
+
+  Elastic training is Pathways-specific and does not work on McJAX.
+  """
+
+  elastic_enabled: bool = Field(False, description="Whether to enable elastic training.")
+  elastic_timeout_seconds: int = Field(
+      300,
+      description=(
+          "The maximum number of seconds to wait for `elastic_minimum_slice_count` slices to become active. If this"
+          " timeout is reached during any retry attempt, a `TimeoutError` is raised and training fails."
+      ),
+  )
+  elastic_max_retries: int = Field(
+      10,
+      description="The maximum number of times to retry training when a slice failure occurs or when scaling up.",
+  )
+
+
 class GcpMonitoring(BaseModel):
   """Configuration for GCP-specific workload monitoring."""
 
@@ -1947,6 +1967,7 @@ class MaxTextConfig(
     Checkpointing,
     OrbaxStorage,
     EmergencyCheckpointing,
+    ElasticTraining,
     # Data Types and Quantization
     DataTypes,
     Quantization,
@@ -2456,6 +2477,8 @@ def calculate_global_batch_sizes(per_device_batch_size, expansion_factor, num_de
     # H. RUN ALL CROSS-FIELD VALIDATIONS
     if self.load_parameters_path and self.load_full_state_path:
       raise ValueError("At most one of `load_parameters_path` or `load_full_state_path` should be set.")
+    if self.elastic_enabled and not self.enable_single_controller:
+      raise ValueError("Elastic training is only supported with Pathways (`enable_single_controller=True`).")
     if (self.load_parameters_path or self.load_full_state_path) and not self.enable_checkpointing:
       raise ValueError("You must set enable_checkpointing=True to load a checkpoint.")
     if self.enable_multi_tier_checkpointing:

src/maxtext/trainers/pre_train/train.py

@@ -41,6 +41,7 @@
 from maxtext.configs import pyconfig
 from maxtext.common.common_types import ShardMode
 from maxtext.utils.globals import EPS
+from maxtext.utils import elastic_utils
 # Placeholder: internal
 
 # pylint: disable=too-many-positional-arguments
@@ -678,8 +679,30 @@ def run(config, recorder, diagnostic_config):
 def main(argv: Sequence[str]) -> None:
   config, recorder, diagnostic_config = initialize(argv)
   record_goodput(recorder, RECORD_JOB_START_TIME)
+
+  if config.elastic_enabled:
+    max_logging.log("Elastic utils: Elastic training enabled.")
+
+    def elastic_train_func():
+      """Train function for elastic training.
+
+      Initializes variables and runs the train loop.
+      """
+      elastic_config, elastic_recorder, elastic_diagnostic_config = initialize(argv)
+      run(
+          elastic_config,
+          elastic_recorder,
+          elastic_diagnostic_config,
+      )
+
+    train_func = elastic_utils.elastic_retry(config)(elastic_train_func)
+  else:
+    # Use the already initialized variables
+    def train_func():
+      run(config, recorder, diagnostic_config)
+
   with maybe_monitor_goodput(config):
-    run(config, recorder, diagnostic_config)
+    train_func()
 
 
 if __name__ == "__main__":

src/maxtext/utils/elastic_utils.py

@@ -0,0 +1,129 @@
+# Copyright 2023–2025 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
+#
+#    https://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.
+
+"""Utility functions for Elastic Training."""
+
+import functools
+import jax
+from maxtext.utils import gcs_utils
+from maxtext.utils import max_logging
+import pathwaysutils
+from pathwaysutils.elastic import manager
+
+
+elastic_manager: manager.Manager | None = None
+
+
+def elastic_enabled(config) -> bool:
+  """Returns whether elastic mode is enabled."""
+  return pathwaysutils.is_pathways_backend_used() and config.elastic_enabled
+
+
+def clean_up_checkpoints(checkpoint_dir: str):
+  """Cleans up incomplete checkpoints after an elastic event."""
+  max_logging.log("Elastic utils: Checking for incomplete checkpoint after an elastic event...")
+  checkpoint_dir = gcs_utils.add_trailing_slash(checkpoint_dir)
+
+  # 1. List the "directories" (steps)
+  checkpoints = gcs_utils.gcs_list_directories(checkpoint_dir)
+
+  # 2. Filter for directories that are numbers
+  checkpoints = [cp for cp in checkpoints if cp.isdigit()]
+
+  if not checkpoints:
+    max_logging.log("Found no existing checkpoints. Continuing")
+    return
+
+  # Sort naturally (numerical sort) and get the last one
+  checkpoints.sort(key=int)
+  latest_checkpoint_name = checkpoints[-1]
+  latest_checkpoint_path = f"{checkpoint_dir}{latest_checkpoint_name}/"
+
+  max_logging.log(f"Checking latest checkpoint: {latest_checkpoint_path}")
+
+  # 3. Check for commit_success file
+  success_markers = gcs_utils.gcs_glob_pattern(f"{latest_checkpoint_path}commit_success*")
+
+  if not success_markers:
+    max_logging.log(f"No commit_success file found. Deleting {latest_checkpoint_path}...")
+    gcs_utils.gcs_delete_directory(latest_checkpoint_path)
+  else:
+    max_logging.log(f"Found commit_success file. Keeping {latest_checkpoint_path}.")
+
+
+def live_devices():
+  """Returns the list of live devices."""
+  global elastic_manager
+  # If pathways is not used or elastic_manager is not initialized, return all devices
+  if pathwaysutils.is_pathways_backend_used():
+    if elastic_manager is None:
+      elastic_manager = manager.Manager()
+    # Filter devices that are in active slices
+    return [d for d in jax.devices() if d.slice_index in elastic_manager.active_slice_indices]
+  return jax.devices()
+
+
+def chain_callbacks(*funcs):
+  """Helper function to chain callbacks."""
+
+  def wrapper():
+    for func in funcs:
+      func()
+
+  return wrapper
+
+
+def elastic_retry(config, callback_fn=None):
+  """Decorator for elastic retry.
+
+  If an elastic event occurs, the decorator will retry the decorated function
+  up to `config.elastic_max_retries` times.
+  Before each retry, it cleans up partial checkpoints by calling
+  `clean_up_checkpoints`. If `callback_fn` is provided, it is
+  called after `clean_up_checkpoints`.
+
+  Args:
+    config: Config object.
+    callback_fn: Optional callback function to be called after
+      `clean_up_checkpoints` on an elastic event.
+
+  Returns:
+    A decorator for elastic retry.
+  """
+  global elastic_manager
+  if not elastic_enabled(config):
+    msg = (
+        "Elastic training requires the Pathways backend, and elastic_enabled"
+        " must be set to True: current config.elastic_enabled:"
+        f" {config.elastic_enabled}, pathways backend used:"
+        f" {pathwaysutils.is_pathways_backend_used()}"
+    )
+    raise ValueError(msg)
+
+  max_logging.log("Elastic Retry Enabled")
+  if elastic_manager is None:
+    elastic_manager = manager.Manager()
+
+  cleanup_partial = functools.partial(clean_up_checkpoints, config.checkpoint_dir)
+
+  if callback_fn is None:
+    effective_callback = cleanup_partial
+  else:
+    effective_callback = chain_callbacks(cleanup_partial, callback_fn)
+
+  return elastic_manager.elastic_retry(
+      max_retries=config.elastic_max_retries,
+      timeout=config.elastic_timeout_seconds,
+      on_elastic_event_callback=effective_callback,
+  )

src/maxtext/utils/gcs_utils.py

@@ -12,14 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-""" Common GCS Utils needed by multiple modules"""
+"""Common GCS Utils needed by multiple modules"""
 import shutil
 import json
 import os
 import socket
 from pathlib import Path
 from etils import epath
 import uuid
+from concurrent.futures import ThreadPoolExecutor
 
 import yaml
 
@@ -168,6 +169,35 @@ def gcs_list_directories(directory_path):
   return directories
 
 
+def gcs_delete_directory(directory_path: str):
+  """Deletes a "directory" (all blobs with the prefix) from GCS.
+
+  Args:
+      directory_path: The GCS path (gs://...) representing the "directory" to delete.
+  """
+  if not _gcs_guard("gcs_delete_directory"):
+    return
+  storage_client = storage.Client()
+  bucket_name, directory_prefix = parse_gcs_bucket_and_prefix(directory_path)
+  bucket = storage_client.bucket(bucket_name)
+
+  # Ensures the prefix has a trailing slash to avoid deleting more than intended.
+  if not directory_prefix.endswith("/"):
+    directory_prefix += "/"
+
+  blobs = list(bucket.list_blobs(prefix=directory_prefix))
+  if blobs:
+    # Uses a ThreadPoolExecutor to delete blobs in parallel to match gsutil -m performance.
+    def _delete_blob(blob):
+      try:
+        blob.delete()
+      except Exception as e:  # pylint: disable=broad-except
+        max_logging.log(f"Error deleting blob {blob.name}: {e}")
+
+    with ThreadPoolExecutor(max_workers=32) as executor:
+      executor.map(_delete_blob, blobs)
+
+
 def gcs_glob_pattern(pattern):
   """
   Globs GCS files and returns a list of full GCS paths.