support grain checkpoint (#133)

Author: aireenmei

docs/data_README.md

@@ -6,7 +6,8 @@ Currently MaxDiffusion supports 3 data input pipelines, controlled by the flag `
 | -------- | ---------------- | --------------- | ----------------------- |
 | HuggingFace (hf) | datasets in HuggingFace Hub or local/Cloud Storage | Formats supported in HF Hub: parquet, arrow, json, csv, txt | data are not loaded in memory but streamed from the saved location, good for big dataset | 
 | tf | dataset will be downaloaded form HuggingFace Hub to disk | Formats supported in HF Hub: parquet, arrow, json, csv, txt | Will read the whole dataset into memory, works for small dataset |
-| tfrecord | local/Cloud Storage | tfrecord | data are not loaded in memory but streamed from the saved location, good for big dataset |
+| tfrecord | local/Cloud Storage | TFRecord | data are not loaded in memory but streamed from the saved location, good for big dataset |
+| Grain | local/Cloud Storage | ArrayRecord (or any random access format) | data are not loaded in memory but streamed from the saved location, good for big dataset, supports global shuffle and data iterator checkpoint for determinism (see details in [doc](https://github.com/AI-Hypercomputer/maxtext/blob/main/getting_started/Data_Input_Pipeline.md#grain-pipeline---for-determinism)) |
 
 ## Usage examples
 
@@ -45,6 +46,12 @@ dataset_type: tfrecord
 train_data_dir: gs://<bucket>/<folder>  # will use all TFRecord files under the directory
 ```
 
+### Grain (dataset_type=grain)
+```
+dataset_type: grain
+grain_train_files: gs://<bucket>/<folder>/*.arrayrecord  # match the file pattern
+```
+
 ## Best Practice
 ### Multihost Dataloading
 In multihost environment, if use a streaming type of input pipeline and the data format only supports sequential reads (dataset_type in (hf, tfrecord in MaxDiffusion)), the most performant way is to have each data file only accessed by one host, and each host access a subset of data files (shuffle is within the subset of files). This requires (# of data files) > (# of hosts loading data). We recommand users to reshard the dataset if this requirement is not met.

src/maxdiffusion/checkpointing/base_stable_diffusion_checkpointer.py

@@ -22,6 +22,7 @@
 import jax
 from jax.sharding import Mesh
 import orbax.checkpoint as ocp
+import grain.python as grain
 from maxdiffusion import (
     max_utils,
     FlaxStableDiffusionPipeline,
@@ -57,7 +58,11 @@ def __init__(self, config, checkpoint_type):
     self.total_train_batch_size = self.config.total_train_batch_size
 
     self.checkpoint_manager = create_orbax_checkpoint_manager(
-        self.config.checkpoint_dir, enable_checkpointing=True, save_interval_steps=1, checkpoint_type=checkpoint_type
+        self.config.checkpoint_dir,
+        enable_checkpointing=True,
+        save_interval_steps=1,
+        checkpoint_type=checkpoint_type,
+        dataset_type=config.dataset_type,
     )
 
   def _create_optimizer(self, config, learning_rate):
@@ -157,6 +162,22 @@ def create_text_encoder_2_state(self, pipeline, params, checkpoint_item_name, is
         training=is_training,
     )
 
+  def restore_data_iterator_state(self, data_iterator):
+    if (
+        self.config.dataset_type == "grain"
+        and data_iterator is not None
+        and (self.checkpoint_manager.directory / str(self.checkpoint_manager.latest_step()) / "iter").exists()
+    ):
+      max_logging.log("Restoring data iterator from checkpoint")
+      restored = self.checkpoint_manager.restore(
+          self.checkpoint_manager.latest_step(),
+          args=ocp.args.Composite(iter=grain.PyGrainCheckpointRestore(data_iterator.local_iterator)),
+      )
+      data_iterator.local_iterator = restored["iter"]
+    else:
+      max_logging.log("data iterator checkpoint not found")
+    return data_iterator
+
   def _get_pipeline_class(self):
     if self.checkpoint_type == STABLE_DIFFUSION_CHECKPOINT:
       pipeline_class = FlaxStableDiffusionPipeline
@@ -212,7 +233,7 @@ def load_diffusers_checkpoint(self):
       params = jax.tree_util.tree_map(lambda x: x.astype(self.config.weights_dtype), params)
     return pipeline, params
 
-  def save_checkpoint(self, train_step, pipeline, params, train_states):
+  def save_checkpoint(self, train_step, pipeline, params, train_states, data_iterator=None):
     def config_to_json(model_or_config):
       return json.loads(model_or_config.to_json_string())
 
@@ -233,7 +254,8 @@ def config_to_json(model_or_config):
 
     tokenizer_config = {"path": self.config.tokenizer_model_name_or_path}
     items["tokenizer_config"] = ocp.args.JsonSave(tokenizer_config)
-
+    if self.config.dataset_type == "grain":
+      items["iter"] = grain.PyGrainCheckpointSave(data_iterator.local_iterator)
     self.checkpoint_manager.save(train_step, args=ocp.args.Composite(**items))
 
   def load_params(self, step=None):

src/maxdiffusion/checkpointing/checkpointing_utils.py

@@ -40,6 +40,7 @@ def create_orbax_checkpoint_manager(
     enable_checkpointing: bool,
     save_interval_steps,
     checkpoint_type: str,
+    dataset_type: str = "tf",
     use_async: bool = True,
     orbax_logger: Optional[abstract_logger.AbstractLogger] = None,
 ):
@@ -70,6 +71,8 @@ def create_orbax_checkpoint_manager(
         "text_encoder_2_state",
         "text_encoder_2_config",
     )
+  if dataset_type == "grain":
+    item_names += ("iter",)
 
   print("item_names: ", item_names)
 

src/maxdiffusion/trainers/base_stable_diffusion_trainer.py

@@ -128,6 +128,8 @@ def start_training(self):
 
     # Load dataset
     data_iterator = self.load_dataset(pipeline, params, train_states)
+    if self.config.dataset_type == "grain":
+      data_iterator = self.restore_data_iterator_state(data_iterator)
 
     data_shardings = self.get_data_shardings()
     # Compile train_step

src/maxdiffusion/trainers/stable_diffusion_trainer.py

@@ -15,6 +15,7 @@
  """
 
 import os
+import sys
 from functools import partial
 import datetime
 import time
@@ -211,7 +212,6 @@ def training_loop(self, p_train_step, pipeline, params, train_states, data_itera
         unet_state, text_encoder_state, train_metric, train_rngs = p_train_step(
             unet_state, vae_state, text_encoder_state, example_batch, train_rngs
         )
-      samples_count = self.total_train_batch_size * (step + 1)
       new_time = datetime.datetime.now()
 
       train_utils.record_scalar_metrics(
@@ -221,11 +221,15 @@ def training_loop(self, p_train_step, pipeline, params, train_states, data_itera
         train_utils.write_metrics(writer, local_metrics_file, running_gcs_metrics, train_metric, step, self.config)
       last_step_completion = new_time
 
-      if step != 0 and self.config.checkpoint_every != -1 and samples_count % self.config.checkpoint_every == 0:
+      if step != 0 and self.config.checkpoint_every != -1 and step % self.config.checkpoint_every == 0:
         train_states["unet_state"] = unet_state
         train_states["vae_state"] = vae_state
         train_states["text_encoder"] = text_encoder_state
-        self.save_checkpoint(step, pipeline, params, train_states)
+        self.save_checkpoint(step, pipeline, params, train_states, data_iterator)
+
+      if self.checkpoint_manager.reached_preemption(step):
+        self.checkpoint_manager.wait_until_finished()
+        sys.exit()
 
       if self.config.enable_profiler and step == last_profiling_step:
         max_utils.deactivate_profiler(self.config)
@@ -239,7 +243,7 @@ def training_loop(self, p_train_step, pipeline, params, train_states, data_itera
     train_states["vae_state"] = vae_state
     train_states["text_encoder"] = text_encoder_state
     # save the inference states of the last checkpoint so they can be easily loaded during gen.
-    self.save_checkpoint(self.config.max_train_steps - 1, pipeline, params, train_states)
+    self.save_checkpoint(self.config.max_train_steps - 1, pipeline, params, train_states, data_iterator)
     self.checkpoint_manager.wait_until_finished()