Refactor: Recording and logging training and evaluation metrics in all trainers

Author: SurbhiJainUSC

MaxText/elastic_train.py

@@ -38,12 +38,10 @@
 for more details about the elastic manager.
 """
 from collections.abc import Sequence
-import datetime
 import logging
 import os
 import sys
 import time
-import queue
 
 from absl import app
 
@@ -68,13 +66,11 @@
 from MaxText import max_logging
 from MaxText import profiler
 from MaxText import pyconfig
-from MaxText.gcp_workload_monitor import GCPWorkloadMonitor
 from MaxText.input_pipeline.input_pipeline_interface import create_data_iterator
 from MaxText.metric_logger import MetricLogger
 from MaxText.train import check_example_batch
 from MaxText.train import get_first_step
 from MaxText.train import load_next_batch
-from MaxText.train import record_scalar_metrics
 from MaxText.train import save_checkpoint
 from MaxText.train import setup_mesh_and_model
 from MaxText.train import setup_train_loop
@@ -120,7 +116,7 @@ def elastic_handler(
     checkpoint_manager.close()
 
   with jax.default_device(elastic_manager.default_device):
-    init_rng, writer, checkpoint_manager, mesh, model, learning_rate_schedule, tx = setup_mesh_and_model(
+    init_rng, checkpoint_manager, mesh, model, learning_rate_schedule, tx = setup_mesh_and_model(
         config, elastic_manager.good_devices
     )
     with mesh:
@@ -171,7 +167,7 @@ def elastic_handler(
       )
 
       example_batch = None
-      metric_logger = MetricLogger(writer, config)
+      metric_logger = MetricLogger(config=config, learning_rate_schedule=learning_rate_schedule)
 
       jax.block_until_ready(state)
 
@@ -186,15 +182,13 @@ def elastic_handler(
       example_batch,
       learning_rate_schedule,
       metric_logger,
-      writer,
   )
 
 
 def train_loop(config, elastic_manager, recorder, state=None):
   """Main Training loop."""
   (
       init_rng,
-      writer,
       checkpoint_manager,
       state_mesh_shardings,
       model,
@@ -214,25 +208,13 @@ def train_loop(config, elastic_manager, recorder, state=None):
       donate_argnums_train,
   ) = maxtext_utils.get_functional_train_with_signature(train_step, mesh, state_mesh_shardings, model, config)
 
-  num_model_parameters = max_utils.calculate_num_params_from_pytree(state.params)
-  max_logging.log(f"number parameters: {num_model_parameters/1e9:.3f} billion")
-  per_device_tflops, _, _ = maxtext_utils.calculate_tflops_training_per_device(config)
-  per_device_tokens = maxtext_utils.calculate_tokens_training_per_device(config)
-
-  # Write train config params, num model params, and XLA flags to tensorboard
-  max_utils.add_text_to_summary_writer("num_model_parameters", str(num_model_parameters), writer)
-  max_utils.add_text_to_summary_writer("libtpu_init_args", os.environ["LIBTPU_INIT_ARGS"], writer)
-  maxtext_utils.add_config_to_summary_writer(config, writer)
-
   p_train_step = jax.jit(
       functional_train,
       in_shardings=in_shard_train,
       out_shardings=out_shard_train,
       static_argnums=static_argnums_train,
       donate_argnums=donate_argnums_train,
   )
-  running_gcs_metrics = [] if config.gcs_metrics else None
-  metrics = None
 
   start_step = get_first_step(state)  # this is the start_step for training
   prof = profiler.Profiler(config, offset_step=start_step)
@@ -242,18 +224,6 @@ def train_loop(config, elastic_manager, recorder, state=None):
   last_profiling_step = prof.finished_initial_profile_step
 
   example_batch = None
-  last_step_completion = datetime.datetime.now()
-
-  performance_metric_queue = None
-  if config.report_heartbeat_metric_for_gcp_monitoring or config.report_performance_metric_for_gcp_monitoring:
-    gcp_workload_monitor = GCPWorkloadMonitor(config.run_name)
-    if config.report_heartbeat_metric_for_gcp_monitoring:
-      gcp_workload_monitor.start_heartbeat_reporting_thread(config.heartbeat_reporting_interval_in_seconds)
-    if config.report_performance_metric_for_gcp_monitoring:
-      performance_metric_queue = queue.Queue()
-      gcp_workload_monitor.start_performance_reporting_thread(performance_metric_queue)
-
-  metric_logger = MetricLogger(writer, config)
   step = start_step
 
   elastic_manager.maybe_snapshot(
@@ -267,6 +237,12 @@ def train_loop(config, elastic_manager, recorder, state=None):
   )
 
   input_data_shardings = maxtext_utils.get_input_data_sharding(config, mesh)
+
+  metric_logger = MetricLogger(config=config, learning_rate_schedule=learning_rate_schedule)
+
+  # Write train config params, num model params, and XLA flags to tensorboard
+  metric_logger.write_setup_info_to_tensorboard(state.params)
+
   # Using while loop instead of a for loop because with elasticity
   # the step is restored back to the latest snapshot when a slice is lost
   while step < config.steps:
@@ -292,11 +268,7 @@ def train_loop(config, elastic_manager, recorder, state=None):
           with maybe_record_goodput(recorder, GoodputEvent.STEP, step):
             state, metrics = p_train_step(state, example_batch, nextrng)
 
-        step_time_delta = datetime.datetime.now() - last_step_completion
-        last_step_completion = datetime.datetime.now()
-        record_scalar_metrics(metrics, step_time_delta, per_device_tflops, learning_rate_schedule(step), per_device_tokens)
-        if performance_metric_queue:
-          performance_metric_queue.put(step_time_delta.total_seconds())
+        metric_logger.record_train_metrics(metrics, step)
 
         if checkpoint_manager is not None:
           state_to_save = state
@@ -308,8 +280,6 @@ def train_loop(config, elastic_manager, recorder, state=None):
             checkpoint_manager.wait_until_finished()
             sys.exit()
 
-        metric_logger.write_metrics(running_gcs_metrics, metrics, step)
-
         if step == last_profiling_step or prof.should_deactivate_periodic_profile(step):
           prof.deactivate(blocking_object=state)
 
@@ -347,7 +317,6 @@ def train_loop(config, elastic_manager, recorder, state=None):
             example_batch,
             learning_rate_schedule,
             metric_logger,
-            writer,
         ) = ret
 
       if step == start_step:
@@ -377,7 +346,6 @@ def train_loop(config, elastic_manager, recorder, state=None):
             example_batch,
             learning_rate_schedule,
             metric_logger,
-            writer,
         ) = ret
 
   if checkpoint_manager is not None:
@@ -398,8 +366,7 @@ def train_loop(config, elastic_manager, recorder, state=None):
         max_logging.log(f"Checkpoint is already saved for step {int(state.step)-1}.")
 
     checkpoint_manager.wait_until_finished()
-  metric_logger.write_metrics(running_gcs_metrics, metrics, config.steps - 1)  # final step metrics
-  max_utils.close_summary_writer(writer)
+  metric_logger.flush_metrics_and_cleanup()
 
   if example_batch:
     with mesh, nn_partitioning.axis_rules(config.logical_axis_rules):

MaxText/experimental/rl/grpo_trainer.py

@@ -60,13 +60,13 @@
 from MaxText.common_types import Array
 from MaxText.experimental.rl import grpo_input_pipeline
 from MaxText.gcp_workload_monitor import GCPWorkloadMonitor
+from MaxText.globals import EPS
 from MaxText.layers import models
 from MaxText.metric_logger import MetricLogger
 from MaxText.train import (
     validate_train_config,
     get_first_step,
     load_next_batch,
-    record_scalar_metrics,
     save_checkpoint,
     check_example_batch,
     setup_mesh_and_model,
@@ -82,7 +82,6 @@
 # pylint: disable=too-many-positional-arguments
 
 Transformer = models.Transformer
-EPS = 1e-8
 
 
 # -----------------------------------------------------------------------------
@@ -676,7 +675,6 @@ def train_loop(config, config_inference, recorder, state=None):
   """Main Training loop."""
   (
       init_rng,
-      writer,
       checkpoint_manager,
       state_mesh_shardings,
       model,
@@ -727,15 +725,10 @@ def train_loop(config, config_inference, recorder, state=None):
     ) = maxtext_utils.get_functional_eval_with_signature(eval_step, mesh, state_mesh_shardings, model, config)
 
   # TODO: fix tflops calculations for grpo setting
-  num_model_parameters = max_utils.calculate_num_params_from_pytree(state.params)
-  max_logging.log(f"number parameters: {num_model_parameters/1e9:.3f} billion")
-  per_device_tflops, _, _ = maxtext_utils.calculate_tflops_training_per_device(config)
-  per_device_tokens = maxtext_utils.calculate_tokens_training_per_device(config)
+  metric_logger = MetricLogger(config=config, learning_rate_schedule=learning_rate_schedule)
 
   # Write train config params, num model params, and XLA flags to tensorboard
-  max_utils.add_text_to_summary_writer("num_model_parameters", str(num_model_parameters), writer)
-  max_utils.add_text_to_summary_writer("libtpu_init_args", os.environ["LIBTPU_INIT_ARGS"], writer)
-  maxtext_utils.add_config_to_summary_writer(config, writer)
+  metric_logger.write_setup_info_to_tensorboard(state.params)
 
   # Define the compilation of functional_train, either by loading the compiled version or wrapping a new one in a jit
   if config.compiled_trainstep_file != "":
@@ -774,9 +767,6 @@ def train_loop(config, config_inference, recorder, state=None):
       donate_argnums=(0,),
   )
 
-  running_gcs_metrics = [] if config.gcs_metrics else None
-  metrics = None
-
   start_step = get_first_step(state)  # this is the start_step for training
   prof = profiler.Profiler(config, offset_step=start_step)
   first_profiling_step = prof.start_initial_profile_step
@@ -785,19 +775,14 @@ def train_loop(config, config_inference, recorder, state=None):
   last_profiling_step = prof.finished_initial_profile_step
 
   example_batch = None
-  last_step_completion = datetime.datetime.now()
-
-  performance_metric_queue = None
-  if config.report_heartbeat_metric_for_gcp_monitoring or config.report_performance_metric_for_gcp_monitoring:
-    gcp_workload_monitor = GCPWorkloadMonitor(config.run_name)
-    if config.report_heartbeat_metric_for_gcp_monitoring:
-      gcp_workload_monitor.start_heartbeat_reporting_thread(config.heartbeat_reporting_interval_in_seconds)
-    if config.report_performance_metric_for_gcp_monitoring:
-      performance_metric_queue = queue.Queue()
-      gcp_workload_monitor.start_performance_reporting_thread(performance_metric_queue)
-
-  metric_logger = MetricLogger(writer, config)
+
   input_data_shardings = maxtext_utils.get_input_data_sharding(config, mesh)
+
+  metric_logger = MetricLogger(config=config, learning_rate_schedule=learning_rate_schedule)
+
+  # Write train config params, num model params, and XLA flags to tensorboard
+  metric_logger.write_setup_info_to_tensorboard(state.params)
+
   for step in np.arange(start_step, config.steps):
     if step == first_profiling_step or prof.should_activate_periodic_profile(step):
       optional_postfix = f"step_{step}" if config.profile_periodically_period > 0 else ""
@@ -823,11 +808,7 @@ def train_loop(config, config_inference, recorder, state=None):
         with mesh, nn_partitioning.axis_rules(config.logical_axis_rules):
           state, metrics = p_train_step(state, example_batch, rng)
 
-    step_time_delta = datetime.datetime.now() - last_step_completion
-    last_step_completion = datetime.datetime.now()
-    record_scalar_metrics(metrics, step_time_delta, per_device_tflops, learning_rate_schedule(step), per_device_tokens)
-    if performance_metric_queue:
-      performance_metric_queue.put(step_time_delta.total_seconds())
+    metric_logger.record_train_metrics(metrics, step)
 
     if checkpoint_manager is not None:
       state_to_save = state if not config.use_dpo else _split_grpo_state(state)[0]
@@ -839,8 +820,6 @@ def train_loop(config, config_inference, recorder, state=None):
         checkpoint_manager.wait_until_finished()
         sys.exit()
 
-    metric_logger.write_metrics(running_gcs_metrics, metrics, step)
-
     if config.dump_hlo and step == start_step:
       jax.block_until_ready(state)  # Ensure compilation has finished.
       max_utils.upload_dump(
@@ -853,43 +832,18 @@ def train_loop(config, config_inference, recorder, state=None):
 
     if config.eval_interval > 0 and step > start_step and (step + 1) % config.eval_interval == 0:
       assert eval_data_iterator
-      cumulative_eval_metrics = {
-          "scalar": {
-              "eval/total_loss": 0.0,
-              "eval/total_weights": 0.0,
-              "eval/avg_loss": 0.0,
-              "eval/moe_lb_loss": 0.0,
-          }
-      }
-      eval_dpo_reward_accuracy = 0.0
       eval_step_count = 0
       # pylint: disable=not-callable
       for eval_batch in eval_data_iterator:
         if config.eval_steps > 0 and eval_step_count >= config.eval_steps:
           break
         with mesh, nn_partitioning.axis_rules(config.logical_axis_rules):
           eval_metrics = p_eval_step(state, eval_batch, rng)
-        cumulative_eval_metrics["scalar"]["eval/total_loss"] += float(eval_metrics["scalar"]["evaluation/total_loss"])
-        cumulative_eval_metrics["scalar"]["eval/total_weights"] += float(eval_metrics["scalar"]["evaluation/total_weights"])
-        cumulative_eval_metrics["scalar"]["eval/moe_lb_loss"] += float(eval_metrics["scalar"]["evaluation/moe_lb_loss"])
-        eval_dpo_reward_accuracy += float(eval_metrics["scalar"].get("evaluation/dpo_reward_accuracy", 0.0))  # for dpo only
+        metric_logger.record_eval_metrics(metrics=eval_metrics, eval_step_count=None, step=step)
         max_logging.log(f"Completed eval step {eval_step_count}")
         eval_step_count += 1
-      eval_loss = cumulative_eval_metrics["scalar"]["eval/total_loss"] / (
-          cumulative_eval_metrics["scalar"]["eval/total_weights"] + EPS
-      )
-      cumulative_eval_metrics["scalar"]["eval/avg_loss"] = eval_loss
-      cumulative_eval_metrics["scalar"]["eval/avg_moe_lb_loss"] = (
-          cumulative_eval_metrics["scalar"]["eval/moe_lb_loss"] / eval_step_count
-      )
-      if config.use_dpo:
-        cumulative_eval_metrics["scalar"]["eval/dpo_reward_accuracy"] = eval_dpo_reward_accuracy / eval_step_count
-      metric_logger.write_metrics(running_gcs_metrics, cumulative_eval_metrics, step, is_training=False)
-      max_logging.log(
-          f"average loss after {step=}: {eval_step_count=}, {eval_loss=},"
-          f" total_weights={cumulative_eval_metrics['scalar']['eval/total_weights']}"
-      )
-      if eval_loss <= config.target_eval_loss:
+      metric_logger.record_eval_metrics(metrics=None, eval_step_count=eval_step_count, step=step)
+      if metric_logger.cumulative_eval_metrics["scalar"]["eval/avg_loss"] <= config.target_eval_loss:
         max_logging.log(f"Early stop and exit loop after reaching {config.target_eval_loss=}")
         prof.deactivate()
         break
@@ -911,8 +865,8 @@ def train_loop(config, config_inference, recorder, state=None):
         max_logging.log(f"Checkpoint already saved for step {int(state.step)-1}.")
 
     checkpoint_manager.wait_until_finished()
-  metric_logger.write_metrics(running_gcs_metrics, metrics, config.steps - 1)  # final step metrics
-  max_utils.close_summary_writer(writer)
+  metric_logger.flush_metrics_and_cleanup()
+
   if example_batch:
     with mesh, nn_partitioning.axis_rules(config.logical_axis_rules):
       compiled = p_train_step.lower(state, example_batch, rng).compile()

MaxText/globals.py

@@ -16,6 +16,8 @@
 
 import os.path
 
-PKG_DIR = os.path.dirname(os.path.abspath(__file__))
+PKG_DIR = os.path.dirname(os.path.abspath(__file__))  # MaxText directory path
+EPS = 1e-8  # Epsilon to calculate loss
+DEFAULT_OCDBT_TARGET_DATA_FILE_SIZE = 2 * 1024**3  # Default checkpoint file size
 
-__all__ = ["PKG_DIR"]
+__all__ = ["DEFAULT_OCDBT_TARGET_DATA_FILE_SIZE", "EPS", "PKG_DIR"]