merge

Author: priyakasimbeg

algoperf/pytorch_utils.py

@@ -27,7 +27,6 @@ def pytorch_setup() -> Tuple[bool, int, torch.device, int]:
   # torch.backends.cudnn.fp32_precision = "ieee"
   # torch.backends.cudnn.conv.fp32_precision = "tf32"
   # torch.backends.cudnn.rnn.fp32_precision = "tf32"
-  
 
   use_pytorch_ddp = 'LOCAL_RANK' in os.environ
   rank = int(os.environ['LOCAL_RANK']) if use_pytorch_ddp else 0

algoperf/workloads/finewebedu_lm/workload.py

@@ -85,11 +85,11 @@ def train_stddev(self):
 
   @property
   def max_allowed_runtime_sec(self) -> int:
-    return 31_967 # 8.9 hours
+    return 31_967  # 8.9 hours
 
   @property
   def eval_period_time_sec(self) -> int:
-    return 2_571 # approximately 25 evals
+    return 2_571  # approximately 25 evals
 
   @property
   def step_hint(self) -> int:

scoring/score_submissions.py

@@ -70,7 +70,7 @@
 flags.DEFINE_string(
   'include_submissions',
   '',
-  'Optional comma seperated list of names of submissions to include from scoring.'
+  'Optional comma seperated list of names of submissions to include from scoring.',
 )
 FLAGS = flags.FLAGS
 
@@ -128,12 +128,14 @@ def get_summary_df(workload, workload_df, include_test_split=False):
   # compute the step times
   def delta(series):
     return series.shift(1, fill_value=0) - series
-  accumulated_time_intervals = delta(workload_df['accumulated_submission_time'])  # exclude first step
-  step_intervals = delta(workload_df['global_step']) # exclude time up to first step
+  accumulated_time_intervals = delta(workload_df['accumulated_submission_time'])
+  step_intervals = delta(workload_df['global_step'])
   if len(accumulated_time_intervals) < 2:
     print(f"WARNING: The number of evals may be too low to calculate reliable step time for {workload}")
 
   summary_df['step_time (s)'] = np.median((accumulated_time_intervals / step_intervals).iloc[0])
+
+
   summary_df['step_hint'] = scoring_utils.get_workload_stephint(workload)
 
   # test metrics
@@ -224,7 +226,9 @@ def main(_):
 
     for submission in all_submission_dirs:
       print(submission)
-      if submission not in FLAGS.exclude_submissions.split(',') and (submission in include_submissions):
+      if submission not in FLAGS.exclude_submissions.split(',') and (
+        submission in include_submissions
+      ):
         experiment_path = os.path.join(FLAGS.submission_directory, submission)
         df = scoring_utils.get_experiment_df(experiment_path)
         results[submission] = df