revert +1 in steps logging

Elnifio · Elnifio · commit 253e440867e7 · 2025-03-10T14:25:41.000-07:00
diff --git a/large_language_model_pretraining/nemo/callbacks.py b/large_language_model_pretraining/nemo/callbacks.py
@@ -148,7 +148,7 @@ def version(self):
 ### MLPerf callbacks
 def compute_consumed_mllog_tokens(trainer, init_global_step, global_batch_size, seq_length):
     consumed_samples = (
-        (trainer.global_step + 1) * global_batch_size # global steps are 0-indexed
+        trainer.global_step * global_batch_size
     )
     return int(consumed_samples) # we log the epoch numbers in sequences, not tokens
 

Original file line number	Diff line number	Diff line change
`@@ -148,7 +148,7 @@ def version(self):`
`148`	`148`	`### MLPerf callbacks`
`149`	`149`	`def compute_consumed_mllog_tokens(trainer, init_global_step, global_batch_size, seq_length):`
`150`	`150`	`consumed_samples = (`
`151`		`- (trainer.global_step + 1) * global_batch_size # global steps are 0-indexed`
	`151`	`+ trainer.global_step * global_batch_size`
`152`	`152`	`)`
`153`	`153`	`return int(consumed_samples) # we log the epoch numbers in sequences, not tokens`
`154`	`154`