[TorchTrain] Minor fix for pytorch#197 (pytorch#204)

The changes made in github editor didn't go in when doing ghstack land.

torchtrain/config_manager.py

@@ -200,10 +200,10 @@ def __init__(self):
         self.parser.add_argument(
             "--checkpoint.interval",
             type=int,
-            default=3600,
+            default=500,
             help=(
                 "Checkpointing interval. The unit of measurement is in seconds or "
-                "steps depending on --checkpoint.interval."
+                "steps depending on --checkpoint.interval_type."
             ),
         )
         self.parser.add_argument(

train_configs/debug_model.toml

@@ -38,6 +38,7 @@ compile = false
 dataset = "alpaca"   # supported datasets = alpaca (52K), minipile (1M), c4 (177M)
 
 [checkpoint]
+interval = 500
 interval_type = "steps"
 folder = ""
 

train_configs/llama_13b.toml

@@ -38,6 +38,7 @@ compile = false
 dataset = "openwebtext"
 
 [checkpoint]
+interval = 500
 interval_type = "steps"
 folder = ""
 

train_configs/llama_70b.toml

@@ -38,6 +38,7 @@ compile = false
 dataset = "openwebtext"
 
 [checkpoint]
+interval = 500
 interval_type = "steps"
 folder = ""
 

train_configs/llama_7b.toml

@@ -38,5 +38,6 @@ compile = false
 dataset = "alpaca"
 
 [checkpoint]
+interval = 500
 interval_type = "steps"
 folder = ""