update all toml files to use experimental section (#392)

Author: wanchaol

train_configs/llama2_13b.toml

@@ -33,11 +33,13 @@ max_norm = 1.0  # grad norm clipping
 steps = 1000
 data_parallel_degree = -1
 tensor_parallel_degree = 1
-pipeline_parallel_degree = 1
 fp8_linear = ""
 compile = false
 dataset = "c4"
 
+[experimental]
+pipeline_parallel_degree = 1
+
 [checkpoint]
 enable_checkpoint = false
 folder = "checkpoint"

train_configs/llama2_70b.toml

@@ -33,11 +33,13 @@ max_norm = 1.0  # grad norm clipping
 steps = 1000
 data_parallel_degree = -1
 tensor_parallel_degree = 8  # 8-way TP
-pipeline_parallel_degree = 1
 fp8_linear = ""
 compile = false
 dataset = "c4"
 
+[experimental]
+pipeline_parallel_degree = 1
+
 [checkpoint]
 enable_checkpoint = false
 folder = "checkpoint"

train_configs/llama2_7b.toml

@@ -32,11 +32,13 @@ max_norm = 1.0  # grad norm clipping
 steps = 1000
 data_parallel_degree = -1
 tensor_parallel_degree = 1  # dp-only would be sufficient for 7B
-pipeline_parallel_degree = 1
 fp8_linear = ""
 compile = false
 dataset = "c4"
 
+[experimental]
+pipeline_parallel_degree = 1
+
 [checkpoint]
 enable_checkpoint = false
 folder = "checkpoint"

train_configs/llama3_70b.toml

@@ -33,11 +33,13 @@ max_norm = 1.0  # grad norm clipping
 steps = 1000
 data_parallel_degree = -1
 tensor_parallel_degree = 8  # 8-way TP
-pipeline_parallel_degree = 1
 fp8_linear = ""
 compile = false
 dataset = "c4"
 
+[experimental]
+pipeline_parallel_degree = 1
+
 [checkpoint]
 enable_checkpoint = false
 folder = "checkpoint"

train_configs/llama3_8b.toml

@@ -33,11 +33,13 @@ max_norm = 1.0  # grad norm clipping
 steps = 1000
 data_parallel_degree = -1
 tensor_parallel_degree = 1
-pipeline_parallel_degree = 1
 fp8_linear = ""
 compile = false
 dataset = "c4"
 
+[experimental]
+pipeline_parallel_degree = 1
+
 [checkpoint]
 enable_checkpoint = false
 folder = "checkpoint"