[CICD] Update aquila tp2-pp2 ci for evaluation (flagos-ai#1110)

### PR Category
CICD

### PR Types
Improvements

### PR Description
Update aquila tp2_pp2 ci for evaluation, to detect and validate the
evaluation-related functionalities of FlagScale.

Author: lxd-cumt

flagscale/train/megatron/training/training.py

@@ -2931,6 +2931,7 @@ def evaluate(
     verbose=False,
     non_loss_data_func=None,
     extra_valid_index=None,
+    eval_iters=None,
 ):
     """Evaluation."""
     args = get_args()

tests/functional_tests/train/aquila/conf/train/tp2_pp2.yaml

@@ -1,6 +1,3 @@
-defaults:
-  - data
-
 system:
   tensor_model_parallel_size: 2
   pipeline_model_parallel_size: 2
@@ -45,6 +42,8 @@ model:
   weight_decay: 0.1
   clip_grad: 1.0
   train_iters: 10
+  eval_iters: 1
+  eval_interval: 5
   micro_batch_size: 4
   global_batch_size: 1024
   seed: 42
@@ -59,3 +58,14 @@ model:
       lr_warmup_samples: 0
       lr_warmup_fraction: 0.01
       lr_decay_style: cosine
+
+data:
+  data_path: /home/gitlab-runner/data/pile_wikipedia_demo/pile_wikipedia_demo
+  split: 80, 10, 10
+  tokenizer:
+    legacy_tokenizer: true
+    tokenizer_type: AquilaTokenizerFS
+    vocab_file: ./examples/aquila/tokenizer/vocab.json
+    merge_file: ./examples/aquila/tokenizer/merges.txt
+    special_tokens_file: ./examples/aquila/tokenizer/special_tokens.txt
+    vocab_size: 100008

tests/functional_tests/train/aquila/gold_values/tp2_pp2.json

@@ -1 +1 @@
-{"lm loss:": {"values": [11.63682, 11.63587, 11.49153, 11.38188, 11.28651, 11.21933, 11.17029, 11.12899, 11.10917, 11.08776]}}
+{"lm loss:": {"values": [11.63839, 11.63632, 11.49138, 11.38161, 11.28878, 11.21583, 74297.01, 11.22827, 11.16404, 11.1232, 11.10253, 11.07966, 11.07366, 64451.08, 11.07513, 64546.05, 11.07297, 64406.66]}}