[fix] Add working A100 config for deberta-xxlarge (deepspeed stuck during evaluation, deadlock?)

model/reward/instructor/configs/deberta-v2-xxlarge-a100.yaml

@@ -0,0 +1,17 @@
+model_name: microsoft/deberta-v2-xxlarge
+learning_rate: 2e-6
+scheduler: cosine
+gradient_checkpointing: false
+gradient_accumulation_steps: 12
+per_device_train_batch_size: 2
+per_device_eval_batch_size: 4
+warmup_steps: 600
+eval_steps: 1000000
+save_steps: 1000
+max_length: 400
+num_train_epochs: 3
+datasets:
+  - webgpt
+  - hfsummary
+  - anthropic_rlhf
+  - gptsynthetic

model/reward/instructor/utils.py

@@ -103,6 +103,7 @@ def argument_parsing(parser):
         "gradient_accumulation_steps",
         "num_train_epochs",
         "save_steps",
+        "eval_steps",
         "per_device_train_batch_size",
         "per_device_eval_batch_size",
     ]:
@@ -142,11 +143,3 @@ def get_datasets(dataset_list: List[AnyStr], tokenizer):
             evals["anthropic_rlhf"] = eval
     train = ConcatDataset(train_datasets)
     return train, evals
-
-
-if __name__ == "__main__":
-    from transformers import AutoModelForSequenceClassification
-
-    model = AutoModelForSequenceClassification.from_pretrained("bigscience/bloomz-560m")
-    freeze_top_n_layers(model, 10)
-    print(model.state_dict().keys())