FlashAttention support

config.json

@@ -12,38 +12,27 @@
       "AutoModelForQuestionAnswering": "modeling_gptbert.GptBertForQuestionAnswering",
       "AutoModelForMultipleChoice": "modeling_gptbert.GptBertForMultipleChoice"
   },
-  "attention_dropout": 0.0,
-  "attention_output_dropout_p": 0.0,
-  "attention_inter_norm_affine": false,
-  "attention_inter_norm_eps": 1e-07,
-  "attention_pre_norm_affine": false,
-  "attention_pre_norm_eps": 1e-07,
-  "attention_probabilities_dropout_p": 0.0,
-  "classifier_post_norm_affine": false,
-  "classifier_post_norm_eps": 1e-07,
-  "classifier_pre_norm_affine": false,
-  "classifier_pre_norm_eps": 1e-07,
-  "d_qk": 64,
-  "d_v": 64,
-  "embedding_dropout_p": 0.1,
-  "feed_forward_dropout_p": 0.0,
-  "feed_forward_inter_norm_affine": false,
-  "feed_forward_inter_norm_eps": 1e-07,
-  "feed_forward_pre_norm_affine": false,
-  "feed_forward_pre_norm_eps": 1e-07,
+    "unk_token_id": 0,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 3,
+  "mask_token_id": 4,
   "hidden_size": 192,
   "intermediate_size": 512,
   "max_sequence_length": 16384,
-  "num_attention_heads": 3,
-  "num_kv_heads": 3,
   "num_layers": 16,
+  "attention_dropout": 0.0,
+  "hidden_dropout": 0.0,
+  "embedding_dropout": 0.1,
+  "classifier_dropout": 0.2,
+  "layer_norm_eps": 1e-07,
+  "query_key_head_size": 64,
+  "value_head_size": 64,
+  "num_attention_heads": 3,
   "rope_theta": 160000,
   "vocab_size": 51200,
-  "word_norm_affine": true,
-  "word_norm_eps": 1e-07,
-  "short_long_ratio": 4,
-  "window_length": 8192,
-  "is_decoder": false,
-  "not_flex": true,
-  "hidden_dropout_prob": 0.2
-}
+  "local_global_ratio": 4,
+  "global_window_length": 8192,
+  "local_window_length": 256,
+  "deterministic_flash_attn": false
+}