[BUG] The text generated by the hybrid engine does not meet expectations, the model is GPTNeoX

huggingface model config:
{
"activation_function": "gelu",
"architectures": [
"GPTNeoXForCausalLM"
],
"bos_token_id": 0,
"eos_token_id": 2,
"hidden_act": "gelu",
"hidden_size": 1536,
"initializer_range": 0.02,
"intermediate_size": 6144,
"layer_norm_eps": 1e-05,
"max_position_embeddings": 4096,
"model_type": "gpt_neox",
"num_attention_heads": 12,
"num_hidden_layers": 12,
"rotary_emb_base": 10000,
"rotary_pct": 1,
"tie_word_embeddings": false,
"transformers_version": "4.28.1",
"use_cache": true,
"use_parallel_residual": false,
"vocab_size": 30336
}

When hybrid_engine.enabled=False, the response is fine. The implementation of GPTNeoX Container in hybrid_engine looks inconsistent with huggingface.