phi-1_5模型导出来，推理不正常。

[aa03146]

原生的只支持 phi-2，理论上对 phi-1_5也是支持的，下面是我参考 phi-2的写了load_model函数，模型倒是能正常导出来，但好像是一些参数对不上，导出来 的模型推理也不正常，麻烦帮忙看下，非常感谢！这是模型地址：https://huggingface.co/microsoft/phi-1_5 下面是我写的load_model函数： def load_model(self): transformer = self.model.model
self.lm = self.model.lm_head
self.embed_ = transformer.embed_tokens self.hidden_size = self.embed_.weight.shape[-1] self.blocks_ = transformer.layers self.final_layernorm_ = transformer.final_layernorm
# Some wrapper self.stop_ids.append(self.tokenizer.eos_token_id) self.block_nums = len(self.blocks_) self.embed = Embedding(self.embed_, self.embed_bf16) self.lm = Lm(self.lm) self.blocks = [PHI2Block(self.blocks_[i], i, self.hidden_size) for i in range(self.block_nums)]

    # Some config for export
    self.past_kv_shape = [len(self.blocks), 1, 0, 2, 32, 80]
    self.block_dynamic_axes = {
        "inputs_embeds" : { 0: "seq_len" },
        "attention_mask" : { 2: "seq_len", 3: "seq_len" },
        "position_ids" : { 0: "seq_len" },
        "past_key_values" : { 1: "history_len" }
    }
    self.model_dynamic_axes = {
        "input_ids" : { 0: "seq_len" },
        "attention_mask" : { 2: "seq_len", 3: "seq_len" },
        "position_ids" : { 0: "seq_len" },
        "past_key_values" : { 2: "history_len" }
    }