1.0

README.md

@@ -102,10 +102,73 @@ ChatGLM-6B 是清华大学开源的开源的、支持中英双语的对话语言
 运行如下命令：
 
     python cli_demo.py
-    
+
 程序会在命令行中进行交互式的对话，在命令行中输入指示并回车即可生成回复，输入 clear 可以清空对话历史，输入 stop 终止程序。
 
+## 重新 pretrain
+
+由于当前的 [GLM-130B](https://github.com/THUDM/GLM-130B#news) 与 ChatGLM 的模型结构非常类似，所以对于有训练 GLM-130B 的用户来说，可以通过修改 ChatGLM 的 config.json 使用堆参数的方式将参数量达到130B。该项目为了满足用户对 ChatGLM 重新 pretrain 的需求，继续添加了 simple-pretrain 目录，旨在提供一种改动最小的 pretrain 示例。pretrain步骤如下：
+
+1. 将 simple-pretrain/ptuning 下的文件移到本 ptuning 目录下，替换相关文件
+
+2. 将 modeling_chatglm.py 移到[ ChatGLM 模型](https://huggingface.co/THUDM/chatglm-6b)所在目录替换原始 modeling_chatglm.py 
+
+3. 在本 ptuning 目录下：
+
+   ```
+   bash ds_pretrain.sh
+   ```
+
+说明：convert.py 可以将原始的txt数据转换成 chatglm 可用的 json 形式的数据集格式。该示例使用指环王1书籍作为预训练数据集。
+
+### 实验设置
+
+```
+LR=1e-5
+
+MASTER_PORT=$(shuf -n 1 -i 10000-65535)
+
+HIP_VISIBLE_DEVICES=0,1,2,3 deepspeed --num_gpus=4 --master_port $MASTER_PORT main.py \
+    --deepspeed deepspeed.json \
+    --do_train \
+    --train_file The-Lord-of-the-Rings-1.json \
+    --prompt_column prompt \
+    --response_column response \
+    --overwrite_cache \
+    --model_name_or_path THUDM/chatglm-6b \
+    --output_dir ./output/pretrain \
+    --overwrite_output_dir \
+    --max_source_length 8 \
+    --max_target_length 128 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 4 \
+    --predict_with_generate \
+    --max_steps 2000 \
+    --logging_steps 5 \
+    --save_steps 1000 \
+    --learning_rate $LR \
+    --fp16
+```
+
+### 训练loss收敛情况
+
+由于该示例预训练数据集较小，loss会降的至较低水平到0.1左右。
+
+![img](http://hpczentao.sugon.com/file-read-106608.jpeg)
+
+## 强化学习(RLHF)微调方案
+
+目前在 DCU 上 ChatGLM 使用强化学习微调有两种方案：
+
+- 使用 Lora，只更新低秩适配层，可以直接参考项目：https://github.com/hiyouga/ChatGLM-Efficient-Tuning/blob/main/examples/covid_doctor.md
+- 使用 DeepSpeed-Chat 方案全参微调，目前已经适配完成，欢迎尝试：https://github.com/yuguo-Jack/ChatGLM-6B-in-DeepSpeed-Chat
+
+## 源码仓库及问题反馈
+
+https://developer.hpccube.com/codes/modelzoo/chatglm
+
 ## 参考
-[THUDM/ChatGLM-6B](https://github.com/THUDM/ChatGLM-6B/tree/main)
 
+[THUDM/ChatGLM-6B](https://github.com/THUDM/ChatGLM-6B/tree/main)
 

model.properties

+# 模型名称
+modelName=ChatGLM-6B
+# 模型描述
+modelDescription=基于Pytorch框架的ChatGLM-6B
+# 应用场景(多个标签以英文逗号分割)
+appScenario=训练,推理,train,inference,nlp,智能聊天助手
+# 框架类型(多个标签以英文逗号分割)
+frameType=Pytorch

simple-pretrain/convert.py

+import json
+
+def convert_txt_to_json(txt_file):
+    json_data = []
+    with open(txt_file, 'r', encoding='gb18030') as file:
+        for line in file:
+            line = line.strip()
+            if line:
+                prompt = line[0]
+                response = line[1:]
+                json_entry = {'prompt': prompt, 'response': response}
+                json_data.append(json_entry)
+    
+    return json_data
+
+txt_file = 'The-Lord-of-the-Rings-1.txt'  # 替换成你的文本文件路径
+json_data = convert_txt_to_json(txt_file)
+
+# 将JSON数据写入文件
+json_file = 'The-Lord-of-the-Rings-1.json'  # 输出的JSON文件路径
+with open(json_file, 'w', encoding='utf-8') as file:
+    for entry in json_data:
+        json.dump(entry, file, ensure_ascii=False)
+        file.write('\n')
+

simple-pretrain/modeling_chatglm.py

+""" PyTorch ChatGLM model. """
+
+import math
+import copy
+import os
+import warnings
+import re
+import sys
+
+import torch
+import torch.utils.checkpoint
+import torch.nn.functional as F
+from torch import nn
+from torch.nn import CrossEntropyLoss, LayerNorm
+from torch.nn.utils import skip_init
+from typing import Optional, Tuple, Union, List, Callable, Dict, Any
+
+from transformers.utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+)
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+    BaseModelOutputWithPastAndCrossAttentions,
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import logging
+from transformers.generation.logits_process import LogitsProcessor
+from transformers.generation.utils import LogitsProcessorList, StoppingCriteriaList, GenerationConfig, ModelOutput
+
+from .configuration_chatglm import ChatGLMConfig
+
+# flags required to enable jit fusion kernels
+
+if sys.platform != 'darwin':
+    torch._C._jit_set_profiling_mode(False)
+    torch._C._jit_set_profiling_executor(False)
+    torch._C._jit_override_can_fuse_on_cpu(True)
+    torch._C._jit_override_can_fuse_on_gpu(True)
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "THUDM/ChatGLM-6B"
+_CONFIG_FOR_DOC = "ChatGLM6BConfig"
+
+CHATGLM_6B_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "THUDM/chatglm-6b",
+    # See all ChatGLM-6B models at https://huggingface.co/models?filter=chatglm
+]
+
+
+class InvalidScoreLogitsProcessor(LogitsProcessor):
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        if torch.isnan(scores).any() or torch.isinf(scores).any():
+            scores.zero_()
+            scores[..., 5] = 5e4
+        return scores
+
+
+def load_tf_weights_in_chatglm_6b(model, config, tf_checkpoint_path):
+    """Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    arrays = []
+    for name, shape in init_vars:
+        logger.info(f"Loading TF weight {name} with shape {shape}")
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        arrays.append(array)
+
+    for name, array in zip(names, arrays):
+        name = name.split("/")
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(
+                n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"]
+                for n in name
+        ):
+            logger.info(f"Skipping {'/'.join(name)}")
+            continue
+        pointer = model
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            if scope_names[0] == "kernel" or scope_names[0] == "gamma":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "output_weights":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "squad":
+                pointer = getattr(pointer, "classifier")
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    logger.info(f"Skipping {'/'.join(name)}")
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+        if m_name[-11:] == "_embeddings":
+            pointer = getattr(pointer, "weight")
+        elif m_name == "kernel":
+            array = np.transpose(array)
+        try:
+            assert (
+                    pointer.shape == array.shape
+            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info(f"Initialize PyTorch weight {name}")
+        pointer.data = torch.from_numpy(array)
+    return model
+
+
+class PrefixEncoder(torch.nn.Module):
+    """
+    The torch.nn model to encode the prefix
+    Input shape: (batch-size, prefix-length)
+    Output shape: (batch-size, prefix-length, 2*layers*hidden)
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.prefix_projection = config.prefix_projection
+        if self.prefix_projection:
+            # Use a two-layer MLP to encode the prefix
+            self.embedding = torch.nn.Embedding(config.pre_seq_len, config.hidden_size)
+            self.trans = torch.nn.Sequential(
+                torch.nn.Linear(config.hidden_size, config.hidden_size),
+                torch.nn.Tanh(),
+                torch.nn.Linear(config.hidden_size, config.num_layers * config.hidden_size * 2)
+            )
+        else:
+            self.embedding = torch.nn.Embedding(config.pre_seq_len, config.num_layers * config.hidden_size * 2)
+
+    def forward(self, prefix: torch.Tensor):
+        if self.prefix_projection:
+            prefix_tokens = self.embedding(prefix)
+            past_key_values = self.trans(prefix_tokens)
+        else:
+            past_key_values = self.embedding(prefix)
+        return past_key_values
+
+
+@torch.jit.script
+def gelu_impl(x):
+    """OpenAI's gelu implementation."""
+    return 0.5 * x * (1.0 + torch.tanh(0.7978845608028654 * x *
+                                       (1.0 + 0.044715 * x * x)))
+
+
+def gelu(x):
+    return gelu_impl(x)
+
+
+class RotaryEmbedding(torch.nn.Module):
+    def __init__(self, dim, base=10000, precision=torch.half, learnable=False):
+        super().__init__()
+        inv_freq = 1. / (base ** (torch.arange(0, dim, 2).float() / dim))
+        inv_freq = inv_freq.half()
+        self.learnable = learnable
+        if learnable:
+            self.inv_freq = torch.nn.Parameter(inv_freq)
+            self.max_seq_len_cached = None
+        else:
+            self.register_buffer('inv_freq', inv_freq)
+            self.max_seq_len_cached = None
+            self.cos_cached = None
+            self.sin_cached = None
+        self.precision = precision
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys,
+                              error_msgs):
+        pass
+
+    def forward(self, x, seq_dim=1, seq_len=None):
+        if seq_len is None:
+            seq_len = x.shape[seq_dim]
+        if self.max_seq_len_cached is None or (seq_len > self.max_seq_len_cached):
+            self.max_seq_len_cached = None if self.learnable else seq_len
+            t = torch.arange(seq_len, device=x.device, dtype=self.inv_freq.dtype)
+            freqs = torch.einsum('i,j->ij', t, self.inv_freq)
+            # Different from paper, but it uses a different permutation in order to obtain the same calculation
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            if self.precision == torch.bfloat16:
+                emb = emb.float()
+
+            # [sx, 1 (b * np), hn]
+            cos_cached = emb.cos()[:, None, :]
+            sin_cached = emb.sin()[:, None, :]
+            if self.precision == torch.bfloat16:
+                cos_cached = cos_cached.bfloat16()
+                sin_cached = sin_cached.bfloat16()
+            if self.learnable:
+                return cos_cached, sin_cached
+            self.cos_cached, self.sin_cached = cos_cached, sin_cached
+        return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...]
+
+    def _apply(self, fn):
+        if self.cos_cached is not None:
+            self.cos_cached = fn(self.cos_cached)
+        if self.sin_cached is not None:
+            self.sin_cached = fn(self.sin_cached)
+        return super()._apply(fn)
+
+
+def rotate_half(x):
+    x1, x2 = x[..., :x.shape[-1] // 2], x[..., x.shape[-1] // 2:]
+    return torch.cat((-x2, x1), dim=x1.ndim - 1)  # dim=-1 triggers a bug in earlier torch versions
+
+
+@torch.jit.script
+def apply_rotary_pos_emb_index(q, k, cos, sin, position_id):
+    # position_id: [sq, b], q, k: [sq, b, np, hn], cos: [sq, 1, hn] -> [sq, b, 1, hn]
+    cos, sin = F.embedding(position_id, cos.squeeze(1)).unsqueeze(2), \
+        F.embedding(position_id, sin.squeeze(1)).unsqueeze(2)
+    q, k = (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)
+    return q, k
+
+
+def attention_fn(
+        self,
+        query_layer,
+        key_layer,
+        value_layer,
+        attention_mask,
+        hidden_size_per_partition,
+        layer_id,
+        layer_past=None,
+        scaling_attention_score=True,
+        use_cache=False,
+):
+    if layer_past is not None:
+        past_key, past_value = layer_past[0], layer_past[1]
+        key_layer = torch.cat((past_key, key_layer), dim=0)
+        value_layer = torch.cat((past_value, value_layer), dim=0)
+
+    # seqlen, batch, num_attention_heads, hidden_size_per_attention_head
+    seq_len, b, nh, hidden_size = key_layer.shape
+
+    if use_cache:
+        present = (key_layer, value_layer)
+    else:
+        present = None
+
+    query_key_layer_scaling_coeff = float(layer_id + 1)
+    if scaling_attention_score:
+        query_layer = query_layer / (math.sqrt(hidden_size) * query_key_layer_scaling_coeff)
+
+    # ===================================
+    # Raw attention scores. [b, np, s, s]
+    # ===================================
+
+    # [b, np, sq, sk]
+    output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0))
+
+    # [sq, b, np, hn] -> [sq, b * np, hn]
+    query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1)
+    # [sk, b, np, hn] -> [sk, b * np, hn]
+    key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1)
+
+    matmul_result = torch.zeros(
+        1, 1, 1,
+        dtype=query_layer.dtype,
+        device=query_layer.device,
+    )
+
+    matmul_result = torch.baddbmm(
+        matmul_result,
+        query_layer.transpose(0, 1),  # [b * np, sq, hn]
+        key_layer.transpose(0, 1).transpose(1, 2),  # [b * np, hn, sk]
+        beta=0.0,
+        alpha=1.0,
+    )
+
+    # change view to [b, np, sq, sk]
+    attention_scores = matmul_result.view(*output_size)
+
+    if self.scale_mask_softmax:
+        self.scale_mask_softmax.scale = query_key_layer_scaling_coeff
+        attention_probs = self.scale_mask_softmax(attention_scores, attention_mask.contiguous())
+    else:
+        if not (attention_mask == 0).all():
+            # if auto-regressive, skip
+            attention_scores.masked_fill_(attention_mask, -10000.0)
+        dtype = attention_scores.dtype
+        attention_scores = attention_scores.float()
+        attention_scores = attention_scores * query_key_layer_scaling_coeff
+
+        attention_probs = F.softmax(attention_scores, dim=-1)
+
+        attention_probs = attention_probs.type(dtype)
+
+    # =========================
+    # Context layer. [sq, b, hp]
+    # =========================
+
+    # value_layer -> context layer.
+    # [sk, b, np, hn] --> [b, np, sq, hn]
+
+    # context layer shape: [b, np, sq, hn]
+    output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3))
+
+    # change view [sk, b * np, hn]
+    value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1)
+
+    # change view [b * np, sq, sk]
+    attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1)
+
+    # matmul: [b * np, sq, hn]
+    context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1))
+
+    # change view [b, np, sq, hn]
+    context_layer = context_layer.view(*output_size)
+
+    # [b, np, sq, hn] --> [sq, b, np, hn]
+    context_layer = context_layer.permute(2, 0, 1, 3).contiguous()
+
+    # [sq, b, np, hn] --> [sq, b, hp]
+    new_context_layer_shape = context_layer.size()[:-2] + (hidden_size_per_partition,)
+    context_layer = context_layer.view(*new_context_layer_shape)
+
+    outputs = (context_layer, present, attention_probs)
+
+    return outputs
+
+
+def default_init(cls, *args, **kwargs):
+    return cls(*args, **kwargs)
+
+
+class SelfAttention(torch.nn.Module):
+    def __init__(self, hidden_size, num_attention_heads,
+                 layer_id, hidden_size_per_attention_head=None, bias=True,
+                 params_dtype=torch.float, position_encoding_2d=True, empty_init=True):
+        if empty_init:
+            init_method = skip_init
+        else:
+            init_method = default_init
+        super(SelfAttention, self).__init__()
+
+        self.layer_id = layer_id
+        self.hidden_size = hidden_size
+        self.hidden_size_per_partition = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.num_attention_heads_per_partition = num_attention_heads
+        self.position_encoding_2d = position_encoding_2d
+        self.rotary_emb = RotaryEmbedding(
+            self.hidden_size // (self.num_attention_heads * 2)
+            if position_encoding_2d
+            else self.hidden_size // self.num_attention_heads,
+            base=10000,
+            precision=torch.half,
+            learnable=False,
+        )
+
+        self.scale_mask_softmax = None
+
+        if hidden_size_per_attention_head is None:
+            self.hidden_size_per_attention_head = hidden_size // num_attention_heads
+        else:
+            self.hidden_size_per_attention_head = hidden_size_per_attention_head
+
+        self.inner_hidden_size = num_attention_heads * self.hidden_size_per_attention_head
+
+        # Strided linear layer.
+        self.query_key_value = init_method(
+            torch.nn.Linear,
+            hidden_size,
+            3 * self.inner_hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+
+        self.dense = init_method(
+            torch.nn.Linear,
+            self.inner_hidden_size,
+            hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+
+    @staticmethod
+    def attention_mask_func(attention_scores, attention_mask):
+        attention_scores.masked_fill_(attention_mask, -10000.0)
+        return attention_scores
+
+    def split_tensor_along_last_dim(self, tensor, num_partitions,
+                                    contiguous_split_chunks=False):
+        """Split a tensor along its last dimension.
+        Arguments:
+            tensor: input tensor.
+            num_partitions: number of partitions to split the tensor
+            contiguous_split_chunks: If True, make each chunk contiguous
+                                    in memory.
+        """
+        # Get the size and dimension.
+        last_dim = tensor.dim() - 1
+        last_dim_size = tensor.size()[last_dim] // num_partitions
+        # Split.
+        tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
+        # Note: torch.split does not create contiguous tensors by default.
+        if contiguous_split_chunks:
+            return tuple(chunk.contiguous() for chunk in tensor_list)
+
+        return tensor_list
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            position_ids,
+            attention_mask: torch.Tensor,
+            layer_id,
+            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            use_cache: bool = False,
+            output_attentions: bool = False,
+    ):
+        """
+        hidden_states: [seq_len, batch, hidden_size]
+        attention_mask: [(1, 1), seq_len, seq_len]
+        """
+
+        # [seq_len, batch, 3 * hidden_size]
+        mixed_raw_layer = self.query_key_value(hidden_states)
+
+        # [seq_len, batch, 3 * hidden_size] --> [seq_len, batch, num_attention_heads, 3 * hidden_size_per_attention_head]
+        new_tensor_shape = mixed_raw_layer.size()[:-1] + (
+            self.num_attention_heads_per_partition,
+            3 * self.hidden_size_per_attention_head,
+        )
+        mixed_raw_layer = mixed_raw_layer.view(*new_tensor_shape)
+
+        # [seq_len, batch, num_attention_heads, hidden_size_per_attention_head]
+        (query_layer, key_layer, value_layer) = self.split_tensor_along_last_dim(mixed_raw_layer, 3)
+
+        if self.position_encoding_2d:
+            q1, q2 = query_layer.chunk(2, dim=(query_layer.ndim - 1))
+            k1, k2 = key_layer.chunk(2, dim=(key_layer.ndim - 1))
+            cos, sin = self.rotary_emb(q1, seq_len=position_ids.max() + 1)
+            position_ids, block_position_ids = position_ids[:, 0, :].transpose(0, 1).contiguous(), \
+                position_ids[:, 1, :].transpose(0, 1).contiguous()
+            q1, k1 = apply_rotary_pos_emb_index(q1, k1, cos, sin, position_ids)
+            q2, k2 = apply_rotary_pos_emb_index(q2, k2, cos, sin, block_position_ids)
+            query_layer = torch.concat([q1, q2], dim=(q1.ndim - 1))
+            key_layer = torch.concat([k1, k2], dim=(k1.ndim - 1))
+        else:
+            position_ids = position_ids.transpose(0, 1)
+            cos, sin = self.rotary_emb(value_layer, seq_len=position_ids.max() + 1)
+            # [seq_len, batch, num_attention_heads, hidden_size_per_attention_head]
+            query_layer, key_layer = apply_rotary_pos_emb_index(query_layer, key_layer, cos, sin, position_ids)
+
+        # [seq_len, batch, hidden_size]
+        context_layer, present, attention_probs = attention_fn(
+            self=self,
+            query_layer=query_layer,
+            key_layer=key_layer,
+            value_layer=value_layer,
+            attention_mask=attention_mask,
+            hidden_size_per_partition=self.hidden_size_per_partition,
+            layer_id=layer_id,
+            layer_past=layer_past,
+            use_cache=use_cache
+        )
+
+        output = self.dense(context_layer)
+
+        outputs = (output, present)
+
+        if output_attentions:
+            outputs += (attention_probs,)
+
+        return outputs  # output, present, attention_probs
+
+
+class GEGLU(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.activation_fn = F.gelu
+
+    def forward(self, x):
+        # dim=-1 breaks in jit for pt<1.10
+        x1, x2 = x.chunk(2, dim=(x.ndim - 1))
+        return x1 * self.activation_fn(x2)
+
+
+class GLU(torch.nn.Module):
+    def __init__(self, hidden_size, inner_hidden_size=None,
+                 layer_id=None, bias=True, activation_func=gelu, params_dtype=torch.float, empty_init=True):
+        super(GLU, self).__init__()
+        if empty_init:
+            init_method = skip_init
+        else:
+            init_method = default_init
+        self.layer_id = layer_id
+        self.activation_func = activation_func
+
+        # Project to 4h.
+        self.hidden_size = hidden_size
+        if inner_hidden_size is None:
+            inner_hidden_size = 4 * hidden_size
+        self.inner_hidden_size = inner_hidden_size
+        self.dense_h_to_4h = init_method(
+            torch.nn.Linear,
+            self.hidden_size,
+            self.inner_hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+        # Project back to h.
+        self.dense_4h_to_h = init_method(
+            torch.nn.Linear,
+            self.inner_hidden_size,
+            self.hidden_size,
+            bias=bias,
+            dtype=params_dtype,
+        )
+
+    def forward(self, hidden_states):
+        """
+        hidden_states: [seq_len, batch, hidden_size]
+        """
+
+        # [seq_len, batch, inner_hidden_size]
+        intermediate_parallel = self.dense_h_to_4h(hidden_states)
+
+        intermediate_parallel = self.activation_func(intermediate_parallel)
+
+        output = self.dense_4h_to_h(intermediate_parallel)
+
+        return output
+
+
+class GLMBlock(torch.nn.Module):
+    def __init__(
+            self,
+            hidden_size,
+            num_attention_heads,
+            layernorm_epsilon,
+            layer_id,
+            inner_hidden_size=None,
+            hidden_size_per_attention_head=None,
+            layernorm=LayerNorm,
+            use_bias=True,
+            params_dtype=torch.float,
+            num_layers=28,
+            position_encoding_2d=True,
+            empty_init=True
+    ):
+        super(GLMBlock, self).__init__()
+        # Set output layer initialization if not provided.
+
+        self.layer_id = layer_id
+
+        # Layernorm on the input data.
+        self.input_layernorm = layernorm(hidden_size, eps=layernorm_epsilon)
+
+        self.position_encoding_2d = position_encoding_2d
+
+        # Self attention.
+        self.attention = SelfAttention(
+            hidden_size,
+            num_attention_heads,
+            layer_id,
+            hidden_size_per_attention_head=hidden_size_per_attention_head,
+            bias=use_bias,
+            params_dtype=params_dtype,
+            position_encoding_2d=self.position_encoding_2d,
+            empty_init=empty_init
+        )
+
+        # Layernorm on the input data.
+        self.post_attention_layernorm = layernorm(hidden_size, eps=layernorm_epsilon)
+
+        self.num_layers = num_layers
+
+        # GLU
+        self.mlp = GLU(
+            hidden_size,
+            inner_hidden_size=inner_hidden_size,
+            bias=use_bias,
+            layer_id=layer_id,
+            params_dtype=params_dtype,
+            empty_init=empty_init
+        )
+
+    def forward(
+            self,
+            hidden_states: torch.Tensor,
+            position_ids,
+            attention_mask: torch.Tensor,
+            layer_id,
+            layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+            use_cache: bool = False,
+            output_attentions: bool = False,
+    ):
+        """
+        hidden_states: [seq_len, batch, hidden_size]
+        attention_mask: [(1, 1), seq_len, seq_len]
+        """
+
+        # Layer norm at the begining of the transformer layer.
+        # [seq_len, batch, hidden_size]
+        attention_input = self.input_layernorm(hidden_states)
+
+        # Self attention.
+        attention_outputs = self.attention(
+            attention_input,
+            position_ids,
+            attention_mask=attention_mask,
+            layer_id=layer_id,
+            layer_past=layer_past,
+            use_cache=use_cache,
+            output_attentions=output_attentions
+        )
+
+        attention_output = attention_outputs[0]
+
+        outputs = attention_outputs[1:]
+
+        # Residual connection.
+        alpha = (2 * self.num_layers) ** 0.5
+        hidden_states = attention_input * alpha + attention_output
+
+        mlp_input = self.post_attention_layernorm(hidden_states)
+
+        # MLP.
+        mlp_output = self.mlp(mlp_input)
+
+        # Second residual connection.
+        output = mlp_input * alpha + mlp_output
+
+        if use_cache:
+            outputs = (output,) + outputs
+        else:
+            outputs = (output,) + outputs[1:]
+
+        return outputs  # hidden_states, present, attentions
+
+
+class ChatGLMPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and
+    a simple interface for downloading and loading pretrained models.
+    """
+
+    is_parallelizable = False
+    supports_gradient_checkpointing = True
+    config_class = ChatGLMConfig
+    base_model_prefix = "transformer"
+    _no_split_modules = ["GLMBlock"]
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    def _init_weights(self, module: nn.Module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=0.02)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=0.02)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def get_masks(self, input_ids, device):
+        batch_size, seq_length = input_ids.shape
+        context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids]
+        attention_mask = torch.ones((batch_size, seq_length, seq_length), device=device)
+        attention_mask.tril_()
+        for i, context_length in enumerate(context_lengths):
+            attention_mask[i, :, :context_length] = 1
+        attention_mask.unsqueeze_(1)
+        attention_mask = (attention_mask < 0.5).bool()
+
+        return attention_mask
+
+    def get_position_ids(self, input_ids, mask_positions, device, use_gmasks=None):
+        batch_size, seq_length = input_ids.shape
+        if use_gmasks is None:
+            use_gmasks = [False] * batch_size
+        context_lengths = [seq.tolist().index(self.config.bos_token_id) for seq in input_ids]
+        if self.position_encoding_2d:
+            position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
+            for i, context_length in enumerate(context_lengths):
+                position_ids[i, context_length:] = mask_positions[i]
+            block_position_ids = [torch.cat((
+                torch.zeros(context_length, dtype=torch.long, device=device),
+                torch.arange(seq_length - context_length, dtype=torch.long, device=device) + 1
+            )) for context_length in context_lengths]
+            block_position_ids = torch.stack(block_position_ids, dim=0)
+            position_ids = torch.stack((position_ids, block_position_ids), dim=1)
+        else:
+            position_ids = torch.arange(seq_length, dtype=torch.long, device=device).unsqueeze(0).repeat(batch_size, 1)
+            for i, context_length in enumerate(context_lengths):
+                if not use_gmasks[i]:
+                    position_ids[i, context_length:] = mask_positions[i]
+
+        return position_ids
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, ChatGLMModel):
+            module.gradient_checkpointing = value
+
+
+CHATGLM_6B_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general
+    usage and behavior.
+
+    Parameters:
+        config ([`~ChatGLM6BConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the configuration.
+            Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+CHATGLM_6B_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`ChatGLM6BTokenizer`].
+            See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings.
+            Selected in the range `[0, config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert *input_ids* indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare ChatGLM-6B Model transformer outputting raw hidden-states without any specific head on top.",
+    CHATGLM_6B_START_DOCSTRING,
+)
+class ChatGLMModel(ChatGLMPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well
+    as a decoder, in which case a layer of cross-attention is added between
+    the self-attention layers, following the architecture described in [Attention is
+    all you need](https://arxiv.org/abs/1706.03762) by Ashish Vaswani,
+    Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the
+    `is_decoder` argument of the configuration set to `True`.
+    To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder`
+    argument and `add_cross_attention` set to `True`; an
+    `encoder_hidden_states` is then expected as an input to the forward pass.
+    """
+
+    def __init__(self, config: ChatGLMConfig, empty_init=True):
+        super().__init__(config)
+        if empty_init:
+            init_method = skip_init
+        else:
+            init_method = default_init
+        # recording parameters
+        self.max_sequence_length = config.max_sequence_length
+        self.hidden_size = config.hidden_size
+        self.params_dtype = torch.half
+        self.num_attention_heads = config.num_attention_heads
+        self.vocab_size = config.vocab_size
+        self.num_layers = config.num_layers
+        self.layernorm_epsilon = config.layernorm_epsilon
+        self.inner_hidden_size = config.inner_hidden_size
+        self.hidden_size_per_attention_head = self.hidden_size // self.num_attention_heads
+        self.position_encoding_2d = config.position_encoding_2d
+        self.pre_seq_len = config.pre_seq_len
+        self.prefix_projection = config.prefix_projection
+
+        self.word_embeddings = init_method(
+            torch.nn.Embedding,
+            num_embeddings=self.vocab_size, embedding_dim=self.hidden_size,
+            dtype=self.params_dtype
+        )
+        self.gradient_checkpointing = False
+
+        def get_layer(layer_id):
+            return GLMBlock(
+                self.hidden_size,
+                self.num_attention_heads,
+                self.layernorm_epsilon,
+                layer_id,
+                inner_hidden_size=self.inner_hidden_size,
+                hidden_size_per_attention_head=self.hidden_size_per_attention_head,
+                layernorm=LayerNorm,
+                use_bias=True,
+                params_dtype=self.params_dtype,
+                position_encoding_2d=self.position_encoding_2d,
+                empty_init=empty_init
+            )
+
+        self.layers = torch.nn.ModuleList(
+            [get_layer(layer_id) for layer_id in range(self.num_layers)]
+        )
+
+        # Final layer norm before output.
+        self.final_layernorm = LayerNorm(self.hidden_size, eps=self.layernorm_epsilon)
+
+        if self.pre_seq_len is not None:
+            for param in self.parameters():
+                param.requires_grad = False
+            self.prefix_tokens = torch.arange(self.pre_seq_len).long()
+            self.prefix_encoder = PrefixEncoder(config)
+            self.dropout = torch.nn.Dropout(0.1)
+
+            # total_params = sum(p.numel() for p in self.parameters())
+            # trainable_params = sum(p.numel() for p in self.parameters() if p.requires_grad)
+            # print("Using p-tuning v2: # trainable_params = {} / {}".format(trainable_params, total_params))
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.word_embeddings
+
+    def set_input_embeddings(self, new_embeddings: torch.Tensor):
+        self.word_embeddings = new_embeddings
+
+    def get_prompt(self, batch_size, device, dtype=torch.half):
+        prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(device)
+        past_key_values = self.prefix_encoder(prefix_tokens).type(dtype)
+        past_key_values = past_key_values.view(
+            batch_size,
+            self.pre_seq_len,
+            self.num_layers * 2,
+            self.num_attention_heads,
+            self.hidden_size // self.num_attention_heads
+        )
+        # seq_len, b, nh, hidden_size
+        past_key_values = self.dropout(past_key_values)
+        past_key_values = past_key_values.permute([2, 1, 0, 3, 4]).split(2)
+        # past_key_values = [(v[0], v[1]) for v in past_key_values]
+        return past_key_values
+
+    @add_start_docstrings_to_model_forward(CHATGLM_6B_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+            self,
+            input_ids: Optional[torch.LongTensor] = None,
+            position_ids: Optional[torch.LongTensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
+            inputs_embeds: Optional[torch.LongTensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPast]:
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape[:2]
+        elif inputs_embeds is not None:
+            batch_size, seq_length = inputs_embeds.shape[:2]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+
+        if past_key_values is None:
+            if self.pre_seq_len is not None:
+                past_key_values = self.get_prompt(batch_size=input_ids.shape[0], device=input_ids.device,
+                                                  dtype=inputs_embeds.dtype)
+            else:
+                past_key_values = tuple([None] * len(self.layers))
+
+            if attention_mask is None:
+                attention_mask = self.get_masks(
+                    input_ids,
+                    device=input_ids.device
+                )
+
+
+            if position_ids is None:
+                MASK, gMASK = self.config.mask_token_id, self.config.gmask_token_id
+                seqs = input_ids.tolist()
+
+                mask_positions, use_gmasks = [], []
+                for seq in seqs:
+                    mask_token = gMASK if gMASK in seq else MASK
+                    use_gmask = mask_token == gMASK
+                    mask_positions.append(seq.index(mask_token))
+                    use_gmasks.append(use_gmask)
+
+                position_ids = self.get_position_ids(
+                    input_ids,
+                    mask_positions=mask_positions,
+                    device=input_ids.device,
+                    use_gmasks=use_gmasks
+                )
+
+        if self.pre_seq_len is not None and attention_mask is not None:
+            prefix_attention_mask = torch.ones(batch_size, 1, input_ids.size(-1), self.pre_seq_len).to(
+                attention_mask.device)
+            prefix_attention_mask = (prefix_attention_mask < 0.5).bool()
+            attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=3)
+
+        # [seq_len, batch, hidden_size]
+        hidden_states = inputs_embeds.transpose(0, 1)
+
+        presents = () if use_cache else None
+        all_self_attentions = () if output_attentions else None
+        all_hidden_states = () if output_hidden_states else None
+
+        if attention_mask is None:
+            attention_mask = torch.zeros(1, 1, device=input_ids.device).bool()
+        else:
+            attention_mask = attention_mask.to(hidden_states.device)
+
+        for i, layer in enumerate(self.layers):
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+            layer_past = past_key_values[i]
+
+            if self.gradient_checkpointing and self.training:
+                layer_ret = torch.utils.checkpoint.checkpoint(
+                    layer,
+                    hidden_states,
+                    position_ids,
+                    attention_mask,
+                    torch.tensor(i),
+                    layer_past,
+                    use_cache,
+                    output_attentions
+                )
+            else:
+                layer_ret = layer(
+                    hidden_states,
+                    position_ids=position_ids,
+                    attention_mask=attention_mask,
+                    layer_id=torch.tensor(i),
+                    layer_past=layer_past,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions
+                )
+
+            hidden_states = layer_ret[0]
+
+            if use_cache:
+                presents = presents + (layer_ret[1],)
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_ret[2 if use_cache else 1],)
+
+        # Final layer norm.
+        hidden_states = self.final_layernorm(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=presents,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
+    def __init__(self, config: ChatGLMConfig, empty_init=True):
+        super().__init__(config)
+        if empty_init:
+            init_method = skip_init
+        else:
+            init_method = default_init
+
+        # self.hidden_size = config.hidden_size
+        # self.params_dtype = torch.half
+        # self.vocab_size = config.vocab_size
+        self.max_sequence_length = config.max_sequence_length
+
+        self.position_encoding_2d = config.position_encoding_2d
+
+        self.transformer = ChatGLMModel(config, empty_init=empty_init)
+
+        self.lm_head = init_method(
+            nn.Linear,
+            config.hidden_size,
+            config.vocab_size,
+            bias=False,
+            dtype=torch.half
+        )
+
+        self.config = config
+
+        self.quantized = False
+
+        if self.config.quantization_bit:
+            self.quantize(self.config.quantization_bit, empty_init=True)
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def _update_model_kwargs_for_generation(
+        self,
+        outputs: ModelOutput,
+        model_kwargs: Dict[str, Any],
+        is_encoder_decoder: bool = False,
+        standardize_cache_format: bool = False,
+    ) -> Dict[str, Any]:
+        # update past_key_values
+        model_kwargs["past_key_values"] = self._extract_past_from_model_output(
+            outputs, standardize_cache_format=standardize_cache_format
+        )
+
+        # update attention mask
+        if "attention_mask" in model_kwargs:
+            attention_mask = model_kwargs["attention_mask"]
+            if attention_mask is not None and attention_mask.dtype == torch.bool:
+                attention_mask = torch.cat(
+                    [attention_mask, attention_mask.new_ones((*attention_mask.shape[:3], 1))], dim=3)
+                new_attention_mask = attention_mask[:, :, -1:].clone()
+                new_attention_mask[..., -1] = False
+                model_kwargs["attention_mask"] = torch.cat(
+                    [attention_mask, new_attention_mask], dim=2
+                )
+
+        # update position ids
+        if "position_ids" in model_kwargs:
+            position_ids = model_kwargs["position_ids"]
+            new_position_id = position_ids[..., -1:].clone()
+            new_position_id[:, 1, :] += 1
+            model_kwargs["position_ids"] = torch.cat(
+                [position_ids, new_position_id], dim=-1
+            )
+
+        return model_kwargs
+
+    def prepare_inputs_for_generation(
+            self,
+            input_ids: torch.LongTensor,
+            past: Optional[torch.Tensor] = None,
+            past_key_values: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.Tensor] = None,
+            **kwargs
+    ) -> dict:
+        batch_size, seq_length = input_ids.shape
+        MASK, gMASK = self.config.mask_token_id, self.config.gmask_token_id
+        seqs = input_ids.tolist()
+        mask_positions, use_gmasks = [], []
+        for seq in seqs:
+            mask_token = gMASK if gMASK in seq else MASK
+            use_gmask = mask_token == gMASK
+            mask_positions.append(seq.index(mask_token))
+            use_gmasks.append(use_gmask)
+
+        # only last token for input_ids if past is not None
+        if past is not None or past_key_values is not None:
+            last_token = input_ids[:, -1].unsqueeze(-1)
+            if attention_mask is not None and attention_mask.dtype == torch.bool:
+                attention_mask = attention_mask[:, :, -1:]
+            else:
+                attention_mask = None
+            if position_ids is not None:
+                position_ids = position_ids[..., -1:]
+            else:
+                context_lengths = [seq.index(self.config.bos_token_id) for seq in seqs]
+                if self.position_encoding_2d:
+                    position_ids = torch.tensor(
+                        [[mask_position, seq_length - context_length] for mask_position, context_length in
+                         zip(mask_positions, context_lengths)], dtype=torch.long, device=input_ids.device).unsqueeze(-1)
+                else:
+                    position_ids = torch.tensor([mask_position for mask_position in mask_positions], dtype=torch.long,
+                                                device=input_ids.device).unsqueeze(-1)
+
+            if past is None:
+                past = past_key_values
+            return {
+                "input_ids": last_token,
+                "past_key_values": past,
+                "position_ids": position_ids,
+                "attention_mask": attention_mask
+            }
+        else:
+            if attention_mask is not None and attention_mask.dtype != torch.bool:
+                logger.warning_once(f"The dtype of attention mask ({attention_mask.dtype}) is not bool")
+                attention_mask = None
+            if attention_mask is None:
+                attention_mask = self.get_masks(
+                    input_ids,
+                    device=input_ids.device
+                )
+            if position_ids is None:
+                position_ids = self.get_position_ids(
+                    input_ids,
+                    device=input_ids.device,
+                    mask_positions=mask_positions,
+                    use_gmasks=use_gmasks
+                )
+
+            return {
+                "input_ids": input_ids,
+                "past_key_values": past,
+                "position_ids": position_ids,
+                "attention_mask": attention_mask
+            }
+
+    def forward(
+            self,
+            input_ids: Optional[torch.Tensor] = None,
+            position_ids: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+            past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
+            inputs_embeds: Optional[torch.Tensor] = None,
+            labels: Optional[torch.Tensor] = None,
+            use_cache: Optional[bool] = None,
+            output_attentions: Optional[bool] = None,
+            output_hidden_states: Optional[bool] = None,
+            return_dict: Optional[bool] = None,
+    ):
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = transformer_outputs[0]
+
+        lm_logits = self.lm_head(hidden_states).permute(1, 0, 2).contiguous()
+
+        loss = None
+        if labels is not None:
+            lm_logits = lm_logits.to(torch.float32)
+
+            # Shift so that tokens < n predict n
+            shift_logits = lm_logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss(ignore_index=-100)
+            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+            lm_logits = lm_logits.to(hidden_states.dtype)
+            loss = loss.to(hidden_states.dtype)
+
+        if not return_dict:
+            output = (lm_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @staticmethod
+    def _reorder_cache(
+            past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor
+    ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]:
+        """
+        This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
+        [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
+        beam_idx at every generation step.
+
+        Output shares the same memory storage as `past`.
+        """
+        return tuple(
+            (
+                layer_past[0].index_select(1, beam_idx.to(layer_past[0].device)),
+                layer_past[1].index_select(1, beam_idx.to(layer_past[1].device)),
+            )
+            for layer_past in past
+        )
+
+    def process_response(self, response):
+        response = response.strip()
+        response = response.replace("[[训练时间]]", "2023年")
+        punkts = [
+            [",", "，"],
+            ["!", "！"],
+            [":", "："],
+            [";", "；"],
+            ["\?", "？"],
+        ]
+        for item in punkts:
+            response = re.sub(r"([\u4e00-\u9fff])%s" % item[0], r"\1%s" % item[1], response)
+            response = re.sub(r"%s([\u4e00-\u9fff])" % item[0], r"%s\1" % item[1], response)
+        return response
+
+    @torch.no_grad()
+    def chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 2048, num_beams=1,
+             do_sample=True, top_p=0.7, temperature=0.95, logits_processor=None, **kwargs):
+        if history is None:
+            history = []
+        if logits_processor is None:
+            logits_processor = LogitsProcessorList()
+        logits_processor.append(InvalidScoreLogitsProcessor())
+        gen_kwargs = {"max_length": max_length, "num_beams": num_beams, "do_sample": do_sample, "top_p": top_p,
+                      "temperature": temperature, "logits_processor": logits_processor, **kwargs}
+        if not history:
+            prompt = query
+        else:
+            prompt = ""
+            for i, (old_query, response) in enumerate(history):
+                prompt += "[Round {}]\n问：{}\n答：{}\n".format(i, old_query, response)
+            prompt += "[Round {}]\n问：{}\n答：".format(len(history), query)
+        inputs = tokenizer([prompt], return_tensors="pt")
+        inputs = inputs.to(self.device)
+        outputs = self.generate(**inputs, **gen_kwargs)
+        outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
+        response = tokenizer.decode(outputs)
+        response = self.process_response(response)
+        history = history + [(query, response)]
+        return response, history
+
+    @torch.no_grad()
+    def stream_chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 2048,
+                    do_sample=True, top_p=0.7, temperature=0.95, logits_processor=None, **kwargs):
+        if history is None:
+            history = []
+        if logits_processor is None:
+            logits_processor = LogitsProcessorList()
+        logits_processor.append(InvalidScoreLogitsProcessor())
+        gen_kwargs = {"max_length": max_length, "do_sample": do_sample, "top_p": top_p,
+                      "temperature": temperature, "logits_processor": logits_processor, **kwargs}
+        if not history:
+            prompt = query
+        else:
+            prompt = ""
+            for i, (old_query, response) in enumerate(history):
+                prompt += "[Round {}]\n问：{}\n答：{}\n".format(i, old_query, response)
+            prompt += "[Round {}]\n问：{}\n答：".format(len(history), query)
+        inputs = tokenizer([prompt], return_tensors="pt")
+        inputs = inputs.to(self.device)
+        for outputs in self.stream_generate(**inputs, **gen_kwargs):
+            outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
+            response = tokenizer.decode(outputs)
+            response = self.process_response(response)
+            new_history = history + [(query, response)]
+            yield response, new_history
+
+    @torch.no_grad()
+    def stream_generate(
+            self,
+            input_ids,
+            generation_config: Optional[GenerationConfig] = None,
+            logits_processor: Optional[LogitsProcessorList] = None,
+            stopping_criteria: Optional[StoppingCriteriaList] = None,
+            prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
+            **kwargs,
+    ):
+        batch_size, input_ids_seq_length = input_ids.shape[0], input_ids.shape[-1]
+
+        if generation_config is None:
+            generation_config = self.generation_config
+        generation_config = copy.deepcopy(generation_config)
+        model_kwargs = generation_config.update(**kwargs)
+        bos_token_id, eos_token_id = generation_config.bos_token_id, generation_config.eos_token_id
+
+        if isinstance(eos_token_id, int):
+            eos_token_id = [eos_token_id]
+
+        has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None
+        if has_default_max_length and generation_config.max_new_tokens is None:
+            warnings.warn(
+                f"Using `max_length`'s default ({generation_config.max_length}) to control the generation length. "
+                "This behaviour is deprecated and will be removed from the config in v5 of Transformers -- we"
+                " recommend using `max_new_tokens` to control the maximum length of the generation.",
+                UserWarning,
+            )
+        elif generation_config.max_new_tokens is not None:
+            generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length
+            if not has_default_max_length:
+                logger.warn(
+                    f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(="
+                    f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. "
+                    "Please refer to the documentation for more information. "
+                    "(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)",
+                    UserWarning,
+                )
+
+        if input_ids_seq_length >= generation_config.max_length:
+            input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids"
+            logger.warning(
+                f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to"
+                f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
+                " increasing `max_new_tokens`."
+            )
+
+        # 2. Set generation parameters if not already defined
+        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
+        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
+
+        logits_processor = self._get_logits_processor(
+            generation_config=generation_config,
+            input_ids_seq_length=input_ids_seq_length,
+            encoder_input_ids=input_ids,
+            prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
+            logits_processor=logits_processor,
+        )
+
+        stopping_criteria = self._get_stopping_criteria(
+            generation_config=generation_config, stopping_criteria=stopping_criteria
+        )
+        logits_warper = self._get_logits_warper(generation_config)
+
+        unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)
+        scores = None
+        while True:
+            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
+            # forward pass to get next token
+            outputs = self(
+                **model_inputs,
+                return_dict=True,
+                output_attentions=False,
+                output_hidden_states=False,
+            )
+
+            next_token_logits = outputs.logits[:, -1, :]
+
+            # pre-process distribution
+            next_token_scores = logits_processor(input_ids, next_token_logits)
+            next_token_scores = logits_warper(input_ids, next_token_scores)
+
+            # sample
+            probs = nn.functional.softmax(next_token_scores, dim=-1)
+            if generation_config.do_sample:
+                next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
+            else:
+                next_tokens = torch.argmax(probs, dim=-1)
+
+            # update generated ids, model inputs, and length for next step
+            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
+            model_kwargs = self._update_model_kwargs_for_generation(
+                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
+            )
+            unfinished_sequences = unfinished_sequences.mul((sum(next_tokens != i for i in eos_token_id)).long())
+
+            # stop when each sentence is finished, or if we exceed the maximum length
+            if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores):
+                break
+            yield input_ids
+
+    def quantize(self, bits: int, empty_init=False, **kwargs):
+        if bits == 0:
+            return
+
+        from .quantization import quantize
+
+        if self.quantized:
+            logger.info("Already quantized.")
+            return self
+
+        self.quantized = True
+
+        self.config.quantization_bit = bits
+
+        self.transformer = quantize(self.transformer, bits, empty_init=empty_init, **kwargs)
+        return self

simple-pretrain/ptuning/deepspeed.json

+{
+  "train_micro_batch_size_per_gpu": "auto",
+  "gradient_accumulation_steps": "auto",
+  "zero_allow_untested_optimizer": true,
+  "fp16": {
+    "enabled": "auto",
+    "loss_scale": 0,
+    "initial_scale_power": 16,
+    "loss_scale_window": 1000,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
+  "zero_optimization": {
+    "stage": 3,
+    "offload_optimizer": {
+      "device": "cpu",
+      "pin_memory": true
+    },
+    "offload_param": {
+      "device": "cpu",
+      "pin_memory": true
+    },
+    "allgather_partitions": true,
+    "allgather_bucket_size": 5e8,
+    "overlap_comm": true,
+    "reduce_scatter": true,
+    "reduce_bucket_size": 5e8,
+    "contiguous_gradients" : true,
+    "stage3_gather_16bit_weights_on_model_save": true
+  }
+}

simple-pretrain/ptuning/ds_pretrain.sh

+LR=1e-5
+
+MASTER_PORT=$(shuf -n 1 -i 10000-65535)
+
+HIP_VISIBLE_DEVICES=0,1,2,3 deepspeed --num_gpus=4 --master_port $MASTER_PORT main.py \
+    --deepspeed deepspeed.json \
+    --do_train \
+    --train_file The-Lord-of-the-Rings-1.json \
+    --prompt_column prompt \
+    --response_column response \
+    --overwrite_cache \
+    --model_name_or_path THUDM/chatglm-6b \
+    --output_dir ./output/pretrain \
+    --overwrite_output_dir \
+    --max_source_length 8 \
+    --max_target_length 128 \
+    --per_device_train_batch_size 16 \
+    --per_device_eval_batch_size 4 \
+    --gradient_accumulation_steps 4 \
+    --predict_with_generate \
+    --max_steps 2000 \
+    --logging_steps 5 \
+    --save_steps 1000 \
+    --learning_rate $LR \
+    --fp16
+

simple-pretrain/ptuning/main.py

+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the library models for sequence to sequence.
+"""
+# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
+
+import logging
+import os
+import sys
+import json
+
+import numpy as np
+from datasets import load_dataset
+import jieba 
+from rouge_chinese import Rouge
+from nltk.translate.bleu_score import sentence_bleu, SmoothingFunction
+import torch
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoModel,
+    AutoTokenizer,
+    DataCollatorForSeq2Seq,
+    HfArgumentParser,
+    Seq2SeqTrainingArguments,
+    set_seed,
+)
+from trainer_seq2seq import Seq2SeqTrainer
+
+from arguments import ModelArguments, DataTrainingArguments
+
+logger = logging.getLogger(__name__)
+
+def main():
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+
+    if training_args.should_log:
+        # The default of training_args.log_level is passive, so we set log level at info here to have that default.
+        transformers.utils.logging.set_verbosity_info()
+
+    log_level = training_args.get_process_log_level()
+    logger.setLevel(log_level)
+    # datasets.utils.logging.set_verbosity(log_level)
+    transformers.utils.logging.set_verbosity(log_level)
+    transformers.utils.logging.enable_default_handler()
+    transformers.utils.logging.enable_explicit_format()
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Load dataset
+    data_files = {}
+    if data_args.train_file is not None:
+        data_files["train"] = data_args.train_file
+        extension = data_args.train_file.split(".")[-1]
+    if data_args.validation_file is not None:
+        data_files["validation"] = data_args.validation_file
+        extension = data_args.validation_file.split(".")[-1]
+    if data_args.test_file is not None:
+        data_files["test"] = data_args.test_file
+        extension = data_args.test_file.split(".")[-1]
+
+    raw_datasets = load_dataset(
+        extension,
+        data_files=data_files,
+        cache_dir=model_args.cache_dir,
+        use_auth_token=True if model_args.use_auth_token else None,
+    )
+
+    # Load pretrained model and tokenizer
+    config = AutoConfig.from_pretrained(model_args.model_name_or_path, trust_remote_code=True)
+    config.pre_seq_len = model_args.pre_seq_len
+    config.prefix_projection = model_args.prefix_projection
+
+    tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, trust_remote_code=True)
+
+    if model_args.ptuning_checkpoint is not None:
+        # Evaluation
+        # Loading extra state dict of prefix encoder
+        model = AutoModel.from_pretrained(model_args.model_name_or_path, config=config, trust_remote_code=True)
+        prefix_state_dict = torch.load(os.path.join(model_args.ptuning_checkpoint, "pytorch_model.bin"))
+        new_prefix_state_dict = {}
+        for k, v in prefix_state_dict.items():
+            if k.startswith("transformer.prefix_encoder."):
+                new_prefix_state_dict[k[len("transformer.prefix_encoder."):]] = v
+        model.transformer.prefix_encoder.load_state_dict(new_prefix_state_dict)
+    else:
+        # model = AutoModel.from_pretrained(model_args.model_name_or_path, config=config, trust_remote_code=True)
+        model = AutoModel.from_config(config=config, trust_remote_code=True)
+
+    if model_args.quantization_bit is not None:
+        print(f"Quantized to {model_args.quantization_bit} bit")
+        model = model.quantize(model_args.quantization_bit)
+    if model_args.pre_seq_len is not None:
+        # P-tuning v2
+        model = model.half()
+        model.transformer.prefix_encoder.float()
+    else:
+        # Finetune
+        model = model.float()
+
+    prefix = data_args.source_prefix if data_args.source_prefix is not None else ""
+
+    # Preprocessing the datasets.
+    # We need to tokenize inputs and targets.
+    if training_args.do_train:
+        column_names = raw_datasets["train"].column_names
+    elif training_args.do_eval:
+        column_names = raw_datasets["validation"].column_names
+    elif training_args.do_predict:
+        column_names = raw_datasets["test"].column_names
+    else:
+        logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.")
+        return
+
+    # Get the column names for input/target.
+    prompt_column = data_args.prompt_column
+    response_column = data_args.response_column
+    history_column = data_args.history_column
+    
+    # Temporarily set max_target_length for training.
+    max_target_length = data_args.max_target_length
+
+    def preprocess_function_eval(examples):
+        inputs, targets = [], []
+        for i in range(len(examples[prompt_column])):
+            if examples[prompt_column][i] and examples[response_column][i]:
+                query = examples[prompt_column][i]
+                if history_column is None or len(examples[history_column][i]) == 0:
+                    prompt = query
+                else:
+                    prompt = ""
+                    history = examples[history_column][i]
+                    for turn_idx, (old_query, response) in enumerate(history):
+                        prompt += "[Round {}]\n问：{}\n答：{}\n".format(turn_idx, old_query, response)
+                    prompt += "[Round {}]\n问：{}\n答：".format(len(history), query)
+                inputs.append(prompt)
+                targets.append(examples[response_column][i])
+
+        inputs = [prefix + inp for inp in inputs]
+        model_inputs = tokenizer(inputs, max_length=data_args.max_source_length, truncation=True, padding=True)
+        labels = tokenizer(text_target=targets, max_length=max_target_length, truncation=True)
+
+        if data_args.ignore_pad_token_for_loss:
+            labels["input_ids"] = [
+                [(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
+            ]
+        model_inputs["labels"] = labels["input_ids"]
+
+        return model_inputs
+
+    def preprocess_function_train(examples):
+        max_seq_length = data_args.max_source_length + data_args.max_target_length
+
+        model_inputs = {
+            "input_ids": [],
+            "labels": [],
+        }
+        for i in range(len(examples[prompt_column])):
+            if examples[prompt_column][i] and examples[response_column][i]:
+                query, answer = examples[prompt_column][i], examples[response_column][i]
+
+                if history_column is None:
+                    prompt = query
+                else:
+                    prompt = ""
+                    history = examples[history_column][i]
+                    for turn_idx, (old_query, response) in enumerate(history):
+                        prompt += "[Round {}]\n问：{}\n答：{}\n".format(turn_idx, old_query, response)
+                    prompt += "[Round {}]\n问：{}\n答：".format(len(history), query)
+
+                prompt = prefix + prompt
+                a_ids = tokenizer.encode(text=prompt, add_special_tokens=False)
+                b_ids = tokenizer.encode(text=answer, add_special_tokens=False)
+
+                if len(a_ids) > data_args.max_source_length - 1:
+                    a_ids = a_ids[: data_args.max_source_length - 1]
+
+                if len(b_ids) > data_args.max_target_length - 2:
+                    b_ids = b_ids[: data_args.max_target_length - 2]
+
+                input_ids = tokenizer.build_inputs_with_special_tokens(a_ids, b_ids)
+
+                context_length = input_ids.index(tokenizer.bos_token_id)
+                mask_position = context_length - 1
+                labels = [-100] * context_length + input_ids[mask_position+1:]
+                
+                pad_len = max_seq_length - len(input_ids)
+                input_ids = input_ids + [tokenizer.pad_token_id] * pad_len
+                labels = labels + [tokenizer.pad_token_id] * pad_len
+                if data_args.ignore_pad_token_for_loss:
+                    labels = [(l if l != tokenizer.pad_token_id else -100) for l in labels]
+
+                model_inputs["input_ids"].append(input_ids)
+                model_inputs["labels"].append(labels)
+
+        return model_inputs
+    
+    def print_dataset_example(example):
+        print("input_ids",example["input_ids"])
+        print("inputs", tokenizer.decode(example["input_ids"]))
+        print("label_ids", example["labels"])
+        print("labels", tokenizer.decode(example["labels"]))
+
+    if training_args.do_train:
+        if "train" not in raw_datasets:
+            raise ValueError("--do_train requires a train dataset")
+        train_dataset = raw_datasets["train"]
+        if data_args.max_train_samples is not None:
+            max_train_samples = min(len(train_dataset), data_args.max_train_samples)
+            train_dataset = train_dataset.select(range(max_train_samples))
+        with training_args.main_process_first(desc="train dataset map pre-processing"):
+            train_dataset = train_dataset.map(
+                preprocess_function_train,
+                batched=True,
+                num_proc=data_args.preprocessing_num_workers,
+                remove_columns=column_names,
+                load_from_cache_file=not data_args.overwrite_cache,
+                desc="Running tokenizer on train dataset",
+            )
+        print_dataset_example(train_dataset[0])
+
+    if training_args.do_eval:
+        max_target_length = data_args.val_max_target_length
+        if "validation" not in raw_datasets:
+            raise ValueError("--do_eval requires a validation dataset")
+        eval_dataset = raw_datasets["validation"]
+        if data_args.max_eval_samples is not None:
+            max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
+            eval_dataset = eval_dataset.select(range(max_eval_samples))
+        with training_args.main_process_first(desc="validation dataset map pre-processing"):
+            eval_dataset = eval_dataset.map(
+                preprocess_function_eval,
+                batched=True,
+                num_proc=data_args.preprocessing_num_workers,
+                remove_columns=column_names,
+                load_from_cache_file=not data_args.overwrite_cache,
+                desc="Running tokenizer on validation dataset",
+            )
+        print_dataset_example(eval_dataset[0])
+
+    if training_args.do_predict:
+        max_target_length = data_args.val_max_target_length
+        if "test" not in raw_datasets:
+            raise ValueError("--do_predict requires a test dataset")
+        predict_dataset = raw_datasets["test"]
+        if data_args.max_predict_samples is not None:
+            max_predict_samples = min(len(predict_dataset), data_args.max_predict_samples)
+            predict_dataset = predict_dataset.select(range(max_predict_samples))
+        with training_args.main_process_first(desc="prediction dataset map pre-processing"):
+            predict_dataset = predict_dataset.map(
+                preprocess_function_eval,
+                batched=True,
+                num_proc=data_args.preprocessing_num_workers,
+                remove_columns=column_names,
+                load_from_cache_file=not data_args.overwrite_cache,
+                desc="Running tokenizer on prediction dataset",
+            )
+        print_dataset_example(predict_dataset[0])
+
+    # Data collator
+    label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
+    data_collator = DataCollatorForSeq2Seq(
+        tokenizer,
+        model=model,
+        label_pad_token_id=label_pad_token_id,
+        pad_to_multiple_of=None,
+        padding=False
+    )
+
+    # Metric
+    def compute_metrics(eval_preds):
+        preds, labels = eval_preds
+        if isinstance(preds, tuple):
+            preds = preds[0]
+        decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
+        if data_args.ignore_pad_token_for_loss:
+            # Replace -100 in the labels as we can't decode them.
+            labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
+        decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
+
+        score_dict = {
+            "rouge-1": [],
+            "rouge-2": [],
+            "rouge-l": [],
+            "bleu-4": []
+        }
+        for pred, label in zip(decoded_preds, decoded_labels):
+            hypothesis = list(jieba.cut(pred))
+            reference = list(jieba.cut(label))
+            rouge = Rouge()
+            scores = rouge.get_scores(' '.join(hypothesis) , ' '.join(reference))
+            result = scores[0]
+            
+            for k, v in result.items():
+                score_dict[k].append(round(v["f"] * 100, 4))
+            bleu_score = sentence_bleu([list(label)], list(pred), smoothing_function=SmoothingFunction().method3)
+            score_dict["bleu-4"].append(round(bleu_score * 100, 4))
+
+        for k, v in score_dict.items():
+            score_dict[k] = float(np.mean(v))
+        return score_dict
+
+    # Override the decoding parameters of Seq2SeqTrainer
+    training_args.generation_max_length = (
+        training_args.generation_max_length
+        if training_args.generation_max_length is not None
+        else data_args.val_max_target_length
+    )
+    training_args.generation_num_beams = (
+        data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
+    )
+    # Initialize our Trainer
+    trainer = Seq2SeqTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_dataset if training_args.do_train else None,
+        eval_dataset=eval_dataset if training_args.do_eval else None,
+        tokenizer=tokenizer,
+        data_collator=data_collator,
+        compute_metrics=compute_metrics if training_args.predict_with_generate else None,
+        save_prefixencoder=model_args.pre_seq_len is not None
+    )
+
+    # Training
+    if training_args.do_train:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        # elif last_checkpoint is not None:
+        #     checkpoint = last_checkpoint
+        model.gradient_checkpointing_enable()
+        model.enable_input_require_grads()
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        # trainer.save_model()  # Saves the tokenizer too for easy upload
+
+        metrics = train_result.metrics
+        max_train_samples = (
+            data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
+        )
+        metrics["train_samples"] = min(max_train_samples, len(train_dataset))
+
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # Evaluation
+    results = {}
+    max_seq_length = data_args.max_source_length + data_args.max_target_length + 1
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        metrics = trainer.evaluate(metric_key_prefix="eval", do_sample=True, top_p=0.7, max_length=max_seq_length, temperature=0.95)
+        max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
+        metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    if training_args.do_predict:
+        logger.info("*** Predict ***")
+        predict_results = trainer.predict(predict_dataset, metric_key_prefix="predict", max_length=max_seq_length, do_sample=True, top_p=0.7, temperature=0.95)
+        metrics = predict_results.metrics
+        max_predict_samples = (
+            data_args.max_predict_samples if data_args.max_predict_samples is not None else len(predict_dataset)
+        )
+        metrics["predict_samples"] = min(max_predict_samples, len(predict_dataset))
+
+        trainer.log_metrics("predict", metrics)
+        trainer.save_metrics("predict", metrics)
+
+        if trainer.is_world_process_zero():
+            if training_args.predict_with_generate:
+                predictions = tokenizer.batch_decode(
+                    predict_results.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
+                )
+                predictions = [pred.strip() for pred in predictions]
+                labels = tokenizer.batch_decode(
+                    predict_results.label_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
+                )
+                labels = [label.strip() for label in labels]
+                output_prediction_file = os.path.join(training_args.output_dir, "generated_predictions.txt")
+                with open(output_prediction_file, "w", encoding="utf-8") as writer:
+                    for p, l in zip(predictions, labels):
+                        res = json.dumps({"labels": l, "predict": p}, ensure_ascii=False)
+                        writer.write(f"{res}\n")
+    return results
+
+
+def _mp_fn(index):
+    # For xla_spawn (TPUs)
+    main()
+
+
+if __name__ == "__main__":
+    main()