from transformers import pipeline
classifier = pipeline("sentiment-analysis") # 情感分析
result = classifier("I've been waiting for a HuggingFace course my whole life.")
print(result)

from transformers import pipeline
generator = pipeline("text-generation")     # 文本生成
results = generator("In this course, we will teach you how to")
print(results)

# Building the config
config = BertConfig()
# Building the model from the config,使用随机值对其进行初始化
model = BertModel(config)
print(config)
BertConfig {
  [...]
  "hidden_size": 768,               # 定义了hidden状态向量的大小
  "intermediate_size": 3072,
  "max_position_embeddings": 512,
  "num_attention_heads": 12,
  "num_hidden_layers": 12,          # 定义了Transformer模型的层数
  [...]
}
# 加载已经训练过的Transformers模型
model = BertModel.from_pretrained("bert-base-cased")

# 加载与保存分词器
from transformers import BertTokenizer

tokenizer = BertTokenizer.from_pretrained("bert-base-cased")
tokenizer.save_pretrained("./models/bert-base-cased/")
# 加载与保存模型
from transformers import AutoModel
# 所有存储在 HuggingFace Model Hub 上的模型都可以通过 Model.from_pretrained() 来加载权重，参数可以是 checkpoint 的名称，也可以是本地路径（预先下载的模型目录）
model = AutoModel.from_pretrained("bert-base-cased")
model.save_pretrained("./models/bert-base-cased/") # 保存模型

inputs = tokenizer(["来到美丽的大自然，我们发现"], return_tensors="pt")
# {'input_ids': tensor([[    1, 68846, 68881, 67701, 67668, 98899, 91935]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1]])}

gen_kwargs = {"max_length": 128, "top_p": 0.8, "temperature": 0.8, "do_sample": True, "repetition_penalty": 1.1}
output = model.generate(**inputs, **gen_kwargs)
# decode the new tokens
output = tokenizer.decode(output[0].tolist(), skip_special_tokens=True)
print(output)

from datasets import load_dataset
raw_datasets = load_dataset("glue", "mrpc")
print(raw_datasets)
# 包含训练集、验证集和测试集。每一个集合都包含几个列(sentence1, sentence2, label, and idx)以及一个代表行数的变量
#DatasetDict({
#    train: Dataset({
#        features: ['sentence1', 'sentence2', 'label', 'idx'],
#        num_rows: 3668
#    })
#    validation: Dataset({
#        features: ['sentence1', 'sentence2', 'label', 'idx'],
#        num_rows: 408
#    })
#    test: Dataset({
#        features: ['sentence1', 'sentence2', 'label', 'idx'],
#        num_rows: 1725
#    })
#})

# example 是一个dict，对应数据集的每个元素，并返回一个包含input_ids、attention_mask 和token_type_ids为key的新dict
# 在机器学习任务中，一个example通常定义为模型的输入（也成为特征集合）
def tokenize_function(example):
    return tokenizer(example["sentence1"], example["sentence2"], truncation=True)
tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)
print(raw_datasets)
#DatasetDict({
#    train: Dataset({
#        features: ['attention_mask', 'idx', 'input_ids', 'label', 'sentence1', 'sentence2', 'token_type_ids'],
#        num_rows: 3668
#    })
#    validation: Dataset({
#        features: ['attention_mask', 'idx', 'input_ids', 'label', 'sentence1', 'sentence2', 'token_type_ids'],
#        num_rows: 408
#    })
#    test: Dataset({
#        features: ['attention_mask', 'idx', 'input_ids', 'label', 'sentence1', 'sentence2', 'token_type_ids'],
#        num_rows: 1725
#    })
#})

from datasets import load_dataset
from transformers import AutoTokenizer, DataCollatorWithPadding

raw_datasets = load_dataset("glue", "mrpc")
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
def tokenize_function(example):
    return tokenizer(example["sentence1"], example["sentence2"], truncation=True)
tokenized_datasets = raw_datasets.map(tokenize_function, batched=True)
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)

from datasets import load_dataset
#{
#    "data": [
#        {
#            "title": "Terremoto del Sichuan del 2008",
#            "paragraphs": [{...}]
#        }
#    ],
#     "version": "1.1"
#}
squad_it_dataset = load_dataset("json", data_files="SQuAD_it-train.json", field="data")
print(squad_it_dataset) # 加载本地文件会创建一个带有train的DatasetDict 对象
# DatasetDict({
#    train: Dataset({
#        features: ['title', 'paragraphs'],
#        num_rows: 442
#    })
#})
data_files = {"train": "SQuAD_it-train.json", "test": "SQuAD_it-test.json"}
squad_it_dataset = load_dataset("json", data_files=data_files, field="data")
print(squad_it_dataset) # 包括 train 和 test 的 DatasetDict 对象
#DatasetDict({
#    train: Dataset({
#        features: ['title', 'paragraphs'],
#        num_rows: 442
#    })
#    test: Dataset({
#        features: ['title', 'paragraphs'],
#        num_rows: 48
#    })
#})

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForSequenceClassification.from_pretrained(checkpoint)
sequences = ["I've been waiting for a HuggingFace course my whole life.", "So have I!"]
tokens = tokenizer(sequences, padding=True, truncation=True, return_tensors="pt")
print(tokens)
#{
#    'input_ids': tensor([
#        [  101,  1045,  1005,  2310,  2042,  3403,  2005,  1037, 17662, 12172,2607,  2026,  2878,  2166,  1012,   102],
#        [  101,  2061,  2031,  1045,   999,   102,     0,     0,     0,     0, 0,     0,     0,     0,     0,     0]
#    ]),     
#    'attention_mask': tensor([
#        [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
#        [1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
#    ])
#}
output = model(**tokens)
print(output)
# SequenceClassifierOutput(
#    loss=None, 
#    logits=tensor([[-1.5607,  1.6123],[-3.6183,  3.9137]], grad_fn=<AddmmBackward0>), 
#    hidden_states=None, 
#    attentions=None
#)

class Module:
	__call__ : Callable[..., Any] = _call_impl
	def _call_impl(self, *args, **kwargs):
        forward_call = (self._slow_forward if torch._C._get_tracing_state() else self.forward)
       	...
     	result = forward_call(*args, **kwargs)  # dict 可以作为kwargs参数传入
     	... # 涉及到动态图的构建
     	return result

class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin, PeftAdapterMixin):
	def save_pretrained(...)
	@classmethod
    def from_pretrained(...)
class GenerationMixin:
   def generate(inputs,...):> Union[GenerateOutput, torch.LongTensor]:
        ...
        outputs = self(model_input_ids,attention_mask=model_attn,...)
        ...             
class ChatGLMPreTrainedModel(TorchModel, PreTrainedModel):
    ...
class ChatGLMModel(ChatGLMPreTrainedModel):
	def forward(self,input_ids,attention_mask,...):
        ...
class GPT2PreTrainedModel(PreTrainedModel):
    ...
class GPT2LMHeadModel(GPT2PreTrainedModel):
    def __init__(self, config):
        ...
        self.transformer = GPT2Model(config)
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        ...
    def forward(self,input_ids,attention_mask,labels,...):
       transformer_outputs = self.transformer(input_ids,attention_mask,...)
       hidden_states = transformer_outputs[0]
       lm_logits = self.lm_head(hidden_states)
       loss = None
       if labels is not None:
           labels = labels.to(lm_logits.device) # move labels to correct device to enable model parallelism
           shift_logits = lm_logits[..., :-1, :].contiguous() # Shift so that tokens < n predict n
           shift_labels = labels[..., 1:].contiguous()
           loss_fct = CrossEntropyLoss()
           loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
        return ((loss,) + output) if loss is not None else output     

def generate(self,
        prompt_tokens: List[List[int]],  # 输入的提示
        max_gen_len: int,  # 最大生成长度
        temperature: float = 0.6,  # 影响生成文本的随机性
        top_p: float = 0.9,  # 用于决定采样过程中保留的 token 集合的概率阈值
        logprobs: bool = False,  # 是否返回每个 token 的对数概率
        echo: bool = False,  # 是否返回输入的提示
) -> Tuple[List[List[int]], Optional[List[List[float]]]]:
    # ---------------------------初始化长度为 total_len tokens张量，并填充 pad_id----------------------------------
    params = self.model.params
    bsz = len(prompt_tokens)
    assert bsz <= params.max_batch_size, (bsz, params.max_batch_size)

    min_prompt_len = min(len(t) for t in prompt_tokens)
    max_prompt_len = max(len(t) for t in prompt_tokens)
    assert max_prompt_len <= params.max_seq_len
    total_len = min(params.max_seq_len, max_gen_len + max_prompt_len)

    pad_id = self.tokenizer.pad_id
    tokens = torch.full((bsz, total_len), pad_id, dtype=torch.long, device="cuda")
    # 将prompt_tokens中的token复制到tokens张量中。
    for k, t in enumerate(prompt_tokens):
        tokens[k, : len(t)] = torch.tensor(t, dtype=torch.long, device="cuda")
    if logprobs: # 是否返回每个 token 的对数概率
        # 创建一个与tokens相同形状的token_logprobs张量，并用0填充
        token_logprobs = torch.zeros_like(tokens, dtype=torch.float)

    prev_pos = 0
    eos_reached = torch.tensor([False] * bsz, device="cuda")
    input_text_mask = tokens != pad_id
    # -------------------------------------------------------------

    for cur_pos in range(min_prompt_len, total_len):
        # 调用模型的forward方法获取logits
        logits = self.model.forward(tokens[:, prev_pos:cur_pos], prev_pos)
        if logprobs: # 是否返回每个 token 的对数概率
            # 计算token level的logprobs
            token_logprobs[:, prev_pos + 1: cur_pos + 1] = -F.cross_entropy(
                input=logits.transpose(1, 2),
                target=tokens[:, prev_pos + 1: cur_pos + 1],
                reduction="none",
                ignore_index=pad_id,
            )
        # 根据温度参数和top_p参数对logits进行softmax和采样，得到下一个token
        if temperature > 0:
            # sample_top_p函数对probs进行采样
            probs = torch.softmax(logits[:, -1] / temperature, dim=-1)
            next_token = sample_top_p(probs, top_p)
        else:
            # 将logits中概率最大的token作为下一个token。
            next_token = torch.argmax(logits[:, -1], dim=-1)

        next_token = next_token.reshape(-1)
        # only replace token if prompt has already been generated
        next_token = torch.where(
            input_text_mask[:, cur_pos], tokens[:, cur_pos], next_token
        )
        # tokens张量更新
        tokens[:, cur_pos] = next_token
        eos_reached |= (~input_text_mask[:, cur_pos]) & (
                next_token == self.tokenizer.eos_id
        )
        prev_pos = cur_pos
        # 检查是否已经生成了所有的eos token，如果是则停止生成
        if all(eos_reached):
            break

    if logprobs:
        # token_logprobs列表化
        token_logprobs = token_logprobs.tolist()
    out_tokens, out_logprobs = [], []
    for i, toks in enumerate(tokens.tolist()):
        # cut to max gen len
        # 对于 tokens 张量中的每一行（即每一个生成的序列），如果 echo 参数为假，则去掉提示部分
        start = 0 if echo else len(prompt_tokens[i])
        toks = toks[start: len(prompt_tokens[i]) + max_gen_len]
        probs = None
        if logprobs:
            probs = token_logprobs[i][start: len(prompt_tokens[i]) + max_gen_len]
        # cut to eos tok if any
        # 存在结束标记，则去掉结束标记之后的部分
        if self.tokenizer.eos_id in toks:
            eos_idx = toks.index(self.tokenizer.eos_id)
            toks = toks[:eos_idx]
            probs = probs[:eos_idx] if logprobs else None
        out_tokens.append(toks)
        out_logprobs.append(probs)
    # 返回生成的tokens和对数概率（如果logprobs参数为真）
    return (out_tokens, out_logprobs if logprobs else None)

def run_qwen2():
    qwen2config = Qwen2Config(vocab_size=151936,...)
    qwen2model = Qwen2Model(config=qwen2config)
    input_ids = torch.randint(0,qwen2config.vocab_size,(4,30))
    res = qwen2model(input_ids)
    print(type(res))

class Qwen2Model(Qwen2PreTrainedModel):
    def __init__(self, config: Qwen2Config):
        super().__init__(config)
        self.padding_idx = config.pad_token_id
        self.vocab_size = config.vocab_size

        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
        self.layers = nn.ModuleList(
            [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

        self.gradient_checkpointing = False
        # Initialize weights and apply final processing
        self.post_init()
    def forward(input_ids,attention_mask,position_ids,...) -> Union[...]:
        inputs_embeds = self.embed_tokens(input_ids)
        # embed positions
        hidden_states = inputs_embeds
        for idx, decoder_layer in enumerate(self.layers):
            # 将所有的hidden_states保存成tuple
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            # 将hs送入每一层decoder_layer
            layer_outputs = decoder_layer(
                hidden_states,
                attention_mask=attention_mask,
                position_ids=position_ids,
                past_key_value=past_key_value,
                output_attentions=output_attentions,
                use_cache=use_cache,
            )
            # 取出上一层decoder_输出的hs,再传入下一个layer
            # 只要第一个,第二个是cache的一个类，然后进入下一个layer
            hidden_states = layer_outputs[0]
        # 将最后layers输出后的hidden_states进行标准化  
        hidden_states = self.norm(hidden_states)   
        # 加上最后一层的hidden_states
        if output_hidden_states:
            all_hidden_states += (hidden_states,) 

# 为了提高性能，ATTENTION layer 有一些优化实现，比如flash attention，sdpa
QWEN2_ATTENTION_CLASSES = {
    "eager": Qwen2Attention,  # 一般情况下是这个
    "flash_attention_2": Qwen2FlashAttention2,
    "sdpa": Qwen2SdpaAttention,
}
class Qwen2DecoderLayer(nn.Module):
    def __init__(self, config: Qwen2Config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = QWEN2_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)

        self.mlp = Qwen2MLP(config)
        # input_layernorm和post_attention_layernorm内容是一样的，只是应用的顺序不一样。
        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
    def forward(hidden_states,attention_mask,position_ids,...) -> Union[...]:
        residual = hidden_states 
        #  标准化后送入attn
        hidden_states = self.input_layernorm(hidden_states)  # RMSNorm标准化
        # Self Attention
        hidden_states, self_attn_weights, present_key_value = self.self_attn(  
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_value=past_key_value,
            output_attentions=output_attentions,
            use_cache=use_cache,
            **kwargs,
        )
        # 残差与新的hidden_states相加
        hidden_states = residual + hidden_states
        # Fully Connected
        residual = hidden_states
        # 同样的RMSNorm标准化
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        outputs = (hidden_states,)
        return outputs

class Qwen2Attention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""
    def __init__(self, config: Qwen2Config):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.hidden_size // self.num_heads
        self.num_key_value_heads = config.num_key_value_heads # 键值对的头数
        self.num_key_value_groups = self.num_heads // self.num_key_value_heads # 键值对的组数
        self.max_position_embeddings = config.max_position_embeddings
        self.rope_theta = config.rope_theta
        self.is_causal = True
        self.attention_dropout = config.attention_dropout

        if (self.head_dim * self.num_heads) != self.hidden_size:
            raise ValueError(
                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
                f" and `num_heads`: {self.num_heads})."
            )
        # 后续LoRa也基本都对q_proj/k_proj/v_proj/o_proj 四个Linear操作动的刀子
        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias)
        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias)
        
        self.rotary_emb = Qwen2RotaryEmbedding(
            self.head_dim,
            max_position_embeddings=self.max_position_embeddings,
            base=self.rope_theta,
        )
    def forward(hidden_states,attention_mask,position_ids,...) -> Union[...]:
        # 获取形状信息,hidden_states输入的为(bs,T,hd)
        bsz, q_len, _ = hidden_states.size()
        # 对hidden_states进行Linear生成query、key、value
        query_states = self.q_proj(hidden_states)
        key_states = self.k_proj(hidden_states)
        value_states = self.v_proj(hidden_states)
        # reshape多头处理--分块--(bs,T,heads,hd_d)
        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
        # 将旋转位置嵌入应用于查询和键张量。使用了旋转位置嵌入的余弦和正弦部分，将它们与查询和键张量相乘，并将结果相加，从而实现旋转位置嵌入的效果
        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
        # 先将key_states和value_states重复了num_key_value_groups次
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
        # 使用dot attn实现q*kT/hd_d^0.5
        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) # # 多维矩阵乘法
        # 然后 attn_weights 加上 attention_mask，实现读取顺序
        attn_weights = attn_weights + attention_mask
        # softmax + dropout + values_states相乘
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
        attn_output = torch.matmul(attn_weights, value_states)
        # 转置，修改形状等reshape操作
        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
        # 最后在进行一次o_proj
        attn_output = self.o_proj(attn_output)
        # 返回结果
        return attn_output, attn_weights, past_key_value

class Qwen2MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        # 这俩不必多说
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size

        # 三个全连接层
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = ACT2FN[config.hidden_act]

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj

class Qwen2RMSNorm(nn.Module):  # 标准化层
    def __init__(self, hidden_size, eps=1e-6):
        """
        Qwen2RMSNorm is equivalent to T5LayerNorm
        """
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.variance_epsilon = eps

    def forward(self, hidden_states):
        input_dtype = hidden_states.dtype
        hidden_states = hidden_states.to(torch.float32)
        variance = hidden_states.pow(2).mean(-1, keepdim=True)
        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
        return self.weight * hidden_states.to(input_dtype)

import torch

my_model = ModelClass(...)
state_dict = torch.load(checkpoint_file)
my_model.load_state_dict(state_dict)

import torch
from accelerate import init_empty_weights, load_checkpoint_and_dispatch

with init_empty_weights():
    model = MyModel(...)

model = load_checkpoint_and_dispatch(
    model, checkpoint=checkpoint_file, device_map="auto" # device_map=auto会根据可用资源自动确定模型每一层的放置位置
)

input = torch.randn(2,3)
input = input.to("cuda")
output = model(input)

class Opt(GPT):
    """Opt"""
    @classmethod
    def _create_config(cls, ckpt_path: str):
        offset = 2
        config_dict = get_config_from_path(ckpt_path)
        config = GptInitModelParameters(
            head_num=config_dict['num_attention_heads'],
            size_per_head=config_dict['hidden_size'] // config_dict['num_attention_heads'], # 每个注意力头的维度大小
            layer_num=config_dict.get('num_hidden_layers', 12),
            vocab_size=config_dict['vocab_size'],   # 词表的大小
            max_seq_len=config_dict['max_position_embeddings'] + offset # 模型能够处理的最长序列长度
        )
        config.layernorm_type = 'pre_layernorm'
        config.norm_type = "layernorm"
        config.has_post_decoder_layernorm = True
        config.hidden_size = config_dict['hidden_size']
        config.inter_size = config_dict["ffn_dim"]
        config.has_positional_encoding = True
        config.activation_type = 'relu'
        config.add_special_tokens = True
        config.special_tokens.eos_token_id = config_dict.get('eos_token_id', 2)
        config.special_tokens.pad_token_id = config_dict.get('pad_token_id', 1)
        config.special_tokens.bos_token_id = config_dict.get('bos_token_id', 2)
        config.head_num_kv = config.head_num
        return config

class OptWeightInfo(ModelDeployWeightInfo):
    def _get_weight_info(self):
        layer_weights = [
            # * Attention之前 layer_norm部分
            WeightInfo(W.pre_ln_gamma, [CkptWeightInfo('model.decoder.layers.{i}.self_attn_layer_norm.weight', identity)], identity),
            WeightInfo(W.pre_ln_beta, [CkptWeightInfo('model.decoder.layers.{i}.self_attn_layer_norm.bias', identity)], identity),

            # *  Attention
            WeightInfo(W.attn_qkv_w, [
                CkptWeightInfo('model.decoder.layers.{i}.self_attn.q_proj.weight', identity),
                CkptWeightInfo('model.decoder.layers.{i}.self_attn.k_proj.weight', identity),
                CkptWeightInfo('model.decoder.layers.{i}.self_attn.v_proj.weight', identity),
                ], functools.partial(merge_qkv_hf)),
            ...