class LlavaMetaModel:
    def __init__(self, config):
        super(LlavaMetaModel, self).__init__(config)

        if hasattr(config, "mm_vision_tower"):
            self.vision_tower = build_vision_tower(config, delay_load=True)
            self.mm_projector = build_vision_projector(config)

    def get_vision_tower(self):
        vision_tower = getattr(self, 'vision_tower', None)
        if type(vision_tower) is list:
            vision_tower = vision_tower[0]
        return vision_tower

    def initialize_vision_modules(self, model_args, fsdp=None):
        vision_tower = model_args.vision_tower
        mm_vision_select_layer = model_args.mm_vision_select_layer
        mm_vision_select_feature = model_args.mm_vision_select_feature
        pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter

        self.config.mm_vision_tower = vision_tower

        if self.get_vision_tower() is None:
            vision_tower = build_vision_tower(model_args)

            if fsdp is not None and len(fsdp) > 0:
                self.vision_tower = [vision_tower]
            else:
                self.vision_tower = vision_tower
        else:
            if fsdp is not None and len(fsdp) > 0:
                vision_tower = self.vision_tower[0]
            else:
                vision_tower = self.vision_tower
            vision_tower.load_model()

        self.config.use_mm_proj = True
        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
        self.config.mm_hidden_size = vision_tower.hidden_size
        self.config.mm_vision_select_layer = mm_vision_select_layer
        self.config.mm_vision_select_feature = mm_vision_select_feature

        if getattr(self, 'mm_projector', None) is None:
            self.mm_projector = build_vision_projector(self.config)
        else:
            # In case it is frozen by LoRA
            for p in self.mm_projector.parameters():
                p.requires_grad = True

        if pretrain_mm_mlp_adapter is not None:
            mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
            def get_w(weights, keyword):
                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}

            self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'))

class LlavaMetaForCausalLM(ABC):

    @abstractmethod # 装饰器
    def get_model(self): # 由子类给出
        pass

    def get_vision_tower(self):
        return self.get_model().get_vision_tower()

    def encode_images(self, images): # 用 vision_tower 和 mm_projector 对 image 编码
        image_features = self.get_model().get_vision_tower()(images)
        image_features = self.get_model().mm_projector(image_features)
        return image_features

    def prepare_inputs_labels_for_multimodal(
        self, input_ids, position_ids, attention_mask, past_key_values, labels, images
    ): # 
        vision_tower = self.get_vision_tower()
        if vision_tower is None or images is None or input_ids.shape[1] == 1:
            if past_key_values is not None and vision_tower is not None and images is not None and input_ids.shape[1] == 1:
                target_shape = past_key_values[-1][-1].shape[-2] + 1
                attention_mask = torch.cat((attention_mask, torch.ones(
                    (attention_mask.shape[0], target_shape - attention_mask.shape[1]),
                    dtype=attention_mask.dtype,
                    device=attention_mask.device
                )), dim=1)
                position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
            return input_ids, position_ids, attention_mask, past_key_values, None, labels

        if type(images) is list or images.ndim == 5:
            concat_images = torch.cat([image for image in images], dim=0)
            image_features = self.encode_images(concat_images)
            split_sizes = [image.shape[0] for image in images]
            image_features = torch.split(image_features, split_sizes, dim=0)
            image_features = [x.flatten(0, 1).to(self.device) for x in image_features]
        else:
            image_features = self.encode_images(images).to(self.device)

        # TODO: image start / end is not implemented here to support pretraining.
        if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
            raise NotImplementedError

        # Let's just add dummy tensors if they do not exist,
        # it is a headache to deal with None all the time.
        # But it is not ideal, and if you have a better idea,
        # please open an issue / submit a PR, thanks.
        _labels = labels
        _position_ids = position_ids
        _attention_mask = attention_mask
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
        else:
            attention_mask = attention_mask.bool()
        if position_ids is None:
            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
        if labels is None:
            labels = torch.full_like(input_ids, IGNORE_INDEX)

        # remove the padding using attention_mask -- TODO: double check
        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]

        new_input_embeds = []
        new_labels = []
        cur_image_idx = 0
        for batch_idx, cur_input_ids in enumerate(input_ids):
            num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
            if num_images == 0:
                cur_image_features = image_features[cur_image_idx]
                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
                new_input_embeds.append(cur_input_embeds)
                new_labels.append(labels[batch_idx])
                cur_image_idx += 1
                continue

            image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
            cur_input_ids_noim = []
            cur_labels = labels[batch_idx]
            cur_labels_noim = []
            for i in range(len(image_token_indices) - 1):
                cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
                cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
            split_sizes = [x.shape[0] for x in cur_labels_noim]
            cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
            cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
            cur_new_input_embeds = []
            cur_new_labels = []

            for i in range(num_images + 1):
                cur_new_input_embeds.append(cur_input_embeds_no_im[i])
                cur_new_labels.append(cur_labels_noim[i])
                if i < num_images:
                    cur_image_features = image_features[cur_image_idx]
                    cur_image_idx += 1
                    cur_new_input_embeds.append(cur_image_features)
                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))

            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
            cur_new_labels = torch.cat(cur_new_labels)

            new_input_embeds.append(cur_new_input_embeds)
            new_labels.append(cur_new_labels)
        

        # Truncate sequences to max length as image embeddings can make the sequence longer
        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
        if tokenizer_model_max_length is not None:
            new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]
            

        # Combine them
        max_len = max(x.shape[0] for x in new_input_embeds)
        batch_size = len(new_input_embeds)

        new_input_embeds_padded = []
        new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
        attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
        position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)

        for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
            cur_len = cur_new_embed.shape[0]
            if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
                new_input_embeds_padded.append(torch.cat((
                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
                    cur_new_embed
                ), dim=0))
                if cur_len > 0:
                    new_labels_padded[i, -cur_len:] = cur_new_labels
                    attention_mask[i, -cur_len:] = True
                    position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
            else:
                new_input_embeds_padded.append(torch.cat((
                    cur_new_embed,
                    torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
                ), dim=0))
                if cur_len > 0:
                    new_labels_padded[i, :cur_len] = cur_new_labels
                    attention_mask[i, :cur_len] = True
                    position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)

        new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)

        if _labels is None:
            new_labels = None
        else:
            new_labels = new_labels_padded

        if _attention_mask is None:
            attention_mask = None
        else:
            attention_mask = attention_mask.to(dtype=_attention_mask.dtype)

        if _position_ids is None:
            position_ids = None

        return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels

    def initialize_vision_tokenizer(self, model_args, tokenizer):
        if model_args.mm_use_im_patch_token:
            tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

        if model_args.mm_use_im_start_end:
            num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
            self.resize_token_embeddings(len(tokenizer))

            if num_new_tokens > 0:
                input_embeddings = self.get_input_embeddings().weight.data
                output_embeddings = self.get_output_embeddings().weight.data

                input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
                    dim=0, keepdim=True)
                output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
                    dim=0, keepdim=True)

                input_embeddings[-num_new_tokens:] = input_embeddings_avg
                output_embeddings[-num_new_tokens:] = output_embeddings_avg

            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = True
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False

            if model_args.pretrain_mm_mlp_adapter:
                mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
                embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
                assert num_new_tokens == 2
                if input_embeddings.shape == embed_tokens_weight.shape:
                    input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
                elif embed_tokens_weight.shape[0] == num_new_tokens:
                    input_embeddings[-num_new_tokens:] = embed_tokens_weight
                else:
                    raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
        elif model_args.mm_use_im_patch_token:
            if model_args.tune_mm_mlp_adapter:
                for p in self.get_input_embeddings().parameters():
                    p.requires_grad = False
                for p in self.get_output_embeddings().parameters():
                    p.requires_grad = False

from ..llava_arch import LlavaMetaModel, LlavaMetaForCausalLM

class LlavaLlamaModel(LlavaMetaModel, LlamaModel):
    config_class = LlavaConfig

    def __init__(self, config: LlamaConfig):
        super(LlavaLlamaModel, self).__init__(config)

# transformer 库中的类 LlamaForCausalLM
class LlamaForCausalLM(LlamaPreTrainedModel):
    _tied_weights_keys = ["lm_head.weight"]

    def __init__(self, config):
        super().__init__(config)
        self.model = LlamaModel(config)
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

class LlavaLlamaForCausalLM(LlamaForCausalLM, LlavaMetaForCausalLM):
    config_class = LlavaConfig

    def __init__(self, config):
        super(LlamaForCausalLM, self).__init__(config) # LlamaForCausalLM 是 transformer 库中的类, 这里显式调用了 LlamaForCausalLM 类的 __init__ 方法, 也就是说这里的 __init__ 调用结束之后, 有两个 self.model
        self.model = LlavaLlamaModel(config)
        self.pretraining_tp = config.pretraining_tp
        self.vocab_size = config.vocab_size
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) # LLM 的头, 输出维度为词表大小

        # Initialize weights and apply final processing
        self.post_init()

    def get_model(self):
        return self.model

    def forward( # TODO: 补充每个参数的含义
        self,
        input_ids: torch.LongTensor = None, # 
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None, # 用于加速
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None, # 用于加速
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        images: Optional[torch.FloatTensor] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        
        if inputs_embeds is None:
            (
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                inputs_embeds,
                labels
            ) = self.prepare_inputs_labels_for_multimodal(
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                labels,
                images
            ) # prepare_inputs_labels_for_multimodal() 函数是 llava_arch.py 文件中类 LlavaMetaForCausalLM 中的方法, 作用的 vision encoder + projector + text encoder + vision-text tokens concat

        return super().forward(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            labels=labels,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict
        )
    # 这里的 forward 函数返回的是父类的 forward() 函数, 根据 Python 的方法解析顺序 (Method Resolution Order, MRO), MRO 是确定多重继承中方法调用顺序的算法。在 Python 中，MRO 是由 C3 线性化算法来计算的, 也就是说这里是按照 LlamaForCausalLM, LlavaMetaForCausalLM 的先后顺序解析的, 所以这里调用的实际上是类 LlamaForCausalLM 的 forward() 函数. 除此之外, 第二个父类 LlavaMetaForCausalLM 里没有 forward() 方法.

    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
        '''
        
        '''
        images = kwargs.pop("images", None)
        _inputs = super().prepare_inputs_for_generation(
            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
        )
        if images is not None:
            _inputs['images'] = images
        return _inputs

def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: bool = False,
        use_cache: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        bsz, q_len, _ = hidden_states.size()

        if self.pretraining_tp > 1:
            key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.pretraining_tp
            query_slices = self.q_proj.weight.split((self.num_heads * self.head_dim) // self.pretraining_tp, dim=0)
            key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
            value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)

            query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.pretraining_tp)]
            query_states = torch.cat(query_states, dim=-1)

            key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.pretraining_tp)]
            key_states = torch.cat(key_states, dim=-1)

            value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.pretraining_tp)]
            value_states = torch.cat(value_states, dim=-1)

        else:
            query_states = self.q_proj(hidden_states)
            key_states = self.k_proj(hidden_states)
            value_states = self.v_proj(hidden_states)

        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)

        kv_seq_len = key_states.shape[-2]
        if past_key_value is not None:
            kv_seq_len += past_key_value[0].shape[-2]
        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)

        if past_key_value is not None:
            # reuse k, v, self_attention
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)

        past_key_value = (key_states, value_states) if use_cache else None

        # repeat k/v heads if n_kv_heads < n_heads
        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)

        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)

        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
            raise ValueError(
                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
                f" {attn_weights.size()}"
            )

        if attention_mask is not None:
            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
                )
            attn_weights = attn_weights + attention_mask

        # upcast attention to fp32
        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
        attn_output = torch.matmul(attn_weights, value_states)

        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
            raise ValueError(
                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
            )

        attn_output = attn_output.transpose(1, 2).contiguous()
        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)

        if self.pretraining_tp > 1:
            attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
            o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.pretraining_tp, dim=1)
            attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.pretraining_tp)])
        else:
            attn_output = self.o_proj(attn_output)

        if not output_attentions:
            attn_weights = None

        return attn_output, attn_weights, past_key_value

# 这个在 1541 行左右
return self.greedy_search(
    input_ids,
    logits_processor=logits_processor,
    stopping_criteria=stopping_criteria,
    pad_token_id=generation_config.pad_token_id,
    eos_token_id=generation_config.eos_token_id,
    output_scores=generation_config.output_scores,
    return_dict_in_generate=generation_config.return_dict_in_generate,
    synced_gpus=synced_gpus,
    streamer=streamer,
    **model_kwargs,
)

def greedy_search(
    self,
    input_ids: torch.LongTensor,
    logits_processor: Optional[LogitsProcessorList] = None,
    stopping_criteria: Optional[StoppingCriteriaList] = None,
    max_length: Optional[int] = None,
    pad_token_id: Optional[int] = None,
    eos_token_id: Optional[Union[int, List[int]]] = None,
    output_attentions: Optional[bool] = None, # 传入该参数为 True 
    output_hidden_states: Optional[bool] = None,
    output_scores: Optional[bool] = None,
    return_dict_in_generate: Optional[bool] = None, # 传入该参数为 True
    synced_gpus: bool = False,
    streamer: Optional["BaseStreamer"] = None,
    **model_kwargs,
) -> Union[GreedySearchOutput, torch.LongTensor]:

if return_dict_in_generate:
    if output_scores:
        scores += (next_tokens_scores,)
        if output_attentions:
            decoder_attentions += (
                (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
            )

# 2433 行左右
if return_dict_in_generate:
    if self.config.is_encoder_decoder:
		...
    else: # 由于是 decoder only 结构, 因此返回的是下面的内容
        return GreedySearchDecoderOnlyOutput(
            sequences=input_ids,
            scores=scores,
            attentions=decoder_attentions,
            hidden_states=decoder_hidden_states,
        )

@dataclass
class GreedySearchDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using greedy search. 
	"""

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None

while True:
    if synced_gpus:
        # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
        # The following logic allows an early break if all peers finished generating their sequence
        this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
        # send 0.0 if we finished, 1.0 otherwise
        dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
        # did all peers finish? the reduced sum will be 0.0 then
        if this_peer_finished_flag.item() == 0.0:
            break

            # prepare model inputs
            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=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                continue  # don't waste resources running the code we don't need

                next_token_logits = outputs.logits[:, -1, :]

                # pre-process distribution
                next_tokens_scores = logits_processor(input_ids, next_token_logits)

                # Store scores, attentions and hidden_states when required
                if return_dict_in_generate:
                    if output_scores:
                        scores += (next_tokens_scores,)
                        if output_attentions:
                            decoder_attentions += (
                                (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                            )
                            if self.config.is_encoder_decoder:
                                cross_attentions += (outputs.cross_attentions,)

                                if output_hidden_states:
                                    decoder_hidden_states += (
                                        (outputs.decoder_hidden_states,)
                                        if self.config.is_encoder_decoder
                                        else (outputs.hidden_states,)
                                    )

                                    # argmax
                                    next_tokens = torch.argmax(next_tokens_scores, dim=-1)

                                    # finished sentences should have their next token be a padding token
                                    if eos_token_id is not None:
                                        if pad_token_id is None:
                                            raise ValueError("If `eos_token_id` is defined, make sure that `pad_token_id` is defined.")
                                            next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences)

                                            # update generated ids, model inputs, and length for next step
                                            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
                                            if streamer is not None:
                                                streamer.put(next_tokens.cpu())
                                                model_kwargs = self._update_model_kwargs_for_generation(
                                                    outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
                                                )

                                                # if eos_token was found in one sentence, set sentence to finished
                                                if eos_token_id_tensor is not None:
                                                    unfinished_sequences = unfinished_sequences.mul(
                                                        next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
                                                    )

                                                    # stop when each sentence is finished
                                                    if unfinished_sequences.max() == 0:
                                                        this_peer_finished = True

                                                        # stop if we exceed the maximum length
                                                        if stopping_criteria(input_ids, scores):
                                                            this_peer_finished = True

                                                            if this_peer_finished and not synced_gpus:
                                                                break

CausalLMOutputWithPast(
    loss=loss,
    logits=logits,
    past_key_values=outputs.past_key_values,
    hidden_states=outputs.hidden_states,
    attentions=outputs.attentions,
)

outputs = self.model(
    input_ids=input_ids,
    attention_mask=attention_mask,
    position_ids=position_ids,
    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 = outputs[0]
if self.pretraining_tp > 1:
    lm_head_slices = self.lm_head.weight.split(self.vocab_size // self.pretraining_tp, dim=0)
    logits = [F.linear(hidden_states, lm_head_slices[i]) for i in range(self.pretraining_tp)]
    logits = torch.cat(logits, dim=-1)
    else:
        logits = self.lm_head(hidden_states)
        logits = logits.float()

self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

# 这里是输出 token 之间的 attention 矩阵
attn_matrices=[]
out_token_num=len(out.attentions)
layer_num=len(out.attentions[0])
input_token_num=out.attentions[0][0].shape[2]
for i in range(layer_num):
    attn_mat=torch.zeros((out_token_num,out_token_num)).cuda()
    # 每一层得到最终的 attn_matrix
    for j,attn in enumerate(out.attentions):
        if j>0:
            # attn : 32*([1,32,1,627+j])
            attn_max=attn[i].max(1).values.data.squeeze()
            attn_max=attn_max / attn_max.sum(-1, keepdim=True)
            # attn_max : 32*([627+j])
            attn_mat[j-1,:j]=attn_max[input_token_num:] # 获取后 j 个元素, 赋值给 attn_mat 所以为j-1的行的前j-1个元素
            # print(attn_mat)
            attn_matrices.append(attn_mat)

            for i in range(layer_num):
                vis(attn_matrices[i].cpu(), i)