Qwen系列源码解析

原创已于 2025-10-22 17:09:58 修改 · 296 阅读

5 ·

CC 4.0 BY-SA版权

文章标签：

#python #qwen

于 2025-10-21 13:12:26 首次发布

Qwen

Qwen-72B模型

Hyperparameter	Value
n_layers	80
n_heads	64
d_model	8192
vocab_size	151851
sequence_length	32768

Qwen的tokenizer

Qwen-VL

LLM Backbone：Qwen-7B Base 模型
Visual Encoder：ViT架构使用OpenCLIP预训练好的ViT-bigG模型（经过2B的训练数据训出来的ViT模型）。
Adapter随机初始化的单层cross-attention，使用一组可学习的query向量将vit的图像特征作为kv，将视觉特征压缩到固定的256长度。
+二维绝对位置编码（三角位置编码）

输入数据：图片<img></img>，bounding box用“左上-右下”表示，统一归一化道(0,1000)，描述bounding box的文本用<ref></ref>表示。
在这里插入图片描述
训练：
第一阶段：单任务大规模预训练（Pre-training ），主要使用大量网上抓取和内部的图文pair数据做预训练，训练数据有1.4B，英文数据占比77.3%，中文占比22.7%，训练数据的图片统一处理成224*224的尺寸。该阶段LLM模型参数是frozen的，ViT和Cross-Attention层的参数是激活更新的，这个阶段主要通过大规模数据训练模型的vision模态对齐语言模型的能力。

第二阶段：多任务预训练（Multi-task Pre-training），这个阶段使用了更高分辨率、更高质量的数据，同时引入图文混排的数据。该阶段是个多任务的预训练阶段，包括7个任务，其中有6个Vision任务（包括Captioning ，VQA，grounding等）和1个文本生成任务，这个阶段模型是全参数激活的。该阶段之所以引入文本生成任务，主要是为了保证模型的通用文本处理能力。该阶段的训练数据，Vision数据的分辨率从224*224提升到448*448，数据做了精选处理，包括多模态数据69M和文本数据7.8M。第二阶段的数据量比第一阶段少了2个量级。该阶段训练完成后，最终产出了Qwen-VL base模型。

第三阶段：指令微调（Supervised Fine-tuning)，主要提升模型的指令遵循能力和对话能力。在这个阶段作者对数据做了些数据增强，通过人工标注、模型生成和策略拼接等方式构造多模态的多轮会话数据。该阶段指令集数据共收集了350K。

VIT

在这里插入图片描述

标准的ViT实现输入图片先被调整成长宽比1:1的正方形，再分割成固定大小的patch块。（分辨率和pathch的size都是固定的）
卷积操作=(输入长度+2*padding-卷积核大小)/步长+1
Patch Embedding：使用卷积核大小为16×16，步长(stride)为16，卷积核个数为768，卷积后再展平，size变化为：[224, 224, 3] -> [14, 14, 768] -> [196, 768]。
+[cls]+使用可学习的绝对位置编码

Image预处理：训练阶段，随机裁剪输入图像并将裁剪后的图像调整为224*224，以一定概率水平翻转，再转化为Tensor格式进行标准化处理。
预测阶段，将输入图像调整为256*256，对调整后的图像进行中心裁剪保留中心区域224*224，再转化为Tensor格式进行标准化处理。

如何想要改变微调图像的分辨率大小，需要对预训练的位置嵌入执行 2D 插值，以扩展到微调尺寸：获取新位置token数目，计算原始patch个数和新位置patch个数，将patch位置嵌入转为2D维度，使用双线性插值，改为新的patch个数，将cls和patch位置嵌入合并。

Qwen-VL参考
 ViT源码细节
 Qwen2VL流程

Qwen2

configuration_qwen2.py

# Default tensor parallel plan for base model `Qwen2`
    base_model_tp_plan = {
        "layers.*.self_attn.q_proj": "colwise",
        "layers.*.self_attn.k_proj": "colwise",
        "layers.*.self_attn.v_proj": "colwise",
        "layers.*.self_attn.o_proj": "rowwise",
        "layers.*.mlp.gate_proj": "colwise",
        "layers.*.mlp.up_proj": "colwise",
        "layers.*.mlp.down_proj": "rowwise",
    }
    base_model_pp_plan = {
        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
        "norm": (["hidden_states"], ["hidden_states"]),
    }

张量并行TP：将模型权重矩阵拆分到多个GPUGPU上，"colwise"按列拆分，"rowwise"按行拆分。
流水线管道并行PP：将模型的不同层分配到不同GPU上，GPU0将input_ids输出为inputs_embeds传给下一个GPU，GPU1输入[“hidden_states”, “attention_mask”]输出[“hidden_states”]，以此类推。

modular_qwen2.py

class Qwen2MLP(LlamaMLP):
    def __init__(self, config):
        super().__init__(config)
        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)

class LlamaMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.hidden_size = config.hidden_size
        self.intermediate_size = config.intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
        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

常规MLP代码 pass

class Qwen2Attention(LlamaAttention):
    def __init__(self, config: Qwen2Config, layer_idx: int):
        super().__init__(config, layer_idx)
        self.q_proj = nn.Linear(config.hidden_size, config.num_attention_heads * self.head_dim, bias=True)
        self.k_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=True)
        self.v_proj = nn.Linear(config.hidden_size, config.num_key_value_heads * self.head_dim, bias=True)
        self.o_proj = nn.Linear(config.num_attention_heads * self.head_dim, config.hidden_size, bias=False)
        self.sliding_window = config.sliding_window if config.layer_types[layer_idx] == "sliding_attention" else None

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_values: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[FlashAttentionKwargs],
    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_values is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            sliding_window=self.sliding_window,  # main diff with Llama
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights
def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
    
    cos = cos.unsqueeze(unsqueeze_dim)
    sin = sin.unsqueeze(unsqueeze_dim)
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed

class LlamaAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(self, config: LlamaConfig, layer_idx: int):
        super().__init__()
        self.config = config
        self.layer_idx = layer_idx
        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
        self.scaling = self.head_dim**-0.5
        self.attention_dropout = config.attention_dropout
        self.is_causal = True

        self.q_proj = nn.Linear(
            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
        )
        self.k_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.v_proj = nn.Linear(
            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
        )
        self.o_proj = nn.Linear(
            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
        )

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        position_embeddings: tuple[torch.Tensor, torch.Tensor],
        attention_mask: Optional[torch.Tensor],
        past_key_values: Optional[Cache] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> tuple[torch.Tensor, torch.Tensor]:
        input_shape = hidden_states.shape[:-1]
        hidden_shape = (*input_shape, -1, self.head_dim)

        query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
        value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)

        cos, sin = position_embeddings
        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

        if past_key_values is not None:
            # sin and cos are specific to RoPE models; cache_position needed for the static cache
            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)

        attention_interface: Callable = eager_attention_forward
        if self.config._attn_implementation != "eager":
            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]

        attn_output, attn_weights = attention_interface(
            self,
            query_states,
            key_states,
            value_states,
            attention_mask,
            dropout=0.0 if not self.training else self.attention_dropout,
            scaling=self.scaling,
            **kwargs,
        )

        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
        attn_output = self.o_proj(attn_output)
        return attn_output, attn_weights
    
def eager_attention_forward(
    module: nn.Module,
    query: torch.Tensor,
    key: torch.Tensor,
    value: torch.Tensor,
    attention_mask: Optional[torch.Tensor],
    scaling: float,
    dropout: float = 0.0,
    **kwargs: Unpack[TransformersKwargs],
):
    key_states = repeat_kv(key, module.num_key_value_groups)
    value_states = repeat_kv(value, module.num_key_value_groups)

    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
    if attention_mask is not None:
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
    attn_output = torch.matmul(attn_weights, value_states)
    attn_output = attn_output.transpose(1, 2).contiguous()

    return attn_output, attn_weights
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)

这里提到sliding_window，具体滑动窗口计算注意力的实现代码没找到，但应该是仅计算其前后sliding_window//2范围内的注意力分数（超出窗口的位置被掩码屏蔽）。Qwen2应该就是对query和key加上rope位置编码使用滑动窗口的MHA。

class Qwen2RMSNorm(nn.Module):
        def __init__(self, hidden_size, eps: float = 1e-6) -> None:
            """
            Qwen2RMSNorm is equivalent to T5LayerNorm
            """
            super().__init__()
            self.weight = nn.Parameter(torch.ones(hidden_size))
            self.variance_epsilon = eps

        def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
            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)

        def extra_repr(self):
            return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"

常规RMSNorm

class Qwen2Model(MistralModel):
    def __init__(self, config: Qwen2Config):
        super().__init__(config)
        self.has_sliding_layers = "sliding_attention" in self.config.layer_types

    @check_model_inputs
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        # It may already have been prepared by e.g. `generate`
        if not isinstance(causal_mask_mapping := attention_mask, dict):
            # Prepare mask arguments
            mask_kwargs = {
                "config": self.config,
                "input_embeds": inputs_embeds,
                "attention_mask": attention_mask,
                "cache_position": cache_position,
                "past_key_values": past_key_values,
                "position_ids": position_ids,
            }
            # Create the masks
            causal_mask_mapping = {
                "full_attention": create_causal_mask(**mask_kwargs),
            }
            # The sliding window alternating layers are not always activated depending on the config
            if self.has_sliding_layers:
                causal_mask_mapping["sliding_attention"] = create_sliding_window_causal_mask(**mask_kwargs)

        hidden_states = inputs_embeds

        # create position embeddings to be shared across the decoder layers
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask_mapping[decoder_layer.attention_type],
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )

        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
        )
class MistralModel(MistralPreTrainedModel):
    def __init__(self, config: MistralConfig):
        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(
            [MistralDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
        )
        self.norm = MistralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.rotary_emb = MistralRotaryEmbedding(config=config)
        self.gradient_checkpointing = False

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

    @check_model_inputs
    @auto_docstring
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        cache_position: Optional[torch.LongTensor] = None,
        **kwargs: Unpack[TransformersKwargs],
    ) -> BaseModelOutputWithPast:
        if (input_ids is None) ^ (inputs_embeds is not None):
            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids)

        if use_cache and past_key_values is None:
            past_key_values = DynamicCache(config=self.config)

        if cache_position is None:
            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
            cache_position = torch.arange(
                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
            )

        if position_ids is None:
            position_ids = cache_position.unsqueeze(0)

        mask_function = create_causal_mask if self.config.sliding_window is None else create_sliding_window_causal_mask
        causal_mask = mask_function(
            config=self.config,
            input_embeds=inputs_embeds,
            attention_mask=attention_mask,
            cache_position=cache_position,
            past_key_values=past_key_values,
            position_ids=position_ids,
        )

        hidden_states = inputs_embeds
        position_embeddings = self.rotary_emb(hidden_states, position_ids)

        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
            hidden_states = decoder_layer(
                hidden_states,
                attention_mask=causal_mask,
                position_ids=position_ids,
                past_key_values=past_key_values,
                use_cache=use_cache,
                cache_position=cache_position,
                position_embeddings=position_embeddings,
                **kwargs,
            )
        hidden_states = self.norm(hidden_states)
        return BaseModelOutputWithPast(
            last_hidden_state=hidden_states,
            past_key_values=past_key_values if use_cache else None,
        )
class MistralDecoderLayer(GradientCheckpointingLayer):
    def __init__(self, config: MistralConfig, layer_idx: int):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.self_attn = MistralAttention(config=config, layer_idx=layer_idx)
        self.mlp = MistralMLP(config)
        self.input_layernorm = MistralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.post_attention_layernorm = MistralRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        past_key_values: Optional[Cache] = None,
        use_cache: Optional[bool] = False,
        cache_position: Optional[torch.LongTensor] = None,
        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
        **kwargs: Unpack[TransformersKwargs],
    ) -> torch.Tensor:
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        # Self Attention
        hidden_states, _ = self.self_attn(
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            cache_position=cache_position,
            position_embeddings=position_embeddings,
            **kwargs,
        )
        hidden_states = residual + hidden_states

        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states

Qwen2模型架构，
补一下Mistral 7B模型好像是第一个使用Sliding Window Attention的，

Qwen2-VL

动态分辨率

保留原始图片宽高比，再resize到适当大小，满足[min_pixel,max_pixel]区间再做卷积patch处理。将不同图片放到一个序列里计算，根据每个图片的起止token位置计算attention mask进行隔离。
引入2D-rope旋转位置编码：对(x,y)将维度为d对输入向量分成两半，前一半用x对一维rope矩阵，后一半用y对一维rope矩阵处理，然后拼接在一起就是二维。
补代码实现
投影层用压缩Vision token + MLP Adapter：
concat空间位置临近的patch特征，再经过2层的MLP线性变换，将n压缩为n/4。引入<|vision_start>|和<|vision_end>|标记视觉token。
在这里插入图片描述

Q：为什么不使用Cross-Attention的架构？
A：cross-attn适合处理固定长度的kv，Qwen2-vl动态分辨率导致每个图片的token序列变长，kv长度变化导致计算效率降低（padding浪费）以及（大图和小图都变成query长度难以学习稳定的特征压缩映射），所以无法使用cross-attn做特征压缩处理。

M-rope

文本：三个维度值保持一致
图像：x,y,T时间维度保持固定
视频：对(x,y,z)将维度d的输入向量分成三份，第一份用x的一维rope矩阵处理，第二份用y的一维rope矩阵处理，第三份用z的一维rope矩阵处理，再拼接在一起。

def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
    mrope_section = mrope_section * 2
    cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
        unsqueeze_dim
    )
    sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
        unsqueeze_dim
    )

    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    return q_embed, k_embed

统一图像和视频：图片复制一份变为时序为2的帧序列，视频以每秒两帧的速率对视频进行采样，最终可采样偶数个帧序列。

Qwen2.5-VL

在这里插入图片描述
Qwen-VL将box的坐标点做(0,1000)的规范化处理，Qwen2.5-VL不进行坐标归一化，而是使用实际的像素点来表示坐标，这样能是模型学习到图像的真实尺寸信息。
Qwen2-VL的M-rope对于一个模态的起始位置的position ID，是相对于前面模态三维ID中最大的ID再加1得到。这对于时序维度的ID处理其实是不合理的，视频的时间是有绝对含义的，所以2.5对时间维度的位置ID，采用了绝对位置编码。同时也引入了动态帧的技术，每秒随机动态采集帧序列，使得模型能够通过不同时间ID的间隔，来学习时间的节奏。
从头训练原声动态分辨率的ViT，引入窗口注意力机制，只有四层是全注意力层，其余层最大窗口大小为8*8，小于8*8部分不填充保持原始尺度。

源码

Qwen3

modular_qwen3_moe.py

class Qwen3MoeSparseMoeBlock(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.num_experts = config.num_experts
        self.top_k = config.num_experts_per_tok
        self.norm_topk_prob = config.norm_topk_prob

        # gating
        self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
        self.experts = nn.ModuleList(
            [Qwen3MoeMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
        )

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        """ """
        batch_size, sequence_length, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
        # router_logits: (batch * sequence_length, n_experts)
        router_logits = self.gate(hidden_states)

        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
        if self.norm_topk_prob:  # only diff with mixtral sparse moe block!
            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
        # we cast back to the input dtype
        routing_weights = routing_weights.to(hidden_states.dtype)

        final_hidden_states = torch.zeros(
            (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
        )

        # One hot encode the selected experts to create an expert mask
        # this will be used to easily index which expert is going to be sollicitated
        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)

        # Loop over all available experts in the model and perform the computation on each expert
        expert_hit = torch.greater(expert_mask.sum(dim=(-1, -2)), 0).nonzero()
        for expert_idx in expert_hit:
            expert_layer = self.experts[expert_idx]
            idx, top_x = torch.where(expert_mask[expert_idx].squeeze(0))

            # Index the correct hidden states and compute the expert hidden state for
            # the current expert. We need to make sure to multiply the output hidden
            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]

            # However `index_add_` only support torch tensors for indexing so we'll use
            # the `top_x` tensor here.
            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
        final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
        return final_hidden_states, router_logits