求解释Saliency分数得到的tensor

[Patrick-Ni]

您好！我参考您的代码，将应用于GPT2的Attentioner Manager应用到Llama上，然后得到了saliency分数，每一层都是[1,1,seq_len,seq_len]，部分具体数值如下：

我想知道这里每一层的saliency分数的具体含义？
我的代码如下：

class LlamaAttentionManager(AttentionerManagerBase):
    def __init__(self, model: PreTrainedModel):
        super().__init__(model)

    def register_attentioner_to_model(self):
        attention_adapters = []
        for layer in self.model.modules():
            if isinstance(layer, LlamaDecoderLayer):  # 假设解码器层被称为 TransformerDecoderLayer
                # 假设每个解码器层中的注意力模块叫做 self_attention
                attention_adapter = AttentionAdapter()
                layer.self_attn.forward = partial(llama_attn, layer.self_attn, attention_adapter=attention_adapter)
                attention_adapters.append(attention_adapter)
        return attention_adapters

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

    if self.config.pretraining_tp > 1:
        key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
        query_slices = self.q_proj.weight.split(
            (self.num_heads * self.head_dim) // self.config.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.config.pretraining_tp)]
        query_states = torch.cat(query_states, dim=-1)

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

        value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.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)

    cos, sin = self.rotary_emb(value_states, position_ids)
    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)

    if past_key_value 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_value.update(key_states, value_states, self.layer_idx, cache_kwargs)

    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 attention_mask is not None:  # no matter the length, we just slice it
        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
        attn_weights = attn_weights + causal_mask

    # print(attn_weights.shape)

    if attention_adapter is not None:
        attn_weights = attention_adapter(attn_weights)

    # upcast attention to fp32
    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)

    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, -1)

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

[leanwang326]

不好意思忘记回了：这里[1,1,i,j]对应于从j到i的attention的重要性 （|attenion_prob_{i,j} * attenion_prob_{i,j}.grad|）(然后我当时为了处理方便就限制bsz为1了，第二个1是因为对所有head取了平均