How is PLoRA used?

[bang123-box]

作者您好您的工作非常吸引我，但是在阅读了您的源码之后我有一些事情想向您请教一下，Coobiw/InternLM-XComposer2_Enhanced的L169行中的Plora是如何使用的呢，我看您的train.py中已经有了lora的实现，感觉这里的PLora好像没有什么用。非常期待您的回复

[Coobiw]

PLoRA是internlm-xc2这个baseline里实现好、已经训练好了的部分

[bang123-box]

好的，确实是这样的IX_2的源码中也有这个PLoRA，感谢作者您的回复

[Coobiw]

嗯嗯 这部分是xc2里转门给visual tokens用的lora，和后面微调llm用的lora不太一样

[bang123-box]

PeftModelForCausalLM(
  (base_model): LoraModel(
    (model): InternLMXComposer2ForCausalLM(
      (model): InternLM2Model(
        (tok_embeddings): Embedding(92544, 4096, padding_idx=2)
        (layers): ModuleList(
          (0-31): 32 x InternLM2DecoderLayer(
            (attention): InternLM2FlashAttention2(
              (wqkv): Linear(
                in_features=4096, out_features=6144, bias=False
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.05, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=4096, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=6144, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
              )
              (wo): Linear(
                in_features=4096, out_features=4096, bias=False
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.05, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=4096, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=4096, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
              )
              (rotary_emb): InternLM2DynamicNTKScalingRotaryEmbedding()
            )
            (feed_forward): InternLM2MLP(
              (w1): Linear(
                in_features=4096, out_features=14336, bias=False
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.05, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=4096, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=14336, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
              )
              (w3): Linear(
                in_features=4096, out_features=14336, bias=False
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.05, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=4096, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=14336, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
              )
              (w2): Linear(
                in_features=14336, out_features=4096, bias=False
                (lora_dropout): ModuleDict(
                  (default): Dropout(p=0.05, inplace=False)
                )
                (lora_A): ModuleDict(
                  (default): Linear(in_features=14336, out_features=8, bias=False)
                )
                (lora_B): ModuleDict(
                  (default): Linear(in_features=8, out_features=4096, bias=False)
                )
                (lora_embedding_A): ParameterDict()
                (lora_embedding_B): ParameterDict()
              )
              (act_fn): SiLUActivation()
            )
            (attention_norm): InternLM2RMSNorm()
            (ffn_norm): InternLM2RMSNorm()
          )
        )
        (norm): InternLM2RMSNorm()
      )
      (output): Linear(in_features=4096, out_features=92544, bias=False)
      (vit): CLIPVisionTower(
        (vision_tower): CLIPVisionModel(
          (vision_model): CLIPVisionTransformer(
            (embeddings): CLIPVisionEmbeddings(
              (patch_embedding): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14), bias=False)
              (position_embedding): Embedding(1226, 1024)
            )
            (pre_layrnorm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
            (encoder): CLIPEncoder(
              (layers): ModuleList(
                (0-23): 24 x CLIPEncoderLayer(
                  (self_attn): CLIPAttention(
                    (k_proj): Linear(in_features=1024, out_features=1024, bias=True)
                    (v_proj): Linear(in_features=1024, out_features=1024, bias=True)
                    (q_proj): Linear(in_features=1024, out_features=1024, bias=True)
                    (out_proj): Linear(in_features=1024, out_features=1024, bias=True)
                  )
                  (layer_norm1): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
                  (mlp): CLIPMLP(
                    (activation_fn): QuickGELUActivation()
                    (fc1): Linear(in_features=1024, out_features=4096, bias=True)
                    (fc2): Linear(in_features=4096, out_features=1024, bias=True)
                  )
                  (layer_norm2): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
                )
              )
            )
            (post_layernorm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
          )
        )
      )
      (vision_proj): Sequential(
        (0): Linear(in_features=1024, out_features=4096, bias=True)
        (1): GELU(approximate='none')
        (2): Linear(in_features=4096, out_features=4096, bias=True)
      )
    )
  )
)

还有一点我想问一下您，我发现SFT阶段对XC2使用lora之后PLora类里面的self.Plora_A和self.Plora_B不存在了，这样的话那这个类的forward函数里面的im_mask应该就失效了，也就是说SFT阶段不存在对visual tokens用lora了，但是再您开源的这个https://huggingface.co/IDEA-FinAI/chartmoe模型权重里面，为什么还能看到类似于
model.layers.0.attention.wqkv.Plora_A.weight: pytorch_model-00001-of-00002.bin,model.layers.0.attention.wqkv.Plora_B.weight": "pytorch_model-00001-of-00002.bin这样的权重。按照SFT阶段的lora微调方式，貌似都不会使用Plora_A和Plora_B,但是最后在generate的时候确传入了im_mask，让其走PLora类里面的的这一步：

if torch.sum(im_mask) > 0:
    part_x = x[im_mask]
    res[im_mask] += self.Plora_B(self.Plora_A(self.lora_dropout(part_x))) * self.lora_scaling

[Coobiw]

sft阶段的lora和PLoRA并没有什么关联，PLoRA是原始base model的一部分，全阶段保留即可

[Coobiw]

至于你说的lora的时候PLoRA“不见了”，我想你可以具体打印出来每个InternLM2MLP，他的w1、w2、w3应该都有PLoRA相应的AB矩阵，我猜测没显示出来是因为PLoRA继承自nn.Linear（而peft中的lora只会加在Linear上），所以可能会这么显示（感觉可以观察验证一下）

可以参考这两段代码：

1️⃣https://huggingface.co/Coobiw/InternLM-XComposer2_Enhanced/blob/main/modeling_internlm2.py#L278

class InternLM2MLP(nn.Module):

def __init__(self, config):
  | super().__init__()
  | self.config = config
  | self.hidden_size = config.hidden_size
  | self.intermediate_size = config.intermediate_size
  |  
  | self.w1 = PLoRA(
  | self.hidden_size,
  | self.intermediate_size,
  | bias=False,
  | lora_r=256,
  | lora_alpha=256,
  | lora_len=1225)
  | self.w3 = PLoRA(
  | self.hidden_size,
  | self.intermediate_size,
  | bias=False,
  | lora_r=256,
  | lora_alpha=256,
  | lora_len=1225)
  | self.w2 = PLoRA(
  | self.intermediate_size,
  | self.hidden_size,
  | bias=False,
  | lora_r=256,
  | lora_alpha=256,
  | lora_len=1225)
  |  
  | self.act_fn = ACT2FN[config.hidden_act]
  |  
def forward(self, x, im_mask):
  | down_proj = self.w2(
  | self.act_fn(self.w1(x, im_mask)) * self.w3(x, im_mask), im_mask)
  |  
  | return down_proj

2️⃣：https://huggingface.co/Coobiw/InternLM-XComposer2_Enhanced/blob/main/build_mlp.py#L169

class PLoRA(nn.Linear):
--
  |  
def __init__(self,
  | in_features: int,
  | out_features: int,
  | bias: bool = True,
  | device=None,
  | dtype=None,
  | lora_r=8,
  | lora_alpha=16,
  | lora_dropout=0.05,
  | lora_len=0,
  | **kwargs) -> None:
  | super().__init__(in_features, out_features, bias, device, dtype)
  | self.lora_r = lora_r
  | self.lora_alpha = lora_alpha
  | self.lora_len = lora_len
  | if lora_dropout > 0.:
  | self.lora_dropout = nn.Dropout(p=lora_dropout)
  | else:
  | self.lora_dropout = lambda x: x
  | self.lora_scaling = self.lora_alpha / self.lora_r
  |  
  | self.Plora_A = nn.Linear(
  | in_features, self.lora_r, bias=False, device=device, dtype=dtype)
  | self.Plora_B = nn.Linear(
  | self.lora_r, out_features, bias=False, device=device, dtype=dtype)
  |  
  | self.reset_parameters()
  |  
def reset_parameters(self):
  | if hasattr(self, 'lora_A'):
  | # initialize A the same way as the default for nn.Linear and B to zero
  | nn.init.kaiming_uniform_(self.lora_A.weight, a=math.sqrt(5))
  | nn.init.zeros_(self.lora_B.weight)
  |  
def forward(self, x, im_mask=None):
  | res = super().forward(x)
  | if im_mask is not None:
  | if torch.sum(im_mask) > 0:
  | part_x = x[im_mask]
  | res[im_mask] += self.Plora_B(
  | self.Plora_A(
  | self.lora_dropout(part_x))) * self.lora_scaling
  | else:
  | part_x = x[:, :1]
  | res[:, :1] += self.Plora_B(
  | self.Plora_A(self.lora_dropout(part_x))) * 0
  | return res

[bang123-box]

再次感谢您的积极回复，我后面再去尝试一下。