optimize lib dependencies

deepseek_vl2/models/modeling_deepseek_vl_v2.py

@@ -1,4 +1,3 @@
-from attrdict import AttrDict
 from dataclasses import dataclass
 import logging
 import gc

deepseek_vl2/models/siglip_vit.py

@@ -13,7 +13,6 @@ from timm.layers import (
 )
 from timm.models._manipulate import named_apply, checkpoint_seq, adapt_input_conv
 from transformers.modeling_utils import is_flash_attn_2_available
-from xformers.ops import memory_efficient_attention
 from functools import partial
 
 
@@ -134,6 +133,8 @@ class Attention(nn.Module):
         self.proj_drop = nn.Dropout(proj_drop) if proj_drop > 0. else nn.Identity()
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
+        from xformers.ops import memory_efficient_attention
+
         B, N, C = x.shape
         qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim)
 

requirements.txt

@@ -4,7 +4,6 @@ xformers>=0.0.21
 timm>=0.9.16
 accelerate
 sentencepiece
-attrdict
 einops
 
 # for gradio demo