[MM Encoder]: Migrate legacy ViT MultiHeadAttention to new MMEncoderAttention interface (#30684)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2025-12-19 02:04:19 +08:00
committed by GitHub
parent 62be3670cb
commit 700a5ad6c6
20 changed files with 182 additions and 266 deletions
+3 -2
View File
@@ -17,7 +17,8 @@ from transformers import BatchFeature, PretrainedConfig, ProcessorMixin, TensorT
from transformers.image_utils import ImageInput
from transformers.tokenization_utils_base import TextInput
from vllm.attention.layer import Attention, MultiHeadAttention
from vllm.attention.layer import Attention
from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.config.multimodal import BaseDummyOptions
@@ -222,7 +223,7 @@ class MultiHeadDotProductAttention(nn.Module):
)
self.scale = self.head_dim**-0.5
self.attn = MultiHeadAttention(
self.attn = MMEncoderAttention(
self.num_heads, self.head_dim, self.scale, num_kv_heads=self.num_kv_heads
)