[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
+2 -2
View File
@@ -8,7 +8,7 @@ from collections.abc import Iterable
import torch
import torch.nn as nn
from vllm.attention.layer import MultiHeadAttention
from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
from vllm.distributed import get_tensor_model_parallel_world_size
from vllm.distributed.utils import divide
from vllm.model_executor.layers.activation import SiluAndMul
@@ -126,7 +126,7 @@ class AIMv2Attention(nn.Module):
self.tp_size = get_tensor_model_parallel_world_size()
self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
self.attn = MultiHeadAttention(
self.attn = MMEncoderAttention(
self.num_heads_per_partition, self.head_dim, self.scale
)