[Models] Replace all nn.Conv2d with vLLM's Conv2dLayer (#28842)

Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
This commit is contained in:
Isotr0py
2025-11-18 18:56:04 +00:00
committed by GitHub
parent c64c0b78de
commit e4bb2684bc
20 changed files with 83 additions and 45 deletions
+2 -1
View File
@@ -12,6 +12,7 @@ from transformers import Blip2VisionConfig, BlipVisionConfig
from vllm.attention.layer import MultiHeadAttention
from vllm.distributed import divide, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.conv import Conv2dLayer
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
@@ -47,7 +48,7 @@ class BlipVisionEmbeddings(nn.Module):
self.class_embedding = nn.Parameter(torch.randn(1, 1, self.embed_dim))
self.patch_embedding = nn.Conv2d(
self.patch_embedding = Conv2dLayer(
in_channels=3,
out_channels=self.embed_dim,
kernel_size=self.patch_size,