[Attention] Refactor AttentionMetadata Preparation for Encoder-only Models (#23154)

Signed-off-by: Chen Zhang <zhangch99@outlook.com>
This commit is contained in:
Chen Zhang
2025-08-22 05:05:59 +00:00
committed by GitHub
parent 5964069367
commit 17373dcd93
12 changed files with 226 additions and 214 deletions
+7 -7
View File
@@ -7,7 +7,7 @@ import torch
from torch import nn
from transformers import ModernBertConfig
from vllm.attention import Attention, AttentionType
from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import VllmConfig
from vllm.distributed import get_tensor_model_parallel_world_size
@@ -104,12 +104,12 @@ class ModernBertAttention(nn.Module):
head_size=self.head_dim,
dim=self.head_dim,
base=rope_theta)
self.attn = Attention(self.num_heads,
self.head_dim,
self.scaling,
prefix=f"{layer_id}.attn",
attn_type=AttentionType.ENCODER_ONLY,
per_layer_sliding_window=sliding_window)
self.attn = EncoderOnlyAttention(
self.num_heads,
self.head_dim,
self.scaling,
prefix=f"{layer_id}.attn",
per_layer_sliding_window=sliding_window)
self.Wo = RowParallelLinear(config.hidden_size,
config.hidden_size,
bias=config.attention_bias)