[Attention] Refactor AttentionMetadata Preparation for Encoder-only Models (#23154)
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
This commit is contained in:
@@ -7,7 +7,7 @@ import torch
|
||||
from torch import nn
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
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 CacheConfig, VllmConfig
|
||||
from vllm.distributed import (divide, get_tensor_model_parallel_rank,
|
||||
@@ -119,14 +119,13 @@ class BertWithRopeAttention(nn.Module):
|
||||
|
||||
self.rotary_emb = get_rope(**rotary_kwargs)
|
||||
|
||||
self.attn = Attention(num_heads=self.num_heads,
|
||||
head_size=self.head_dim,
|
||||
scale=self.scaling,
|
||||
num_kv_heads=self.num_kv_heads,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.attn",
|
||||
attn_type=AttentionType.ENCODER_ONLY)
|
||||
self.attn = EncoderOnlyAttention(num_heads=self.num_heads,
|
||||
head_size=self.head_dim,
|
||||
scale=self.scaling,
|
||||
num_kv_heads=self.num_kv_heads,
|
||||
cache_config=cache_config,
|
||||
quant_config=quant_config,
|
||||
prefix=f"{prefix}.attn")
|
||||
|
||||
self.out_proj = RowParallelLinear(input_size=hidden_size,
|
||||
output_size=hidden_size,
|
||||
|
||||
Reference in New Issue
Block a user