[Hardware][CPU] Add embedding models support for CPU backend (#10193)

Signed-off-by: Isotr0py <2037008807@qq.com>

vllm/model_executor/models/bert.py

@@ -5,7 +5,6 @@ from torch import nn
 from transformers import BertConfig
 
 from vllm.attention import Attention, AttentionMetadata, AttentionType
-from vllm.attention.backends.xformers import XFormersImpl
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import get_act_fn
@@ -218,11 +217,6 @@ class BertSelfAttention(nn.Module):
                               quant_config=quant_config,
                               prefix=f"{prefix}.attn")
 
-        if not isinstance(self.attn.impl, XFormersImpl):
-            raise ValueError(
-                "Encoder-only models currently require XFORMERS attention "
-                "backend. Set VLLM_ATTENTION_BACKEND=XFORMERS to use BERT.")
-
     def forward(
         self,
         hidden_states: torch.Tensor,