[Model][Mamba] Add selector for mamba attention backend and make it pluggable for other device (#26487)

Signed-off-by: shen-shanshan <467638484@qq.com>

vllm/model_executor/models/plamo2.py

@@ -4,10 +4,6 @@
 
 from collections.abc import Iterable
 from itertools import islice
-from typing import TYPE_CHECKING
-
-if TYPE_CHECKING:
-    from vllm.attention.backends.abstract import AttentionBackend
 
 import torch
 from torch import nn
@@ -467,11 +463,6 @@ class Plamo2MambaMixer(MambaBase, CustomOp):
     def mamba_type(self) -> str:
         return "mamba2"
 
-    def get_attn_backend(self) -> type["AttentionBackend"]:
-        from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionBackend
-
-        return Mamba2AttentionBackend
-
 
 def plamo2_mamba_mixer(
     hidden_states: torch.Tensor,

vllm/model_executor/models/qwen3_next.py

@@ -10,7 +10,7 @@ from einops import rearrange
 from torch import nn
 from transformers.activations import ACT2FN
 
-from vllm.attention import Attention, AttentionBackend, AttentionMetadata
+from vllm.attention import Attention, AttentionMetadata
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import (
     CacheConfig,
@@ -216,12 +216,7 @@ class Qwen3NextSparseMoeBlock(nn.Module):
 class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
     @property
     def mamba_type(self) -> str:
-        return "linear_attention"
-
-    def get_attn_backend(self) -> type["AttentionBackend"]:
-        from vllm.v1.attention.backends.gdn_attn import GDNAttentionBackend
-
-        return GDNAttentionBackend
+        return "gdn_attention"
 
     def get_state_dtype(self) -> tuple[torch.dtype, torch.dtype]:
         return MambaStateDtypeCalculator.gated_delta_net_state_dtype(