[Attention] Remove imports from vllm/attention/__init__.py (#29342)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>

docs/contributing/model/basic.md

@@ -29,7 +29,7 @@ The initialization code should look like this:
     ```python
     from torch import nn
     from vllm.config import VllmConfig
-    from vllm.attention import Attention
+    from vllm.attention.layer import Attention
 
     class MyAttention(nn.Module):
         def __init__(self, vllm_config: VllmConfig, prefix: str):

tests/compile/test_fusion_attn.py

@@ -9,8 +9,9 @@ from tests.compile.backend import LazyInitPass, TestBackend
 from tests.utils import flat_product
 from tests.v1.attention.utils import BatchSpec, create_common_attn_metadata
 from vllm._custom_ops import cutlass_scaled_fp4_mm, scaled_fp4_quant
-from vllm.attention import Attention, AttentionMetadata
+from vllm.attention.backends.abstract import AttentionMetadata
 from vllm.attention.backends.registry import AttentionBackendEnum
+from vllm.attention.layer import Attention
 from vllm.attention.selector import global_force_attn_backend_context_manager
 from vllm.compilation.fusion_attn import ATTN_OP, AttnFusionPass
 from vllm.compilation.fx_utils import find_op_nodes

tests/compile/test_qk_norm_rope_fusion.py

@@ -5,7 +5,8 @@ import pytest
 import torch
 
 from tests.compile.backend import TestBackend
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.compilation.matcher_utils import FLASHINFER_ROTARY_OP, RMS_OP, ROTARY_OP
 from vllm.compilation.noop_elimination import NoOpEliminationPass
 from vllm.compilation.post_cleanup import PostCleanupPass

tests/kernels/utils.py

@@ -14,7 +14,7 @@ import torch
 from torch._prims_common import TensorLikeType
 
 from tests.kernels.quant_utils import native_w8a8_block_matmul
-from vllm.attention import AttentionType
+from vllm.attention.backends.abstract import AttentionType
 from vllm.model_executor.layers.activation import SiluAndMul
 from vllm.model_executor.layers.fused_moe.utils import moe_kernel_quantize_input
 from vllm.utils import (

tests/v1/worker/test_gpu_model_runner.py

@@ -5,8 +5,8 @@ import numpy as np
 import pytest
 import torch
 
-from vllm.attention import Attention
 from vllm.attention.backends.abstract import MultipleOf
+from vllm.attention.layer import Attention
 from vllm.config import (
     CacheConfig,
     ModelConfig,

tests/v1/worker/test_utils.py

@@ -7,7 +7,7 @@ from vllm.v1.worker.utils import bind_kv_cache
 
 
 def test_bind_kv_cache():
-    from vllm.attention import Attention
+    from vllm.attention.layer import Attention
 
     ctx = {
         "layers.0.self_attn": Attention(32, 128, 0.1),
@@ -35,7 +35,7 @@ def test_bind_kv_cache():
 
 
 def test_bind_kv_cache_non_attention():
-    from vllm.attention import Attention
+    from vllm.attention.layer import Attention
 
     # example from Jamba PP=2
     ctx = {

vllm/attention/__init__.py

@@ -1,19 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-from vllm.attention.backends.abstract import (
-    AttentionBackend,
-    AttentionMetadata,
-    AttentionType,
-)
-from vllm.attention.layer import Attention
-from vllm.attention.selector import get_attn_backend, get_mamba_attn_backend
-
-__all__ = [
-    "Attention",
-    "AttentionBackend",
-    "AttentionMetadata",
-    "AttentionType",
-    "get_attn_backend",
-    "get_mamba_attn_backend",
-]

vllm/attention/backends/abstract.py

@@ -178,7 +178,7 @@ class AttentionBackend(ABC):
         By default, only supports decoder attention.
         Backends should override this to support other attention types.
         """
-        from vllm.attention import AttentionType
+        from vllm.attention.backends.abstract import AttentionType
 
         return attn_type == AttentionType.DECODER
 

vllm/attention/layer.py

@@ -10,8 +10,11 @@ import torch.nn as nn
 import torch.nn.functional as F
 
 import vllm.envs as envs
-from vllm.attention import AttentionType
+from vllm.attention.backends.abstract import (
-from vllm.attention.backends.abstract import AttentionBackend, MLAAttentionImpl
+    AttentionBackend,
+    AttentionType,
+    MLAAttentionImpl,
+)
 from vllm.attention.backends.registry import AttentionBackendEnum
 from vllm.attention.selector import get_attn_backend
 from vllm.attention.utils.kv_sharing_utils import validate_kv_sharing_target

vllm/compilation/fusion_attn.py

@@ -10,7 +10,7 @@ from torch import fx
 from torch._higher_order_ops.auto_functionalize import auto_functionalized
 from torch._inductor.pattern_matcher import PatternMatcherPass
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.config import VllmConfig, get_layers_from_vllm_config
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization.utils.quant_utils import (

vllm/compilation/qk_norm_rope_fusion.py

@@ -9,7 +9,7 @@ from torch import fx
 from torch._higher_order_ops.auto_functionalize import auto_functionalized
 from torch._inductor.pattern_matcher import PatternMatcherPass
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.config import VllmConfig, get_layers_from_vllm_config
 from vllm.logger import init_logger
 from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding

vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py

@@ -20,7 +20,7 @@ import torch
 import zmq
 
 from vllm import envs
-from vllm.attention import AttentionBackend
+from vllm.attention.backends.abstract import AttentionBackend
 from vllm.attention.backends.registry import AttentionBackendEnum
 from vllm.attention.selector import get_attn_backend
 from vllm.config import VllmConfig

vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py

@@ -8,7 +8,8 @@ from typing import Any, ClassVar
 
 import torch
 
-from vllm.attention import Attention, AttentionBackend, AttentionMetadata
+from vllm.attention.backends.abstract import AttentionBackend, AttentionMetadata
+from vllm.attention.layer import Attention
 from vllm.config import VllmConfig, get_layers_from_vllm_config
 from vllm.distributed.kv_events import BlockRemoved, BlockStored, KVCacheEvent
 from vllm.distributed.kv_transfer.kv_connector.v1 import (

vllm/model_executor/layers/mamba/linear_attn.py

@@ -8,7 +8,7 @@ import torch.nn.functional as F
 from einops import rearrange
 from torch import nn
 
-from vllm.attention import AttentionMetadata
+from vllm.attention.backends.abstract import AttentionMetadata
 from vllm.config import CacheConfig, ModelConfig, get_current_vllm_config
 from vllm.distributed.communication_op import tensor_model_parallel_all_reduce
 from vllm.distributed.parallel_state import (

vllm/model_executor/model_loader/utils.py

@@ -11,8 +11,7 @@ import torch
 from torch import nn
 from typing_extensions import assert_never
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention, MLAAttention
-from vllm.attention.layer import MLAAttention
 from vllm.config import ModelConfig, VllmConfig, set_current_vllm_config
 from vllm.logger import init_logger
 from vllm.model_executor.layers.quantization.base_config import (

vllm/model_executor/models/afmoe.py

@@ -9,7 +9,8 @@ from itertools import islice
 import torch
 from torch import nn
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
 from vllm.distributed import (

vllm/model_executor/models/apertus.py

@@ -32,7 +32,8 @@ import torch
 from torch import nn
 from transformers import ApertusConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig

vllm/model_executor/models/arctic.py

@@ -8,7 +8,7 @@ from itertools import islice
 import torch
 from torch import nn
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/baichuan.py

@@ -29,7 +29,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/bailing_moe.py

@@ -32,7 +32,7 @@ import torch.nn.functional as F
 from torch import nn
 from transformers.configuration_utils import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/bloom.py

@@ -27,7 +27,7 @@ import torch
 from torch import nn
 from transformers import BloomConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/chameleon.py

@@ -16,7 +16,7 @@ from transformers import (
     ChameleonVQVAEConfig,
 )
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.config import CacheConfig, VllmConfig
 from vllm.config.multimodal import BaseDummyOptions
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/chatglm.py

@@ -12,7 +12,7 @@ import torch
 from torch import nn
 from torch.nn import LayerNorm
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/clip.py

@@ -14,8 +14,7 @@ from transformers import (
     CLIPVisionConfig,
 )
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention, MultiHeadAttention
-from vllm.attention.layer import MultiHeadAttention
 from vllm.config import VllmConfig
 from vllm.config.multimodal import BaseDummyOptions
 from vllm.distributed import divide, get_tensor_model_parallel_world_size

vllm/model_executor/models/commandr.py

@@ -30,7 +30,7 @@ import torch
 from torch import nn
 from transformers import Cohere2Config, CohereConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/dbrx.py

@@ -8,7 +8,7 @@ import torch
 import torch.nn as nn
 from transformers import DbrxConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (
     get_pp_group,

vllm/model_executor/models/deepseek_v2.py

@@ -33,8 +33,8 @@ from torch import nn
 from transformers import DeepseekV2Config, DeepseekV3Config
 
 from vllm._aiter_ops import rocm_aiter_ops
-from vllm.attention import Attention
 from vllm.attention.backends.abstract import AttentionBackend
+from vllm.attention.layer import Attention
 from vllm.attention.ops.common import pack_seq_triton, unpack_seq_triton
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, ParallelConfig, VllmConfig, get_current_vllm_config

vllm/model_executor/models/dots1.py

@@ -32,7 +32,7 @@ import torch
 from torch import nn
 from transformers import Dots1Config
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/ernie45_moe.py

@@ -32,7 +32,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
 from vllm.distributed import (

vllm/model_executor/models/ernie45_vl_moe.py

@@ -31,7 +31,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 
 # from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig

vllm/model_executor/models/exaone.py

@@ -32,7 +32,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/exaone4.py

@@ -28,7 +28,7 @@ import torch
 from torch import nn
 from transformers import Exaone4Config
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/falcon.py

@@ -30,7 +30,7 @@ from torch import nn
 from torch.nn import LayerNorm
 from transformers import FalconConfig as HF_FalconConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/gemma.py

@@ -26,7 +26,7 @@ import torch
 from torch import nn
 from transformers import GemmaConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/gemma2.py

@@ -23,7 +23,7 @@ import torch
 from torch import nn
 from transformers import Gemma2Config
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/gemma3.py

@@ -23,7 +23,8 @@ import torch.nn.functional as F
 from torch import nn
 from transformers import Gemma3TextConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/gemma3n.py

@@ -21,7 +21,7 @@ import torch
 from torch import nn
 from transformers.models.gemma3n.configuration_gemma3n import Gemma3nTextConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_tensor_model_parallel_world_size

vllm/model_executor/models/glm4.py

@@ -29,7 +29,8 @@ import torch
 from torch import nn
 from transformers import Glm4Config
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/glm4_moe.py

@@ -31,7 +31,7 @@ import torch
 from torch import nn
 from transformers.models.glm4_moe import Glm4MoeConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
 from vllm.distributed import (

vllm/model_executor/models/gpt2.py

@@ -27,7 +27,7 @@ import torch
 from torch import nn
 from transformers import GPT2Config
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed.parallel_state import (

vllm/model_executor/models/gpt_bigcode.py

@@ -28,7 +28,7 @@ import torch
 from torch import nn
 from transformers import GPTBigCodeConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/gpt_j.py

@@ -26,7 +26,7 @@ import torch
 from torch import nn
 from transformers import GPTJConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/gpt_neox.py

@@ -26,7 +26,7 @@ import torch
 from torch import nn
 from transformers import GPTNeoXConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/gpt_oss.py

@@ -7,7 +7,8 @@ import torch.distributed as dist
 from torch import nn
 from transformers import GptOssConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/granite.py

@@ -31,7 +31,7 @@ import torch
 from torch import nn
 from transformers import GraniteConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/granitemoe.py

@@ -31,7 +31,7 @@ from typing import Any
 import torch
 from torch import nn
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/grok1.py

@@ -31,7 +31,7 @@ import torch
 import torch.nn.functional as F
 from torch import nn
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/hunyuan_v1.py

@@ -33,7 +33,8 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
 from vllm.distributed import (

vllm/model_executor/models/internlm2.py

@@ -10,7 +10,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/jais.py

@@ -28,7 +28,7 @@ from itertools import islice
 import torch
 from torch import nn
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/lfm2.py

@@ -7,7 +7,7 @@ import torch
 import torch.nn as nn
 from transformers import Lfm2Config
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/lfm2_moe.py

@@ -6,7 +6,7 @@ from itertools import islice
 import torch
 import torch.nn as nn
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, ModelConfig, VllmConfig, get_current_vllm_config
 from vllm.distributed import (

vllm/model_executor/models/llama.py

@@ -31,7 +31,8 @@ import torch
 from torch import nn
 from transformers import LlamaConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig

vllm/model_executor/models/llama4.py

@@ -24,7 +24,7 @@ import torch
 from torch import nn
 from transformers import Llama4TextConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.attention.layers.chunked_local_attention import ChunkedLocalAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig

vllm/model_executor/models/minicpm.py

@@ -33,7 +33,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/minicpm3.py

@@ -29,7 +29,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.layernorm import RMSNorm

vllm/model_executor/models/minimax_m2.py

@@ -30,7 +30,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/minimax_text_01.py

@@ -14,7 +14,8 @@ import torch
 from torch import nn
 from transformers import MiniMaxConfig
 
-from vllm.attention import Attention, AttentionMetadata
+from vllm.attention.backends.abstract import AttentionMetadata
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.distributed.parallel_state import (

vllm/model_executor/models/mixtral.py

@@ -32,7 +32,7 @@ import torch
 from torch import nn
 from transformers import MixtralConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
 from vllm.distributed import (

vllm/model_executor/models/molmo.py

@@ -17,8 +17,7 @@ from transformers import BatchFeature, PretrainedConfig, ProcessorMixin, TensorT
 from transformers.image_utils import ImageInput
 from transformers.tokenization_utils_base import TextInput
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention, MultiHeadAttention
-from vllm.attention.layer import MultiHeadAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.config.multimodal import BaseDummyOptions

vllm/model_executor/models/mpt.py

@@ -10,7 +10,7 @@ import torch
 import torch.nn as nn
 from transformers import MptConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/nemotron.py

@@ -30,7 +30,7 @@ from itertools import islice
 import torch
 from torch import nn
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/nemotron_nas.py

@@ -31,7 +31,7 @@ import torch
 from torch import nn
 from transformers import LlamaConfig
 
-from vllm.attention import AttentionType
+from vllm.attention.backends.abstract import AttentionType
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group

vllm/model_executor/models/olmo.py

@@ -31,7 +31,7 @@ import torch
 from torch import nn
 from transformers import OlmoConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/olmo2.py

@@ -32,7 +32,7 @@ import torch
 from torch import nn
 from transformers import Olmo2Config
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/olmoe.py

@@ -21,7 +21,7 @@ from itertools import islice
 import torch
 from torch import nn
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/openpangu.py

@@ -29,7 +29,8 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, ParallelConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/opt.py

@@ -27,7 +27,7 @@ import torch
 from torch import nn
 from transformers import OPTConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/orion.py

@@ -15,7 +15,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/ouro.py

@@ -33,7 +33,8 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_tensor_model_parallel_world_size

vllm/model_executor/models/persimmon.py

@@ -30,7 +30,7 @@ import torch
 from torch import nn
 from transformers import PersimmonConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/phi.py

@@ -45,7 +45,7 @@ import torch
 from torch import nn
 from transformers import PhiConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/phimoe.py

@@ -31,7 +31,7 @@ import torch
 from torch import nn
 from transformers.configuration_utils import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/qwen.py

@@ -16,7 +16,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/qwen2.py

@@ -33,7 +33,8 @@ import torch
 from torch import nn
 from transformers import Qwen2Config
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig

vllm/model_executor/models/qwen2_moe.py

@@ -34,7 +34,7 @@ import torch.nn.functional as F
 from torch import nn
 from transformers import Qwen2MoeConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/qwen3.py

@@ -30,7 +30,8 @@ import torch
 from torch import nn
 from transformers import Qwen3Config
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/qwen3_moe.py

@@ -31,7 +31,7 @@ from typing import Any
 import torch
 from torch import nn
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig, get_current_vllm_config
 from vllm.distributed import (

vllm/model_executor/models/qwen3_next.py

@@ -10,7 +10,8 @@ from einops import rearrange
 from torch import nn
 from transformers.activations import ACT2FN
 
-from vllm.attention import Attention, AttentionMetadata
+from vllm.attention.backends.abstract import AttentionMetadata
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import (
     CacheConfig,

vllm/model_executor/models/seed_oss.py

@@ -30,7 +30,8 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig as SeedOssConfig
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/solar.py

@@ -30,7 +30,7 @@ import torch
 from torch import nn
 from transformers import PretrainedConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/stablelm.py

@@ -29,7 +29,7 @@ import torch
 from torch import nn
 from transformers import StableLmConfig
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.activation import SiluAndMul

vllm/model_executor/models/starcoder2.py

@@ -28,7 +28,7 @@ import torch
 from torch import nn
 from transformers import Starcoder2Config
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size

vllm/model_executor/models/step3_text.py

@@ -9,7 +9,7 @@ from typing import Any
 import torch
 from torch import nn
 
-from vllm.attention import Attention
+from vllm.attention.layer import Attention
 from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, ModelConfig, VllmConfig
 from vllm.distributed import (

vllm/model_executor/models/transformers/base.py

@@ -27,7 +27,8 @@ from torch import nn
 from transformers import AutoModel
 from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
+from vllm.attention.layer import Attention
 from vllm.attention.layers.encoder_only_attention import EncoderOnlyAttention
 from vllm.config.utils import getattr_iter
 from vllm.distributed import get_pp_group, get_tp_group

vllm/model_executor/models/whisper.py

@@ -16,8 +16,8 @@ from transformers import (
 )
 from transformers.models.whisper.modeling_whisper import sinusoids
 
-from vllm.attention import Attention, AttentionType
+from vllm.attention.backends.abstract import AttentionType
-from vllm.attention.layer import MultiHeadAttention
+from vllm.attention.layer import Attention, MultiHeadAttention
 from vllm.attention.layers.cross_attention import CrossAttention
 from vllm.config import CacheConfig, ModelConfig, SpeechToTextConfig, VllmConfig
 from vllm.config.multimodal import BaseDummyOptions

vllm/platforms/cuda.py

@@ -335,7 +335,7 @@ class CudaPlatformBase(Platform):
         use_sparse: bool,
         attn_type: str | None = None,
     ) -> str:
-        from vllm.attention import AttentionType
+        from vllm.attention.backends.abstract import AttentionType
 
         if attn_type is None:
             attn_type = AttentionType.DECODER

vllm/v1/attention/backends/cpu_attn.py

@@ -51,7 +51,7 @@ class CPUAttentionBackend(AttentionBackend):
     @classmethod
     def supports_attn_type(cls, attn_type: str) -> bool:
         """CPU attention supports decoder and encoder-only attention."""
-        from vllm.attention import AttentionType
+        from vllm.attention.backends.abstract import AttentionType
 
         return attn_type in (
             AttentionType.DECODER,

vllm/v1/attention/backends/flash_attn.py

@@ -84,7 +84,7 @@ class FlashAttentionBackend(AttentionBackend):
     @classmethod
     def supports_attn_type(cls, attn_type: str) -> bool:
         """FlashAttention supports all attention types."""
-        from vllm.attention import AttentionType
+        from vllm.attention.backends.abstract import AttentionType
 
         return attn_type in (
             AttentionType.DECODER,

vllm/v1/attention/backends/flex_attention.py

@@ -87,7 +87,7 @@ class FlexAttentionBackend(AttentionBackend):
     @classmethod
     def supports_attn_type(cls, attn_type: str) -> bool:
         """FlexAttention supports both decoder and encoder-only attention."""
-        from vllm.attention import AttentionType
+        from vllm.attention.backends.abstract import AttentionType
 
         return attn_type in (AttentionType.DECODER, AttentionType.ENCODER_ONLY)
 

vllm/v1/kv_offload/cpu.py

@@ -4,7 +4,7 @@ from collections.abc import Iterator
 
 import torch
 
-from vllm.attention import AttentionBackend
+from vllm.attention.backends.abstract import AttentionBackend
 from vllm.config import VllmConfig
 from vllm.platforms import current_platform
 from vllm.v1.kv_offload.abstract import LoadStoreSpec, OffloadingManager

vllm/v1/kv_offload/spec.py

@@ -11,7 +11,7 @@ from vllm.v1.kv_offload.abstract import LoadStoreSpec, OffloadingManager
 from vllm.v1.kv_offload.worker.worker import OffloadingHandler
 
 if TYPE_CHECKING:
-    from vllm.attention import AttentionBackend
+    from vllm.attention.backends.abstract import AttentionBackend
     from vllm.config import VllmConfig
 
 logger = init_logger(__name__)

vllm/v1/kv_offload/worker/cpu_gpu.py

@@ -5,7 +5,7 @@ import numpy as np
 import torch
 
 from vllm import _custom_ops as ops
-from vllm.attention import AttentionBackend
+from vllm.attention.backends.abstract import AttentionBackend
 from vllm.logger import init_logger
 from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.v1.kv_offload.mediums import CPULoadStoreSpec, GPULoadStoreSpec

vllm/v1/worker/gpu_model_runner.py

@@ -19,12 +19,13 @@ import torch.nn as nn
 from tqdm import tqdm
 
 import vllm.envs as envs
-from vllm.attention import Attention, AttentionType
 from vllm.attention.backends.abstract import (
     AttentionBackend,
     AttentionMetadata,
+    AttentionType,
     MultipleOf,
 )
+from vllm.attention.layer import Attention
 from vllm.compilation.counter import compilation_counter
 from vllm.compilation.cuda_graph import CUDAGraphWrapper
 from vllm.compilation.monitor import set_cudagraph_capturing_enabled

vllm/v1/worker/kv_connector_model_runner_mixin.py

@@ -13,7 +13,7 @@ from typing import (
 
 import torch
 
-from vllm.attention import AttentionBackend
+from vllm.attention.backends.abstract import AttentionBackend
 from vllm.config import VllmConfig
 from vllm.config.cache import CacheDType
 from vllm.distributed.kv_transfer import (

vllm/v1/worker/tpu_model_runner.py

@@ -17,9 +17,8 @@ import torch_xla.distributed.spmd as xs
 import torch_xla.runtime as xr
 
 import vllm.envs as envs
-from vllm.attention import Attention
 from vllm.attention.backends.abstract import AttentionType
-from vllm.attention.layer import MLAAttention
+from vllm.attention.layer import Attention, MLAAttention
 from vllm.attention.layers.chunked_local_attention import ChunkedLocalAttention
 from vllm.compilation.wrapper import TorchCompileWithNoGuardsWrapper
 from vllm.config import (