Skip to content

vllm.model_executor.layers.attention_layer_base

Base class for attention-like layers.

AttentionLayerBase

Bases: ABC

Base class for attention-like layers (Attention, Mamba, etc.) that support the v1 engine.

This provides a common interface for getting attention backends from different layer types.

Source code in vllm/model_executor/layers/attention_layer_base.py 
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
class AttentionLayerBase(ABC):
    """
    Base class for attention-like layers (Attention, Mamba, etc.)
    that support the v1 engine.

    This provides a common interface for getting attention backends
    from different layer types.
    """

    @abstractmethod
    def get_attn_backend(self) -> type["AttentionBackend"]:
        """Get the attention backend class for this layer."""
        pass

    @abstractmethod
    def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec | None:
        """
        Get the KV cache spec for this layer.
        May be None if the layer does not need KV cache.
        """
        pass

get_attn_backend abstractmethod 

get_attn_backend() -> type[AttentionBackend]

Get the attention backend class for this layer.

Source code in vllm/model_executor/layers/attention_layer_base.py 
24
25
26
27
@abstractmethod
def get_attn_backend(self) -> type["AttentionBackend"]:
    """Get the attention backend class for this layer."""
    pass

get_kv_cache_spec abstractmethod 

get_kv_cache_spec(
    vllm_config: VllmConfig,
) -> KVCacheSpec | None

Get the KV cache spec for this layer. May be None if the layer does not need KV cache.

Source code in vllm/model_executor/layers/attention_layer_base.py 
29
30
31
32
33
34
35
@abstractmethod
def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec | None:
    """
    Get the KV cache spec for this layer.
    May be None if the layer does not need KV cache.
    """
    pass