vllm/vllm/model_executor/layers/attention_layer_base.py
Nicolò Lucchesi b26b70bec4
[Misc] Refactor get_kv_cache_spec into AttentionLayerBase (#26587)
Signed-off-by: NickLucche <nlucches@redhat.com>
2025-10-18 13:51:21 +00:00

36 lines
1.0 KiB
Python

# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Base class for attention-like layers."""
from abc import ABC, abstractmethod
from typing import TYPE_CHECKING
from vllm.config import VllmConfig
from vllm.v1.kv_cache_interface import KVCacheSpec
if TYPE_CHECKING:
from vllm.attention.backends.abstract import AttentionBackend
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