| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Base class for attention-like layers.""" |
| 4 | |
| 5 | from abc import ABC, abstractmethod |
| 6 | |
| 7 | import torch |
| 8 | |
| 9 | from vllm.config import VllmConfig |
| 10 | from vllm.v1.attention.backend import AttentionBackend, AttentionImpl |
| 11 | from vllm.v1.kv_cache_interface import KVCacheSpec |
| 12 | |
| 13 | |
| 14 | class AttentionLayerBase(ABC): |
| 15 | """Base class for attention-like layers (Attention, Mamba, etc.) |
| 16 | that support the v1 engine. |
| 17 | |
| 18 | This provides a common interface for getting attention backends |
| 19 | from different layer types. |
| 20 | """ |
| 21 | |
| 22 | impl: "AttentionImpl" |
| 23 | supports_dcp: bool = True |
| 24 | # Set by the speculator on its draft model's layers. |
| 25 | is_draft_layer: bool = False |
| 26 | |
| 27 | def bind_kv_cache(self, kv_cache: torch.Tensor) -> None: |
| 28 | """Bind the allocated KV cache tensor to this layer. |
| 29 | |
| 30 | The default stores the cache view as-is; subclasses (e.g. Mamba) |
| 31 | override this to unpack the raw buffer into per-state views. |
| 32 | """ |
| 33 | self.kv_cache = kv_cache |
| 34 | |
| 35 | @abstractmethod |
| 36 | def get_attn_backend(self) -> type[AttentionBackend]: |
| 37 | """Get the attention backend class for this layer.""" |
| 38 | pass |
| 39 | |
| 40 | @abstractmethod |
| 41 | def get_kv_cache_spec(self, vllm_config: VllmConfig) -> KVCacheSpec | None: |
| 42 | """Get the KV cache spec for this layer. |
| 43 | May be None if the layer does not need KV cache. |
| 44 | """ |
| 45 | pass |