[3/N][KV-Cache Layout Refactor] Standardize Mamba cache; drop get_transfer_cache_regions (#44456)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Lucas Wilkinson
2026-07-21 09:16:15 +00:00
committed by GitHub
co-authored by Claude OpenAI Codex
parent eb44b3aaa4
commit 6700813f86
10 changed files with 115 additions and 225 deletions
@@ -4,6 +4,8 @@
from abc import ABC, abstractmethod
import torch
from vllm.config import VllmConfig
from vllm.v1.attention.backend import AttentionBackend, AttentionImpl
from vllm.v1.kv_cache_interface import KVCacheSpec
@@ -21,6 +23,14 @@ class AttentionLayerBase(ABC):
impl: "AttentionImpl"
supports_dcp: bool = True
def bind_kv_cache(self, kv_cache: torch.Tensor) -> None:
"""Bind the allocated KV cache tensor to this layer.
The default stores the cache view as-is; subclasses (e.g. Mamba)
override this to unpack the raw buffer into per-state views.
"""
self.kv_cache = kv_cache
@abstractmethod
def get_attn_backend(self) -> type[AttentionBackend]:
"""Get the attention backend class for this layer."""
@@ -2,11 +2,13 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from abc import abstractmethod
from collections.abc import Iterable
from math import prod
import torch
from vllm.config import VllmConfig
from vllm.model_executor.layers.attention_layer_base import AttentionLayerBase
from vllm.utils.torch_utils import get_dtype_size
from vllm.v1.attention.backend import AttentionBackend
from vllm.v1.attention.backends.registry import MambaAttentionBackendEnum
from vllm.v1.attention.selector import get_mamba_attn_backend
@@ -24,6 +26,22 @@ class MambaBase(AttentionLayerBase):
kv_cache: tuple[torch.Tensor, ...]
supports_dcp: bool = False
def bind_kv_cache(self, kv_cache: torch.Tensor) -> None:
"""Unpack a raw ``[B, 1, 1, C]`` int8 page view into per-state views.
Each block's ``C`` bytes hold the layer's states (e.g. conv, ssm)
packed contiguously; slice them out and reinterpret per dtype/shape.
"""
pages = kv_cache.squeeze(dim=(1, 2))
states: list[torch.Tensor] = []
offset = 0
for shape, dtype in zip(self.get_state_shape(), self.get_state_dtype()):
nbytes = prod(shape) * get_dtype_size(dtype)
state = pages[:, offset : offset + nbytes].view(dtype)
states.append(state.view(-1, *shape))
offset += nbytes
self.kv_cache = tuple(states)
@abstractmethod
def get_state_shape(self) -> Iterable[tuple[int, ...]]:
"""