Skip to content

vllm.v1.attention.backends.mamba1_attn

Mamba1AttentionBackend

Bases: AttentionBackend

Source code in vllm/v1/attention/backends/mamba1_attn.py 
13
14
15
16
class Mamba1AttentionBackend(AttentionBackend):
    @staticmethod
    def get_builder_cls() -> type["Mamba1AttentionMetadataBuilder"]:
        return Mamba1AttentionMetadataBuilder

get_builder_cls staticmethod 

get_builder_cls() -> type[Mamba1AttentionMetadataBuilder]
Source code in vllm/v1/attention/backends/mamba1_attn.py 
14
15
16
@staticmethod
def get_builder_cls() -> type["Mamba1AttentionMetadataBuilder"]:
    return Mamba1AttentionMetadataBuilder

Mamba1AttentionMetadata dataclass

Bases: BaseMambaAttentionMetadata

Source code in vllm/v1/attention/backends/mamba1_attn.py 
19
20
21
@dataclass
class Mamba1AttentionMetadata(BaseMambaAttentionMetadata):
    pass

__init__ 

__init__(
    num_prefills: int,
    num_prefill_tokens: int,
    num_decodes: int,
    num_decode_tokens: int,
    num_reqs: int,
    has_initial_states_p: Tensor | None,
    query_start_loc_p: Tensor | None,
    num_computed_tokens_p: Tensor | None,
    state_indices_tensor: Tensor,
    block_idx_last_scheduled_token: Tensor | None,
    block_idx_first_scheduled_token_p: Tensor | None,
    block_idx_last_computed_token: Tensor | None,
    nums_dict: dict | None = None,
    batch_ptr: Tensor | None = None,
    token_chunk_offset_ptr: Tensor | None = None,
) -> None

Mamba1AttentionMetadataBuilder

Bases: BaseMambaAttentionMetadataBuilder[Mamba1AttentionMetadata]

Source code in vllm/v1/attention/backends/mamba1_attn.py 
24
25
26
27
28
class Mamba1AttentionMetadataBuilder(
    BaseMambaAttentionMetadataBuilder[Mamba1AttentionMetadata]
):
    metadata_cls = Mamba1AttentionMetadata
    supports_update_block_table: bool = False

metadata_cls class-attribute instance-attribute 

metadata_cls = Mamba1AttentionMetadata

supports_update_block_table class-attribute instance-attribute 

supports_update_block_table: bool = False