Skip to content

vllm.v1.attention.backends.recoverssm_metadata

Classes:

RecoverSSMPostprocessMetadata dataclass

Metadata used during postprocessing for align-mode prefix caching.

Source code in vllm/v1/attention/backends/recoverssm_metadata.py 
10
11
12
13
14
15
16
17
18
@dataclass(frozen=True)
class RecoverSSMPostprocessMetadata:
    """Metadata used during postprocessing for align-mode prefix caching."""

    num_spec_decodes: int
    request_indices: torch.Tensor | None
    block_table: torch.Tensor
    num_computed_tokens: torch.Tensor
    block_size: int