Skip to content

vllm.models.deepseek_v41.nvidia.ops.candidate_indexer

DSV4.1 candidate-only prefill and decode over the paged indexer cache.

Classes:

  • CandidatePlan –

    One step's row bounds and sorted candidates, shared by consumer layers.

CandidatePlan dataclass

One step's row bounds and sorted candidates, shared by consumer layers.

Source code in vllm/models/deepseek_v41/nvidia/ops/candidate_indexer.py
@dataclass
class CandidatePlan:
    """One step's row bounds and sorted candidates, shared by consumer layers."""

    block_table: torch.Tensor
    row_indices: torch.Tensor
    row_lens: torch.Tensor
    row_starts: torch.Tensor
    sparse_indices: torch.Tensor
    physical_blocks: torch.Tensor
    valid_end: torch.Tensor
    columns: torch.Tensor
    logits: torch.Tensor
    sparse_block_size: int
    prepared: bool = False