FreeToken brings datacenter-scale model serving to your desktop. Run massive models locally, fast and efficiently.
from dataclasses import dataclass
import torch
@dataclass
class BaseCaptureData:
seq_lens: torch.Tensor
positions: torch.Tensor
cu_seqlens_k: torch.Tensor
cu_seqlens_q: torch.Tensor
page_table: torch.Tensor
@classmethod
def create(cls, max_bs: int, max_seq_len: int, device: torch.device, rows_per_req: int = 1, **kwargs):
return cls(
seq_lens=torch.ones((max_bs,), dtype=torch.int32, device=device),
positions=torch.zeros((max_bs * rows_per_req,), dtype=torch.int32, device=device),
cu_seqlens_k=torch.arange(0, max_bs + 1, dtype=torch.int32, device=device),
# constant under a graph: every request has exactly `rows_per_req` query rows
cu_seqlens_q=torch.arange(0, max_bs + 1, dtype=torch.int32, device=device) * rows_per_req,
page_table=torch.zeros((max_bs, max_seq_len), dtype=torch.int32, device=device),
**kwargs,
)