2023-05-14 22:32:38 -07:00
|
|
|
"""Token blocks."""
|
2024-06-27 13:34:55 -07:00
|
|
|
from typing import List
|
2023-02-09 11:26:21 +00:00
|
|
|
|
2023-06-17 03:07:40 -07:00
|
|
|
from vllm.utils import Device
|
2023-02-09 11:26:21 +00:00
|
|
|
|
2024-03-02 03:50:01 -05:00
|
|
|
DEFAULT_LAST_ACCESSED_TIME = -1
|
|
|
|
|
|
2023-02-09 11:26:21 +00:00
|
|
|
|
|
|
|
|
class PhysicalTokenBlock:
|
2023-05-14 22:32:38 -07:00
|
|
|
"""Represents the state of a block in the KV cache."""
|
2023-02-09 11:26:21 +00:00
|
|
|
|
|
|
|
|
def __init__(
|
|
|
|
|
self,
|
|
|
|
|
device: Device,
|
|
|
|
|
block_number: int,
|
|
|
|
|
block_size: int,
|
2024-03-02 03:50:01 -05:00
|
|
|
block_hash: int,
|
|
|
|
|
num_hashed_tokens: int,
|
2023-02-09 11:26:21 +00:00
|
|
|
) -> None:
|
|
|
|
|
self.device = device
|
|
|
|
|
self.block_number = block_number
|
|
|
|
|
self.block_size = block_size
|
2024-03-02 03:50:01 -05:00
|
|
|
self.block_hash = block_hash
|
|
|
|
|
self.num_hashed_tokens = num_hashed_tokens
|
2023-02-09 11:26:21 +00:00
|
|
|
|
|
|
|
|
self.ref_count = 0
|
2024-03-02 03:50:01 -05:00
|
|
|
self.last_accessed = DEFAULT_LAST_ACCESSED_TIME
|
|
|
|
|
|
|
|
|
|
self.computed = False
|
2023-02-14 09:34:07 +00:00
|
|
|
|
|
|
|
|
def __repr__(self) -> str:
|
|
|
|
|
return (f'PhysicalTokenBlock(device={self.device}, '
|
|
|
|
|
f'block_number={self.block_number}, '
|
2024-03-02 03:50:01 -05:00
|
|
|
f'num_hashed_tokens={self.num_hashed_tokens}, '
|
|
|
|
|
f'ref_count={self.ref_count}, '
|
|
|
|
|
f'last_accessed={self.last_accessed}, '
|
|
|
|
|
f'computed={self.computed})')
|
2024-01-17 16:32:10 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
# Mapping: logical block number -> physical block.
|
|
|
|
|
BlockTable = List[PhysicalTokenBlock]
|