This repository was archived by the owner on Sep 13, 2026. It is now read-only.
Repository navigation
Expand file tree
/
Copy pathllama_cache.py
More file actions
155 lines (126 loc) · 4.89 KB
/
Copy pathllama_cache.py
File metadata and controls
155 lines (126 loc) · 4.89 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
import sys
from abc import ABC, abstractmethod
from typing import (
Optional,
Sequence,
Tuple,
)
from collections import OrderedDict
import diskcache
import llama_cpp.llama
from .llama_types import *
class BaseLlamaCache(ABC):
"""Base cache class for a llama.cpp model."""
def __init__(self, capacity_bytes: int = (2 << 30)):
self.capacity_bytes = capacity_bytes
@property
@abstractmethod
def cache_size(self) -> int:
raise NotImplementedError
def _find_longest_prefix_key(
self,
key: Tuple[int, ...],
) -> Optional[Tuple[int, ...]]:
pass
@abstractmethod
def __getitem__(self, key: Sequence[int]) -> "llama_cpp.llama.LlamaState":
raise NotImplementedError
@abstractmethod
def __contains__(self, key: Sequence[int]) -> bool:
raise NotImplementedError
@abstractmethod
def __setitem__(
self, key: Sequence[int], value: "llama_cpp.llama.LlamaState"
) -> None:
raise NotImplementedError
class LlamaRAMCache(BaseLlamaCache):
"""Cache for a llama.cpp model using RAM."""
def __init__(self, capacity_bytes: int = (2 << 30)):
super().__init__(capacity_bytes)
self.capacity_bytes = capacity_bytes
self.cache_state: OrderedDict[
Tuple[int, ...], "llama_cpp.llama.LlamaState"
] = OrderedDict()
@property
def cache_size(self):
return sum([state.llama_state_size for state in self.cache_state.values()])
def _find_longest_prefix_key(
self,
key: Tuple[int, ...],
) -> Optional[Tuple[int, ...]]:
min_len = 0
min_key = None
keys = (
(k, llama_cpp.llama.Llama.longest_token_prefix(k, key))
for k in self.cache_state.keys()
)
for k, prefix_len in keys:
if prefix_len > min_len:
min_len = prefix_len
min_key = k
return min_key
def __getitem__(self, key: Sequence[int]) -> "llama_cpp.llama.LlamaState":
key = tuple(key)
_key = self._find_longest_prefix_key(key)
if _key is None:
raise KeyError("Key not found")
value = self.cache_state[_key]
self.cache_state.move_to_end(_key)
return value
def __contains__(self, key: Sequence[int]) -> bool:
return self._find_longest_prefix_key(tuple(key)) is not None
def __setitem__(self, key: Sequence[int], value: "llama_cpp.llama.LlamaState"):
key = tuple(key)
if key in self.cache_state:
del self.cache_state[key]
self.cache_state[key] = value
while self.cache_size > self.capacity_bytes and len(self.cache_state) > 0:
self.cache_state.popitem(last=False)
# Alias for backwards compatibility
LlamaCache = LlamaRAMCache
class LlamaDiskCache(BaseLlamaCache):
"""Cache for a llama.cpp model using disk."""
def __init__(
self, cache_dir: str = ".cache/llama_cache", capacity_bytes: int = (2 << 30)
):
super().__init__(capacity_bytes)
self.cache = diskcache.Cache(cache_dir)
@property
def cache_size(self):
return int(self.cache.volume()) # type: ignore
def _find_longest_prefix_key(
self,
key: Tuple[int, ...],
) -> Optional[Tuple[int, ...]]:
min_len = 0
min_key: Optional[Tuple[int, ...]] = None
for k in self.cache.iterkeys(): # type: ignore
prefix_len = llama_cpp.llama.Llama.longest_token_prefix(k, key)
if prefix_len > min_len:
min_len = prefix_len
min_key = k # type: ignore
return min_key
def __getitem__(self, key: Sequence[int]) -> "llama_cpp.llama.LlamaState":
key = tuple(key)
_key = self._find_longest_prefix_key(key)
if _key is None:
raise KeyError("Key not found")
value: "llama_cpp.llama.LlamaState" = self.cache.pop(_key) # type: ignore
# NOTE: This puts an integer as key in cache, which breaks,
# Llama.longest_token_prefix(k, key) above since k is not a tuple of ints/tokens
# self.cache.push(_key, side="front") # type: ignore
return value
def __contains__(self, key: Sequence[int]) -> bool:
return self._find_longest_prefix_key(tuple(key)) is not None
def __setitem__(self, key: Sequence[int], value: "llama_cpp.llama.LlamaState"):
print("LlamaDiskCache.__setitem__: called", file=sys.stderr)
key = tuple(key)
if key in self.cache:
print("LlamaDiskCache.__setitem__: delete", file=sys.stderr)
del self.cache[key]
self.cache[key] = value
print("LlamaDiskCache.__setitem__: set", file=sys.stderr)
while self.cache_size > self.capacity_bytes and len(self.cache) > 0:
key_to_remove = next(iter(self.cache))
del self.cache[key_to_remove]
print("LlamaDiskCache.__setitem__: trim", file=sys.stderr)