Repository navigation
Expand file tree
/
Copy pathllama_cache.py
More file actions
1014 lines (879 loc) · 39.9 KB
/
Copy pathllama_cache.py
File metadata and controls
1014 lines (879 loc) · 39.9 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
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
from abc import ABC, abstractmethod
import array
from collections import OrderedDict
import ctypes
from dataclasses import dataclass
import diskcache
import hashlib
import heapq
import sys
from typing import (
Any,
List,
NamedTuple,
Optional,
Sequence,
Tuple,
)
import llama_cpp.llama as llama_core
import llama_cpp.llama_cpp as llama_cpp_lib
from .llama_types import *
def _common_prefix_length(
cached_tokens: Tuple[int, ...], query_tokens: Tuple[int, ...]
) -> int:
if cached_tokens is query_tokens:
return len(cached_tokens)
length = min(len(cached_tokens), len(query_tokens))
# Avoid allocations for early mismatches. Compare long matching runs in C
# using bounded tuple slices, without converting entire histories to NumPy.
probe_length = min(length, 32)
for index in range(probe_length):
if cached_tokens[index] != query_tokens[index]:
return index
for start in range(probe_length, length, 256):
stop = min(start + 256, length)
if cached_tokens[start:stop] != query_tokens[start:stop]:
for index in range(start, stop):
if cached_tokens[index] != query_tokens[index]:
return index
return length
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
def _get_matching_key(
self, key: Sequence[int], min_prefix_length: int
) -> Tuple[int, ...]:
if len(key) <= min_prefix_length:
raise KeyError("No better prefix in cache")
query_key = tuple(key)
cached_key = self._find_longest_prefix_key(query_key)
if cached_key is None or _common_prefix_length(cached_key, query_key) <= min_prefix_length:
raise KeyError("No better prefix in cache")
return cached_key
def get_matching_state(
self, key: Sequence[int], min_prefix_length: int = 0
) -> "llama_core.LlamaState":
"""Read the best match only if its common prefix exceeds min_prefix_length.
Custom caches retain their dictionary interface. Built-in caches also
check key metadata before materializing a snapshot.
"""
state = self[key]
prefix_length = llama_core.Llama.longest_token_prefix(
state.input_ids[:state.n_tokens], key
)
if prefix_length <= min_prefix_length:
raise KeyError("No better prefix in cache")
return state
@abstractmethod
def __getitem__(self, key: Sequence[int]) -> "llama_core.LlamaState":
raise NotImplementedError
@abstractmethod
def __contains__(self, key: Sequence[int]) -> bool:
raise NotImplementedError
@abstractmethod
def __setitem__(
self, key: Sequence[int], value: "llama_core.LlamaState"
) -> None:
raise NotImplementedError
class LlamaDiskCache(BaseLlamaCache):
"""
Disk cache for a llama.cpp model.
Delegates eviction and size management to the SQLite-backed `diskcache` library.
"""
def __init__(
self, cache_dir: str = ".cache/llama_cache", capacity_bytes: int = (2 << 30), verbose: bool = False
):
super().__init__(capacity_bytes)
self.cache_dir = cache_dir
# Keep diskcache's configured eviction policy.
self.cache = diskcache.Cache(cache_dir, size_limit=capacity_bytes)
self.verbose = verbose
@property
def cache_size(self):
# Native O(1) volume calculation
return self.cache.volume() # type: ignore
def _find_longest_prefix_key(
self,
key: Tuple[int, ...],
) -> Optional[Tuple[int, ...]]:
# Early exit if cache is empty
if len(self.cache) == 0:
return None
best_prefix_length = 0
best_key: Optional[Tuple[int, ...]] = None
query_length = len(key)
for cached_key in self.cache.iterkeys(): # type: ignore
prefix_length = _common_prefix_length(cached_key, key)
if prefix_length > best_prefix_length:
best_prefix_length = prefix_length
best_key = cached_key # type: ignore
# Perfect match found, break to prevent full-table disk scan
if best_prefix_length == query_length:
break
return best_key
def __getitem__(self, key: Sequence[int]) -> "llama_core.LlamaState":
return self.get_matching_state(key)
def get_matching_state(
self, key: Sequence[int], min_prefix_length: int = 0
) -> "llama_core.LlamaState":
return self.cache[self._get_matching_key(key, min_prefix_length)]
def __contains__(self, key: Sequence[int]) -> bool:
if len(self.cache) == 0:
return False
return self._find_longest_prefix_key(tuple(key)) is not None
def __setitem__(self, key: Sequence[int], value: "llama_core.LlamaState"):
# diskcache natively handles capacity check and eviction upon assignment
self.cache[tuple(key)] = value
class LlamaRAMCache(BaseLlamaCache):
"""
RAM cache for a llama.cpp model.
Maintains an LRU eviction policy with O(1) size tracking.
"""
def __init__(self, capacity_bytes: int = (2 << 30), verbose: bool = False):
super().__init__(capacity_bytes)
self.capacity_bytes = capacity_bytes
self.cache_state: OrderedDict[
Tuple[int, ...], "llama_core.LlamaState"
] = OrderedDict()
self._current_size = 0
self.verbose = verbose
self._insertion_order: dict[Tuple[int, ...], int] = {}
self._next_insertion_order = 0
@property
def cache_size(self):
return self._current_size
def _find_longest_prefix_key(
self,
key: Tuple[int, ...],
) -> Optional[Tuple[int, ...]]:
best_prefix_length = 0
best_key = None
for cached_key in self.cache_state:
prefix_length = _common_prefix_length(cached_key, key)
if prefix_length > best_prefix_length or (
prefix_length > 0 and prefix_length == best_prefix_length and best_key is not None
and self._insertion_order[cached_key] < self._insertion_order[best_key]
):
best_prefix_length = prefix_length
best_key = cached_key
return best_key
def __getitem__(self, key: Sequence[int]) -> "llama_core.LlamaState":
return self.get_matching_state(key)
def get_matching_state(
self, key: Sequence[int], min_prefix_length: int = 0
) -> "llama_core.LlamaState":
cached_key = self._get_matching_key(key, min_prefix_length)
value = self.cache_state[cached_key]
self.cache_state.move_to_end(cached_key)
return value
def __contains__(self, key: Sequence[int]) -> bool:
if not self.cache_state:
return False
return self._find_longest_prefix_key(tuple(key)) is not None
def __setitem__(self, key: Sequence[int], value: "llama_core.LlamaState"):
if getattr(value, "nbytes", value.llama_state_size) > self.capacity_bytes:
return
key = tuple(key)
if key in self.cache_state:
previous = self.cache_state.pop(key)
self._current_size -= getattr(previous, "nbytes", previous.llama_state_size)
self.cache_state[key] = value
if key not in self._insertion_order:
self._insertion_order[key] = self._next_insertion_order
self._next_insertion_order += 1
self._current_size += getattr(value, "nbytes", value.llama_state_size)
while self._current_size > self.capacity_bytes and len(self.cache_state) > 0:
popped_key, popped_state = self.cache_state.popitem(last=False)
del self._insertion_order[popped_key]
self._current_size -= getattr(popped_state, "nbytes", popped_state.llama_state_size)
self._current_size = max(0, self._current_size)
if len(self.cache_state) == 0:
self._current_size = 0
class _TrieCacheEntry(NamedTuple):
insertion_order: int
cache_key: Tuple[int, ...]
class TrieNode:
"""A node in the prefix tree (Trie)."""
__slots__ = (
"tokens", "children", "state", "cache_key", "insertion_order", "earliest_entry",
"_child_entries",
)
def __init__(
self, earliest_entry: Optional[_TrieCacheEntry] = None,
tokens: Tuple[int, ...] = (),
):
self.tokens = tokens # Compressed incoming edge; empty only at the root.
# Child nodes: {token_id: TrieNode}
self.children: Dict[int, "TrieNode"] = {}
# Stores the LlamaState if this node marks the end of a cached sequence.
self.state: Optional["llama_core.LlamaState"] = None
self.cache_key: Optional[Tuple[int, ...]] = None
self.insertion_order = 0
# Earliest inserted cache entry in this node's subtree, including itself.
self.earliest_entry = earliest_entry
# Allocate a heap only for wide branches that need repeated reselection.
self._child_entries: Optional[List[Tuple[int, int]]] = None
class LlamaTrieCache(BaseLlamaCache):
"""
A Llama cache implementation using a Trie for O(K) prefix lookup
and an OrderedDict for O(1) LRU eviction.
- K = length of the query key (number of tokens)
- N = total number of items in the cache
This solves the O(N*K) lookup bottleneck of the linear scan cache.
Unbranched paths are compressed into token-tuple edges, so insertion and
eviction do not allocate or discard a Python node per token.
Wide branches use bounded lazy heaps when eviction changes their earliest
entry; ordinary single-child paths need no heap allocation.
"""
def __init__(self, capacity_bytes: int = (2 << 30)):
super().__init__(capacity_bytes)
self.root = TrieNode() # The root node of the Trie
self._current_size = 0 # O(1) tracking of cache size in bytes
self._next_insertion_order = 0
self._last_inserted_node: Optional[TrieNode] = None
# LRU Tracker:
# Key: Cached token sequence (Tuple[int, ...])
# Value: The *terminal* TrieNode for that key
self.lru_tracker: OrderedDict[
Tuple[int, ...], TrieNode
] = OrderedDict()
@property
def cache_size(self) -> int:
"""Returns the current total size of the cache in bytes (O(1))."""
return self._current_size
def _find_longest_prefix_node(
self, key: Tuple[int, ...]
) -> Tuple[Optional[TrieNode], Optional[Tuple[int, ...]]]:
"""
Find a state with the longest common prefix, including descendants.
Returns: (The matching TrieNode, The matching key)
"""
node = self.root
depth = 0
query_length = len(key)
while depth < query_length:
child = node.children.get(key[depth])
if child is None:
break
edge_length = len(child.tokens)
if edge_length == 1: # The dictionary lookup already matched it.
node = child
depth += 1
continue
matched = _common_prefix_length(child.tokens, key[depth:depth + edge_length])
node = child
depth += matched
if matched < edge_length:
break
entry = node.earliest_entry if depth > 0 else None
if entry is None:
return None, None
return self.lru_tracker[entry.cache_key], entry.cache_key
def _find_longest_prefix_key(self, key: Tuple[int, ...]) -> Optional[Tuple[int, ...]]:
return self._find_longest_prefix_node(key)[1]
def __getitem__(self, key: Sequence[int]) -> "llama_core.LlamaState":
"""
Retrieves the state for the longest matching prefix in O(K) time.
Updates the LRU status.
"""
return self.get_matching_state(key)
def get_matching_state(
self, key: Sequence[int], min_prefix_length: int = 0
) -> "llama_core.LlamaState":
cached_key = self._get_matching_key(key, min_prefix_length)
node = self.lru_tracker[cached_key]
self.lru_tracker.move_to_end(cached_key)
assert node.state is not None
return node.state
def __contains__(self, key: Sequence[int]) -> bool:
"""Check for a nonempty common prefix without reading the state."""
node, _ = self._find_longest_prefix_node(tuple(key))
return node is not None
@staticmethod
def _rebuild_child_entries(node: TrieNode) -> None:
node._child_entries = [
(child.earliest_entry.insertion_order, token)
for token, child in node.children.items()
if child.earliest_entry is not None
]
heapq.heapify(node._child_entries)
@classmethod
def _record_child_entry(cls, node: TrieNode, token: int) -> None:
if node._child_entries is None:
return
if len(node.children) <= 8:
node._child_entries = None
return
child = node.children.get(token)
if child is not None and child.earliest_entry is not None:
heapq.heappush(node._child_entries, (child.earliest_entry.insertion_order, token))
# Bound stale heap entries even if the subtree's oldest entry stays live.
if len(node._child_entries) > 2 * len(node.children):
cls._rebuild_child_entries(node)
@classmethod
def _update_earliest_entry(cls, node: TrieNode) -> None:
if len(node.children) > 8:
if node._child_entries is None:
cls._rebuild_child_entries(node)
while node._child_entries:
insertion_order, token = node._child_entries[0]
child = node.children.get(token)
if (
child is not None and child.earliest_entry is not None
and child.earliest_entry.insertion_order == insertion_order
):
entry = child.earliest_entry
break
heapq.heappop(node._child_entries)
else:
entry = None
else:
node._child_entries = None
if not node.children:
entry = None
elif len(node.children) == 1:
entry = next(iter(node.children.values())).earliest_entry
else:
entry = min(
(child.earliest_entry for child in node.children.values()
if child.earliest_entry is not None),
key=lambda item: item.insertion_order, default=None,
)
if node.state is not None:
assert node.cache_key is not None
if entry is None or node.insertion_order < entry.insertion_order:
entry = _TrieCacheEntry(node.insertion_order, node.cache_key)
node.earliest_entry = entry
def _prune(self, key: Tuple[int, ...]):
"""
(Helper) Removes a key and its state from the Trie.
Also removes empty parent nodes (branch pruning).
"""
path: List[Tuple[TrieNode, int]] = []
node = self.root
depth = 0
key_length = len(key)
while depth < key_length:
token = key[depth]
child = node.children.get(token)
if child is None:
return
edge_length = len(child.tokens)
if edge_length > 1 and key[depth:depth + edge_length] != child.tokens:
return
path.append((node, token))
depth += edge_length
node = child
# 2. Remove the state
if node.state is None:
return # Node has no state
self._current_size -= getattr(node.state, "nbytes", node.state.llama_state_size)
if node is self._last_inserted_node:
self._last_inserted_node = None
removed_order = node.insertion_order
node.state = None
node.cache_key = None
# Removing a later terminal entry cannot change any ancestor's best match.
if node.earliest_entry is None or node.earliest_entry.insertion_order != removed_order:
if path:
self._compress_child(*path[-1])
return
if node.children:
self._update_earliest_entry(node)
if path:
self._compress_child(*path[-1])
else:
node.earliest_entry = None
if not path:
return
parent, token = path[-1]
del parent.children[token]
# Only surviving ancestors can need a new prefix candidate.
for index in range(len(path) - 1, -1, -1):
parent, token = path[index]
if parent._child_entries is not None:
self._record_child_entry(parent, token)
candidate_changed = (
parent.earliest_entry is not None
and parent.earliest_entry.insertion_order == removed_order
)
if candidate_changed:
self._update_earliest_entry(parent)
if index > 0:
self._compress_child(*path[index - 1])
if not candidate_changed:
break
@staticmethod
def _compress_child(parent: TrieNode, token: int) -> None:
node = parent.children[token]
if node.state is None:
if not node.children:
del parent.children[token]
elif len(node.children) == 1:
child = next(iter(node.children.values()))
child.tokens = node.tokens + child.tokens
parent.children[token] = child
def __setitem__(self, key: Sequence[int], value: "llama_core.LlamaState"):
"""
Adds a (key, state) pair, maintaining prefix candidates and LRU order.
Eviction updates only affected ancestors and wide-branch candidate heaps.
"""
state_size = getattr(value, "nbytes", value.llama_state_size)
if state_size > self.capacity_bytes:
return
key_tuple = tuple(key)
# 1. Find a terminal node, splitting compressed edges only at a branch
# or at a newly cached prefix. Existing candidates precede new entries.
node = self.root
entry = None
depth = 0
key_length = len(key_tuple)
previous = self._last_inserted_node
if previous is not None and previous.cache_key is not None:
previous_length = len(previous.cache_key)
if key_length > previous_length and key_tuple[:previous_length] == previous.cache_key:
node, depth = previous, previous_length
while depth < key_length:
token = key_tuple[depth]
child = node.children.get(token)
if child is None:
if entry is None:
entry = _TrieCacheEntry(self._next_insertion_order, key_tuple)
self._next_insertion_order += 1
child = node.children[token] = TrieNode(entry, key_tuple[depth:])
# Insertion orders only increase, so a new entry cannot replace
# an existing subtree candidate. New nodes start with it directly.
if node.earliest_entry is None:
node.earliest_entry = entry
if node._child_entries is not None:
self._record_child_entry(node, token)
node = child
break
edge_length = len(child.tokens)
if edge_length == 1:
node = child
depth += 1
continue
matched = _common_prefix_length(child.tokens, key_tuple[depth:depth + edge_length])
if matched < edge_length:
assert matched > 0
branch = TrieNode(child.earliest_entry, child.tokens[:matched])
child.tokens = child.tokens[matched:]
branch.children[child.tokens[0]] = child
node.children[token] = branch
node = branch
depth += matched
continue
depth += edge_length
node = child
# 2. Check if updating an existing item
if node.state is not None:
self._current_size -= getattr(node.state, "nbytes", node.state.llama_state_size)
self.lru_tracker.move_to_end(key_tuple)
else:
if entry is None:
entry = _TrieCacheEntry(self._next_insertion_order, key_tuple)
self._next_insertion_order += 1
node.insertion_order = entry.insertion_order
if node.earliest_entry is None:
node.earliest_entry = entry
self.lru_tracker[key_tuple] = node
# 3. Set new state and update O(1) size
node.state = value
node.cache_key = key_tuple
self._current_size += state_size
# 5. Eviction logic
while self._current_size > self.capacity_bytes and self.lru_tracker:
# Get the least recently used item in O(1)
evicted_key, _ = self.lru_tracker.popitem(last=False)
# Remove the evicted item from the Trie
self._prune(evicted_key)
self._last_inserted_node = node
# Alias for backwards compatibility
LlamaCache = LlamaTrieCache
@dataclass
class HybridCheckpoint:
"""
Represents a single snapshot of the Hybrid/Recurrent model state.
Notes:
- When on_device=False, `data` contains the full host-side serialized state.
- When on_device=True, `data` contains only the host-visible portion of the
serialized state. The tensor payload is stored in llama_context-owned
device buffers by llama.cpp, keyed by seq_id.
"""
pos: int # The token position (cursor) where this snapshot was taken.
data: bytes # The raw binary RNN state data.
hash_val: str # SHA-256 hash of the token prefix to ensure exact sequence matching.
size: int # Number of bytes written by llama_state_seq_get_data_ext().
seq_id: int # Sequence id used by llama.cpp state APIs.
pos_min: int = -1 # Minimum backend memory position covered by the snapshot.
pos_max: Optional[int] = None # Maximum backend memory position covered by the snapshot.
class HybridCheckpointCache(BaseLlamaCache):
"""
Checkpoint manager for Hybrid/Recurrent model states.
This cache is designed for models whose memory cannot be safely truncated like
a regular Transformer KV cache. For recurrent/hybrid architectures, rollback is
implemented by saving and restoring sequence state snapshots.
Two operating modes are supported:
1. Host mode: on_device=False
- Full checkpoint payload is materialized as Python bytes.
- Multiple checkpoints per seq_id are safe.
- This mode is suitable for multi-turn rollback and longer conversation reuse.
2. Device mode: on_device=True
- LLAMA_STATE_SEQ_FLAGS_ON_DEVICE is forwarded to llama.cpp.
- Tensor payloads are stored in llama_context-owned device buffers.
- The device buffers are created per seq_id in llama.cpp.
- Therefore only one active checkpoint per seq_id is safe.
- This mode is suitable for fast speculative / branch rollback where avoiding
device-to-host tensor copies is more important than keeping many historical
checkpoints.
Important:
Do not treat on_device=True as "Python owns a VRAM checkpoint". Python only
owns the host-visible serialized portion. The tensor payload lives inside the
llama_context and is keyed by seq_id.
"""
def __init__(
self,
ctx: llama_cpp_lib.llama_context_p,
max_checkpoints: int = 16,
on_device: bool = False,
verbose: bool = False
):
"""
Args:
ctx (llama_context_p):
Borrowed llama.cpp context pointer used by the state sequence APIs.
This cache does not own the context and must not free it.
max_checkpoints(int): Maximum number of Python-side checkpoint entries to keep.
- Host mode: This is the maximum number of historical checkpoints across all seq_ids.
- Device mode: This is still a global upper bound for Python-side metadata entries,
but this class also enforces at most one active checkpoint per seq_id,
because llama.cpp stores device tensor payloads per seq_id.
on_device(bool): Whether to request llama.cpp to keep tensor checkpoint payloads in
context-owned device buffers via LLAMA_STATE_SEQ_FLAGS_ON_DEVICE.
verbose(bool): Enables diagnostic logging to stderr for checkpoint save/restore/eviction.
"""
if ctx is None:
raise ValueError("HybridCheckpointCache(__init__): Failed to create HybridCheckpointCache with a null model context")
self._ctx = ctx
self._context_ref = None
self.on_device = on_device
self.verbose = verbose
# In host mode, max_checkpoints means "maximum number of Python-owned
# checkpoints across all seq_ids".
#
# In device mode, llama.cpp stores tensor payloads in device buffers keyed
# by seq_id. Multiple Python checkpoint metadata entries for the same seq_id
# would point to the same mutable device-side slot, so only one checkpoint
# per seq_id is safe.
self.max_checkpoints = max_checkpoints
# Python-side checkpoint registry.
#
# Host mode:
# Each HybridCheckpoint owns a full serialized checkpoint payload.
#
# Device mode:
# Each HybridCheckpoint owns only the host-visible serialized portion.
# The corresponding tensor payload is owned by llama_context.
self.checkpoints: list[HybridCheckpoint] = []
# Total Python-tracked checkpoint size in bytes.
#
# Host mode:
# Roughly equals the total serialized checkpoint payload size.
#
# Device mode:
# Tracks only the host-visible part returned by llama.cpp, not the
# context-owned device tensor storage.
self._current_size = 0
# Cache C API function pointers for faster repeated calls.
self._get_size_ext = llama_cpp_lib.llama_state_seq_get_size_ext
self._get_data_ext = llama_cpp_lib.llama_state_seq_get_data_ext
self._set_data_ext = llama_cpp_lib.llama_state_seq_set_data_ext
self._get_memory = llama_cpp_lib.llama_get_memory
self._memory_seq_pos_min = llama_cpp_lib.llama_memory_seq_pos_min
self._memory_seq_pos_max = llama_cpp_lib.llama_memory_seq_pos_max
self._memory_seq_rm = llama_cpp_lib.llama_memory_seq_rm
# State serialization flags forwarded to llama.cpp.
#
# LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY:
# Save only the sequence-specific / partial state needed for recurrent
# rollback instead of a full context state.
#
# LLAMA_STATE_SEQ_FLAGS_ON_DEVICE:
# Ask llama.cpp to store tensor payloads in context-owned device buffers.
self._flags = llama_cpp_lib.LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY
if on_device:
self._flags |= llama_cpp_lib.LLAMA_STATE_SEQ_FLAGS_ON_DEVICE
if self.max_checkpoints <= 0 and self.verbose:
print("HybridCheckpointCache(__init__): Cache is DISABLED (max_checkpoints <= 0). "
"Rollback capabilities are turned off. This is optimal for single-turn workflows.",
file=sys.stderr)
if self.on_device and self.max_checkpoints > 1 and self.verbose:
print(
"HybridCheckpointCache(__init__): on_device=True stores tensor payloads "
"in llama_context-owned device buffers keyed by seq_id. Multiple "
"historical checkpoints for the same seq_id are unsafe, so this cache "
"will keep only one checkpoint per seq_id.",
file=sys.stderr,
)
@property
def cache_size(self) -> int:
"""
Returns the host-visible checkpoint size tracked by Python.
In host mode, this is close to the full serialized checkpoint payload size.
In device mode, this is only the host-visible metadata/payload size returned
by llama.cpp. Device-side tensor storage is owned by llama_context and is not
fully represented by this number.
"""
return self._current_size
def clear(self):
"""
Clears Python-side checkpoint metadata.
This does not explicitly release llama_context-owned device buffers. The
device buffers are managed by llama.cpp and are associated with the context.
"""
if not self.checkpoints:
# Empty Checkpoint: Return immediately, no need to clear.
return
self.checkpoints.clear()
self._current_size = 0
if self.verbose:
print("HybridCheckpointCache(clear): cleared", file=sys.stderr)
def close(self):
self.clear()
self._context_ref = None
self._ctx = None
self._get_size_ext = None
self._get_data_ext = None
self._set_data_ext = None
self._get_memory = None
self._memory_seq_pos_min = None
self._memory_seq_pos_max = None
self._memory_seq_rm = None
def __del__(self) -> None:
self.close()
# Helper tools
def _invalidate_related(self, seq_id: int, suffix_start=None, device_only=False):
owner = self._context_ref() if self._context_ref is not None else None
if owner is not None:
owner._invalidate_checkpoints(
seq_id, suffix_start=suffix_start, device_only=device_only
)
else:
self._invalidate_memory(seq_id, suffix_start)
def _invalidate_memory(
self, seq_id: int = -1, suffix_start: Optional[int] = None,
keep_seq_id: Optional[int] = None,
) -> None:
"""Drop partial snapshots whose native prefix is no longer available."""
self.checkpoints[:] = [
cp for cp in self.checkpoints
if (seq_id >= 0 and cp.seq_id != seq_id)
or cp.seq_id == keep_seq_id
or (suffix_start is not None and cp.pos_max is not None
and 0 <= cp.pos_max < suffix_start)
]
self._current_size = sum(cp.size for cp in self.checkpoints)
def _hash_prefix(self, tokens: List[int], length: int) -> str:
"""
Computes a SHA-256 hash for a sequence of tokens up to the specified length.
This ensures that checkpoints are only restored for the EXACT same conversation history.
"""
if length <= 0:
return "empty"
length = min(length, len(tokens))
data = array.array('i', tokens[:length]).tobytes()
return hashlib.sha256(data).hexdigest()[:32]
def _evict_checkpoints_if_needed(self) -> None:
"""
Evicts old checkpoints if needed
Host mode:
This evicts full Python-owned checkpoint payloads, so FIFO historical
checkpoints are safe and useful.
Device mode:
This evicts Python-side metadata only. The device tensor payload is owned
by llama_context and is keyed by seq_id.
"""
while len(self.checkpoints) > self.max_checkpoints:
old_cp = self.checkpoints.pop(0)
self._current_size -= old_cp.size
if self._current_size < 0:
self._current_size = 0
if self.verbose:
print(
f"HybridCheckpointCache: evicted checkpoint "
f"seq_id={old_cp.seq_id}, pos={old_cp.pos}",
file=sys.stderr,
)
def find_best_checkpoint(self, tokens: List[int], seq_id: int = 0) -> Optional[HybridCheckpoint]:
"""
Finds the longest valid checkpoint that perfectly matches the provided token prefix.
The hash check prevents restoring a checkpoint that has the same length but
belongs to a different prompt/history.
Returns None if no matching checkpoint is found.
"""
# Empty Checkpoint: Instant return, no hash calculation needed.
if self.max_checkpoints <= 0 or len(self.checkpoints) == 0:
return None
best_cp: Optional[HybridCheckpoint] = None
best_pos = -1
for cp in self.checkpoints:
if cp.seq_id != seq_id or cp.pos > len(tokens):
# Skip if sequence ID mismatches or checkpoint is longer than the current prompt
continue
# Verify cryptographic integrity of the prompt history
if self._hash_prefix(tokens, cp.pos) == cp.hash_val:
if cp.pos > best_pos:
# Keep the checkpoint with the longest matching prefix (highest pos)
best_pos = cp.pos
best_cp = cp
return best_cp
def save_checkpoint(
self,
current_pos: int,
tokens: List[int],
seq_id: int = 0
) -> bool:
"""
Extracts the RNN hidden state from the C++ backend and saves it as a checkpoint.
Manages eviction (FIFO) if the maximum number of checkpoints is exceeded.
"""
# 0. Early Exit / Feature Toggle
# If the user disables checkpoints (max_checkpoints <= 0), we immediately return.
# This absolutely critical bypass prevents massive (e.g., 150MB+) synchronous
# VRAM-to-RAM copies over the PCIe bus, eliminating multi-second delays at the
# end of generation for single-turn workflows.
# This is more friendly to the single-call ComfyUI ecosystem. :)
if self.max_checkpoints <= 0:
if self.verbose:
print("HybridCheckpointCache(save_checkpoint): Cache is DISABLED (max_checkpoints <= 0). "
"Operating in single-turn conversation mode. Skipping state extraction to optimize generation latency.",
file=sys.stderr)
return False
# In on-device mode, remove old Python metadata for this seq_id before saving
# the new checkpoint. The underlying llama.cpp device buffer for this seq_id
# will be overwritten by the get_data_ext() call.
if self.on_device:
self._invalidate_related(seq_id, device_only=True)
flags = self._flags
# 1. Query the required host-visible buffer size.
# In on_device mode this may exclude the large tensor payload
# that stays in device memory.
size = self._get_size_ext(self._ctx, seq_id, flags)
if size == 0:
if self.verbose:
print("HybridCheckpointCache(save_checkpoint): size=0, skip")
return False
# 2. Allocate buffer and extract raw state data
buffer = (ctypes.c_uint8 * size)()
n_written = self._get_data_ext(self._ctx, buffer, size, seq_id, flags)
if n_written != size:
if self.verbose:
print(
f"HybridCheckpointCache(save_checkpoint): get_data_ext failed "
f"({n_written}/{size})",
file=sys.stderr,
)
return False
# Note: This deep copy isolates the state from subsequent C++ backend mutations
data_bytes = bytes(buffer)
hash_val = self._hash_prefix(tokens, current_pos)
# Token count and backend memory positions are not interchangeable for all
# models. Multimodal inputs, custom position IDs, and SWA can make pos_max
# differ from current_pos - 1, so retain the native range with the snapshot.
memory = self._get_memory(self._ctx)
pos_min = self._memory_seq_pos_min(memory, seq_id)
pos_max = self._memory_seq_pos_max(memory, seq_id)
# Replace the same cursor only after the new capture succeeds.
self.checkpoints[:] = [
cp for cp in self.checkpoints
if cp.seq_id != seq_id or cp.pos != current_pos
]
self._current_size = sum(cp.size for cp in self.checkpoints)
# 3. Store the newly extracted checkpoint
self.checkpoints.append(HybridCheckpoint(
pos=current_pos,
data=data_bytes,
hash_val=hash_val,
size=n_written,
seq_id=seq_id,
pos_min=pos_min,
pos_max=pos_max)
)
self._current_size += n_written
# 4. Evicts old checkpoints if needed
self._evict_checkpoints_if_needed()
if self.verbose:
mode = "device" if self.on_device else "host"
print(
f"HybridCheckpointCache(save_checkpoint): saved {mode} checkpoint "
f"seq_id={seq_id}, pos={current_pos}, size={size / 1024 / 1024:.2f} MiB, "
f"hcc_count={len(self.checkpoints)}, "
f"hcc_mem_used={self._current_size / 1024 / 1024:.2f} MiB",
file=sys.stderr,
)
return True
def restore_checkpoint(self, cp: HybridCheckpoint, seq_id: int = 0) -> bool:
"""
Injects a previously saved RNN state checkpoint back into the C++ backend memory.
"""
# 1. Verify sequence ID matches to prevent cross-sequence contamination
if cp.seq_id != seq_id:
if self.verbose:
print(f"HybridCheckpointCache(restore_checkpoint): [Error] Sequence ID mismatch: checkpoint has {cp.seq_id}, requested {seq_id}", file=sys.stderr)
return False
# Host partial snapshots also borrow the current attention prefix.
if self._ctx is None or not any(item is cp for item in self.checkpoints):
if self.verbose:
print(
"HybridCheckpointCache(restore_checkpoint): stale checkpoint; refusing restore.",
file=sys.stderr,
)
return False
flags = self._flags
if cp.size <= 0 or cp.size != len(cp.data):
return False
# Native consumes const bytes synchronously. Keep their owner alive.
state_bytes = bytes(cp.data)
state_pointer = ctypes.cast(ctypes.c_char_p(state_bytes), ctypes.POINTER(ctypes.c_uint8))
suffix_start = cp.pos_max + 1 if cp.pos_max is not None else cp.pos
try:
success = self._set_data_ext(
self._ctx, state_pointer, cp.size, seq_id, flags
) == cp.size
if success:
success = bool(self._memory_seq_rm(
self._get_memory(self._ctx), seq_id, suffix_start, -1
))
except BaseException:
self._invalidate_related(seq_id)
raise
# Restore and suffix cleanup form one transaction. Partial failure
# invalidates the old history even when the caller retained objects.
self._invalidate_related(seq_id, suffix_start if success else None)
if self.verbose:
mode = "device" if self.on_device else "host"
print(
f"HybridCheckpointCache(restore_checkpoint): restore "
f"{'OK' if success else 'FAIL'} "
f"mode={mode}, seq_id={seq_id}, pos={cp.pos}",