Repository navigation
Expand file tree
/
Copy pathbf16_cache.py
More file actions
418 lines (382 loc) · 16.9 KB
/
Copy pathbf16_cache.py
File metadata and controls
418 lines (382 loc) · 16.9 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
"""Unquantized Cross-Loop KV storage used by component-efficiency ablations.
The production cache packs the same logical layout to 4 bit. This reference
backend keeps Loop 1/2 in BF16 and stores only the selected absolute Loop 3/4
rows in BF16, allowing the quantization component to be isolated cleanly.
"""
from __future__ import annotations
from typing import Sequence
import torch
def _tensor_bytes(value: torch.Tensor) -> int:
return int(value.numel() * value.element_size())
def _rank_map(tokens: int, positions: torch.Tensor, device: torch.device) -> torch.Tensor:
result = torch.full((tokens,), -1, dtype=torch.int32, device=device)
if positions.numel():
result[positions.long()] = torch.arange(
positions.numel(), dtype=torch.int32, device=device
)
return result
def _normalize_indices(
indices: torch.Tensor,
*,
batch: int,
heads: int,
tokens: int,
device: torch.device,
validate_bounds: bool,
) -> torch.Tensor:
if indices.ndim == 1:
indices = indices[None, None, :].expand(batch, heads, -1)
elif indices.ndim == 4 and indices.shape[-2] == 1:
indices = indices.squeeze(-2)
if indices.ndim != 3 or tuple(indices.shape[:2]) != (batch, heads):
raise ValueError("indices must have shape [selected] or [batch, heads, selected]")
logical = indices.to(device=device, dtype=torch.int32).contiguous()
if validate_bounds and logical.numel() and (
int(logical.min().item()) < 0 or int(logical.max().item()) >= tokens
):
raise IndexError("cache index is outside the logical sequence")
return logical
class BF16CrossLoopKVLayerBuilder:
"""Incrementally build a BF16 base-plus-sparse-override cache layer."""
def __init__(self) -> None:
self._loop1_key: torch.Tensor | None = None
self._loop1_value: torch.Tensor | None = None
self._loop2_key: torch.Tensor | None = None
self._loop2_value: torch.Tensor | None = None
@classmethod
def from_loop1(
cls,
loop1_key: torch.Tensor,
loop1_value: torch.Tensor,
**_: object,
) -> "BF16CrossLoopKVLayerBuilder":
if loop1_key.ndim != 4 or loop1_key.shape[0] != 1:
raise ValueError("Loop-1 K/V must match [1, heads, tokens, channels]")
if loop1_value.shape != loop1_key.shape:
raise ValueError("Loop-1 K/V shapes must match")
if loop1_value.dtype != loop1_key.dtype or loop1_value.device != loop1_key.device:
raise ValueError("Loop-1 K/V must share dtype and device")
result = cls()
result._loop1_key = loop1_key.clone()
result._loop1_value = loop1_value.clone()
return result
@property
def retains_dense_loop1_prefix(self) -> bool:
return True
def add_loop2(self, loop2_key: torch.Tensor, loop2_value: torch.Tensor) -> None:
if self._loop1_key is None or self._loop1_value is None:
raise RuntimeError("Loop 1 is not installed")
if self._loop2_key is not None:
raise RuntimeError("Loop 2 is already installed")
if loop2_key.shape != self._loop1_key.shape or loop2_value.shape != loop2_key.shape:
raise ValueError("Loop-2 K/V shape differs from Loop 1")
if loop2_key.dtype != self._loop1_key.dtype or loop2_key.device != self._loop1_key.device:
raise ValueError("Loop-2 K/V dtype or device differs from Loop 1")
self._loop2_key = loop2_key.clone()
self._loop2_value = loop2_value.clone()
def finalize(
self,
*,
loop3_positions: torch.Tensor,
loop3_keys: torch.Tensor,
loop3_values: torch.Tensor,
loop4_positions: torch.Tensor,
loop4_keys: torch.Tensor,
loop4_values: torch.Tensor,
) -> "BF16CrossLoopKVLayer":
if any(value is None for value in (
self._loop1_key, self._loop1_value, self._loop2_key, self._loop2_value
)):
raise RuntimeError("Loop 1/2 must be installed before finalization")
return BF16CrossLoopKVLayer.from_prefill_updates(
loop1_key=self._loop1_key, # type: ignore[arg-type]
loop1_value=self._loop1_value, # type: ignore[arg-type]
loop2_key=self._loop2_key, # type: ignore[arg-type]
loop2_value=self._loop2_value, # type: ignore[arg-type]
loop3_positions=loop3_positions,
loop3_keys=loop3_keys,
loop3_values=loop3_values,
loop4_positions=loop4_positions,
loop4_keys=loop4_keys,
loop4_values=loop4_values,
)
class BF16CrossLoopKVLayer:
"""Two dense BF16 anchors plus nested BF16 late-loop overrides."""
storage_kind = "bf16_cross_loop"
reader_backend = "bf16_cross_loop"
def __init__(self) -> None:
self._base_keys: tuple[torch.Tensor, torch.Tensor]
self._base_values: tuple[torch.Tensor, torch.Tensor]
self._late_keys: tuple[torch.Tensor, torch.Tensor]
self._late_values: tuple[torch.Tensor, torch.Tensor]
self.override_indices: tuple[torch.Tensor, torch.Tensor]
self.rank_maps: tuple[torch.Tensor, torch.Tensor]
self._decode_keys: list[torch.Tensor]
self._decode_values: list[torch.Tensor]
self._prompt_tokens = 0
self._shape_prefix = (0, 0, 0)
self._dtype = torch.bfloat16
self._device = torch.device("cpu")
@classmethod
def from_prefill_updates(
cls,
*,
loop1_key: torch.Tensor,
loop1_value: torch.Tensor,
loop2_key: torch.Tensor,
loop2_value: torch.Tensor,
loop3_positions: torch.Tensor,
loop3_keys: torch.Tensor,
loop3_values: torch.Tensor,
loop4_positions: torch.Tensor,
loop4_keys: torch.Tensor,
loop4_values: torch.Tensor,
) -> "BF16CrossLoopKVLayer":
dense = (loop1_key, loop1_value, loop2_key, loop2_value)
if loop1_key.ndim != 4 or loop1_key.shape[0] != 1:
raise ValueError("dense prefill K/V must have shape [1, heads, tokens, channels]")
if any(tensor.shape != loop1_key.shape for tensor in dense):
raise ValueError("Loop-1/2 K/V tensors must share a shape")
if any(tensor.dtype != loop1_key.dtype or tensor.device != loop1_key.device for tensor in dense):
raise ValueError("Loop-1/2 K/V tensors must share dtype and device")
batch, heads, tokens, channels = map(int, loop1_key.shape)
positions = []
for logical, key, value in (
(loop3_positions, loop3_keys, loop3_values),
(loop4_positions, loop4_keys, loop4_values),
):
logical = logical.to(device=loop1_key.device, dtype=torch.int32).contiguous()
if logical.ndim != 1 or (
logical.numel() > 1 and not bool(torch.all(logical[1:] > logical[:-1]).item())
):
raise ValueError("late-loop positions must be strictly sorted")
if logical.numel() and (
int(logical.min().item()) < 0 or int(logical.max().item()) >= tokens
):
raise IndexError("late-loop update is outside the prompt")
expected = (batch, heads, int(logical.numel()), channels)
if tuple(key.shape) != expected or value.shape != key.shape:
raise ValueError("late-loop compact K/V shape differs from its positions")
if key.dtype != loop1_key.dtype or key.device != loop1_key.device:
raise ValueError("late-loop K/V dtype or device differs from Loop 1")
positions.append(logical)
positions3, positions4 = positions
if positions4.numel():
ranks4 = torch.searchsorted(positions3, positions4)
valid = ranks4 < positions3.numel()
matched = torch.zeros_like(valid)
matched[valid] = positions3[ranks4[valid]] == positions4[valid]
if not bool(matched.all().item()):
raise ValueError("Loop-4 positions must be nested in Loop-3 positions")
result = cls()
result._prompt_tokens = tokens
result._shape_prefix = (batch, heads, channels)
result._dtype = loop1_key.dtype
result._device = loop1_key.device
result._base_keys = (loop1_key.clone(), loop2_key.clone())
result._base_values = (loop1_value.clone(), loop2_value.clone())
result._late_keys = (loop3_keys.clone(), loop4_keys.clone())
result._late_values = (loop3_values.clone(), loop4_values.clone())
result.override_indices = (positions3, positions4)
result.rank_maps = (
_rank_map(tokens, positions3, result._device),
_rank_map(tokens, positions4, result._device),
)
empty = torch.empty(
(batch, heads, 0, channels), dtype=result._dtype, device=result._device
)
result._decode_keys = [empty.clone() for _ in range(4)]
result._decode_values = [empty.clone() for _ in range(4)]
return result
@property
def token_count(self) -> int:
return self._prompt_tokens + int(self._decode_keys[0].shape[2])
@property
def tail_tokens(self) -> int:
return int(self._decode_keys[0].shape[2])
@property
def quantized_tokens(self) -> int:
return 0
@property
def storage_bytes(self) -> int:
tensors = (
*self._base_keys,
*self._base_values,
*self._late_keys,
*self._late_values,
*self._decode_keys,
*self._decode_values,
*self.override_indices,
*self.rank_maps,
)
return sum(_tensor_bytes(tensor) for tensor in tensors)
@property
def dense_bf16_bytes(self) -> int:
batch, heads, channels = self._shape_prefix
element_size = torch.empty((), dtype=self._dtype).element_size()
return 4 * 2 * batch * heads * self.token_count * channels * element_size
@property
def storage_ratio(self) -> float:
return self.storage_bytes / self.dense_bf16_bytes
def _selected(self, loop: int, logical: torch.Tensor, *, values: bool) -> torch.Tensor:
batch, heads, channels = self._shape_prefix
prompt = logical < self._prompt_tokens
safe_prompt = torch.where(prompt, logical, torch.zeros_like(logical)).long()
bases = self._base_values if values else self._base_keys
late = self._late_values if values else self._late_keys
base = bases[0 if loop == 0 else 1]
selected = torch.gather(
base,
2,
safe_prompt.unsqueeze(-1).expand(batch, heads, -1, channels),
)
for stream in range(max(0, loop - 1)):
ranks = self.rank_maps[stream].index_select(0, safe_prompt.flatten()).view_as(logical)
active = prompt & (ranks >= 0)
safe_rank = torch.where(active, ranks, torch.zeros_like(ranks)).long()
update = torch.gather(
late[stream],
2,
safe_rank.unsqueeze(-1).expand(batch, heads, -1, channels),
)
selected = torch.where(active.unsqueeze(-1), update, selected)
if self.tail_tokens:
tail = ~prompt
safe_tail = torch.where(
tail, logical - self._prompt_tokens, torch.zeros_like(logical)
).long()
decode = self._decode_values[loop] if values else self._decode_keys[loop]
tail_selected = torch.gather(
decode,
2,
safe_tail.unsqueeze(-1).expand(batch, heads, -1, channels),
)
selected = torch.where(tail.unsqueeze(-1), tail_selected, selected)
return selected
def gather(self, loop: int, indices: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
loop = int(loop)
if loop not in range(4):
raise ValueError("loop must be in [0, 3]")
if indices.ndim != 1:
raise ValueError("indices must be one-dimensional")
logical = _normalize_indices(
indices,
batch=self._shape_prefix[0],
heads=self._shape_prefix[1],
tokens=self.token_count,
device=self._device,
validate_bounds=True,
)
return self._selected(loop, logical, values=False), self._selected(loop, logical, values=True)
def qk(
self,
loop: int,
query: torch.Tensor,
indices: torch.Tensor,
*,
validate_indices: bool = True,
) -> torch.Tensor:
batch, heads, channels = self._shape_prefix
if tuple(query.shape) != (batch, heads, 1, channels):
raise ValueError("query shape differs from this cache layer")
logical = _normalize_indices(
indices,
batch=batch,
heads=heads,
tokens=self.token_count,
device=self._device,
validate_bounds=validate_indices,
)
selected = self._selected(int(loop), logical, values=False)
return query.float() @ selected.float().transpose(-1, -2)
def pv(
self,
loop: int,
weights: torch.Tensor,
indices: torch.Tensor,
*,
validate_indices: bool = True,
) -> torch.Tensor:
batch, heads, _ = self._shape_prefix
logical = _normalize_indices(
indices,
batch=batch,
heads=heads,
tokens=self.token_count,
device=self._device,
validate_bounds=validate_indices,
)
if tuple(weights.shape) != (batch, heads, 1, logical.shape[-1]):
raise ValueError("weights must match [batch, heads, 1, selected]")
selected = self._selected(int(loop), logical, values=True)
return weights.float() @ selected.float()
def cached_mass_sparse_attention_delta(
self,
loop: int,
*,
query: torch.Tensor,
current_key: torch.Tensor,
current_value: torch.Tensor,
current_position: int,
indices: torch.Tensor,
valid: torch.Tensor,
source_selected_output: torch.Tensor,
global_mass: torch.Tensor,
scaling: float,
) -> torch.Tensor:
logical = _normalize_indices(
indices,
batch=self._shape_prefix[0],
heads=self._shape_prefix[1],
tokens=max(self.token_count, int(current_position) + 1),
device=self._device,
validate_bounds=False,
)
valid = valid.to(device=self._device, dtype=torch.bool)
is_current = valid & (logical == int(current_position))
safe = torch.where(is_current, torch.zeros_like(logical), logical)
logits = self.qk(loop, query, safe, validate_indices=False)
current_logits = query.float() @ current_key.float().transpose(-1, -2)
logits = torch.where(is_current[:, :, None, :], current_logits, logits)
logits = (logits * float(scaling)).masked_fill(~valid[:, :, None, :], float("-inf"))
weights = torch.softmax(logits, dim=-1, dtype=torch.float32)
target = self.pv(
loop,
weights * (~is_current)[:, :, None, :],
safe,
validate_indices=False,
)
target += torch.sum(weights * is_current[:, :, None, :], dim=-1, keepdim=True) * current_value.float()
return global_mass.float() * (target - source_selected_output.float())
def dense_source_attention(
self,
loop: int,
*,
query: torch.Tensor,
current_key: torch.Tensor,
current_value: torch.Tensor,
scaling: float,
return_logits: bool,
) -> tuple[torch.Tensor, torch.Tensor | None]:
logical = torch.arange(self.token_count, device=self._device, dtype=torch.int32)
logical = logical[None, None].expand(self._shape_prefix[0], self._shape_prefix[1], -1)
past_logits = self.qk(loop, query, logical, validate_indices=False)
current_logits = query.float() @ current_key.float().transpose(-1, -2)
logits = torch.cat((past_logits, current_logits), dim=-1) * float(scaling)
weights = torch.softmax(logits, dim=-1, dtype=torch.float32)
output = self.pv(loop, weights[..., :-1], logical, validate_indices=False)
output += weights[..., -1:] * current_value.float()
return output, logits if return_logits else None
def append_decode_token(
self,
keys: Sequence[torch.Tensor],
values: Sequence[torch.Tensor],
) -> None:
if len(keys) != 4 or len(values) != 4:
raise ValueError("exactly four loop K/V tensors are required")
expected = (*self._shape_prefix[:2], 1, self._shape_prefix[2])
for index, (key, value) in enumerate(zip(keys, values, strict=True)):
if tuple(key.shape) != expected or value.shape != key.shape:
raise ValueError("decode K/V shape does not match the cache")
self._decode_keys[index] = torch.cat((self._decode_keys[index], key), dim=2)
self._decode_values[index] = torch.cat((self._decode_values[index], value), dim=2)