Repository navigation
Expand file tree
/
Copy pathengine.py
More file actions
607 lines (563 loc) · 23.7 KB
/
Copy pathengine.py
File metadata and controls
607 lines (563 loc) · 23.7 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
"""Batch-one Ouro inference orchestration for FlashLoop."""
from __future__ import annotations
import math
from dataclasses import dataclass
from typing import Any, Callable
import torch
from .cache import CrossLoopKVCache, CrossLoopKVLayerBuilder
from .bf16_cache import BF16CrossLoopKVLayerBuilder
from .config import FlashLoopConfig, validate_ouro_model
from .prefill import PrefillReceipt, dense_prefill_loop, sparse_prefill_loop
from .selection import select_prefill_tokens
from .weights import fused_mlp, project_qkv, project_value, repack_projection_weights
@dataclass(frozen=True)
class EngineState:
cache: CrossLoopKVCache
next_position: int
mask_loop3: torch.Tensor
mask_loop4: torch.Tensor
prefill_receipts: tuple[PrefillReceipt, ...]
@dataclass(frozen=True)
class PrefillOutput:
logits: torch.Tensor
state: EngineState
@dataclass(frozen=True)
class DecodeReceipt:
dense_attention_calls: int
sparse_attention_calls: int
dense_late_loop_calls: int
selected_keys_per_layer: tuple[int, ...]
@dataclass(frozen=True)
class DecodeOutput:
logits: torch.Tensor
state: EngineState
receipt: DecodeReceipt
@dataclass(frozen=True)
class _LayerSource:
projected_output: torch.Tensor
indices: torch.Tensor
valid: torch.Tensor
global_mass: torch.Tensor
selected_weights: torch.Tensor
selected_output: torch.Tensor
def _rotate_half(values: torch.Tensor) -> torch.Tensor:
first, second = values.chunk(2, dim=-1)
return torch.cat((-second, first), dim=-1)
def _apply_rope(
query: torch.Tensor,
key: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
cos = cos.unsqueeze(1)
sin = sin.unsqueeze(1)
return (
query * cos + _rotate_half(query) * sin,
key * cos + _rotate_half(key) * sin,
)
def _hidden_delta_score(current: torch.Tensor, previous: torch.Tensor) -> torch.Tensor:
numerator = torch.linalg.vector_norm(current.float() - previous.float(), dim=-1)
denominator = torch.linalg.vector_norm(current.float(), dim=-1).clamp_min(1e-12)
return numerator / denominator
def _pad_decode_indices(
indices: torch.Tensor,
*,
block_size: int = 64,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Pad decode indices to a stable capacity and return their validity mask.
One spare block is reserved at exact boundaries so an autoregressive append
does not immediately change the tensor shape and trigger a new GPU kernel
specialization.
"""
block_size = int(block_size)
if block_size <= 0:
raise ValueError("block_size must be positive")
count = int(indices.shape[-1])
capacity = (count // block_size + 1) * block_size
padded = torch.zeros(
(*indices.shape[:-1], capacity),
dtype=indices.dtype,
device=indices.device,
)
padded[..., :count] = indices
valid = torch.zeros_like(padded, dtype=torch.bool)
valid[..., :count] = True
return padded, valid
class FlashLoopEngine:
"""Narrow adapter that reuses official Ouro weights without copying them."""
def __init__(
self,
model: torch.nn.Module,
*,
config: FlashLoopConfig | None = None,
validate_model: bool = True,
) -> None:
if validate_model:
validate_ouro_model(model)
elif int(getattr(model.config, "total_ut_steps", 0)) != 4:
raise ValueError("FlashLoop requires exactly four recurrent loops")
self.model = model.eval()
self.config = config or FlashLoopConfig()
self.layers = self.model.model.layers
self.backbone = self.model.model
if validate_model and self.config.fuse_projection_weights:
repack_projection_weights(self.layers)
self._state: EngineState | None = None
@classmethod
def from_pretrained(
cls,
model_path: str,
*,
config: FlashLoopConfig | None = None,
torch_dtype: torch.dtype = torch.bfloat16,
device: str | torch.device = "cuda",
) -> "FlashLoopEngine":
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
model_path,
trust_remote_code=True,
torch_dtype=torch_dtype,
).to(device)
return cls(model, config=config, validate_model=True)
@property
def state(self) -> EngineState:
if self._state is None:
raise RuntimeError("engine cache is not initialized")
return self._state
def reset(self) -> None:
self._state = None
@torch.inference_mode()
def prefill(
self,
input_ids: torch.Tensor,
attention_mask: torch.Tensor | None = None,
*,
stage_observer: Callable[[str], None] | None = None,
) -> PrefillOutput:
def stage(name: str) -> None:
if stage_observer is not None:
stage_observer(name)
stage("start")
if self._state is not None:
raise RuntimeError("engine cache is already initialized; call reset()")
if input_ids.ndim != 2 or input_ids.shape[0] != 1:
raise ValueError("the first FlashLoop engine supports batch size one")
if input_ids.shape[1] <= 0:
raise ValueError("prefill requires at least one prompt token")
if attention_mask is not None:
if attention_mask.shape != input_ids.shape or not bool(attention_mask.bool().all().item()):
raise ValueError("padding is not supported by the first batch-one engine")
hidden = self.backbone.embed_tokens(input_ids)
tokens = int(input_ids.shape[1])
positions = torch.arange(tokens, device=input_ids.device).unsqueeze(0)
rope = self.backbone.rotary_emb(hidden, positions)
# Dense SDPA uses ``is_causal=True`` and sparse prefill constructs only
# its active-query mask, so retaining an [tokens, tokens] FP32 mask is
# unnecessary (256 MiB at an 8K prompt).
causal = None
stage("inputs_ready")
loop1 = dense_prefill_loop(
self.layers,
hidden,
position_embeddings=rope,
attention_mask=causal,
final_norm=self.backbone.norm,
loop=1,
)
hidden1 = loop1.hidden_states
cache1 = loop1.composite_cache
updates1: list[Any] = list(loop1.updates)
receipt1 = loop1.receipt
del loop1
stage("loop1_dense")
builders: list[CrossLoopKVLayerBuilder | BF16CrossLoopKVLayerBuilder] = []
builder_type = (
CrossLoopKVLayerBuilder
if self.config.quantize_cross_loop_kv
else BF16CrossLoopKVLayerBuilder
)
for layer_index in range(len(self.layers)):
update1 = updates1[layer_index]
builders.append(
builder_type.from_loop1(
update1.keys,
update1.values,
group_size=self.config.group_size,
residual_length=self.config.residual_length,
absolute_late_overrides=self.config.absolute_late_overrides,
absolute_override_dtype=self.config.absolute_override_dtype,
reader_backend=self.config.kv_reader_backend,
)
)
updates1[layer_index] = None
del updates1, cache1
stage("loop1_packed")
loop2 = dense_prefill_loop(
self.layers,
hidden1,
position_embeddings=rope,
attention_mask=causal,
final_norm=self.backbone.norm,
loop=2,
)
hidden2 = loop2.hidden_states
cache2 = loop2.composite_cache
updates2: list[Any] = list(loop2.updates)
receipt2 = loop2.receipt
del loop2
stage("loop2_dense")
for layer_index, builder in enumerate(builders):
update2 = updates2[layer_index]
builder.add_loop2(update2.keys, update2.values)
stage("loop2_delta_packed")
fraction3, fraction4 = self.config.effective_prefill_fractions
mask3 = select_prefill_tokens(
_hidden_delta_score(hidden2, hidden1),
fraction=fraction3,
)
del hidden1
loop3 = sparse_prefill_loop(
self.layers,
hidden2,
active_mask=mask3,
composite_cache=cache2,
position_embeddings=rope,
attention_mask=causal,
final_norm=self.backbone.norm,
loop=3,
consume_inputs=True,
retain_updates=False,
)
hidden3 = loop3.hidden_states
cache3 = loop3.composite_cache
receipt3 = loop3.receipt
del loop3, cache2
stage("loop3_sparse")
mask4 = select_prefill_tokens(
_hidden_delta_score(hidden3, hidden2),
fraction=fraction4,
eligible=mask3,
)
positions3 = torch.nonzero(mask3[0], as_tuple=False).flatten().to(torch.int32)
positions4 = torch.nonzero(mask4[0], as_tuple=False).flatten().to(torch.int32)
positions3_long = positions3.long()
positions4_long = positions4.long()
ranks4_in3 = torch.searchsorted(positions3, positions4).long()
loop3_at4_keys: list[torch.Tensor | None] = []
loop3_at4_values: list[torch.Tensor | None] = []
for key, value in zip(cache3.keys, cache3.values, strict=True):
loop3_at4_keys.append(key.index_select(2, positions4_long))
loop3_at4_values.append(value.index_select(2, positions4_long))
del hidden2
loop4 = sparse_prefill_loop(
self.layers,
hidden3,
active_mask=mask4,
eligible_mask=mask3,
composite_cache=cache3,
position_embeddings=rope,
attention_mask=causal,
final_norm=self.backbone.norm,
loop=4,
consume_inputs=True,
retain_updates=False,
)
hidden4 = loop4.hidden_states
cache4 = loop4.composite_cache
receipt4 = loop4.receipt
del loop4, cache3, hidden3
stage("loop4_sparse")
physical = CrossLoopKVCache(num_layers=len(self.layers))
for layer_index in range(len(self.layers)):
loop3_key = cache4.keys[layer_index].index_select(2, positions3_long)
loop3_value = cache4.values[layer_index].index_select(2, positions3_long)
if positions4.numel():
snapshot_key = loop3_at4_keys[layer_index]
snapshot_value = loop3_at4_values[layer_index]
assert snapshot_key is not None and snapshot_value is not None
loop3_key.index_copy_(2, ranks4_in3, snapshot_key)
loop3_value.index_copy_(2, ranks4_in3, snapshot_value)
loop4_key = cache4.keys[layer_index].index_select(2, positions4_long)
loop4_value = cache4.values[layer_index].index_select(2, positions4_long)
physical.set_layer(
layer_index,
builders[layer_index].finalize(
loop3_positions=positions3,
loop3_keys=loop3_key,
loop3_values=loop3_value,
loop4_positions=positions4,
loop4_keys=loop4_key,
loop4_values=loop4_value,
),
)
updates2[layer_index] = None
loop3_at4_keys[layer_index] = None
loop3_at4_values[layer_index] = None
cache4.keys[layer_index] = torch.empty(0, device=hidden4.device)
cache4.values[layer_index] = torch.empty(0, device=hidden4.device)
stage("physical_cache_finalized")
self._state = EngineState(
cache=physical,
next_position=tokens,
mask_loop3=mask3,
mask_loop4=mask4,
prefill_receipts=(receipt1, receipt2, receipt3, receipt4),
)
logits = self.model.lm_head(hidden4[:, -1]).float()
stage("logits_ready")
return PrefillOutput(logits=logits, state=self._state)
def _selected_value_aggregate(
self,
*,
layer_index: int,
loop: int,
selected_indices: torch.Tensor,
selected_weights: torch.Tensor,
current_value: torch.Tensor,
past_tokens: int,
) -> torch.Tensor:
is_current = selected_indices == past_tokens
safe = torch.where(is_current, torch.zeros_like(selected_indices), selected_indices)
past_weights = selected_weights * (~is_current)[:, :, None, :]
output = self.state.cache.pv(
layer_index,
loop,
past_weights,
safe,
validate_indices=False,
)
output += torch.sum(
selected_weights * is_current[:, :, None, :],
dim=-1,
keepdim=True,
) * current_value.float()
return output
@torch.inference_mode()
def decode(self, token_ids: torch.Tensor) -> DecodeOutput:
state = self.state
if token_ids.shape != (1, 1):
raise ValueError("decode requires exactly one token at batch size one")
hidden = self.backbone.embed_tokens(token_ids)
position_ids = torch.tensor(
[[state.next_position]], device=token_ids.device, dtype=torch.long
)
rope = self.backbone.rotary_emb(hidden, position_ids)
past_tokens = state.next_position
current_keys: list[list[torch.Tensor | None]] = [
[None] * 4 for _ in range(len(self.layers))
]
current_values: list[list[torch.Tensor | None]] = [
[None] * 4 for _ in range(len(self.layers))
]
sources: list[_LayerSource | None] = [None] * len(self.layers)
selected_counts: list[int] = []
for loop in range(4):
for layer_index, layer in enumerate(self.layers):
residual = hidden
normalized = layer.input_layernorm(hidden)
attention = layer.self_attn
head_dim = int(attention.head_dim)
query_heads = int(attention.q_proj.out_features // head_dim)
kv_heads = int(attention.k_proj.out_features // head_dim)
if loop >= 2 and self.config.reuse_selected_probabilities:
value_raw = project_value(attention, normalized)
value = value_raw.view(
1, 1, kv_heads, head_dim
).transpose(1, 2)
key = current_keys[layer_index][1]
assert key is not None
query = None
else:
query_raw, key_raw, value_raw = project_qkv(attention, normalized)
query = query_raw.view(
1, 1, query_heads, head_dim
).transpose(1, 2)
key = key_raw.view(1, 1, kv_heads, head_dim).transpose(1, 2)
value = value_raw.view(
1, 1, kv_heads, head_dim
).transpose(1, 2)
query, key = _apply_rope(query, key, *rope)
current_keys[layer_index][loop] = key
current_values[layer_index][loop] = value
if loop < 2 or not self.config.enable_sparse_decode:
assert query is not None
unprojected, source_logits = state.cache.dense_source_attention(
layer_index,
loop,
query=query,
current_key=key,
current_value=value,
scaling=float(attention.scaling),
return_logits=loop == 1 and self.config.enable_sparse_decode,
)
projected = attention.o_proj(
unprojected.to(query.dtype)
.transpose(1, 2)
.reshape(1, 1, -1)
)
if loop == 1 and self.config.enable_sparse_decode:
assert source_logits is not None
weights = torch.softmax(
source_logits, dim=-1, dtype=torch.float32
)
budget = min(
past_tokens + 1,
max(
1,
math.ceil(
(past_tokens + 1)
* self.config.effective_decode_key_fraction
),
),
)
top = torch.topk(weights, k=budget, dim=-1)
selected_positions = torch.sort(top.indices, dim=-1).values
logical_positions = torch.arange(
past_tokens + 1,
device=hidden.device,
dtype=torch.int32,
)[None, None].expand(1, query_heads, -1)
selected_indices = torch.gather(
logical_positions,
-1,
selected_positions.squeeze(-2).long(),
)
selected_global = torch.gather(
weights,
-1,
selected_positions,
)
selected_indices, selected_valid = _pad_decode_indices(
selected_indices
)
padded_global = torch.zeros(
(*selected_global.shape[:-1], selected_indices.shape[-1]),
device=selected_global.device,
dtype=selected_global.dtype,
)
padded_global[..., :budget] = selected_global
selected_global = padded_global
mass = selected_global.sum(dim=-1, keepdim=True)
selected_weights = selected_global / mass
selected_output = self._selected_value_aggregate(
layer_index=layer_index,
loop=1,
selected_indices=selected_indices,
selected_weights=selected_weights,
current_value=value,
past_tokens=past_tokens,
)
sources[layer_index] = _LayerSource(
projected_output=projected,
indices=selected_indices,
valid=selected_valid,
global_mass=mass,
selected_weights=selected_weights,
selected_output=selected_output,
)
selected_counts.append(budget)
else:
source = sources[layer_index]
assert source is not None
if self.config.reuse_selected_probabilities:
target = self._selected_value_aggregate(
layer_index=layer_index,
loop=loop,
selected_indices=source.indices,
selected_weights=source.selected_weights,
current_value=value,
past_tokens=past_tokens,
)
delta = source.global_mass.float() * (
target - source.selected_output.float()
)
else:
assert query is not None
delta = state.cache.cached_mass_sparse_attention_delta(
layer_index,
loop,
query=query,
current_key=key,
current_value=value,
current_position=past_tokens,
indices=source.indices,
valid=source.valid,
source_selected_output=source.selected_output,
global_mass=source.global_mass,
scaling=float(attention.scaling),
)
projected = source.projected_output + attention.o_proj(
delta.to(value.dtype).transpose(1, 2).reshape(1, 1, -1)
)
hidden = residual + layer.input_layernorm_2(projected)
residual = hidden
hidden = residual + layer.post_attention_layernorm_2(
fused_mlp(layer.mlp, layer.post_attention_layernorm(hidden))
)
hidden = self.backbone.norm(hidden)
for layer_index, physical_layer in enumerate(state.cache.layers):
assert physical_layer is not None
layer_keys = current_keys[layer_index]
layer_values = current_values[layer_index]
assert all(value is not None for value in (*layer_keys, *layer_values))
physical_layer.append_decode_token(layer_keys, layer_values) # type: ignore[arg-type]
self._state = EngineState(
cache=state.cache,
next_position=past_tokens + 1,
mask_loop3=state.mask_loop3,
mask_loop4=state.mask_loop4,
prefill_receipts=state.prefill_receipts,
)
logits = self.model.lm_head(hidden[:, -1]).float()
return DecodeOutput(
logits=logits,
state=self._state,
receipt=DecodeReceipt(
dense_attention_calls=(
(2 if self.config.enable_sparse_decode else 4) * len(self.layers)
),
sparse_attention_calls=(
(2 if self.config.enable_sparse_decode else 0) * len(self.layers)
),
dense_late_loop_calls=(
(0 if self.config.enable_sparse_decode else 2) * len(self.layers)
),
selected_keys_per_layer=tuple(selected_counts),
),
)
@torch.inference_mode()
def generate(
self,
input_ids: torch.Tensor,
*,
max_new_tokens: int,
eos_token_id: int | None = None,
stop_token_sequences: tuple[tuple[int, ...], ...] = (),
) -> torch.Tensor:
if int(max_new_tokens) <= 0:
raise ValueError("max_new_tokens must be positive")
if eos_token_id is None:
eos_token_id = getattr(self.model.config, "eos_token_id", None)
prefill = self.prefill(input_ids)
generated = input_ids
logits = prefill.logits
completion_tokens: list[int] = []
for step in range(int(max_new_tokens)):
next_token = torch.argmax(logits, dim=-1, keepdim=True)
generated = torch.cat([generated, next_token], dim=1)
completion_tokens.append(int(next_token.item()))
reached_eos = (
eos_token_id is not None
and completion_tokens[-1] == int(eos_token_id)
)
reached_suffix = any(
sequence
and len(completion_tokens) >= len(sequence)
and tuple(completion_tokens[-len(sequence) :]) == sequence
for sequence in stop_token_sequences
)
if reached_eos or reached_suffix:
break
if step + 1 < int(max_new_tokens):
logits = self.decode(next_token).logits
return generated