Spaces:
Paused
Paused
File size: 22,231 Bytes
751ad26 | 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 | from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any, Sequence
import numpy as np
from ...config import DotCacheConfig
from ...encode import encode_page
from ...model_kv_cache import ModelPagedKVCache, default_q_head_to_kv_head
from ...page_cache import PreparedPageCache
from ...tracing import ExecutionTrace
from ...types import EncodedPage
@dataclass(frozen=True, slots=True)
class VllmBlockKey:
layer_id: int
kv_head_id: int
block_id: int
kind: str
@dataclass(slots=True)
class VllmBlockEntry:
key: VllmBlockKey
page: EncodedPage
finalized: bool
token_count: int
token_start: int
@dataclass(slots=True)
class _LiveBlockState:
block_id: int | None = None
token_start: int | None = None
key_rows: list[np.ndarray] = field(default_factory=list)
value_rows: list[np.ndarray] = field(default_factory=list)
def clear(self) -> None:
self.block_id = None
self.token_start = None
self.key_rows.clear()
self.value_rows.clear()
def _normalize_block_tensor(
values: np.ndarray,
*,
num_key_value_heads: int,
block_size: int,
head_dim: int,
name: str,
) -> np.ndarray:
array = np.asarray(values, dtype=np.float32)
if array.ndim == 5:
if array.shape[0] != 1:
raise ValueError(f"{name} batch dimension must be 1")
array = array[0]
if array.ndim != 4:
raise ValueError(f"{name} must have shape [kv_heads, block_count, block_size, head_dim]")
if int(array.shape[0]) != num_key_value_heads:
raise ValueError(f"{name} must contain {num_key_value_heads} KV heads")
if int(array.shape[2]) != block_size:
raise ValueError(f"{name} block size must equal {block_size}")
if int(array.shape[3]) != head_dim:
raise ValueError(f"{name} head_dim must equal {head_dim}")
return array
def _normalize_step_tensor(
values: np.ndarray,
*,
num_key_value_heads: int,
head_dim: int,
name: str,
) -> np.ndarray:
array = np.asarray(values, dtype=np.float32)
if array.ndim == 4:
if array.shape[0] != 1:
raise ValueError(f"{name} batch dimension must be 1")
array = array[0]
if array.ndim != 3:
raise ValueError(f"{name} must have shape [kv_heads, token_count, head_dim]")
if int(array.shape[0]) != num_key_value_heads:
raise ValueError(f"{name} must contain {num_key_value_heads} KV heads")
if int(array.shape[2]) != head_dim:
raise ValueError(f"{name} head_dim must equal {head_dim}")
return array
class VllmPagedKVCache:
def __init__(
self,
*,
config: DotCacheConfig,
num_hidden_layers: int,
num_attention_heads: int,
num_key_value_heads: int,
block_size: int,
backend: str = "torch_cuda",
cache: PreparedPageCache | None = None,
) -> None:
if config.tokens_per_page != block_size:
raise ValueError("DotCache tokens_per_page must equal the vLLM block_size for this phase")
self.config = config
self.block_size = int(block_size)
self.num_hidden_layers = int(num_hidden_layers)
self.num_attention_heads = int(num_attention_heads)
self.num_key_value_heads = int(num_key_value_heads)
self.backend = backend
self.cache = cache if cache is not None else PreparedPageCache()
self.model_kv_cache = ModelPagedKVCache(
config=config,
num_hidden_layers=num_hidden_layers,
num_attention_heads=num_attention_heads,
num_key_value_heads=num_key_value_heads,
backend=backend,
cache=self.cache,
)
self.default_q_head_to_kv_head = default_q_head_to_kv_head(num_attention_heads, num_key_value_heads)
self._blocks: dict[VllmBlockKey, VllmBlockEntry] = {}
self._live_states: dict[tuple[int, int], _LiveBlockState] = {}
@property
def resident_bytes(self) -> int:
return self.model_kv_cache.resident_bytes
def clear(self) -> None:
self._blocks.clear()
self._live_states.clear()
self.model_kv_cache.clear()
def block_entry(self, layer_id: int, kv_head_id: int, block_id: int, kind: str) -> VllmBlockEntry:
key = VllmBlockKey(int(layer_id), int(kv_head_id), int(block_id), kind)
return self._blocks[key]
def block_entries_for_layer(self, layer_id: int, *, kind: str) -> list[VllmBlockEntry]:
return sorted(
[entry for key, entry in self._blocks.items() if key.layer_id == layer_id and key.kind == kind],
key=lambda entry: (entry.key.kv_head_id, entry.key.block_id),
)
def _remove_layer_blocks(self, layer_id: int) -> None:
stale_keys = [key for key in self._blocks if key.layer_id == layer_id]
for key in stale_keys:
del self._blocks[key]
stale_live = [key for key in self._live_states if key[0] == layer_id]
for key in stale_live:
del self._live_states[key]
def sync_layer_blocks(
self,
layer_id: int,
key_blocks: np.ndarray,
value_blocks: np.ndarray,
*,
block_ids: Sequence[int] | None = None,
live_block_token_count: int = 0,
trace: ExecutionTrace | None = None,
) -> None:
keys = _normalize_block_tensor(
key_blocks,
num_key_value_heads=self.num_key_value_heads,
block_size=self.block_size,
head_dim=self.config.head_dim,
name="key_blocks",
)
values = _normalize_block_tensor(
value_blocks,
num_key_value_heads=self.num_key_value_heads,
block_size=self.block_size,
head_dim=self.config.head_dim,
name="value_blocks",
)
if keys.shape[1] != values.shape[1]:
raise ValueError("key_blocks and value_blocks must contain the same number of blocks")
block_count = int(keys.shape[1])
resolved_block_ids = tuple(range(block_count)) if block_ids is None else tuple(int(block_id) for block_id in block_ids)
if len(resolved_block_ids) != block_count:
raise ValueError("block_ids must align with the number of blocks")
if len(set(resolved_block_ids)) != len(resolved_block_ids):
raise ValueError("block_ids must be unique")
if live_block_token_count < 0 or live_block_token_count > self.block_size:
raise ValueError("live_block_token_count must be in [0, block_size]")
self._remove_layer_blocks(layer_id)
self.model_kv_cache.clear_layer(layer_id)
if block_count == 0:
return
finalized_block_count = block_count if live_block_token_count in (0, self.block_size) else block_count - 1
full_tokens = finalized_block_count * self.block_size
dense_keys = keys[:, :finalized_block_count].reshape(
self.num_key_value_heads,
full_tokens,
self.config.head_dim,
) if finalized_block_count > 0 else np.zeros((self.num_key_value_heads, 0, self.config.head_dim), dtype=np.float32)
dense_values = values[:, :finalized_block_count].reshape(
self.num_key_value_heads,
full_tokens,
self.config.head_dim,
) if finalized_block_count > 0 else np.zeros((self.num_key_value_heads, 0, self.config.head_dim), dtype=np.float32)
if finalized_block_count < block_count:
live_keys = keys[:, finalized_block_count, :live_block_token_count]
live_values = values[:, finalized_block_count, :live_block_token_count]
dense_keys = np.concatenate([dense_keys, live_keys], axis=1)
dense_values = np.concatenate([dense_values, live_values], axis=1)
self.model_kv_cache.ingest_prefill_cache(layer_id, dense_keys, dense_values, trace=trace)
self.model_kv_cache.prepare_static_pages(trace=trace)
for block_index, block_id in enumerate(resolved_block_ids[:finalized_block_count]):
token_start = block_index * self.block_size
for kv_head_id in range(self.num_key_value_heads):
key_page = encode_page(
keys[kv_head_id, block_index],
self.config,
kind="K",
layer_id=layer_id,
kv_head_id=kv_head_id,
token_start=token_start,
mode=self.config.resolve_page_mode(kind="K", layer_id=layer_id, kv_head_id=kv_head_id),
build_runtime_metadata=False,
)
value_page = encode_page(
values[kv_head_id, block_index],
self.config,
kind="V",
layer_id=layer_id,
kv_head_id=kv_head_id,
token_start=token_start,
mode=self.config.resolve_page_mode(kind="V", layer_id=layer_id, kv_head_id=kv_head_id),
build_runtime_metadata=False,
)
self._blocks[VllmBlockKey(layer_id, kv_head_id, block_id, "K")] = VllmBlockEntry(
key=VllmBlockKey(layer_id, kv_head_id, block_id, "K"),
page=key_page,
finalized=True,
token_count=self.block_size,
token_start=token_start,
)
self._blocks[VllmBlockKey(layer_id, kv_head_id, block_id, "V")] = VllmBlockEntry(
key=VllmBlockKey(layer_id, kv_head_id, block_id, "V"),
page=value_page,
finalized=True,
token_count=self.block_size,
token_start=token_start,
)
if finalized_block_count < block_count:
live_block_id = resolved_block_ids[finalized_block_count]
live_token_start = finalized_block_count * self.block_size
for kv_head_id in range(self.num_key_value_heads):
live_key_rows = keys[kv_head_id, finalized_block_count, :live_block_token_count]
live_value_rows = values[kv_head_id, finalized_block_count, :live_block_token_count]
key_page = encode_page(
live_key_rows,
self.config,
kind="K",
layer_id=layer_id,
kv_head_id=kv_head_id,
token_start=live_token_start,
mode="M3",
build_runtime_metadata=False,
)
value_page = encode_page(
live_value_rows,
self.config,
kind="V",
layer_id=layer_id,
kv_head_id=kv_head_id,
token_start=live_token_start,
mode="M3",
build_runtime_metadata=False,
)
self._blocks[VllmBlockKey(layer_id, kv_head_id, live_block_id, "K")] = VllmBlockEntry(
key=VllmBlockKey(layer_id, kv_head_id, live_block_id, "K"),
page=key_page,
finalized=False,
token_count=live_block_token_count,
token_start=live_token_start,
)
self._blocks[VllmBlockKey(layer_id, kv_head_id, live_block_id, "V")] = VllmBlockEntry(
key=VllmBlockKey(layer_id, kv_head_id, live_block_id, "V"),
page=value_page,
finalized=False,
token_count=live_block_token_count,
token_start=live_token_start,
)
state = _LiveBlockState(
block_id=live_block_id,
token_start=live_token_start,
key_rows=[np.asarray(row, dtype=np.float32) for row in live_key_rows],
value_rows=[np.asarray(row, dtype=np.float32) for row in live_value_rows],
)
self._live_states[(layer_id, kv_head_id)] = state
def append_step(
self,
layer_id: int,
key_step: np.ndarray,
value_step: np.ndarray,
token_index: int,
*,
trace: ExecutionTrace | None = None,
) -> None:
keys = _normalize_step_tensor(
key_step,
num_key_value_heads=self.num_key_value_heads,
head_dim=self.config.head_dim,
name="key_step",
)
values = _normalize_step_tensor(
value_step,
num_key_value_heads=self.num_key_value_heads,
head_dim=self.config.head_dim,
name="value_step",
)
self.model_kv_cache.append_step(layer_id, keys, values, token_index, trace=trace)
self._update_block_entries_from_steps(layer_id, keys, values, token_index)
def append_step_torch(
self,
layer_id: int,
key_step,
value_step,
token_index: int,
*,
trace: ExecutionTrace | None = None,
) -> None:
self.model_kv_cache.append_step_torch(layer_id, key_step, value_step, token_index, trace=trace)
keys = key_step.detach().to(dtype=key_step.dtype).cpu().numpy().astype(np.float32, copy=False)
values = value_step.detach().to(dtype=value_step.dtype).cpu().numpy().astype(np.float32, copy=False)
self._update_block_entries_from_steps(layer_id, keys, values, token_index)
def _store_block_entry(
self,
layer_id: int,
kv_head_id: int,
block_id: int,
kind: str,
rows: np.ndarray,
*,
token_start: int,
) -> None:
token_count = int(rows.shape[0])
finalized = token_count == self.block_size
page = encode_page(
np.asarray(rows, dtype=np.float32),
self.config,
kind=kind,
layer_id=layer_id,
kv_head_id=kv_head_id,
token_start=token_start,
mode=(
self.config.resolve_page_mode(kind=kind, layer_id=layer_id, kv_head_id=kv_head_id)
if finalized
else "M3"
),
build_runtime_metadata=False,
)
self._blocks[VllmBlockKey(layer_id, kv_head_id, block_id, kind)] = VllmBlockEntry(
key=VllmBlockKey(layer_id, kv_head_id, block_id, kind),
page=page,
finalized=finalized,
token_count=token_count,
token_start=token_start,
)
def _update_block_entries_from_chunk(
self,
layer_id: int,
key_rows_by_head: np.ndarray,
value_rows_by_head: np.ndarray,
token_index: int,
) -> None:
token_count = int(key_rows_by_head.shape[1])
if token_count == 0:
return
block_id = int(token_index) // self.block_size
block_offset = int(token_index) % self.block_size
token_start = block_id * self.block_size
if block_offset + token_count > self.block_size:
raise ValueError("chunk update cannot span multiple blocks")
for kv_head_id in range(self.num_key_value_heads):
state = self._live_states.setdefault((layer_id, kv_head_id), _LiveBlockState())
if block_offset == 0 or state.block_id != block_id:
state.clear()
state.block_id = block_id
state.token_start = token_start
state.key_rows.extend(np.asarray(row, dtype=np.float32) for row in key_rows_by_head[kv_head_id])
state.value_rows.extend(np.asarray(row, dtype=np.float32) for row in value_rows_by_head[kv_head_id])
key_rows = np.stack(state.key_rows, axis=0).astype(np.float32, copy=False)
value_rows = np.stack(state.value_rows, axis=0).astype(np.float32, copy=False)
self._store_block_entry(
layer_id,
kv_head_id,
block_id,
"K",
key_rows,
token_start=token_start,
)
self._store_block_entry(
layer_id,
kv_head_id,
block_id,
"V",
value_rows,
token_start=token_start,
)
if key_rows.shape[0] == self.block_size:
state.clear()
def append_tokens_torch(
self,
layer_id: int,
key_rows,
value_rows,
token_positions,
*,
trace: ExecutionTrace | None = None,
) -> None:
positions = token_positions.reshape(-1).detach().cpu().numpy().astype(np.int64, copy=False)
if positions.size == 0:
return
expected = np.arange(int(positions[0]), int(positions[0]) + positions.size, dtype=np.int64)
if not np.array_equal(positions, expected):
raise ValueError("Phase 6 vLLM adapter requires contiguous batch=1 token positions")
keys_by_head = key_rows.transpose(0, 1)
values_by_head = value_rows.transpose(0, 1)
offset = 0
while offset < positions.size:
token_index = int(positions[offset])
block_offset = token_index % self.block_size
chunk_size = min(self.block_size - block_offset, positions.size - offset)
key_chunk = keys_by_head[:, offset : offset + chunk_size]
value_chunk = values_by_head[:, offset : offset + chunk_size]
self.model_kv_cache.append_step_torch(
layer_id,
key_chunk,
value_chunk,
token_index,
trace=trace,
)
key_chunk_cpu = key_chunk.detach().cpu().numpy().astype(np.float32, copy=False)
value_chunk_cpu = value_chunk.detach().cpu().numpy().astype(np.float32, copy=False)
self._update_block_entries_from_chunk(layer_id, key_chunk_cpu, value_chunk_cpu, token_index)
offset += chunk_size
def _update_block_entries_from_steps(
self,
layer_id: int,
key_rows_by_head: np.ndarray,
value_rows_by_head: np.ndarray,
token_index: int,
) -> None:
token_count = int(key_rows_by_head.shape[1])
for offset in range(token_count):
absolute_token = int(token_index) + offset
block_id = absolute_token // self.block_size
block_offset = absolute_token % self.block_size
token_start = block_id * self.block_size
for kv_head_id in range(self.num_key_value_heads):
state = self._live_states.setdefault((layer_id, kv_head_id), _LiveBlockState())
if block_offset == 0 or state.block_id != block_id:
state.clear()
state.block_id = block_id
state.token_start = token_start
state.key_rows.append(np.asarray(key_rows_by_head[kv_head_id, offset], dtype=np.float32))
state.value_rows.append(np.asarray(value_rows_by_head[kv_head_id, offset], dtype=np.float32))
key_rows = np.stack(state.key_rows, axis=0).astype(np.float32, copy=False)
value_rows = np.stack(state.value_rows, axis=0).astype(np.float32, copy=False)
finalized = key_rows.shape[0] == self.block_size
key_mode = self.config.resolve_page_mode(kind="K", layer_id=layer_id, kv_head_id=kv_head_id) if finalized else "M3"
value_mode = self.config.resolve_page_mode(kind="V", layer_id=layer_id, kv_head_id=kv_head_id) if finalized else "M3"
key_page = encode_page(
key_rows,
self.config,
kind="K",
layer_id=layer_id,
kv_head_id=kv_head_id,
token_start=token_start,
mode=key_mode,
build_runtime_metadata=False,
)
value_page = encode_page(
value_rows,
self.config,
kind="V",
layer_id=layer_id,
kv_head_id=kv_head_id,
token_start=token_start,
mode=value_mode,
build_runtime_metadata=False,
)
self._blocks[VllmBlockKey(layer_id, kv_head_id, block_id, "K")] = VllmBlockEntry(
key=VllmBlockKey(layer_id, kv_head_id, block_id, "K"),
page=key_page,
finalized=finalized,
token_count=int(key_rows.shape[0]),
token_start=token_start,
)
self._blocks[VllmBlockKey(layer_id, kv_head_id, block_id, "V")] = VllmBlockEntry(
key=VllmBlockKey(layer_id, kv_head_id, block_id, "V"),
page=value_page,
finalized=finalized,
token_count=int(value_rows.shape[0]),
token_start=token_start,
)
if finalized:
state.clear()
def decode_layer(
self,
layer_id: int,
query_step: np.ndarray,
q_head_to_kv_head: Sequence[int] | np.ndarray,
*,
query_scale: float = 1.0,
trace: ExecutionTrace | None = None,
) -> np.ndarray:
return self.model_kv_cache.decode_layer(
layer_id,
query_step,
q_head_to_kv_head,
query_scale=query_scale,
trace=trace,
)
def decode_layer_torch(
self,
layer_id: int,
query_step,
q_head_to_kv_head: Sequence[int] | np.ndarray,
*,
query_scale: float = 1.0,
trace: ExecutionTrace | None = None,
):
return self.model_kv_cache.decode_layer_torch(
layer_id,
query_step,
q_head_to_kv_head,
query_scale=query_scale,
trace=trace,
)
|