File size: 42,941 Bytes
03c3dff | 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 | # coding=utf-8
# Copyright 2025 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Any, Callable, Optional, Union
import torch
from transformers.models.qwen3.configuration_qwen3 import Qwen3Config
from .runtime_config import _get_shared_batch_mask, _parse_config_pattern
from .layer_norm_rope import rotate_half
class Qwen3RingBufferCache:
"""
Ring buffer KV cache with summary support.
Two strategies based on per-layer sliding_chunk_num:
- Large window layers (is_large_window=True): append-only buffer storing only text KV.
Summary KV is NOT stored since text tokens attend to all text KV directly.
- Small window layers (is_large_window=False):
Three buffers:
1. key_cache: [ring(ws) | old_summaries(growing) | chunk_mirror(≤C)]
→ attention input, steady state is a single contiguous slice
2. new_summary_buf: ring buffer of size scn, stores summaries whose text
is still in the window (not needed for attention)
3. chunk_buf: size C, holds current chunk's text KV
RoPE position information is baked into KV, so physical order doesn't matter.
"""
is_compileable = False
_SUMMARY_INIT_CAP = 512
_APPEND_HEADROOM = 1024
def __init__(self, config: Qwen3Config, sliding_chunk_nums: list[int]):
super().__init__()
self.summary_chunk_size = getattr(config, "summary_chunk_size", 0)
self.summary_token_num = getattr(config, "summary_token_num", 0)
self.num_hidden_layers = config.num_hidden_layers
self.sliding_chunk_nums = sliding_chunk_nums
large_window_threshold = min(sliding_chunk_nums) * self.summary_chunk_size
self.is_large_window = [sv * self.summary_chunk_size > large_window_threshold for sv in sliding_chunk_nums]
self.window_sizes = [sv * self.summary_chunk_size for sv in sliding_chunk_nums]
pool_mode = getattr(config, "summary_meta_pool_mode", "none") or "none"
pool_groups = _parse_config_pattern(
getattr(config, "summary_meta_pool_group_size", 0) or 0
)
if not isinstance(pool_groups, list):
pool_groups = [int(pool_groups)] * config.num_hidden_layers
self.meta_pool_group_sizes = [int(v) for v in pool_groups]
self.meta_pool_enabled = [
pool_mode in ("mean", "scalar", "dimension_wise")
and group_size > 1
for group_size in self.meta_pool_group_sizes
]
self.meta_pool_mode = pool_mode
self.key_cache = [None for _ in range(config.num_hidden_layers)]
self.value_cache = [None for _ in range(config.num_hidden_layers)]
# Full-layer Lightning Indexer keys are not main-attention K/V. The
# summary-indexer variant stores one key per raw chunk; token-indexer
# mode stores one key per raw text token.
self._sksa_index_keys = [None for _ in range(config.num_hidden_layers)]
self._sparse_gather_storage = {}
# Independent chunk-Q/K mode needs the auxiliary K states for the
# currently accumulating raw chunk. They are separate from the
# backbone attention cache and are discarded once its summary token is
# emitted.
self._sksa_chunk_k = [None for _ in range(config.num_hidden_layers)]
self._sksa_chunk_v = [None for _ in range(config.num_hidden_layers)]
# Large window: append-only
self._text_len = [0] * config.num_hidden_layers
self._capacity = [0] * config.num_hidden_layers
# Small window: ring buffer + summary
self._window_write_ptr = [0] * config.num_hidden_layers
self._n_valid_window = [0] * config.num_hidden_layers
self._old_summary_len = [0] * config.num_hidden_layers # old summaries in key_cache
self._old_summary_cap = [0] * config.num_hidden_layers
self._old_summary_total = [0] * config.num_hidden_layers
# Hierarchical old-summary state. Complete groups are stored as post-RoPE
# meta KV; fewer than group_size evicted summaries remain residual KV.
self._meta_key_cache = [None for _ in range(config.num_hidden_layers)]
self._meta_value_cache = [None for _ in range(config.num_hidden_layers)]
self._meta_len = [0] * config.num_hidden_layers
self._meta_cap = [0] * config.num_hidden_layers
self._residual_pre_key_buf = [None for _ in range(config.num_hidden_layers)]
self._residual_score_buf = [None for _ in range(config.num_hidden_layers)]
self._residual_cos_buf = [None for _ in range(config.num_hidden_layers)]
self._residual_sin_buf = [None for _ in range(config.num_hidden_layers)]
self._residual_position_buf = [None for _ in range(config.num_hidden_layers)]
# New summary ring buffer (small window only): summaries whose text is still in window
self._new_sum_key_buf = [None for _ in range(config.num_hidden_layers)]
self._new_sum_value_buf = [None for _ in range(config.num_hidden_layers)]
self._new_sum_pre_key_buf = [None for _ in range(config.num_hidden_layers)]
self._new_sum_score_buf = [None for _ in range(config.num_hidden_layers)]
self._new_sum_cos_buf = [None for _ in range(config.num_hidden_layers)]
self._new_sum_sin_buf = [None for _ in range(config.num_hidden_layers)]
self._new_sum_position_buf = [None for _ in range(config.num_hidden_layers)]
self._new_sum_len = [0] * config.num_hidden_layers # how many filled (≤ scn)
self._new_sum_write_ptr = [0] * config.num_hidden_layers # ring write pointer
# Mirror region length (small window only): partial chunk text KV count in key_cache
self._chunk_buf_len = [0] * config.num_hidden_layers
# Common
self.cur_chunk_sizes = [0] * config.num_hidden_layers
self.true_tokens = [0] * config.num_hidden_layers
self._total_chunks = [0] * config.num_hidden_layers # completed chunks count
self._reorganized = [False] * config.num_hidden_layers
self._prefill_text_positions = None
self._prefill_summary_positions = None
def __len__(self):
return self.num_hidden_layers
def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
"""Returns nonzero only when this layer's cache is in post-reorganize state."""
if layer_idx >= self.num_hidden_layers:
return 0
if not self._reorganized[layer_idx]:
return 0
if self.is_large_window[layer_idx]:
return self._text_len[layer_idx]
else:
return (self._n_valid_window[layer_idx] + self._chunk_buf_len[layer_idx]
+ self._old_summary_len[layer_idx] + self._new_sum_len[layer_idx])
def get_cur_chunk_size(self, layer_idx: Optional[int] = None) -> int:
if layer_idx is None:
layer_idx = self.num_hidden_layers - 1
return self.cur_chunk_sizes[layer_idx]
def get_true_token_num(self, layer_idx: Optional[int] = None) -> int:
if layer_idx is None:
layer_idx = self.num_hidden_layers - 1
return self.true_tokens[layer_idx]
def set_sksa_index_keys(
self, layer_idx: int, index_keys: torch.Tensor
) -> None:
if index_keys.ndim != 3:
raise ValueError("S-KSA index keys must use [B,M,D] layout")
self._sksa_index_keys[layer_idx] = index_keys.contiguous()
def append_sksa_index_keys(
self, layer_idx: int, index_keys: torch.Tensor
) -> None:
if index_keys.ndim != 3:
raise ValueError("S-KSA index keys must use [B,M,D] layout")
current = self._sksa_index_keys[layer_idx]
if current is None:
self._sksa_index_keys[layer_idx] = index_keys.contiguous()
else:
if current.shape[0] != index_keys.shape[0] or current.shape[2] != index_keys.shape[2]:
raise ValueError("S-KSA index key cache geometry changed during decode")
self._sksa_index_keys[layer_idx] = torch.cat(
(current, index_keys), dim=1
)
def get_sksa_index_keys(self, layer_idx: int) -> torch.Tensor:
index_keys = self._sksa_index_keys[layer_idx]
if index_keys is None:
raise RuntimeError(f"S-KSA index key cache for layer {layer_idx} is uninitialized")
return index_keys
def set_sksa_chunk_k_from_prefill(
self,
layer_idx: int,
chunk_key_states: Optional[torch.Tensor],
summary_mask: torch.Tensor,
) -> None:
"""Retain auxiliary K for the incomplete raw chunk after prefill."""
if chunk_key_states is None:
self._sksa_chunk_k[layer_idx] = None
return
if chunk_key_states.ndim != 4:
raise ValueError("S-KSA chunk K must use [B,H,S,D] layout")
shared = _get_shared_batch_mask(summary_mask, "summary_mask")
positions = self._prefill_text_positions
text_key = (chunk_key_states.index_select(2, positions) if positions is not None
else chunk_key_states[:, :, ~shared, :])
partial = text_key.shape[2] % self.summary_chunk_size
if partial:
self._sksa_chunk_k[layer_idx] = text_key[:, :, -partial:, :].contiguous()
else:
self._sksa_chunk_k[layer_idx] = None
def append_sksa_chunk_k(
self, layer_idx: int, chunk_key_states: Optional[torch.Tensor]
) -> None:
"""Append one decoded raw token's independent auxiliary K state."""
if chunk_key_states is None:
return
if chunk_key_states.ndim != 4 or chunk_key_states.shape[2] != 1:
raise ValueError("decoded S-KSA chunk K must use [B,H,1,D] layout")
current = self._sksa_chunk_k[layer_idx]
self._sksa_chunk_k[layer_idx] = (
chunk_key_states.contiguous()
if current is None
else torch.cat((current, chunk_key_states), dim=2).contiguous()
)
def get_current_sksa_chunk_k(self, layer_idx: int) -> Optional[torch.Tensor]:
return self._sksa_chunk_k[layer_idx]
def clear_sksa_chunk_k(self, layer_idx: int) -> None:
self._sksa_chunk_k[layer_idx] = None
def set_sksa_chunk_v_from_prefill(
self,
layer_idx: int,
chunk_value_states: Optional[torch.Tensor],
summary_mask: torch.Tensor,
) -> None:
"""Retain auxiliary V for the incomplete raw chunk after prefill."""
if chunk_value_states is None:
self._sksa_chunk_v[layer_idx] = None
return
if chunk_value_states.ndim != 4:
raise ValueError("S-KSA chunk V must use [B,H,S,D] layout")
shared = _get_shared_batch_mask(summary_mask, "summary_mask")
positions = self._prefill_text_positions
text_value = (chunk_value_states.index_select(2, positions) if positions is not None
else chunk_value_states[:, :, ~shared, :])
partial = text_value.shape[2] % self.summary_chunk_size
if partial:
self._sksa_chunk_v[layer_idx] = text_value[:, :, -partial:, :].contiguous()
else:
self._sksa_chunk_v[layer_idx] = None
def append_sksa_chunk_v(
self, layer_idx: int, chunk_value_states: Optional[torch.Tensor]
) -> None:
"""Append one decoded raw token's independent auxiliary V state."""
if chunk_value_states is None:
return
if chunk_value_states.ndim != 4 or chunk_value_states.shape[2] != 1:
raise ValueError("decoded S-KSA chunk V must use [B,H,1,D] layout")
current = self._sksa_chunk_v[layer_idx]
self._sksa_chunk_v[layer_idx] = (
chunk_value_states.contiguous()
if current is None
else torch.cat((current, chunk_value_states), dim=2).contiguous()
)
def get_current_sksa_chunk_v(self, layer_idx: int) -> Optional[torch.Tensor]:
return self._sksa_chunk_v[layer_idx]
def clear_sksa_chunk_v(self, layer_idx: int) -> None:
self._sksa_chunk_v[layer_idx] = None
# ── Prefill: per-layer one-shot init into ring buffer layout ──
def update(
self,
key_states: torch.Tensor,
value_states: torch.Tensor,
layer_idx: int,
cache_kwargs: Optional[dict[str, Any]] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Deprecated for prefill — use init_layer_from_prefill instead.
Kept only to surface accidental calls; the prefill path now skips this
intermediate full-length buffer entirely.
"""
raise NotImplementedError(
"Qwen3RingBufferCache.update() is no longer used; prefill goes through "
"init_layer_from_prefill, decode goes through update_text/update_summary."
)
def prepare_prefill(self, text_positions, summary_positions):
self._prefill_text_positions = text_positions
self._prefill_summary_positions = summary_positions
def init_layer_from_prefill(
self,
key_states: torch.Tensor,
value_states: torch.Tensor,
summary_mask: torch.Tensor,
layer_idx: int,
*,
pre_rope_key_states: Optional[torch.Tensor] = None,
summary_scores: Optional[torch.Tensor] = None,
summary_cos: Optional[torch.Tensor] = None,
summary_sin: Optional[torch.Tensor] = None,
summary_position_ids: Optional[torch.Tensor] = None,
meta_key_states: Optional[torch.Tensor] = None,
meta_value_states: Optional[torch.Tensor] = None,
sksa_chunk_key_states: Optional[torch.Tensor] = None,
):
"""Allocate this layer's cache directly into the post-reorganize ring buffer
layout from prefill K/V (already projected + RoPE applied), with NO intermediate
full-length buffer. Replaces the old update() + reorganize_after_prefill() pair
for the prefill path.
key_states / value_states: [bsz, kv_heads, prefill_len_expanded, head_dim]
summary_mask: [bsz, prefill_len_expanded] (bool, batch-uniform)
"""
assert not self._reorganized[layer_idx], f"layer {layer_idx} already reorganized"
prefill_k = key_states
prefill_v = value_states
bsz, heads, prefill_len, head_dim = prefill_k.shape
device, dtype = prefill_k.device, prefill_k.dtype
sm = _get_shared_batch_mask(summary_mask, "summary_mask")
text_mask = ~sm
text_positions = self._prefill_text_positions
summary_positions = self._prefill_summary_positions
if text_positions is None:
text_positions = torch.nonzero(text_mask, as_tuple=False).flatten()
summary_positions = torch.nonzero(sm, as_tuple=False).flatten()
n_text = text_positions.numel()
block = self.summary_chunk_size + self.summary_token_num
# Match what update() + reorganize_after_prefill() would have set:
# update(): cur_chunk_sizes += prefill_len; true_tokens += n_text
# reorganize: cur_chunk_sizes %= block
self.cur_chunk_sizes[layer_idx] = prefill_len % block
self.true_tokens[layer_idx] = n_text
if self.is_large_window[layer_idx]:
text_k = prefill_k.index_select(2, text_positions)
text_v = prefill_v.index_select(2, text_positions)
cap = n_text + self._APPEND_HEADROOM
new_k = torch.empty(bsz, heads, cap, head_dim, dtype=dtype, device=device)
new_v = torch.empty(bsz, heads, cap, head_dim, dtype=dtype, device=device)
if n_text > 0:
new_k[:, :, :n_text, :].copy_(text_k)
new_v[:, :, :n_text, :].copy_(text_v)
self.key_cache[layer_idx] = new_k
self.value_cache[layer_idx] = new_v
self._text_len[layer_idx] = n_text
self._capacity[layer_idx] = cap
self._total_chunks[layer_idx] = n_text // self.summary_chunk_size
else:
all_summary_k = prefill_k.index_select(2, summary_positions)
all_summary_v = prefill_v.index_select(2, summary_positions)
n_summary = all_summary_k.shape[2]
C = self.summary_chunk_size
ws = self.window_sizes[layer_idx]
scn = self.sliding_chunk_nums[layer_idx]
n_complete_chunks = n_text // C
n_partial = n_text % C
n_complete_text = n_complete_chunks * C
n_window_chunks = min(scn, n_complete_chunks)
n_window_text = n_window_chunks * C
window_start = n_complete_text - n_window_text
# Only retain the live window and partial chunk, avoiding a
# full-context text KV temporary on compressed layers.
live_positions = text_positions[window_start:]
text_k = prefill_k.index_select(2, live_positions)
text_v = prefill_v.index_select(2, live_positions)
n_old_total = max(0, n_summary - n_window_chunks)
n_new = n_summary - n_old_total
hierarchical = self.meta_pool_enabled[layer_idx]
group_size = self.meta_pool_group_sizes[layer_idx]
n_meta = n_old_total // group_size if hierarchical else 0
n_old = n_old_total % group_size if hierarchical else n_old_total
old_begin = n_meta * group_size if hierarchical else 0
old_s_cap = max(self._SUMMARY_INIT_CAP, (n_old + 1) * 2)
total_cap = ws + old_s_cap + C
new_k = torch.empty(bsz, heads, total_cap, head_dim, dtype=dtype, device=device)
new_v = torch.empty(bsz, heads, total_cap, head_dim, dtype=dtype, device=device)
if n_window_text > 0:
new_k[:, :, :n_window_text, :].copy_(text_k[:, :, :n_window_text, :])
new_v[:, :, :n_window_text, :].copy_(text_v[:, :, :n_window_text, :])
self._n_valid_window[layer_idx] = n_window_text
self._window_write_ptr[layer_idx] = n_window_text % ws
if n_old > 0:
new_k[:, :, ws:ws + n_old, :].copy_(
all_summary_k[:, :, old_begin:n_old_total, :]
)
new_v[:, :, ws:ws + n_old, :].copy_(
all_summary_v[:, :, old_begin:n_old_total, :]
)
self._old_summary_len[layer_idx] = n_old
self._old_summary_total[layer_idx] = n_old_total
self._old_summary_cap[layer_idx] = old_s_cap
if hierarchical:
required = (
pre_rope_key_states,
summary_scores,
summary_cos,
summary_sin,
summary_position_ids,
meta_key_states,
meta_value_states,
)
if any(item is None for item in required):
raise ValueError("hierarchical prefill cache requires pre-RoPE and meta state")
all_pre_k = pre_rope_key_states.index_select(2, summary_positions)
meta_cap = max(self._SUMMARY_INIT_CAP, (n_meta + 1) * 2)
meta_k_buf = torch.empty(
bsz, heads, meta_cap, head_dim, dtype=dtype, device=device
)
meta_v_buf = torch.empty_like(meta_k_buf)
if n_meta:
meta_k_buf[:, :, :n_meta, :].copy_(meta_key_states[:, :, :n_meta, :])
meta_v_buf[:, :, :n_meta, :].copy_(meta_value_states[:, :, :n_meta, :])
self._meta_key_cache[layer_idx] = meta_k_buf
self._meta_value_cache[layer_idx] = meta_v_buf
self._meta_len[layer_idx] = n_meta
self._meta_cap[layer_idx] = meta_cap
residual_pre = torch.empty(
bsz, heads, group_size, head_dim, dtype=dtype, device=device
)
score_width = (
heads * head_dim
if self.meta_pool_mode == "dimension_wise"
else 1
)
residual_score = torch.empty(
bsz,
group_size,
score_width,
dtype=summary_scores.dtype,
device=device,
)
residual_cos = torch.empty(
bsz, group_size, head_dim, dtype=summary_cos.dtype, device=device
)
residual_sin = torch.empty_like(residual_cos)
residual_position = torch.empty(
bsz, group_size, dtype=torch.long, device=device
)
if n_old:
residual_pre[:, :, :n_old, :].copy_(
all_pre_k[:, :, old_begin:n_old_total, :]
)
residual_score[:, :n_old, :].copy_(
summary_scores[:, old_begin:n_old_total]
)
residual_cos[:, :n_old, :].copy_(
summary_cos[:, old_begin:n_old_total, :]
)
residual_sin[:, :n_old, :].copy_(
summary_sin[:, old_begin:n_old_total, :]
)
residual_position[:, :n_old].copy_(
summary_position_ids[:, old_begin:n_old_total]
)
self._residual_pre_key_buf[layer_idx] = residual_pre
self._residual_score_buf[layer_idx] = residual_score
self._residual_cos_buf[layer_idx] = residual_cos
self._residual_sin_buf[layer_idx] = residual_sin
self._residual_position_buf[layer_idx] = residual_position
if n_partial > 0:
mirror_start = ws + n_old
new_k[:, :, mirror_start:mirror_start + n_partial, :].copy_(
text_k[:, :, n_window_text:, :])
new_v[:, :, mirror_start:mirror_start + n_partial, :].copy_(
text_v[:, :, n_window_text:, :])
self.key_cache[layer_idx] = new_k
self.value_cache[layer_idx] = new_v
self._capacity[layer_idx] = total_cap
self._text_len[layer_idx] = 0
ns_buf_k = torch.empty(bsz, heads, scn, head_dim, dtype=dtype, device=device)
ns_buf_v = torch.empty(bsz, heads, scn, head_dim, dtype=dtype, device=device)
if n_new > 0:
ns_buf_k[:, :, :n_new, :].copy_(all_summary_k[:, :, n_old_total:, :])
ns_buf_v[:, :, :n_new, :].copy_(all_summary_v[:, :, n_old_total:, :])
self._new_sum_key_buf[layer_idx] = ns_buf_k
self._new_sum_value_buf[layer_idx] = ns_buf_v
if hierarchical:
ns_pre = torch.empty(bsz, heads, scn, head_dim, dtype=dtype, device=device)
score_width = (
heads * head_dim
if self.meta_pool_mode == "dimension_wise"
else 1
)
ns_score = torch.empty(
bsz,
scn,
score_width,
dtype=summary_scores.dtype,
device=device,
)
ns_cos = torch.empty(bsz, scn, head_dim, dtype=summary_cos.dtype, device=device)
ns_sin = torch.empty_like(ns_cos)
ns_position = torch.empty(
bsz, scn, dtype=torch.long, device=device
)
if n_new:
ns_pre[:, :, :n_new, :].copy_(all_pre_k[:, :, n_old_total:, :])
ns_score[:, :n_new, :].copy_(summary_scores[:, n_old_total:])
ns_cos[:, :n_new, :].copy_(summary_cos[:, n_old_total:, :])
ns_sin[:, :n_new, :].copy_(summary_sin[:, n_old_total:, :])
ns_position[:, :n_new].copy_(
summary_position_ids[:, n_old_total:]
)
self._new_sum_pre_key_buf[layer_idx] = ns_pre
self._new_sum_score_buf[layer_idx] = ns_score
self._new_sum_cos_buf[layer_idx] = ns_cos
self._new_sum_sin_buf[layer_idx] = ns_sin
self._new_sum_position_buf[layer_idx] = ns_position
self._new_sum_len[layer_idx] = n_new
self._new_sum_write_ptr[layer_idx] = n_new % scn
self._chunk_buf_len[layer_idx] = n_partial
self._total_chunks[layer_idx] = n_old_total + n_new
self._reorganized[layer_idx] = True
self.set_sksa_chunk_k_from_prefill(
layer_idx, sksa_chunk_key_states, summary_mask
)
# ── Reorganize after prefill ──
def reorganize_after_prefill(self, summary_mask: torch.Tensor):
"""No-op sanity check.
Prefill path now writes each layer directly into the ring buffer layout via
init_layer_from_prefill, so by the time Qwen3Model.forward calls this,
all layers must already be reorganized.
"""
for layer_idx in range(self.num_hidden_layers):
assert self._reorganized[layer_idx], (
f"layer {layer_idx} was not initialized via init_layer_from_prefill"
)
self._prefill_text_positions = None
self._prefill_summary_positions = None
# ── Decode: text token update ──
def update_text(
self,
key_states: torch.Tensor,
value_states: torch.Tensor,
layer_idx: int,
*,
sksa_chunk_key_states: Optional[torch.Tensor] = None,
):
"""Write a single text token KV during decode."""
if self.is_large_window[layer_idx]:
cur = self._text_len[layer_idx]
new_len = cur + 1
if new_len > self._capacity[layer_idx]:
cap = max(new_len + self._APPEND_HEADROOM, self._capacity[layer_idx] * 2)
old_k, old_v = self.key_cache[layer_idx], self.value_cache[layer_idx]
bsz, heads, _, head_dim = old_k.shape
new_k = torch.empty(bsz, heads, cap, head_dim, dtype=old_k.dtype, device=old_k.device)
new_v = torch.empty(bsz, heads, cap, head_dim, dtype=old_v.dtype, device=old_v.device)
new_k[:, :, :cur, :].copy_(old_k[:, :, :cur, :])
new_v[:, :, :cur, :].copy_(old_v[:, :, :cur, :])
self.key_cache[layer_idx] = new_k
self.value_cache[layer_idx] = new_v
self._capacity[layer_idx] = cap
self.key_cache[layer_idx][:, :, cur:new_len, :].copy_(key_states)
self.value_cache[layer_idx][:, :, cur:new_len, :].copy_(value_states)
self._text_len[layer_idx] = new_len
else:
# Write only to key_cache mirror region (chunk_buf eliminated)
ws = self.window_sizes[layer_idx]
n_old = self._old_summary_len[layer_idx]
pos = self._chunk_buf_len[layer_idx]
mirror_pos = ws + n_old + pos
self.key_cache[layer_idx][:, :, mirror_pos:mirror_pos+1, :].copy_(key_states)
self.value_cache[layer_idx][:, :, mirror_pos:mirror_pos+1, :].copy_(value_states)
self._chunk_buf_len[layer_idx] = pos + 1
self.cur_chunk_sizes[layer_idx] += 1
self.true_tokens[layer_idx] += 1
self.append_sksa_chunk_k(layer_idx, sksa_chunk_key_states)
# ── Decode: summary token update ──
def update_summary(
self,
key_states: torch.Tensor,
value_states: torch.Tensor,
layer_idx: int,
*,
pre_rope_key_states: Optional[torch.Tensor] = None,
summary_scores: Optional[torch.Tensor] = None,
summary_cos: Optional[torch.Tensor] = None,
summary_sin: Optional[torch.Tensor] = None,
summary_position_ids: Optional[torch.Tensor] = None,
position_bias: Optional[torch.Tensor] = None,
):
"""Write summary token KV during decode (chunk boundary).
Large window: skip.
Small window (order matters — flush mirror before evict to avoid clobbering):
1. Flush mirror region → ring
2. Evict oldest new_summary → old_summary in key_cache (if full)
3. Write new summary → new_summary_buf
"""
n_summary = key_states.shape[2]
if self.is_large_window[layer_idx]:
self.cur_chunk_sizes[layer_idx] += n_summary
self._total_chunks[layer_idx] += n_summary
return
C = self.summary_chunk_size
ws = self.window_sizes[layer_idx]
scn = self.sliding_chunk_nums[layer_idx]
hierarchical = self.meta_pool_enabled[layer_idx]
group_size = self.meta_pool_group_sizes[layer_idx]
cbl = self._chunk_buf_len[layer_idx]
ptr = self._window_write_ptr[layer_idx]
n_old = self._old_summary_len[layer_idx]
# Step 1: Flush mirror region → ring (must happen before evict touches mirror[0])
mirror_start = ws + n_old
if ptr + cbl <= ws:
self.key_cache[layer_idx][:, :, ptr:ptr + cbl, :].copy_(
self.key_cache[layer_idx][:, :, mirror_start:mirror_start + cbl, :])
self.value_cache[layer_idx][:, :, ptr:ptr + cbl, :].copy_(
self.value_cache[layer_idx][:, :, mirror_start:mirror_start + cbl, :])
else:
first = ws - ptr
self.key_cache[layer_idx][:, :, ptr:ws, :].copy_(
self.key_cache[layer_idx][:, :, mirror_start:mirror_start + first, :])
self.value_cache[layer_idx][:, :, ptr:ws, :].copy_(
self.value_cache[layer_idx][:, :, mirror_start:mirror_start + first, :])
rest = cbl - first
self.key_cache[layer_idx][:, :, :rest, :].copy_(
self.key_cache[layer_idx][:, :, mirror_start + first:mirror_start + cbl, :])
self.value_cache[layer_idx][:, :, :rest, :].copy_(
self.value_cache[layer_idx][:, :, mirror_start + first:mirror_start + cbl, :])
self._window_write_ptr[layer_idx] = (ptr + cbl) % ws
if self._n_valid_window[layer_idx] < ws:
self._n_valid_window[layer_idx] = min(ws, self._n_valid_window[layer_idx] + cbl)
self._chunk_buf_len[layer_idx] = 0
# Step 2: Evict oldest new_summary → old_summary (now safe — mirror already flushed)
if self._new_sum_len[layer_idx] >= scn:
read_ptr = self._new_sum_write_ptr[layer_idx]
old_dst = ws + n_old # == mirror_start, but mirror data is already in ring
# Check capacity for old_summary growth
needed = old_dst + 1 + C # +1 for new old_sum, +C for future chunk mirror
if needed > self._capacity[layer_idx]:
new_s_cap = max(self._old_summary_cap[layer_idx] * 2, n_old + self._SUMMARY_INIT_CAP)
new_total = ws + new_s_cap + C
old_k, old_v = self.key_cache[layer_idx], self.value_cache[layer_idx]
bsz, heads, _, head_dim = old_k.shape
nk = torch.empty(bsz, heads, new_total, head_dim, dtype=old_k.dtype, device=old_k.device)
nv = torch.empty(bsz, heads, new_total, head_dim, dtype=old_v.dtype, device=old_v.device)
copy_len = ws + n_old
nk[:, :, :copy_len, :].copy_(old_k[:, :, :copy_len, :])
nv[:, :, :copy_len, :].copy_(old_v[:, :, :copy_len, :])
self.key_cache[layer_idx] = nk
self.value_cache[layer_idx] = nv
self._old_summary_cap[layer_idx] = new_s_cap
self._capacity[layer_idx] = new_total
self.key_cache[layer_idx][:, :, old_dst:old_dst+1, :].copy_(
self._new_sum_key_buf[layer_idx][:, :, read_ptr:read_ptr+1, :])
self.value_cache[layer_idx][:, :, old_dst:old_dst+1, :].copy_(
self._new_sum_value_buf[layer_idx][:, :, read_ptr:read_ptr+1, :])
n_old += 1
self._old_summary_total[layer_idx] += 1
if hierarchical:
self._residual_pre_key_buf[layer_idx][:, :, n_old-1:n_old, :].copy_(
self._new_sum_pre_key_buf[layer_idx][:, :, read_ptr:read_ptr+1, :]
)
self._residual_score_buf[layer_idx][:, n_old-1:n_old, :].copy_(
self._new_sum_score_buf[layer_idx][:, read_ptr:read_ptr+1, :]
)
self._residual_cos_buf[layer_idx][:, n_old-1:n_old, :].copy_(
self._new_sum_cos_buf[layer_idx][:, read_ptr:read_ptr+1, :]
)
self._residual_sin_buf[layer_idx][:, n_old-1:n_old, :].copy_(
self._new_sum_sin_buf[layer_idx][:, read_ptr:read_ptr+1, :]
)
self._residual_position_buf[layer_idx][:, n_old-1:n_old].copy_(
self._new_sum_position_buf[layer_idx][
:, read_ptr:read_ptr+1
]
)
if n_old == group_size:
scores = self._residual_score_buf[layer_idx][
:, :group_size, :
]
if self.meta_pool_mode == "dimension_wise":
if position_bias is None:
weights = torch.full_like(scores, 1.0 / group_size)
else:
weights = torch.softmax(
(
scores
+ position_bias.view(
1, group_size, scores.shape[-1]
)
).float(),
dim=1,
).to(key_states.dtype)
bsz, heads, _, head_dim = self._residual_pre_key_buf[
layer_idx
].shape
weights = weights.view(
bsz, group_size, heads, head_dim
).permute(0, 2, 1, 3)
else:
scalar_scores = scores.squeeze(-1)
if position_bias is None:
scalar_weights = torch.full_like(
scalar_scores, 1.0 / group_size
)
else:
scalar_weights = torch.softmax(
(
scalar_scores
+ position_bias.view(1, group_size)
).float(),
dim=1,
).to(key_states.dtype)
weights = scalar_weights[:, None, :, None]
meta_pre_k = (
self._residual_pre_key_buf[layer_idx][:, :, :group_size, :]
* weights
).sum(dim=2)
meta_v = (
self.value_cache[layer_idx][:, :, ws:ws + group_size, :]
* weights
).sum(dim=2)
last_cos = self._residual_cos_buf[layer_idx][:, group_size - 1, :]
last_sin = self._residual_sin_buf[layer_idx][:, group_size - 1, :]
meta_k = (
meta_pre_k * last_cos[:, None, :]
+ rotate_half(meta_pre_k) * last_sin[:, None, :]
)
meta_dst = self._meta_len[layer_idx]
if meta_dst >= self._meta_cap[layer_idx]:
old_meta_k = self._meta_key_cache[layer_idx]
old_meta_v = self._meta_value_cache[layer_idx]
new_cap = max(meta_dst + self._SUMMARY_INIT_CAP, self._meta_cap[layer_idx] * 2)
bsz, heads, _, head_dim = old_meta_k.shape
new_meta_k = torch.empty(
bsz, heads, new_cap, head_dim,
dtype=old_meta_k.dtype, device=old_meta_k.device,
)
new_meta_v = torch.empty_like(new_meta_k)
new_meta_k[:, :, :meta_dst, :].copy_(old_meta_k[:, :, :meta_dst, :])
new_meta_v[:, :, :meta_dst, :].copy_(old_meta_v[:, :, :meta_dst, :])
self._meta_key_cache[layer_idx] = new_meta_k
self._meta_value_cache[layer_idx] = new_meta_v
self._meta_cap[layer_idx] = new_cap
self._meta_key_cache[layer_idx][:, :, meta_dst:meta_dst+1, :].copy_(
meta_k.unsqueeze(2)
)
self._meta_value_cache[layer_idx][:, :, meta_dst:meta_dst+1, :].copy_(
meta_v.unsqueeze(2)
)
self._meta_len[layer_idx] += 1
n_old = 0
self._old_summary_len[layer_idx] = n_old
# Step 3: Write new summary to new_summary_buf (overwrite oldest slot)
w_ptr = self._new_sum_write_ptr[layer_idx]
self._new_sum_key_buf[layer_idx][:, :, w_ptr:w_ptr+1, :].copy_(key_states)
self._new_sum_value_buf[layer_idx][:, :, w_ptr:w_ptr+1, :].copy_(value_states)
if hierarchical:
if any(
item is None
for item in (
pre_rope_key_states,
summary_scores,
summary_cos,
summary_sin,
summary_position_ids,
)
):
raise ValueError("hierarchical decode requires pre-RoPE summary state")
self._new_sum_pre_key_buf[layer_idx][:, :, w_ptr:w_ptr+1, :].copy_(
pre_rope_key_states
)
self._new_sum_score_buf[layer_idx][:, w_ptr:w_ptr+1, :].copy_(
summary_scores
)
self._new_sum_cos_buf[layer_idx][:, w_ptr:w_ptr+1, :].copy_(summary_cos)
self._new_sum_sin_buf[layer_idx][:, w_ptr:w_ptr+1, :].copy_(summary_sin)
self._new_sum_position_buf[layer_idx][
:, w_ptr:w_ptr+1
].copy_(summary_position_ids)
self._new_sum_write_ptr[layer_idx] = (w_ptr + 1) % scn
if self._new_sum_len[layer_idx] < scn:
self._new_sum_len[layer_idx] += 1
self.cur_chunk_sizes[layer_idx] += n_summary
self._total_chunks[layer_idx] += n_summary
# ── Decode: get KV for attention ──
def get_attention_kv(self, layer_idx: int) -> tuple[torch.Tensor, torch.Tensor, int]:
"""Get full KV for text token attention.
Large window: buffer[:text_len]
Small window (steady state): key_cache[:ws + n_old + cbl] — single slice, zero cat
"""
if self.is_large_window[layer_idx]:
tl = self._text_len[layer_idx]
return (self.key_cache[layer_idx][:, :, :tl, :],
self.value_cache[layer_idx][:, :, :tl, :], 0)
ws = self.window_sizes[layer_idx]
nv = self._n_valid_window[layer_idx]
cbl = self._chunk_buf_len[layer_idx]
n_old = self._old_summary_len[layer_idx]
# Steady state: ring full → [ring(ws) | old_sums(n_old) | chunk_mirror(cbl)] contiguous
if nv == ws:
end = ws + n_old + cbl
base_k = self.key_cache[layer_idx][:, :, :end, :]
base_v = self.value_cache[layer_idx][:, :, :end, :]
else:
# Warmup: ring not full, [nv:ws] is gap → cat
parts_k, parts_v = [], []
if nv > 0:
parts_k.append(self.key_cache[layer_idx][:, :, :nv, :])
parts_v.append(self.value_cache[layer_idx][:, :, :nv, :])
if cbl > 0:
mirror_start = ws + n_old
parts_k.append(self.key_cache[layer_idx][:, :, mirror_start:mirror_start + cbl, :])
parts_v.append(self.value_cache[layer_idx][:, :, mirror_start:mirror_start + cbl, :])
if n_old > 0:
parts_k.append(self.key_cache[layer_idx][:, :, ws:ws + n_old, :])
parts_v.append(self.value_cache[layer_idx][:, :, ws:ws + n_old, :])
base_k = parts_k[0] if len(parts_k) == 1 else torch.cat(parts_k, dim=2)
base_v = parts_v[0] if len(parts_v) == 1 else torch.cat(parts_v, dim=2)
meta_len = self._meta_len[layer_idx]
if meta_len:
base_k = torch.cat(
(base_k, self._meta_key_cache[layer_idx][:, :, :meta_len, :]), dim=2
)
base_v = torch.cat(
(base_v, self._meta_value_cache[layer_idx][:, :, :meta_len, :]), dim=2
)
return base_k, base_v, meta_len
def get_current_chunk_kv(self, layer_idx: int) -> tuple[torch.Tensor, torch.Tensor]:
"""Get KV of the current chunk's C text tokens for summary token attention."""
C = self.summary_chunk_size
if self.is_large_window[layer_idx]:
tl = self._text_len[layer_idx]
return (self.key_cache[layer_idx][:, :, tl - C:tl, :],
self.value_cache[layer_idx][:, :, tl - C:tl, :])
else:
ws = self.window_sizes[layer_idx]
n_old = self._old_summary_len[layer_idx]
cbl = self._chunk_buf_len[layer_idx]
mirror_start = ws + n_old
return (self.key_cache[layer_idx][:, :, mirror_start:mirror_start + cbl, :],
self.value_cache[layer_idx][:, :, mirror_start:mirror_start + cbl, :])
def reset_chunk_counter(self):
"""Reset chunk counters after a chunk boundary step completes."""
block = self.summary_chunk_size + self.summary_token_num
for layer_idx in range(self.num_hidden_layers):
if self.cur_chunk_sizes[layer_idx] >= block:
self.cur_chunk_sizes[layer_idx] %= block
|