File size: 61,309 Bytes
b025706 | 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 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 | # SPDX-FileCopyrightText: © 2025 Tenstorrent USA, Inc.
# SPDX-License-Identifier: Apache-2.0
import inspect
import sys
import torch
from loguru import logger
import ttnn
from models.common.lightweightmodule import LightweightModule
from models.common.sampling._utils import (
is_default_value,
is_power_of_2,
topk_would_route_to_large_indices,
upper_power_of_2,
)
from models.common.sampling.tt_log_probs import LogProbsCalculator
from models.common.sampling.vocab_padding import (
build_invalid_vocab_mask,
build_tail_invalid_vocab_mask,
get_vocab_shard_dims,
)
# Greedy tie-break boost (see TTSampling._adjust_values_for_tiebreak).
# bfloat16 keeps an 8-bit mantissa, so the gap to the next representable value at magnitude |x| is
# between |x| * 2^-8 and |x| * 2^-7. Scaling the tied maximum by 2^-6 is therefore at least 2 ULP at
# every magnitude, whereas a fixed constant is not: 1.0 is below one ULP once |logit| >= 256 (bf16
# spacing there is 2.0), so 256 + 1 rounds straight back to 256 and the tie survives. Multiplying by
# a power of two is exact in bf16, so the boost is reproducible.
TIEBREAK_DELTA_SCALE = 2**-6
# Floor so the boost stays strictly positive when the tied maximum is exactly 0.0. Still a normal
# bf16 number (smallest normal ~1.18e-38) and orders of magnitude below one ULP of any real logit,
# so it never perturbs a non-zero maximum.
TIEBREAK_DELTA_FLOOR = 1e-30
# Added to the global index of every non-maximum candidate so the row min over the masked indices can
# only ever land on a tied maximum. A power of two, so the mask that carries it (built in bfloat16)
# holds it exactly; larger than any padded vocabulary size, which __init__ asserts; and small enough
# that sentinel + index cannot overflow the int32 the min reduce runs in.
TIEBREAK_INDEX_SENTINEL = 2**24
# Widest input ttnn.topk accepts in one call; vocabs beyond it must be cut into chunks.
TOPK_MAX_WIDTH = 64 * 1024
class TTSampling(LightweightModule):
"""
On-device sampling module supporting top-k, top-p, and temperature-based sampling.
This class implements high-performance on-device sampling that can work across different
model implementations by accepting configuration parameters rather than assuming specific
args structures.
Multi-device sampling works by partitioning the vocabulary across devices. Each device
computes top-k locally on its vocabulary partition, then all-gather operations combine
the results across devices to perform global top-k selection before final sampling.
Args:
mesh_device: The device or MeshDevice for computations
tt_ccl: CCL object for distributed operations (supports both line_all_gather and tt_all_gather)
vocab_size: Vocabulary size of the model
padded_vocab_size: Padded vocabulary size (must be divisible by num devices)
max_batch_size: Maximum batch size supported
max_top_k: Maximum number of top-k tokens to consider
cluster_shape: Shape of the device cluster (rows, cols)
sampling_all_gather_axis: Axis to all-gather over in 2D meshes (0=rows, 1=cols, default: 0)
sub_core_grids: Sub-core grid configuration for operations
sub_core_grid_topk: Sub-core grid configuration specifically for top-k operations
start_core: Starting core coordinate for sampling operations
num_gather_links: Number of links to use for all-gather operations (optional)
sampling_memory_config: Memory configuration for sampling tensors (optional)
k, p, temp: Initial sampling parameters (tensors of size max_batch_size)
Note:
Uses persistent buffers when CCL supports line_all_gather (llama3_70b_galaxy),
otherwise uses standard all_gather where the CCL API handles memory allocation (tt-transformers).
"""
@classmethod
def num_single_device_vocab_splits(cls, padded_vocab_size):
"""Fewest power-of-two same-device chunks whose width fits ttnn.topk.
Returns None when no tile-aligned cut exists, in which case the caller
must fall back to host sampling instead of constructing TTSampling.
"""
num_splits = 2
while padded_vocab_size // num_splits > TOPK_MAX_WIDTH:
num_splits *= 2
chunk_width = padded_vocab_size // num_splits
if padded_vocab_size % num_splits != 0 or chunk_width % ttnn.TILE_SIZE != 0:
return None
return num_splits
@staticmethod
def _untilize_chunk_count(width):
"""Fewest tile-aligned chunks of at most TOPK_MAX_WIDTH each, or 1 when
the row is narrow enough (<= 2 * TOPK_MAX_WIDTH, the widest row known
to untilize in one program).
Prefer an even cut so every chunk is at least half of TOPK_MAX_WIDTH
wide (the search is bounded to twice the minimum count).
"""
if width <= 2 * TOPK_MAX_WIDTH:
return 1
min_chunks = (width + TOPK_MAX_WIDTH - 1) // TOPK_MAX_WIDTH # ceil(width / TOPK_MAX_WIDTH)
num_chunks = min_chunks
max_chunks = 2 * num_chunks
while num_chunks <= max_chunks:
if width % num_chunks == 0 and (width // num_chunks) % ttnn.TILE_SIZE == 0:
return num_chunks
num_chunks += 1
# INFO: split accepts an uneven trailing chunk.
return min_chunks
@staticmethod
def _untilize_chunk_width(width, num_chunks):
"""Tile-aligned ttnn.split size that cuts `width` into `num_chunks` pieces
(the last one shorter when the row does not divide evenly)."""
per_chunk = (width + num_chunks - 1) // num_chunks # ceil(width / num_chunks)
return (per_chunk + ttnn.TILE_SIZE - 1) // ttnn.TILE_SIZE * ttnn.TILE_SIZE # round up to a tile
def _is_force_argmax_sampling(self, k, p, temp):
"""Detect whether all users request deterministic greedy decoding.
When every user in the batch has k=1 (top-1), p=0.0 or p=1.0 (no top-p filter),
and temp=1.0 (no temperature scaling), we can skip the full top-k / top-p /
temperature / RNG pipeline and use a single all-gather + argmax instead.
This is significantly faster because argmax needs only one all-gather of the
full logits tensor vs. three gathers (values, indices, sampled tokens) in the
normal path.
Note: callers may represent greedy rows with p=1.0, while the
device argmax-style representation uses p=0.0.
The model config must also set allow_force_argmax=True for this to activate.
Changing this state between decode steps invalidates captured traces, so
SamplingGenerator maintains separate trace slots keyed by force_argmax.
"""
calculator = getattr(self, "log_probs_calculator", None)
needs_topk_logprobs = calculator is not None and calculator.enable_log_probs and calculator._use_topk_logprobs
return (
self._allow_force_argmax_sampling
and not needs_topk_logprobs
and is_default_value(k, 1)
and (is_default_value(p, 1.0) or is_default_value(p, 0.0))
and is_default_value(temp, 1.0)
)
@property
def force_argmax_sampling(self) -> bool:
return self._force_argmax_sampling
def _normalize_device_params(self, k, temp):
"""Map k and temp onto the contract ttnn.sampling actually implements.
Mirrors the rewrites format_sampling_params() applies, so the device is guarded even when a
caller reaches reset_params() directly:
* ``k`` outside [1, max_top_k]. ttnn.sampling walks k entries of each user's two-face
candidate row. A k above max_top_k (callers pass vocab_size to mean "no top-k filter")
runs the top-p scan and the draw off the end of that row into unrelated L1, and k < 1
-- the documented "no restriction" encoding -- is worse: from_torch(dtype=uint32) turns
a negative k into ~4.3e9. Both collapse to max_top_k.
* ``temp == 0``. temp multiplies the logits before the softmax, so the greedy encoding
flattens every candidate to an equal probability and turns the draw uniform. Greedy is
the triple (temp=1, k=1), so rewrite k too -- temp alone would leave a caller passing
temp=0 with k=50 sampling from the top 32 instead of taking the argmax.
``k`` and ``temp`` are required; None was only ever tolerated by the force_argmax path.
"""
if k is None or temp is None:
raise ValueError("k and temp are required; pass format_sampling_params() output.")
k = torch.as_tensor(k)
temp = torch.as_tensor(temp, dtype=torch.float32)
k = torch.where(k < 1, torch.full_like(k, self.max_top_k), k).clamp(max=self.max_top_k)
k = torch.where(temp == 0.0, torch.ones_like(k), k)
temp = torch.where(temp == 0.0, torch.ones_like(temp), temp)
return k, temp
def __init__(
self,
mesh_device,
tt_ccl,
args,
k=None,
p=None,
temp=None,
):
super().__init__()
self.mesh_device = mesh_device
# Use the fast unstable top-k; _adjust_values_for_tiebreak guarantees the correctness
# host-side. stable=True (lowest-index tie-break) is only cheap where the
# multi-core factory can pack [bf16 value | u16 index] fused keys: otherwise a multi-step
# path costs ~2.5-3x more, and additionally the BH topk_large_indices route currently
# has no stable mode at all.
self._topk_stable = False
# Multi-step reduction is supported only on single device
self.multi_step_reduction = list(mesh_device.shape) == [1, 1]
self.tt_ccl = tt_ccl
self._line_all_gather = getattr(self.tt_ccl, "line_all_gather", None)
self._line_all_gather_supports_buffer_key = False
self.pad_to_power_of_2 = getattr(args, "pad_logits_to_power_of_2", False)
if callable(self._line_all_gather):
try:
line_all_gather_sig = inspect.signature(self._line_all_gather)
line_all_gather_params = line_all_gather_sig.parameters
self._line_all_gather_supports_buffer_key = "buffer_key" in line_all_gather_params or any(
param.kind == inspect.Parameter.VAR_KEYWORD for param in line_all_gather_params.values()
)
except (TypeError, ValueError):
logger.warning("Unable to inspect line_all_gather signature; assuming no buffer_key support.")
padded_vocab_size = getattr(args, "padded_vocab_size", None)
self.padded_vocab_size = padded_vocab_size if padded_vocab_size is not None else args.vocab_size
self.vocab_size = args.vocab_size
# Single-device top-k runs the vocab in same-device chunks (multi-step reduction).
# ttnn.topk handles at most TOPK_MAX_WIDTH elements per call, so use the fewest
# power-of-two chunks whose width fits: two chunks preserve the historical behavior
# for every vocab up to 128K, and larger vocabs get four -- e.g. Qwen3's 151936
# (#53064) and Gemma-2's 256000. Callers gate on num_single_device_vocab_splits()
# returning non-None (host-sampling fallback), so reaching None here is a caller
# bug. Raise rather than assert: this guards a correctness invariant (misaligned
# chunks silently corrupt global token indices) and must survive python -O.
self._num_vocab_splits = 2
if self.multi_step_reduction:
self._num_vocab_splits = self.num_single_device_vocab_splits(self.padded_vocab_size)
if self._num_vocab_splits is None:
raise ValueError(
f"padded_vocab_size={self.padded_vocab_size} cannot be cut into "
f"tile-aligned single-device top-k chunks of at most {TOPK_MAX_WIDTH}"
)
# Round up to the next tile boundary (32) — device tensors must be tile-aligned.
raw_batch = getattr(args, "max_batch_size", 32)
self.max_batch_size = max(32, ((raw_batch + 31) // 32) * 32)
self.max_top_k = getattr(args, "max_top_k", 32)
self.cluster_shape = args.cluster_shape
self.sampling_all_gather_axis = getattr(args, "sampling_all_gather_axis", 0)
self.sub_core_grids = getattr(args, "sub_core_grids", None)
self.sub_core_grid_topk = getattr(args, "sub_core_grid_topk", None)
self.start_core = getattr(args, "start_core", ttnn.CoreCoord(0, 0))
self._sampling_sub_core_grids = (
ttnn.num_cores_to_corerangeset_in_subcoregrids(
self.start_core, self.max_batch_size, self.sub_core_grids, row_wise=True
)
if self.sub_core_grids is not None
else None
)
# Blackhole galaxy prefetcher (unfused-CCL) keeps a split senders/worker sub-device manager
# loaded during prefill sampling (prefill samples in decode mode). Auto-multicore ops grab the
# full 12-wide compute grid, which includes the dispatch column that no loaded sub-device covers
# ("kernel group cores do not match sub device cores"). Pin the force-argmax untilize/argmax to
# the worker sub-core grids so they stay inside the worker sub-device. Left None elsewhere so no
# other arch's behaviour changes.
self._force_argmax_sub_core_grids = self.sub_core_grids if getattr(args, "use_unfused_ccl", False) else None
# sampling_dp > 1 when multiple mesh groups each sample users independently
# (e.g. GPT-OSS on [4,8]: 4 rows × 32 users; Llama Galaxy on [8,4]: 4 cols × 8 users)
self._sampling_dp = getattr(args, "sampling_dp", 1)
# Shard params along the non-all-gather axis; replicate along the all-gather axis
if self._sampling_dp > 1:
if self.sampling_all_gather_axis == 0:
self._param_dims = (None, 0) # shard along cols
else:
self._param_dims = (0, None) # shard along rows
else:
self._param_dims = (None, None)
if hasattr(args, "model_config") and "GALAXY_NUM_LINKS" in args.model_config:
# Calculate num_gather_links based on model config
max_num_gather_links = args.model_config["GALAXY_NUM_LINKS"]
self.num_gather_links = (
args.max_top_k // 32 if args.max_top_k // 32 <= max_num_gather_links else max_num_gather_links
)
else:
self.num_gather_links = 1
if hasattr(args, "model_config") and "DECODE_SAMPLING_INPUT_MEMCFG" in args.model_config:
self.sampling_memory_config = args.model_config["DECODE_SAMPLING_INPUT_MEMCFG"]
else:
self.sampling_memory_config = ttnn.DRAM_MEMORY_CONFIG
# Force argmax sampling
if hasattr(args, "model_config") and "SAMPLING_AG_CONFIG" in args.model_config:
# The model config may describe the fastest full-size Galaxy path, but
# the actual CCL shape is resolved from the runtime mesh below.
sampling_ag_config = args.model_config["SAMPLING_AG_CONFIG"]
self._allow_force_argmax_sampling = sampling_ag_config["allow_force_argmax"]
self.num_argmax_gather_links = sampling_ag_config["num_links"]
self.argmax_chunks_per_sync = sampling_ag_config.get("chunks_per_sync", 10)
self.argmax_num_workers_per_link = 1
self.ag_topology = sampling_ag_config["topology"]
else:
self._allow_force_argmax_sampling = False
self.num_argmax_gather_links = self.num_gather_links
self.argmax_chunks_per_sync = 10
self.argmax_num_workers_per_link = 1
self.ag_topology = ttnn.Topology.Linear
# Set defaults for sampling parameters if not provided
# Default: k=1 (top-1), p=0 (effectively argmax), temp=1 (no temperature scaling)
# When p=0, the sampling operation will select the token with highest probability (argmax)
total_param_size = self.max_batch_size * self._sampling_dp
if k is None:
k = torch.ones(total_param_size)
if p is None:
p = torch.zeros(total_param_size)
if temp is None:
temp = torch.ones(total_param_size)
k, temp = self._normalize_device_params(k, temp)
self._force_argmax_sampling = self._is_force_argmax_sampling(k, p, temp)
# Create sampling parameter tensors on device
# When _sampling_dp > 1, dims=(0, None) shards the [128] tensor across 4 rows → [32] per row
self.k_tensor = ttnn.from_torch(
k,
device=self.mesh_device,
dtype=ttnn.uint32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=ttnn.ShardTensor2dMesh(self.mesh_device, dims=self._param_dims, mesh_shape=self.cluster_shape),
)
self.p_tensor = ttnn.from_torch(
p,
device=self.mesh_device,
dtype=ttnn.bfloat16,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=ttnn.ShardTensor2dMesh(self.mesh_device, dims=self._param_dims, mesh_shape=self.cluster_shape),
)
self.temp_tensor = ttnn.from_torch(
temp,
device=self.mesh_device,
dtype=ttnn.bfloat16,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=ttnn.ShardTensor2dMesh(self.mesh_device, dims=self._param_dims, mesh_shape=self.cluster_shape),
)
# The tie-break sentinel has to outrank every real global index. Vocabularies this large are
# far beyond anything shipped, but exceeding it would corrupt the greedy token silently rather
# than fail, so check it at construction. Raise rather than assert: this guards a correctness
# invariant and must survive python -O.
if self.padded_vocab_size > TIEBREAK_INDEX_SENTINEL:
raise ValueError(
f"padded_vocab_size {self.padded_vocab_size} exceeds the greedy tie-break sentinel "
f"{TIEBREAK_INDEX_SENTINEL}; raise TIEBREAK_INDEX_SENTINEL (keeping it a power of two)"
)
# Persistent per-user ARGMAX mask [1,1,N,1] (1.0 where k==1), distributed like k_tensor. Used by
# _adjust_values_for_tiebreak to boost the lowest-index tied-max for greedy users only. Built
# host-side and kept in sync in reset_sampling_params (an on-device reshape of the [N] k_tensor
# to [1,1,N,1] is not sub-device-safe). bfloat16 TILE so it broadcasts over the candidate width
# and multiplies the (bfloat16) winner mask without a dtype mix; 0.0/1.0 are exact in bf16.
self._greedy_col = ttnn.from_torch(
(torch.as_tensor(k).reshape(1, 1, -1, 1) == 1).to(torch.bfloat16),
device=self.mesh_device,
dtype=ttnn.bfloat16,
layout=ttnn.TILE_LAYOUT,
mesh_mapper=ttnn.ShardTensor2dMesh(
self.mesh_device, dims=self._greedy_col_dims(), mesh_shape=self.cluster_shape
),
)
# Create device offset indices for global indexing
self._create_indices_tensors()
self._create_invalid_vocab_mask()
# Log-probs tensor to store the log-probs for the batch
self.tt_log_probs = None
self.log_probs_calculator = LogProbsCalculator(
self.mesh_device,
self.sub_core_grids,
self.tt_ccl,
batch_size=self.max_batch_size,
use_topk_logprobs=getattr(args, "use_topk_logprobs", False),
)
# Seeds tensor: one RNG slot per user across all rows.
# When sampling_dp > 1, shard across rows so each row gets its own slice.
# user_ids tensor: core routing only (32 per row, replicated).
self.seeds_tt_tensor = ttnn.from_torch(
torch.arange(total_param_size).to(torch.uint32),
device=self.mesh_device,
dtype=ttnn.uint32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=(
ttnn.ShardTensor2dMesh(self.mesh_device, dims=self._param_dims, mesh_shape=self.cluster_shape)
if self._sampling_dp > 1
else None
),
memory_config=ttnn.DRAM_MEMORY_CONFIG,
)
self.user_ids_tt_tensor = ttnn.as_tensor(
torch.arange(self.max_batch_size).to(torch.uint32),
dtype=ttnn.uint32,
layout=ttnn.ROW_MAJOR_LAYOUT,
device=self.mesh_device,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
)
def _get_num_sampling_shards(self):
if self.multi_step_reduction:
return self._num_vocab_splits
if 1 in self.cluster_shape:
return max(self.cluster_shape[0], self.cluster_shape[1])
if self.sampling_all_gather_axis not in (0, 1):
raise ValueError(
f"sampling_all_gather_axis must be 0 or 1 for 2D meshes, got {self.sampling_all_gather_axis}"
)
return self.cluster_shape[self.sampling_all_gather_axis]
def _create_indices_tensors(self):
"""Create the per-shard index offsets added to the top-k indices after the gather."""
num_devices_in_mesh = self._get_num_sampling_shards()
indices_device_offsets = torch.ones(
1, 1, self.max_batch_size, self.max_top_k * num_devices_in_mesh, dtype=torch.int64
)
padded_per_device = self.padded_vocab_size // num_devices_in_mesh
for device_id in range(num_devices_in_mesh):
indices_device_offsets[:, :, :, device_id * self.max_top_k : (device_id + 1) * self.max_top_k] = (
device_id * padded_per_device
)
self.tt_indices_device_offsets = ttnn.from_torch(
indices_device_offsets,
device=self.mesh_device,
dtype=ttnn.int32,
layout=ttnn.TILE_LAYOUT,
mesh_mapper=ttnn.ShardTensor2dMesh(self.mesh_device, dims=(None, None), mesh_shape=self.cluster_shape),
memory_config=ttnn.DRAM_MEMORY_CONFIG,
)
def _create_invalid_vocab_mask(self):
self.tt_invalid_vocab_mask = None
self.tt_invalid_vocab_tail_mask = None
self._invalid_vocab_tail_width = 0
vocab_shard_dims = get_vocab_shard_dims(self.cluster_shape, self.sampling_all_gather_axis)
# The compact tail-mask path masks only the padded tail, but it has to slice the
# logits and concat them back. ttnn.concat only honours sub_core_grids when the
# input is unsharded and the output is interleaved; otherwise it falls through to
# the "massaged" untilize/transpose path, which is invoked without sub_core_grids
# and so runs on the full Tensix grid. The sampling logits are width-sharded
# (DECODE_LOGITS_MEMCFG), so on a sampling sub-core grid the concat escapes the
# sub-device. Use the plain full-width additive mask (one elementwise add, no
# reassembly) whenever a sub-core grid is in use.
tail_mask = (
build_tail_invalid_vocab_mask(
self.vocab_size,
self.padded_vocab_size,
self.max_batch_size,
self.cluster_shape,
self.sampling_all_gather_axis,
tile_size=ttnn.TILE_SIZE,
)
if self.sub_core_grids is None
else None
)
if tail_mask is not None:
self._invalid_vocab_tail_width = tail_mask.tail_width
self.tt_invalid_vocab_tail_mask = ttnn.from_torch(
tail_mask.mask,
dtype=ttnn.bfloat16,
layout=ttnn.TILE_LAYOUT,
device=self.mesh_device,
mesh_mapper=ttnn.ShardTensor2dMesh(
self.mesh_device,
dims=vocab_shard_dims,
mesh_shape=self.cluster_shape,
),
memory_config=ttnn.DRAM_MEMORY_CONFIG,
)
return
invalid_vocab_mask = build_invalid_vocab_mask(
self.vocab_size,
self.padded_vocab_size,
self.max_batch_size,
)
if invalid_vocab_mask is None:
return
self.tt_invalid_vocab_mask = ttnn.from_torch(
invalid_vocab_mask,
dtype=ttnn.bfloat16,
layout=ttnn.TILE_LAYOUT,
device=self.mesh_device,
mesh_mapper=ttnn.ShardTensor2dMesh(
self.mesh_device,
dims=vocab_shard_dims,
mesh_shape=self.cluster_shape,
),
memory_config=ttnn.DRAM_MEMORY_CONFIG,
)
def _mask_invalid_vocab_logits(self, logits):
if self.tt_invalid_vocab_tail_mask is not None:
return self._mask_invalid_vocab_tail_logits(logits)
if self.tt_invalid_vocab_mask is None:
return logits
return ttnn.add(
logits,
self.tt_invalid_vocab_mask,
memory_config=logits.memory_config(),
sub_core_grids=self.sub_core_grids,
)
def _mask_invalid_vocab_tail_logits(self, logits):
tail_width = self._invalid_vocab_tail_width
local_width = logits.shape[-1]
valid_width = local_width - tail_width
if tail_width <= 0 or valid_width < 0:
return self._mask_invalid_vocab_logits_fallback(logits)
if valid_width == 0:
return ttnn.add(
logits,
self.tt_invalid_vocab_tail_mask,
memory_config=logits.memory_config(),
sub_core_grids=self.sub_core_grids,
)
valid_logits = ttnn.slice(
logits,
[0, 0, 0, 0],
[logits.shape[0], logits.shape[1], logits.shape[2], valid_width],
memory_config=logits.memory_config(),
sub_core_grids=self.sub_core_grids,
)
tail_logits = ttnn.slice(
logits,
[0, 0, 0, valid_width],
[logits.shape[0], logits.shape[1], logits.shape[2], local_width],
memory_config=logits.memory_config(),
sub_core_grids=self.sub_core_grids,
)
masked_tail_logits = ttnn.add(
tail_logits,
self.tt_invalid_vocab_tail_mask,
memory_config=logits.memory_config(),
sub_core_grids=self.sub_core_grids,
)
masked_logits = ttnn.concat(
[valid_logits, masked_tail_logits],
dim=3,
memory_config=logits.memory_config(),
# Match the slices above: run concat on the sampling sub-device cores,
# otherwise the concat program is placed on the full Tensix grid and
# fails with "Kernel group cores do not match sub device cores"
# (TT_FATAL num_intersections == num_cores).
sub_core_grids=self.sub_core_grids,
)
ttnn.deallocate(valid_logits)
ttnn.deallocate(tail_logits)
ttnn.deallocate(masked_tail_logits)
return masked_logits
def _mask_invalid_vocab_logits_fallback(self, logits):
if self.tt_invalid_vocab_mask is None:
return logits
return ttnn.add(
logits,
self.tt_invalid_vocab_mask,
memory_config=logits.memory_config(),
sub_core_grids=self.sub_core_grids,
)
def _can_slice_valid_vocab_for_argmax(self):
return self.vocab_size < self.padded_vocab_size and self.vocab_size % ttnn.TILE_SIZE == 0
def _slice_valid_vocab_for_argmax(self, logits):
if not self._can_slice_valid_vocab_for_argmax() or logits.shape[-1] != self.padded_vocab_size:
return logits
return ttnn.slice(
logits,
[0, 0, 0, 0],
[logits.shape[0], logits.shape[1], logits.shape[2], self.vocab_size],
memory_config=logits.memory_config(),
sub_core_grids=self.sub_core_grids,
)
def _perform_all_gather(self, tensor, dim, cluster_axis, memory_config, num_links, buffer_key=None):
"""
Flexible all-gather that works across different CCL implementations.
- If `tt_ccl` exposes `line_all_gather`, prefer it (enables persistent buffer usage on some stacks).
- Otherwise fall back to `ttnn.all_gather`.
"""
if callable(self._line_all_gather):
# Some implementations accept `buffer_key` (for persistent buffers), others may not.
line_all_gather_kwargs = {
"dim": dim,
"cluster_axis": cluster_axis,
"memory_config": memory_config,
"num_links": num_links,
}
if self._line_all_gather_supports_buffer_key and buffer_key is not None:
line_all_gather_kwargs["buffer_key"] = buffer_key
return self._line_all_gather(tensor, **line_all_gather_kwargs)
return ttnn.all_gather(
tensor,
dim=dim,
memory_config=memory_config,
cluster_axis=cluster_axis,
)
def _get_sampling_cluster_axis(self):
if self.mesh_device.get_num_devices() <= 1:
return None
# 1D submeshes should use the default CCL axis; forcing axis 1 can make
# smaller Galaxy DP groups request routes outside the submesh.
if 1 in self.cluster_shape:
return None
return self.sampling_all_gather_axis
def _get_force_argmax_all_gather_config(self, cluster_axis):
num_links = self.num_argmax_gather_links
if hasattr(self.tt_ccl, "get_num_links"):
# Clamp the tuned config to the links available on the actual submesh.
num_links = min(num_links, self.tt_ccl.get_num_links(cluster_axis))
topology = self.ag_topology
# Ring is available for T3K-like 8-device groups; smaller DP groups need
# linear routing to avoid wraparound routes such as D0 -> D12.
if self.mesh_device.get_num_devices() < 8:
topology = ttnn.Topology.Linear
return max(1, num_links), topology
def reset_params(
self,
k,
p,
temp,
enable_log_probs: bool | list[bool] = None,
num_logprobs: int | list[int] = None,
empty_slots: list[int] | None = None,
):
"""Update sampling parameters (k, p, temperature, logprobs) dynamically."""
k, temp = self._normalize_device_params(k, temp)
# Top-k reporting needs the candidate pipeline, even for greedy tokens.
# Resolve reporting before choosing the path so its parameters update too.
self.log_probs_calculator.set_log_probs_mode(
enable_log_probs, num_logprobs=num_logprobs, empty_slots=empty_slots
)
self._force_argmax_sampling = self._is_force_argmax_sampling(k, p, temp)
if not self._force_argmax_sampling:
# When _sampling_dp > 1, create multi-device host tensors so
# copy_host_to_device_tensor writes per-row shards correctly.
if self._sampling_dp > 1:
mapper = ttnn.ShardTensor2dMesh(self.mesh_device, dims=self._param_dims, mesh_shape=self.cluster_shape)
else:
mapper = None
self.k_tensor_new = ttnn.from_torch(
k,
device=None,
dtype=ttnn.uint32,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
self.p_tensor_new = ttnn.from_torch(
torch.tensor(p),
device=None,
dtype=ttnn.bfloat16,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
self.temp_tensor_new = ttnn.from_torch(
temp,
device=None,
dtype=ttnn.bfloat16,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=mapper,
)
ttnn.copy_host_to_device_tensor(self.k_tensor_new, self.k_tensor)
ttnn.copy_host_to_device_tensor(self.p_tensor_new, self.p_tensor)
ttnn.copy_host_to_device_tensor(self.temp_tensor_new, self.temp_tensor)
# Keep the greedy tie-break mask (1.0 where k==1) in sync with k, distributed like k_tensor.
self._greedy_col_new = ttnn.from_torch(
(k.reshape(1, 1, -1, 1) == 1).to(torch.bfloat16),
device=None,
dtype=ttnn.bfloat16,
layout=ttnn.TILE_LAYOUT,
mesh_mapper=(
ttnn.ShardTensor2dMesh(
self.mesh_device, dims=self._greedy_col_dims(), mesh_shape=self.cluster_shape
)
if self._sampling_dp > 1
else None
),
)
ttnn.copy_host_to_device_tensor(self._greedy_col_new, self._greedy_col)
def _greedy_col_dims(self):
"""Map the 1-D k_tensor shard dims (self._param_dims, batch on dim0) to the [1,1,N,1] greedy
mask's dims (batch on dim2), so self._greedy_col is distributed exactly like self.k_tensor."""
return tuple(2 if d == 0 else d for d in self._param_dims)
def _adjust_values_for_tiebreak(self, gathered_values, gathered_global_indices):
"""Return gathered_values with, for ARGMAX users (k==1) ONLY, the single lowest-GLOBAL-INDEX
candidate among the tied maxima boosted just past the tie, so ttnn.sampling's argmax selects
it deterministically. This fixes ttnn.sampling's array-position tie-break (it breaks exact
value ties by all_gather/device order, which varies run-to-run/slot-to-slot and flips the
greedy token) by correcting the sampling INPUT in the TILE domain -- avoiding an in-place
write into the ROW_MAJOR output buffer, which NO ttnn op supports on a restricted
sub-device. Random users (k>1) get boost==0 => their values are bit-identical => their
sampling is byte-for-byte unchanged. All ops honor self.sub_core_grids.
Stable top-k (`stable=True`) on EVERY local top-k call would make this pass redundant, but
stable is not uniformly cheap: multicore-eligible bf16 calls take the fused-key engine
(packed [bf16 value | u16 index] words on the plain unstable network, where otherwise costs
~2.5-3x on the SFPU sort stage and the BH topk_large_indices route has no stable mode at
all. Sampling keeps the host-side pass as the stable-sort guarantee instead. Removing this
method, `_greedy_col` and `_greedy_col_dims` in favour of `stable=True` becomes a pure win
once the remaining routes implement a performant stable mode.
KNOWN LIMITATION: this picks the lowest global id among the GATHERED candidates. If a single
device shard holds more than `max_top_k` (32) maxima tied at the same value, its top-k
breaks the tie by array position and drops all but 32 of them, so the true lowest-id token
may never reach the gathered set and this pass cannot recover it. Enabling `stable=True`
on every route (at the costs above) would close that window.
is_winner = (value == rowmax) AND (global_index == lowest_index_among_maxima) # exactly one candidate
lowest_index_among_maxima = min(global_index + not_max * SENTINEL) # == idx at maxima, huge else
The value half runs in bfloat16 and the index half in int32, and neither dtype is incidental --
see the comments on each block. Validated on a restricted active sub-device by
tests/ttnn/unit_tests/operations/reduce/test_tiebreak_input_adjust.py.
"""
scg = self.sub_core_grids
# Every intermediate is deallocated as soon as its last reader is issued: this runs once per
# decode step, so leaving them to Python refcounting would hold ~9 extra buffers per step.
# Value domain, bfloat16: gathered_values is bf16, so the max reduce and the comparisons
# against its result are exact -- bf16 survives the FPU's source-register truncation intact.
maxv = ttnn.max(gathered_values, dim=3, keepdim=True, sub_core_grids=scg) # [1,1,B,1] bf16
is_max = ttnn.eq(gathered_values, maxv, sub_core_grids=scg) # 1.0 at the (tied) maxima, exact
not_max = ttnn.lt(gathered_values, maxv, sub_core_grids=scg) # 1.0 strictly below max
# Per-row boost, >= 2 bf16 ULP of that row's maximum whatever its magnitude or sign.
abs_max = ttnn.abs(maxv, sub_core_grids=scg)
ttnn.deallocate(maxv)
delta_scaled = ttnn.multiply(abs_max, TIEBREAK_DELTA_SCALE, sub_core_grids=scg)
ttnn.deallocate(abs_max)
delta = ttnn.add(delta_scaled, TIEBREAK_DELTA_FLOOR, sub_core_grids=scg) # [1,1,B,1] bf16
ttnn.deallocate(delta_scaled)
# Index domain, int32: lowest global index among the maxima. Push the non-maxima above every
# real index by TIEBREAK_INDEX_SENTINEL, then take the row min.
#
# This min MUST run in int32. ttnn.min/ttnn.max on a FLOAT32 tensor go to the FPU -- only
# int32 and the opt-in accurate fp32 mean take the SFPU reduce path (see
# ttnn/cpp/ttnn/operations/reduction/generic/device/common.hpp:use_sfpu_reduce_path) -- and the
# FPU truncates its source registers to a 10-bit mantissa. Every index above 2**11 therefore
# comes back rounded to a value that equals NO actual index, the winner mask comes out empty,
# and the boost silently degrades to a no-op: the tie survives and the greedy token keeps
# flipping. That is exactly how the first version of this pass failed on device. Int32 min/max
# are integer-exact at every index. Keep the reduce, the mask arithmetic and the equality all
# in int32 -- routing any of them through float32 (or uint32, which is on the FPU path too)
# reintroduces the same rounding. forward() hands us uint32, so cast once here.
idx = ttnn.typecast(gathered_global_indices, ttnn.int32, sub_core_grids=scg)
offset = ttnn.multiply(not_max, TIEBREAK_INDEX_SENTINEL, sub_core_grids=scg) # bf16, exact (2**24)
ttnn.deallocate(not_max)
offset_i32 = ttnn.typecast(offset, ttnn.int32, sub_core_grids=scg)
ttnn.deallocate(offset)
masked_idx = ttnn.add(idx, offset_i32, sub_core_grids=scg) # int32
ttnn.deallocate(offset_i32)
greedy_i = ttnn.min(masked_idx, dim=3, keepdim=True, sub_core_grids=scg) # [1,1,B,1] int32
ttnn.deallocate(masked_idx)
is_lowidx_i32 = ttnn.eq(idx, greedy_i, sub_core_grids=scg) # broadcast over W
ttnn.deallocate(idx)
ttnn.deallocate(greedy_i)
is_lowidx = ttnn.typecast(is_lowidx_i32, ttnn.bfloat16, sub_core_grids=scg) # 1/0 -> 1.0/0.0
ttnn.deallocate(is_lowidx_i32)
is_winner = ttnn.multiply(is_max, is_lowidx, sub_core_grids=scg) # 1.0 at exactly one candidate
ttnn.deallocate(is_max)
ttnn.deallocate(is_lowidx)
# gate by k==1 (self._greedy_col [1,1,B,1] bf16); random users get boost 0 => values unchanged
winner_gated = ttnn.multiply(is_winner, self._greedy_col, sub_core_grids=scg)
ttnn.deallocate(is_winner)
boost = ttnn.multiply(winner_gated, delta, sub_core_grids=scg) # delta broadcasts over W
ttnn.deallocate(winner_gated)
ttnn.deallocate(delta)
adjusted = ttnn.add(gathered_values, boost, sub_core_grids=scg)
ttnn.deallocate(boost)
return adjusted
def _untilize_for_argmax(self, x):
"""Untilize a vocabulary row in bounded chunks while preserving its exact width."""
num_untilize_chunks = self._untilize_chunk_count(x.shape[-1])
# DRAM-interleaved ttnn.split/slice does not honor sub_core_grids (same
# senders-column spill as the vocab-trim slice above).
if num_untilize_chunks > 1 and self._force_argmax_sub_core_grids is None:
# Untilizing the full row in one program needs a static circular-buffer
# region proportional to the row width with past around 150K elements it clashes
# with the model's resident L1 buffers at compile.
x_chunks = ttnn.split(x, self._untilize_chunk_width(x.shape[-1], num_untilize_chunks), dim=3)
untilized_chunks = []
for chunk in x_chunks:
# Free each tiled chunk as soon as its row-major copy exists,
# so peak memory holds around 1 full-vocab buffer less than freeing
# after the loop.
untilized_chunks.append(
ttnn.untilize(
chunk,
use_multicore=True,
sub_core_grids=self._force_argmax_sub_core_grids,
)
)
chunk.deallocate()
x_untilized = ttnn.concat(
untilized_chunks,
dim=3,
sub_core_grids=self._force_argmax_sub_core_grids,
)
for chunk in untilized_chunks:
ttnn.deallocate(chunk)
else:
x_untilized = ttnn.untilize(x, use_multicore=True, sub_core_grids=self._force_argmax_sub_core_grids)
return x_untilized
def forward(
self,
x: ttnn.Tensor,
tt_out_tok: ttnn.Tensor = None,
):
"""
Perform on-device sampling on logits tensor.
The logits are sharded over the devices in the cluster.
We perform local top-k on each device, then all-gather the top-k values and indices across all devices.
We then convert the gathered values and indices to the appropriate format, add the device offsets to get the global vocabulary indices,
and perform the actual sampling with top-k, top-p, and temperature.
Args:
x: Input logits tensor
tt_out_tok: Optional output tensor to write results to
Returns:
Sampled token indices tensor
"""
if self._force_argmax_sampling:
logger.info("Forcing argmax sampling")
# Logprobs consume the original vocabulary shards, before argmax's
# gather/slice. Keep token selection identical when reporting toggles.
logits_for_log_probs = x
# BH galaxy prefetcher (unfused-CCL) keeps a split senders/worker sub-device manager
# loaded during decode. The vocab-trim ttnn.slice auto-grids a 32-core block from origin
# (0,0) on the DRAM-interleaved path (it does not honor sub_core_grids there) and spills
# into the uncovered senders-column tail -> "kernel group cores do not match sub device
# cores". Skip that slice here. Still apply the full-width additive invalid-vocab mask
# (already selected whenever sub_core_grids is set; one elementwise add, no slice/concat)
# so a 0-padded logit cannot win argmax when every valid logit is negative. untilize/
# argmax stay pinned to the worker sub-core grid via _force_argmax_sub_core_grids.
force_argmax_skip_vocab_trim = self._force_argmax_sub_core_grids is not None
slice_valid_vocab = self._can_slice_valid_vocab_for_argmax() and not force_argmax_skip_vocab_trim
if not slice_valid_vocab:
x = self._mask_invalid_vocab_logits(x)
# Gather the output across all devices and untilize the tensor (for argmax)
num_devices = self.mesh_device.get_num_devices()
if num_devices > 1:
cluster_axis = self._get_sampling_cluster_axis()
num_links, topology = self._get_force_argmax_all_gather_config(cluster_axis)
logger.debug(
f"Force argmax sampling all-gather: cluster_axis={cluster_axis}, "
f"num_links={num_links}, topology={topology}"
)
# NOTE: do NOT pass a persistent_output_buffer here. The all_gather_async factory
# disables its barrier semaphore when persistent buffers are used, and the barrier
# is what bounds cross-invocation skew: without it a fast device's next-invocation
# writes/increments can reach a peer before that peer's reader has reset its
# out_ready semaphore, leaving residual counts that make later waits pass early
# (argmax then reads the previous step's logits). Under trace the fresh output
# allocation is frozen at capture, so the address is stable anyway.
barrier_accessor = getattr(
self.tt_ccl,
"get_sampling_barrier_semaphore_handle",
self.tt_ccl.get_and_cycle_barrier_semaphore_handle,
)
# Pin the gather to the worker sub-device only on the BH unfused path, where the
# downstream untilize/argmax are themselves confined via _force_argmax_sub_core_grids.
# Without this, all_gather_async defaults to sub-device 0 (prefetcher/senders under
# the galaxy decode manager). Dispatch only serializes within a sub-device, so a
# gather on 0 races the worker-grid untilize and under trace replay returns the
# previous step's tokens. Eager mode masks this via host dispatch latency.
# Left unpinned on Wormhole: untilize/argmax are not sub-core-grid confined there
# and still use the default sub-device 0, matching pre-existing WH behaviour.
ag_sub_device_id = (
getattr(self.tt_ccl, "worker_sub_device_id", None)
if self._force_argmax_sub_core_grids is not None
else None
)
x = ttnn.experimental.all_gather_async(
x,
persistent_output_buffer=None,
dim=3,
multi_device_global_semaphore=self.tt_ccl.get_and_cycle_ag_semaphore_handles(cluster_axis),
num_links=num_links,
memory_config=x.memory_config(),
cluster_axis=cluster_axis,
topology=topology,
barrier_semaphore=barrier_accessor(cluster_axis),
chunks_per_sync=self.argmax_chunks_per_sync,
num_workers_per_link=self.argmax_num_workers_per_link,
num_buffers_per_channel=2,
subdevice_id=ag_sub_device_id,
)
if slice_valid_vocab:
x = self._slice_valid_vocab_for_argmax(x)
x_untilized = self._untilize_for_argmax(x)
tt_out_tok = ttnn.argmax(
x_untilized,
dim=-1,
output_tensor=tt_out_tok,
keepdim=False,
sub_core_grids=self._force_argmax_sub_core_grids,
)
if self.log_probs_calculator.enable_log_probs:
# argmax drops the final axis; the calculator takes the ordinary
# sampler's [1,1,1,B] ROW_MAJOR index layout.
logprob_indices = (
ttnn.reshape(tt_out_tok, (1, 1, 1, tt_out_tok.shape[-1]))
if len(tt_out_tok.shape) == 3
else tt_out_tok
)
self.tt_log_probs = self.log_probs_calculator.calculate_log_probs(logits_for_log_probs, logprob_indices)
else:
self.tt_log_probs = None
return tt_out_tok, self.tt_log_probs
# Decode logits normally arrive in bfloat16 already; a bf16->bf16 ttnn.typecast is
# semantically a no-op but still dispatches a full copy program, so skip it.
x_bf16 = (
x if x.dtype == ttnn.bfloat16 else ttnn.typecast(x, dtype=ttnn.bfloat16, sub_core_grids=self.sub_core_grids)
)
x_bf16 = self._mask_invalid_vocab_logits(x_bf16)
# The single-device split below exists only because the stock top-k factories cap
# out near 64K columns. When ttnn.topk would take the Blackhole topk_large_indices
# composite for the FULL row, the authoritative C++ route query lets one call
# replace the split/per-chunk-topk/offset-add/concat pipeline; its indices are
# already global vocab positions. Calls constrained to a sub-grid never relax.
route_full_row = (
self.multi_step_reduction
and self.sub_core_grid_topk is None
and topk_would_route_to_large_indices(x_bf16, self._num_vocab_splits * self.max_top_k)
)
if route_full_row:
# stable dropped for the same reason as the chunked path below:
# _adjust_values_for_tiebreak guarantees the greedy pick (#33492).
#
# k is multiplied by the split count the chunked path would have used, so the
# candidate set keeps that path's exact width (num_splits x max_top_k): every
# downstream shape is unchanged, and sampling sees a full-row top-(num_splits*k)
# that is a strict quality upgrade over the union of the per-chunk top-ks. At the
# production max_top_k=32 this also keeps the candidate row two tiles wide --
# ttnn.sampling deadlocks on a SINGLE-tile candidate row whose values all tie
# (compute-internal CB deadlock, writer CWFW; tenstorrent/tt-metal#53781).
topk_values_gathered_bf16_interleaved, topk_indices_gathered = ttnn.topk(
x_bf16,
k=self._num_vocab_splits * self.max_top_k,
dim=-1,
stable=False,
)
elif self.multi_step_reduction:
x_bf16_list = ttnn.split(x_bf16, x_bf16.shape[-1] // self._num_vocab_splits, dim=3)
topk_values_list = []
topk_indices_list = []
# Drop stable=True ONLY when the authoritative C++ route query says
# ttnn.topk will take the Blackhole topk_large_indices composite for
# these halves once the custom arguments are absent.
# Sampling opts out of stable topk for decode perf (_topk_stable is
# False): these chunks route to the single-core factory,
# where stable pays ~2-3x on the SFPU sort stage.
# _adjust_values_for_tiebreak is what guarantees the greedy
# pick after the gather, regardless of per-device tie order. Calls
# that would not route keep today's arguments bit-for-bit, and a
# call the model constrained to a sub-grid is never relaxed.
use_routed_topk = self.sub_core_grid_topk is None and topk_would_route_to_large_indices(
x_bf16_list[0], self.max_top_k
)
for i in range(len(x_bf16_list)):
# Chunks are not padded to a power of two here: an A/B on this path
# (PR #53167) measured no end-to-end decode benefit from steering
# ttnn.topk to the multi-core factory, so single-core chunks stay.
topk_values, topk_indices = ttnn.topk(
x_bf16_list[i],
k=self.max_top_k,
dim=-1,
sub_core_grids=self.sub_core_grid_topk,
# _topk_stable is False: the fast unstable network is used and
# _adjust_values_for_tiebreak guarantees the greedy pick. The routed-topk
# relax (unstable where the call would take the BH topk_large_indices
# composite) is kept so re-enabling stable stays a one-line change.
stable=False if use_routed_topk else self._topk_stable,
)
topk_values_list.append(topk_values)
topk_indices_list.append(topk_indices)
x_bf16_list[i].deallocate()
topk_values_gathered_bf16_interleaved = ttnn.concat(topk_values_list, dim=3)
topk_indices_gathered = ttnn.concat(topk_indices_list, dim=3)
for i in range(len(topk_indices_list)):
ttnn.deallocate(topk_values_list[i])
ttnn.deallocate(topk_indices_list[i])
else:
# apply padding to the input tensor if needed
# if number is not power of 2, pad to upper power of 2
# pad only last dimension with float::min value to upper_power_of_2
# This is necessary to use full optimization in the topk operation.
if self.pad_to_power_of_2 and not is_power_of_2(x_bf16.shape[-1]):
padded_value = upper_power_of_2(x_bf16.shape[-1])
x_bf16 = ttnn.pad(
x_bf16,
[(0, 0), (0, 0), (0, 0), (0, padded_value - x_bf16.shape[-1])],
value=-sys.float_info.max,
sub_core_grids=self.sub_core_grids,
)
# Perform local top-k on each device. Drop stable=True ONLY when the
# authoritative C++ route query says the relaxed call will take the
# Blackhole topk_large_indices composite -- sampling keeps every call
# unstable (_topk_stable is False): this multicore-eligible call would
# usually get the ~free fused-key stable engine, but the greedy
# guarantee must hold on ALL routes (see _topk_stable), and
# _adjust_values_for_tiebreak already provides it host-side.
# Sub-grid-constrained calls never relax.
use_routed_topk = self.sub_core_grid_topk is None and topk_would_route_to_large_indices(
x_bf16, self.max_top_k
)
topk_values, topk_indices = ttnn.topk(
x_bf16,
k=self.max_top_k,
dim=-1,
sub_core_grids=self.sub_core_grid_topk,
# _topk_stable is False: the fast unstable network is used and
# _adjust_values_for_tiebreak guarantees the greedy pick. The routed-topk
# relax (unstable where the call would take the BH topk_large_indices
# composite) is kept so re-enabling stable stays a one-line change.
stable=False if use_routed_topk else self._topk_stable,
)
# For 1D meshes use `cluster_axis=None`. For 2D meshes, use the configured gather axis.
sampling_cluster_axis = self._get_sampling_cluster_axis()
# Gather top-k values across all devices
topk_values_gathered = self._perform_all_gather(
topk_values,
dim=3,
cluster_axis=sampling_cluster_axis,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
num_links=self.num_gather_links,
buffer_key="SAMPLING_VALUES",
)
ttnn.deallocate(topk_values)
# Convert gathered values to appropriate format
if self.sampling_memory_config != ttnn.DRAM_MEMORY_CONFIG:
topk_values_gathered_bf16 = ttnn.to_memory_config(
topk_values_gathered,
memory_config=self.sampling_memory_config,
dtype=ttnn.bfloat16,
)
topk_values_gathered_bf16_interleaved = ttnn.to_memory_config(
topk_values_gathered_bf16, memory_config=ttnn.DRAM_MEMORY_CONFIG
)
ttnn.deallocate(topk_values_gathered_bf16)
else:
topk_values_gathered_bf16_interleaved = topk_values_gathered
# Gather top-k indices across all devices
topk_indices_gathered = self._perform_all_gather(
topk_indices,
dim=3,
cluster_axis=sampling_cluster_axis,
memory_config=ttnn.DRAM_MEMORY_CONFIG,
num_links=self.num_gather_links,
buffer_key="SAMPLING_INDICES",
)
ttnn.deallocate(topk_indices)
# Convert indices to appropriate data types
if route_full_row:
# Full-row top-k indices are already global vocab positions; there are no
# per-chunk offsets to add. The routed composite emits uint16/uint32 on the
# stock op's 16-bit width boundary, so widen only when needed.
if topk_indices_gathered.dtype == ttnn.uint32:
topk_global_indices_interleaved = topk_indices_gathered
else:
topk_global_indices_interleaved = ttnn.typecast(
topk_indices_gathered, dtype=ttnn.uint32, sub_core_grids=self.sub_core_grids
)
ttnn.deallocate(topk_indices_gathered)
else:
topk_indices_gathered_int32 = ttnn.typecast(
topk_indices_gathered, dtype=ttnn.int32, sub_core_grids=self.sub_core_grids
)
if self.sampling_memory_config != ttnn.DRAM_MEMORY_CONFIG:
topk_indices_gathered_int32_sharded = ttnn.to_memory_config(
topk_indices_gathered_int32, self.sampling_memory_config
)
ttnn.deallocate(topk_indices_gathered_int32)
else:
topk_indices_gathered_int32_sharded = topk_indices_gathered_int32
# Add device offsets to get global vocabulary indices
topk_global_indices = ttnn.add(
self.tt_indices_device_offsets,
topk_indices_gathered_int32_sharded,
dtype=ttnn.uint32,
memory_config=self.sampling_memory_config,
)
ttnn.deallocate(topk_indices_gathered_int32_sharded)
topk_global_indices_interleaved = ttnn.to_memory_config(topk_global_indices, ttnn.DRAM_MEMORY_CONFIG)
# Untilize indices for sampling operation
topk_global_indices_interleaved_untilised = ttnn.untilize(
topk_global_indices_interleaved, use_multicore=True, sub_core_grids=self.sub_core_grids
)
# Perform the actual sampling with top-k, top-p, and temperature.
# Host-side tie-break (in lieu of stable=True on every top-k route; see _topk_stable):
# for argmax users (k==1) only, boost the single lowest-GLOBAL-INDEX tied maximum in the
# sampling INPUT so ttnn.sampling's argmax picks it regardless of how the top-k network
# ordered the tied candidates. Random users are byte-for-byte unchanged. Correcting the INPUT
# (not the RM output buffer) is required: no ttnn op writes an interleaved ROW_MAJOR tensor
# in-place on a restricted sub-device. See _adjust_values_for_tiebreak for the full rationale
# and its known limitation (>max_top_k maxima tied within one device shard).
sampling_values = self._adjust_values_for_tiebreak(
topk_values_gathered_bf16_interleaved, topk_global_indices_interleaved
)
# Seed immediately before the draw. The tie-break's int32 ops run on the
# SFPU (use_sfpu_reduce_path admits INT32 MIN/MAX/SUM) on the same sub-core grid,
# and rand_tile's PRNG/LREG state is programmed by manual_seed -- so any SFPU work
# between seeding and drawing can perturb the draw. The original's fp32 tie-break
# took the FPU path and never disturbed it.
ttnn.manual_seed(
seeds=self.seeds_tt_tensor,
user_ids=self.user_ids_tt_tensor,
sub_core_grids=self._sampling_sub_core_grids,
)
tt_out_tok = ttnn.sampling(
sampling_values,
topk_global_indices_interleaved_untilised,
k=self.k_tensor,
p=self.p_tensor,
temp=self.temp_tensor,
sub_core_grids=self._sampling_sub_core_grids,
output_tensor=tt_out_tok,
)
# Compute logprobs if enabled
if self.log_probs_calculator.enable_log_probs and self.log_probs_calculator._use_topk_logprobs:
# New path: top-K logprobs for gpt-oss-120b
self.tt_log_probs = self.log_probs_calculator.calculate_topk_log_probs(
logits_tensor=x,
topk_values=topk_values_gathered_bf16_interleaved,
topk_global_indices=topk_global_indices_interleaved,
sub_core_grid_topk=self.sub_core_grid_topk,
)
elif self.log_probs_calculator.enable_log_probs:
# Old path: single sampled-token logprob
self.tt_log_probs = self.log_probs_calculator.calculate_log_probs(x, tt_out_tok)
else:
self.tt_log_probs = None
ttnn.deallocate(sampling_values)
ttnn.deallocate(topk_values_gathered_bf16_interleaved)
ttnn.deallocate(topk_global_indices_interleaved)
ttnn.deallocate(topk_global_indices_interleaved_untilised)
return tt_out_tok, self.tt_log_probs
|